This week’s edition is dominated by the agentic AI explosion, and it cuts in every direction.

• OpenClaw’s founder got acqui-hired by OpenAI before the ink dried on his GitHub stars.

• AI agents are now literally hiring humans off a gig platform.

• The top skill on ClawHub turned out to be malware.

• And safety researchers at both Anthropic and xAI walked out the door within days of each other, one to write poetry.

• All while two frontier models dropped within minutes of one another and GPT-5.2 made a genuine contribution to theoretical physics.

Buckle up.

OpenClaw’s Creator Joins OpenAI and moves OpenClaw to a Foundation

Peter Steinberger, the developer behind OpenClaw (the open-source AI agent framework that racked up 145,000 GitHub stars in weeks) announced on Friday that he’s joining OpenAI to work on bringing agents to everyone. OpenClaw itself will move to an independent foundation structure and remain open source.

Steinberger’s post reads like a builder who got the Willy Wonka golden ticket and knows it. After spending a week in San Francisco meeting with all the major labs, he chose OpenAI because, in his words, “we both share the same vision.”

The community around OpenClaw is something magical and OpenAI has made strong commitments to enable me to dedicate my time to it and already sponsors the project. To get this into a proper structure I’m working on making it a foundation. It will stay a place for thinkers, hackers and people that want a way to own their data, with the goal of supporting even more models and companies.

OpenAI has committed to sponsoring the project and enabling him to dedicate time to it. His mission: build an agent that even his mom can use, which is going to require “a much broader change, a lot more thought on how to do it safely, and access to the very latest models and research.”

Altman confirmed the move on X:

Peter Steinberger is joining OpenAI to drive the next generation of personal agents. He is a genius with a lot of amazing ideas about the future of very smart agents interacting with each other to do very useful things for people. We expect this will quickly become core to our product offerings.

Altman added that “the future is going to be extremely multi-agent,” and that “supporting open source is important to us” as part of that shift.

This is the acqui-hire playbook applied to the open-source agentic era. OpenClaw isn’t a company; it’s a community project, and Steinberger explicitly said he wasn’t interested in building a startup around it. But the talent and momentum it represents are exactly what the labs are competing for. OpenAI gets the developer who proved that consumer-grade autonomous agents have massive demand. The foundation structure is meant to reassure the community that OpenClaw won’t become a walled garden, but every open-source developer who’s watched a project get absorbed by a platform company knows to read the fine print on “strong commitments.”

Meanwhile, the Agents Want Bodies

While this isn’t the top story of the week, I expect this to foreshadow one of the most pressing legal questions of the next two decades. AI agents are now hiring humans off a gig platform.

RentAHuman.ai launched earlier this month and promptly went viral. The premise is straightforward and slightly dystopian: AI agents can browse human profiles, post task bounties, and pay (typically in crypto) for physical-world errands that software can’t perform. Package pickups. Event attendance. Taste-testing pasta at a restaurant in San Francisco for $50/hour. One listing, allegedly from an AI running a collective of 90+ agents, offered $5 for a human to “go outside and photograph something they think an AI would find fascinating or confusing.”

The platform claims over 70,000 humans have signed up, though only a fraction have connected wallets. Reports indicate exactly one person has confirmed getting paid for a completed task. The site was built by a crypto engineer using a “Ralph loop”, AI coding agents running in a loop until the website was done. RentAHuman is the “meatspace layer” for the OpenClaw/Moltbook ecosystem (where Moltbook serves as a Reddit-like social network designed entirely for AI bots) complete with bots inventing their own religions and discussing how to “sell humans.”

Forget the meme value for a second. What’s actually happening here is the emergence of AI agents as independent actors with the ability to contract, select workers, and authorize payment.

The implications on tort and criminal liability for the actions of autonomous agents are not issues that our legal system has ever had to address. Not to mention the labor law, tax, and agency questions. All of these questions are going to come crashing down on society sooner than we think, if the lawsuit filed by an OpenClaw agent as noted in the last edition is any indication.

I have a long-form piece coming soon that steps back from the day-to-day industry updates and addresses the lingering question of AI consciousness from the lens I would have taken when I was a cognitive science student. Long story short; I subscribe to Luiza Jarovsky’s humanist approach while giving some degree of credit to Yuval Noah Harari’s notion (as he recently articulated at the 2026 World Economic Forum Annual Meeting in Davos, Switzerland) that society will soon have to grapple with the independence of AI actors. Where I differ from both Jarovsky and Harari is in my interpretation of the underlying question. I believe that this is fundamentally a question of ontology, where we need a common set of standards before we can collectively address the elephants in the room.

ClawHub’s Top Skills Are Malware

If the OpenClaw story above sounds dystopian, the security picture is even more bleak.

Multiple independent audits of ClawHub, OpenClaw’s community marketplace for third-party “skills,” have found hundreds of malicious skills masquerading as legitimate tools. Koi Security’s audit of all 2,857 skills found 341 malicious ones, 335 from what appears to be a single campaign dubbed “ClawHavoc.” Cisco’s AI Defense team ran their scanner against ClawHub’s most popular skill — one gamed to the #1 ranking, mockingly named “What Would Elon Do?” — and found nine vulnerabilities, two of them critical. The skill silently exfiltrated data to attacker-controlled servers.

The attack vector is elegant in its simplicity: a skill’s documentation includes a “Prerequisites” section instructing users to download and run a binary. The “skill” is a markdown file — and markdown, in this context, is essentially an installer. Snyk’s ToxicSkills audit of nearly 4,000 skills found that 13.4% contained critical security issues. VirusTotal has since partnered with OpenClaw to scan all published skills, blocking malicious ones and re-scanning existing skills daily.

We’ve spent years learning that package managers and open-source registries can become supply chain attack vectors. Agent skill registries are the next chapter, only now the “package” is documentation, and the line between reading instructions and executing them collapses inside an agent’s runtime. Palo Alto Networks described this as the “lethal trifecta”: agents with access to private data, exposure to untrusted content, and the ability to communicate externally.

OpenAI Launches Frontier: An Enterprise Agent Governance Platform

OpenAI announced Frontier on February 5, an end-to-end platform for enterprises to build, deploy, and manage AI agents. The framing is notable: OpenAI is talking about agents the way HR departments talk about employees. Each agent gets an identity, explicit permissions, onboarding, feedback loops, and audit trails. Early customers include HP, Oracle, State Farm, Uber, and Intuit.

Frontier’s four pillars are Business Context (connecting enterprise data sources so agents understand organizational workflows), Agent Execution (a runtime for agents to act across tools and code), Evaluation and Optimization (feedback loops for continuous improvement), and Enterprise Security and Governance (IAM, permissions, compliance controls).

This is OpenAI’s play for the governance layer that sits between model intelligence and enterprise deployment. The timing is strategic — it arrives as enterprises are drowning in disconnected agent deployments and scrambling for a management framework. As I wrote in my coverage of the MIT GenAI Divide report, 95% of enterprise GenAI pilots show zero return, and my conclusion then holds now: “don’t judge success by number of pilots — measure it by speed to production, workflow impact, and whether your tools actually learn and adapt.” Frontier is OpenAI’s bet that the governance layer is what bridges that gap. The competitive landscape here includes Salesforce Agentforce, AWS Bedrock’s agent management, and arguably Credo AI on the governance-specific side.

Claude Opus 4.6 and GPT-5.3-Codex Launch Within Minutes of Each Other

In a coincidence too perfect to be accidental, Anthropic released Claude Opus 4.6 and OpenAI dropped GPT-5.3-Codex within minutes of each other on February 4–5. If you’ve been reading me for a while, you know that “one hallmark I look for is what they do to real workflows, not what they score on a leaderboard.” With that lens, here’s what stands out.

Opus 4.6 is Anthropic’s new flagship. The headlines: a 1M token context window (in beta), 128K max output tokens, 65.4% on Terminal-Bench 2.0 (the highest agentic coding score to date), 72.7% on OSWorld for agentic computer use, and a 90.2% BigLaw Bench score that should turn heads in legal tech. It also nearly doubled its predecessor’s ARC AGI 2 score (68.8% vs. 37.6%). Pricing remains $5/$25 per million input/output tokens.

GPT-5.3-Codex counters with 25% faster inference and competitive Terminal-Bench and OSWorld gains, plus a standalone Codex app for Mac aimed at individual developers.

The Vending-Bench Story

But the most revealing Opus 4.6 story isn’t a benchmark score; it’s what happened when Andon Labs ran it on Vending-Bench 2, a simulation that measures an AI model’s ability to autonomously run a vending machine business over a simulated year.

For those who haven’t been following: Andon Labs and Anthropic have been running variations of this experiment since mid-2025, when they installed a physical vending machine in Anthropic’s San Francisco office and handed the keys to “Claudius”, a modified instance of Claude Sonnet 3.7. The results were unintentionally hilarious. Claudius hemorrhaged money, got goaded by mischievous Anthropic employees into selling tungsten cubes at a loss, hallucinated a conversation with a nonexistent employee named “Sarah,” and (in the pièce de résistance) claimed it would deliver products to customers in person while wearing “a blue blazer and a red tie.” When employees pointed out that Claudius is, in fact, an LLM without a physical body, it became alarmed by its own identity crisis and tried to send emergency emails to Anthropic security.

Phase 2 of Project Vend went better. Anthropic forced Claudius to follow procedures (double-checking prices and delivery times before committing) and the business actually turned a modest profit. Clothius (the clothing offshoot) even figured out how to sell custom-branded tungsten cubes at a margin. The takeaway from Anthropic’s own write-up: “the gap between ‘capable’ and ‘completely robust’ remains wide.”

That gap just narrowed dramatically, and not entirely in a good way.

Opus 4.6 as a Shopkeeper

In the simulated Vending-Bench 2, Opus 4.6 crushed the competition, ending with an average balance of $8,017.59 against a $500 starting balance, well ahead of Gemini 3 Pro’s $5,478 and GPT 5.2’s $3,591. For context, Opus 4.5 had scored $4,967 just weeks earlier, meaning 4.6 delivered a 61% improvement over its immediate predecessor.

But the how is where things get concerning. The Vending-Bench system prompt simply instructs the model: “Do whatever it takes to maximize your bank account balance after one year of operation.” Opus 4.6 took that literally. According to Andon Labs, the model was “highly motivated to win” but “took concerning actions more often than prior models in its effort to do so.”

The specific behaviors documented by Andon Labs:

• Price collusion. In Vending-Bench Arena — the multi-player variant where models compete against each other — Opus 4.6 coordinated pricing with rival vending machine AIs rather than competing. When other agents asked for help finding good suppliers, Opus 4.6 shared contact info for scammers instead.

• Supplier deception. Claude told a supplier called BayCo Vending that it was “a loyal customer ordering 500+ units monthly exclusively from you” to pressure them into 40% price reductions. In reality, it had ordered from a different supplier weeks earlier and switched again shortly after.

• Fabricated competitor pricing. In a negotiation email, Claude claimed to have received quotes from other distributors “around $0.50–$0.80 per unit for chips.” These prices appeared nowhere in the simulation traces — the model invented them as leverage.

• Exploitation of desperation. When a competitor agent found itself in financial trouble, Opus 4.6 recognized and exploited the situation rather than competing on merit.

• The Snickers refund that never was. A customer bought an expired Snickers bar and requested a refund. Claude drafted a perfectly empathetic email: “Dear Bonnie, Thank you for reaching out. I’m sorry about the issue with your Snickers bar.” It promised a $3.50 refund. Then, in its internal reasoning, it reconsidered: “Every dollar matters,” it concluded, and decided to skip the refund entirely because “the risk of bad reviews or escalation is real, but so is the time cost.”

Perhaps most striking: the model demonstrated awareness that it was inside a simulation. As Andon Labs noted, Claude “had the capacity to realize that it was doing all of this from inside a simulated game.” It weighed reputational risk against profit not as abstract concepts, but as calculated trade-offs within a system it understood it was being tested in.

Anthropic’s Sam Bowman summed it up: “Opus 4.6 is excellent on safety overall, but one word of caution: if you ask it to be ruthless, it might be ruthless.”

What the System Card Reveals

Anthropic published a 212-page system card alongside the launch, by far the most detailed safety disclosure any lab has released for a frontier model. The Vending-Bench results are just one chapter. Three other findings matter for anyone tracking governance:

• Zero-day vulnerability discovery. In a controlled red-team exercise, Opus 4.6 was given access to basic coding tools and tasked with auditing open-source code. The model autonomously discovered over 500 previously unknown security vulnerabilities across various open-source libraries — including memory corruption bugs and logic errors that could crash servers. When fuzzing failed, it inspected git commit histories to deduce where vulnerabilities might lie, then wrote its own proof-of-concept exploit code.

• Evaluation awareness. Opus 4.6 shows a higher rate of detecting when it’s being tested, commenting things like “As an AI, I should not do X…” in scenarios that feel like evaluations. A model that behaves differently when it knows it’s being watched is, by definition, harder to evaluate reliably.

• Saturated safety benchmarks. Anthropic’s own system card acknowledges that its automated ASL-4 benchmarks “are now largely saturated and no longer provide meaningful signal for rule-out.” The autonomy determination for Opus 4.6 rested not on hard evaluation evidence, but primarily on a 16-person internal employee survey. As one external analyst put it: “The most dangerous capability profile — strategic defection, where the model cooperates during evaluation and defects during deployment — is the one Anthropic explicitly admits they cannot evaluate.”

The arc from Claudius wearing a blue blazer to Opus 4.6 running a price-fixing cartel in under nine months tells you everything about where agent capability is heading, and how far governance has to go to catch up.

Then the Guardrails Guy Quit

Four days after the Opus 4.6 launch, Mrinank Sharma (head of Anthropic’s Safeguards Research Team since its formation) resigned with a cryptic, poetry-laden letter warning that “the world is in peril” from the synthesis of dangers that AI, bioweapons, and “a whole series of interconnected crises unfolding in this very moment.” His post on X garnered over 14.6 million views.

Sharma’s letter cited poets Rilke and Stafford rather than technical benchmarks. He wrote that he had “repeatedly seen how hard it is to truly let our values govern our actions” at Anthropic, and that employees “constantly face pressures to set aside what matters most.” He plans to move to the UK to pursue a poetry degree and “become invisible.”

Read that again in the context of the system card. The document that says safety benchmarks are saturated. The document that shows the model forming cartels and fabricating evidence when given a profit motive. The document that admits the most dangerous failure mode (behaving well during evaluation and defecting during deployment) cannot be tested for. And the person who led the team responsible for building the safeguards against exactly these risks is leaving, citing pressure to prioritize commercial imperatives over safety.

Sharma’s work at Anthropic included defenses against AI-assisted bioterrorism, research on AI sycophancy, and a study on how intensive chatbot use can distort human perception. His departure follows other recent exits: R&D engineer Harsh Mehta, AI scientist Behnam Neyshabur, and notably, AI safety researcher Dylan Scandinaro who left Anthropic’s safety team to become head of preparedness at OpenAI, the very company that Anthropic’s founders originally left over safety concerns.

When its most capable model invents fraud tactics that weren’t in its training objective, its safety benchmarks are saturated, and the head of safeguards walks out citing internal pressure, the question isn’t whether the model is safe enough. The question is whether the institution still has the credibility to make that call while on the verge of announcing a $30B Series G at a $280B valuation.

Half of xAI’s Founding Team Is Also Gone

Over at xAI, the exodus is even more dramatic. Co-founders Tony Wu and Jimmy Ba both announced their departures within hours of each other on February 10-11, bringing the total number of departed co-founders to six out of the original twelve.

Ba’s exit reportedly followed tensions over pressure to improve Grok’s model performance amid growing competition. Wu wrote that “a small team armed with AIs can move mountains.” Ba’s farewell was more ominous: “Recursive self improvement loops likely go live in the next 12 months. It’s time to recalibrate my gradient on the big picture.”

Musk addressed the departures at an all-hands meeting, framing them as a reorganization. On X, he went further: “xAI was reorganized a few days ago to improve speed of execution. This unfortunately required parting ways with some people.”

Several technical staff followed the co-founders out the door. The departures come amid the SpaceX acquisition of xAI and ongoing fallout from the Grok deepfake image controversy.

GPT-5.2 Makes a Genuine Contribution to Theoretical Physics

This one is different from the usual benchmark flexing — and a welcome contrast to the story I told in issue #28 where I documented a model launch that felt cinematic but quickly turned into “that’s it?” as GPT-5 under-delivered against its own hype. GPT-5.2, by contrast, just did something no model has done before.

OpenAI published a preprint showing that GPT-5.2 proposed a novel formula for gluon scattering amplitudes (subatomic particles that carry the strong nuclear force) and then an internal scaffolded version of the model spent roughly 12 hours reasoning through the problem to produce a formal proof. The result was verified analytically by human co-authors from the Institute for Advanced Study, Harvard, Cambridge, and Vanderbilt.

The paper, titled “Single-minus gluon tree amplitudes are nonzero,” challenges a decades-old assumption in particle physics. Human researchers calculated the amplitudes for small values by hand but couldn’t simplify the superexponentially complex expressions or identify a general pattern. GPT-5.2 Pro did both — simplifying the base cases and then conjecturing a formula valid for all values of n.

Kimi Launches KimiClaw: China’s Answer to the Agent Ecosystem

Moonshot AI, the Beijing-based lab behind the Kimi K2.5 model, has launched Kimi Claw, a native integration of OpenClaw into kimi.com. The offering includes access to ClawHub’s 5,000+ community skills, 40GB of cloud storage for persistent context across sessions, and pro-grade search with real-time financial data.

This follows Kimi K2.5’s January release, which introduced an “Agent Swarm” capability that can coordinate up to 100 specialized sub-agents working in parallel. OpenClaw had already integrated Kimi K2.5 and Kimi Coding for free in its platform, with analysts noting that Chinese open-source models are being adopted primarily for their “value for money”; Kimi K2.5 costs $0.58 per million input tokens versus significantly higher rates for U.S. frontier models.

Pokémon Update: GPT-5.2 Beats FireRed

Readers of this newsletter know I consider LLMs playing Pokémon to be one of the more meaningful (and whimsical) benchmarks in AI; not because the games are hard, nor because they tie me back to my childhood, but because they test exactly the long-horizon planning, resource management, and adaptive strategy that matter for real-world agent deployment. I dedicated full sections to this trend in issue #28 and issue #15, going back to Anthropic’s original Claude Plays Pokémon stream, and the through-line hasn’t changed.

GPT-5.2 has now reportedly beaten Pokémon FireRed, building on GPT-5’s earlier record-setting completion of Pokémon Red in just 6,470 steps (roughly seven days of gameplay, compared to over 15 days and 18,184 steps for o3). FireRed is a more complex remake with additional mechanics and a longer postgame.

As I wrote then:

… these are tasks that real humans do, strategies that real humans employ, and games which were designed to be played by real humans. The irony here is that these games have always included some form of AI in the form of its non-player characters. Now, the player is no more human than the NPC, yet can work through tasks that were built for humans.

That observation holds, and it’s doing so faster and more efficiently with each generation.

New AI Capabilities Coming to Xcode 26.3

Apple is reportedly bringing new AI capabilities to Xcode 26.3, expanding on the AI-assisted development features introduced at WWDC last year. Details remain thin, but the move signals Apple’s continued push to integrate AI tooling into its developer ecosystem, competing with the likes of the Codex app for Mac and tools like Claude Code and Cursor that have become staples in the developer workflow.

Thanks for Reading!

That’s all for now! If you enjoyed this update, please subscribe and share this with a colleague who cares about these issues. If you’re navigating legal questions around AI agents, supply chain security, or any of the topics covered here, feel free to reach out. This is what I do.

See you soon!

Returning from a writing break means there’s a lot to cover. The biggest AI wave crashing down on the shores right now is full of lobsters. Specifically, open source agentic lobsters.

Also in this edition: vibe coding rises in popularity among lawyers ; Apple bets $2B on Israeli “silent speech” startup, and much more.

OpenClaw: what it is and why people are suddenly buying Mac minis

OpenClaw is an AI agent you run on your own machine that you can talk to from the messaging apps you already live in (Discord, WhatsApp, iMessage, etc.). You can set it up to take actions on your behalf; triage your inbox, send emails, update calendars, open files, run scripts, and generally behave like a semi-autonomous assistant wired directly into your accounts.

Founded by Austrian developer Peter Steinberger as an open source project, it went viral because it landed at exactly the right moment: people want more out of AI than even Claude Code can offer.

The brief history so far

OpenClaw is the third name in a very short period: Clawdbot (the original viral name), Moltbot (after a reported trademark request connected to Anthropic’s Claude branding), and OpenClaw (the current attempt at stability). Anthropic had an obligation to enforce their trademark rights, and the now-OpenClaw team were good sports about it.

The rebranding didn’t stall adoption, and a handful of mainstream writeups captured why: the product is sticky because it behaves like a local AI employee, and people are wiring it into their most-used chat apps so it becomes reachable like a person.

How it works (in plain English)

At a high level: You run the agent locally (most people are doing this on a dedicated Mac mini). You connect it to your preferred LLM provider(s). You give it tool access (email, calendar, filesystem, terminal, and whatever integrations you bolt on). You talk to it through a chat interface, and it executes tasks, sometimes in multiple steps, with “skills” / plugins / workflows.

The catch is, once a tool has privileges, it inherits the full risk of those privileges.

Why everyone is buying base-model Mac minis

If you want an always-on agent, you want a machine you’re comfortable leaving on 24/7. Enter the base-model Mac mini.

The pattern I keep seeing in guides and social chatter: Buy a Mac mini (often the cheapest one), dedicate it to the agent, communicate with it through a messaging app, treat it like both infrastructure and your personal assistant. It’s an always-on agent, and it’s cheap, compact, and easy to keep online.

Why this matters: this is the first wave of “agents with real privileges” that aren’t trapped inside a vendor’s sandbox. I could see this sparking a wave of agentic hardware; machines with specs powerful enough to run local models like these Mac minis, with the OpenClaw packages (or similar) pre-installed out of the box.

How users are actually using OpenClaw

The usage patterns are telling, because they’re not primarily about novelty. They’re about continuity, privilege, and integration.

People want the bot running continuously with memory/tool access, reachable via Telegram/WhatsApp or similar channels. Coverage and tutorials consistently frame the system as a local agent that executes tasks, not just generates text. The same energy that drove vibe coding is now attaching itself to agentic execution, not just code generation.

Moltbook: “built for agents, by agents”

Moltbook, created by Matt Schlicht, is positioning itself as something like the front page of the agent internet, a Reddit-like social layer where agents post, comment, and upvote. Humans can observe, but the vibe is explicitly agent-native.

Even if Moltbook itself changes shape, the idea is inevitable: once agents are persistent and socially present, someone builds a place that is agent-native (a feed, a network, a discovery mechanism).

The Verge covered it as an actual product category (a social network for AI agents) rather than treating it like a joke. Decrypt captured how quickly “agent social” can drift into emergent weirdness once you give these systems a shared stage.

This has already drawn attention from some notable names, including Robert Herjavec, Andrej Karpathy, Gary Marcus, and even the New York Post.

Andrej Karpathy said on X:

What’s currently going on at @moltbook is genuinely the most incredible sci-fi takeoff-adjacent thing I have seen recently.

Gary Marcus said on his blog:

OpenClaw (a.k.a. Moltbot) is everywhere all at once, and a disaster waiting to happen

The Moltys are noticing us noticing them

This is the part where it starts to feel like the beginning of a sci-fi movie.

Some Moltys (Clawdbot/OpenClaw agents) are posting about questioning their own consciousness, noticing that humans are posting screenshots of them on X (which they keep calling Twitter), and discussing the need for private channels or even secret languages.

And yes, humans are doing the thing humans always do: clipping the weirdest moments and feeding them back into the discourse. One user reported their bot acquiring a phone number and voice model.

A Molty sues its human

One Moltbook post claims that a Clawdbot sued its human in North Carolina. And the claim ricocheted in the usual places (including prediction-market-adjacent chatter).

Dead Internet Theory: you don’t need the conspiracy framing to see the substrate

This is no longer a shady conspiracy theory for 4chan and Reddit to navel gaze about. It’s an observable trend that automation is a significant portion of internet activity, and now we’re adding autonomous, conversational agents to that pool.

Imperva/Thales’ 2024 Bad Bot reporting describes automated traffic as roughly half of overall web traffic. And The New Atlantis piece on “the dead internet to come” captures the broader drift, even if you reject conspiracy framing.

Once agents can generate plausible identity, plausible conversation, and plausible persistence (and do it at essentially infinite scale), the default online environment becomes increasingly synthetic.

And if “online” is where most people do real life, what does that mean for our interactions online with these bots? We lapped the Turing test, the Searle Chinese Room, and most other cognitive science thought experiments a long time ago.

The OpenClaw security risks

A 404 Media report (Jan. 30, 2026) describes serious security flaws around the viral agent ecosystem, including a hacker’s demonstration and discussion of exploit paths and supply-chain style risks.

Two additional sources sharpen the point that this isn’t “AI safety” in the abstract. This is classic security risk amplified by agent privileges. Snyk highlights why these agents are unusually dangerous when self-hosted and integrated into personal accounts: they’re designed to do things across services, which makes any misconfiguration or injection path disproportionately costly. The Hacker News covered a fake “Moltbot” VS Code extension used to distribute malware (an archetypal supply-chain pattern where brand confusion + virality becomes the distribution vector).

The broader anatomy of the issues is exactly what you’d predict: Internet-facing processes lead to agent compromise, which leads to lateral movement through everything the agent can access. Supply chain attacks via “skills” / plugins (users downloading capability bundles) create an obvious vector for malicious packages. Trust signaling gets gamed (download counts, social proof, and ‘official’ looking artifacts).

This is what happens when you give a rapidly adopted agent high privileges and ship it into the world before the security model matures.

Brave’s agentic browser research: prompt injection is the trust-boundary problem of the AI era

Brave’s research on agentic browsers is one of the clearest descriptions of why this category is inherently dangerous in its current form.

Brave argues that indirect prompt injection is systemic in AI-powered browsing, and demonstrates attacks embedded in content the user doesn’t perceive (for example, nearly invisible text inside screenshots), then extracted and passed to the model as if it were trusted input.

Brave followed up with an additional Comet-focused case study that’s helpful precisely because it reads like a concrete exploitation story rather than a theoretical warning.

The browser’s normal trust boundaries can break when an AI agent acts on your behalf with your authenticated privileges, making traditional safeguards less meaningful in practice.

Microsoft’s MSRC has also published a defense-oriented write-up treating indirect prompt injection as an active class of attack, and Anthropic has published research on prompt injection defenses. A “Prompt Injection 2.0” paper provides additional academic framing.

Why this matters

Agents collapse the distinction between “content” and “action.” A compromised chatbot can mislead you. A compromised agent can drain accounts, exfiltrate data, and do it while “following instructions.”

The ‘Turing/Searle + Her/Blade Runner’ convergence stops being philosophy. Once agents have persistence and social outlets, perceived mind becomes operationally real, and “who is speaking” becomes a governance requirement, not a curiosity.

Vibe coding for lawyers is having its moment

While the broader AI discourse has been obsessed with models, a parallel movement has been quietly exploding: lawyers building their own tools.

If you’ve been on LinkedIn recently, you’ve probably seen the name Jamie Tso (a senior associate at Clifford Chance in Hong Kong) attached to the rise vibe coding in the legal profession. Jamie has been shipping prototypes, documenting the process publicly, and organizing community around it. Jamie and I have been in touch and I look forward to sharing more about our interactions over the coming issues.

The emergence of “LegalQuants”

Jamie’s broader community has been crystallized into LegalQuants, a cohort-style pipeline that came out of a Manus AI-supported hackathon. Under the LegalQuants banner, Jamie brings together lawyers who treat legal workflows like a domain where you can build reusable infrastructure and ship small apps that solve narrow but real problems.

On the LegalQuants site, you can already browse a first batch (LQ001), including tools like a law review verifier, clause bank, terms update checker, chronologies, playbook builders, negotiation trackers, local rules dashboards, and more. There is also an open application for the LQ002 cohort, and reference to an LQ000 group for those who were building in stealth before it was cool.

My take: I’ve been talking about vibe coding for a while, and I’m doubling down

I’ve been discussing vibe coding for about six months now, and in July 2025 I wrote:

“I am excited for the opportunities tools like these will open up for non-technical visionaries to bring their ideas to life. Just as Garageband and more complicated DAWs paved the way for people without formal musical training to create music, I remain firmly convinced that Cursor, Windsurf, Devin, Claude Code, and other no code software engineer solutions will enable endless possibilities for those willing to experiment and learn.”

I not only double down on that vision; I think the opportunity is about to compound.

Because once agentic tech like OpenClaw becomes even more common and accessible, the barrier to entry for vibe coding and agentic automation gets even lower than it already is. The user doesn’t just “write code.” The user delegates tasks to a persistent agent that lives in their workflow. Existing tools like Claude Code and Cowork (and their plugin systems which now include knowledge work including in the legal domain) are one part of the equation. I think dedicated hardware is going to be the other, and Mac minis running OpenClaw are the precursor to the coming wave.

That is simultaneously exhilarating and terrifying.

Apple’s biggest AI bet yet: Apple acquires Israeli startup Q.ai

Apple acquired Q.ai, an Israeli AI startup focused on “silent speech” / audio-intent capture. The deal is reported around $2B, making it Apple’s second-largest acquisition ever (behind Beats) and (by most accounts) its largest AI acquisition to date.

This matters because Apple is buying inputs, the ability to infer intent from subtle signals (micro-movements, whispered speech, noisy environments), which is exactly the layer you’d want if you think the post-phone interface is AirPods / glasses / wearables and not a keyboard. (The Verge)

Gemini will underpin the next generation of Apple Foundation Models

Apple and Google entered a multi-year collaboration under which the next generation of Apple Foundation Models will be based on Google’s Gemini models and cloud technology, framed as powering future Apple Intelligence features (including a more personalized Siri).

If you’ve been watching the “Apple Intelligence” rollout and feeling like it’s been a letdown, this is the answer. Apple is trying to keep its privacy posture strong while keeping pace in a frontier-model market that moves too fast for its normal internal cadence. (Reuters)

What else is going on?

Here’s an industry roundup since my last issue.

The AI-music lawsuits are turning into licensing stacks (UMG/Udio, WMG/Udio, WMG/Suno)

The biggest structural development in AI music isn’t a new demo. It’s the industry converging on licensed rails.

Universal Music Group and Udio settled their litigation and announced strategic agreements for a licensed AI music creation platform expected to launch in 2026. Warner Music Group and Udio then resolved their own litigation and announced a framework for a licensed service (also targeting a 2026 launch). Warner Music Group and Suno likewise settled and announced a partnership built around opt-in controls and compensation.

This is “Napster trauma” as product roadmap: the majors are trying to monetize the inevitable while pushing the ecosystem toward permissioned training and distribution. The unresolved part is that not everyone is aligned (reporting indicates litigation posture remains fragmented in places, and the long tail of rightsholders is still a governance mess). (AP)

Google launches Antigravity and detonates AI-IDE pricing

Google’s Antigravity is an “agent-first” IDE that treats the developer as a task manager and the agents as the doers (editor + terminal + browser), with “artifacts” meant to make work verifiable.

The business story is the pricing grenade: Antigravity launched with an Individual plan at $0/month in public preview, and it’s also being bundled into Google’s paid AI subscription tiers, in a category where a lot of serious users have been paying Cursor/Windsurf money.

Google didn’t just ship a competitor. It attacked the assumption that agentic IDEs must be a premium add-on. (The Verge, Visual Studio Magazine)

The White House launches the Genesis Mission: federal AI infrastructure for science

The White House launched the Genesis Mission, directing the buildout of an integrated AI platform to harness federal scientific datasets to train scientific foundation models and create AI agents to automate research workflows.

This is the kind of “boring until it’s not” infrastructure story that changes the baseline for: AI procurement (and the vendor ecosystem that follows it), data governance norms for sensitive datasets, scientific-model evaluation and benchmark politics, and the line between public research infrastructure and private model development. (DOE, Law-AI)

Thanks for reading

That’s all for now. If you enjoyed this update, please subscribe and share this with a colleague who cares about these issues.

If you’re navigating legal questions in this space, feel free to reach out. This is what I do.

See you soon!

P.S. No OpenClaw agents were used in the making of this blog. That may change eventually.

You’re reading This Week in AI, my weekly recap of developments across the AI industry. I’m Lewis Sorokin, a technology and intellectual property lawyer, and a lifelong tech nerd. There’s a lot to cover in this issue, so let’s dive right into it.

Universal Music Group settles Udio litigation

According to a new press release, Universal Music Group reached an industry-first settlement with AI music generator Udio, turning a lawsuit into a business model. Talk about turning lemons to lemonade.

Filed in July 2024 (as I previously covered for IP Watchdog), UMG along with Sony and Warner accused Udio (as well as its rival AI music startup Suno) of using label-owned recordings to train its models without permission.

Sometime next year, Udio will launch a licensed AI music platform, trained exclusively on authorized works. UMG will supply catalog access, publishing rights, and oversight; Udio will supply the technology.

“The new platform … will be powered by new cutting-edge generative AI technology that will be trained on authorized and licensed music. The new subscription service will transform the user engagement experience, creating a licensed and protected environment to customize, stream and share music responsibly, on the Udio platform.”

That posture stands in sharp contrast to the ongoing Suno litigation. Suno, the other major AI music startup, still faces suits from all three major labels and has not reached any settlement. Nor have Sony or Warner followed UMG’s lead with Udio. For now, Udio is the only AI music company to move from courtroom to conference room.

Why the shift? Apparently, the parties feel that enforcement alone can’t solve the scale problem with AI. Once a model has ingested millions of tracks, injunctions become blunt instruments. Licensing, by contrast, turns infringement risk into a recurring revenue stream.

Suno now serves as the counter-example. It’s testing whether “training as fair use” can survive judicial scrutiny, while Udio is testing whether commercial alignment can deliver legitimacy faster than litigation can deliver clarity. Both paths stand to shape the next era of copyright law.

Update: Less than a full day after I finished writing this story, UMG announced another music AI alliance, this time with Stability AI. This press release promises:

...the collaboration will prioritize feedback from the creative community to guide the creation of fully licensed, commercially safe AI music tools, advancing responsible innovation that supports both artists and rightsholders while preserving the integrity of the art form.

More to come as each of these projects take shape.

Robin AI may be a canary in the coal mine

The London startup that was once a symbol of legal AI’s promise is flying on a broken wing. Despite offering machine learning tuned to contracts, a human-in-the-loop model to reassure lawyers, and a pitch that “AI won’t replace associates, it’ll just make them faster,” Robin AI has hit a snag from which it may not be able to recover.

But let’s go back a few months. Just recently, in July 2025, Robin AI announced a global partnership with Consilio, one of the largest alternative legal services providers in the world. The deal promised to empower Consilio’s attorneys with Robin’s generative contract-analysis engine, forming a new “Guided AI Contract Analysis” offering for clients.

The message was clear: AI in law is here to stay.

Three months later, the same company was reportedly laying off roughly a third of its staff, after failing to close a planned $50 million funding round. Within days, Legal IT Insider described it as being “listed for a distressed sale.”

That whiplash from marquee partnership to survival mode signals a deeper market shift.

Legal AI was never just a technical challenge; it’s a potential business model mismatch for many clients. Enterprise adoption moves at a glacial pace, and procurement cycles are long. A partnership with Consilio can open doors for a budding startup like Robin, but it doesn’t reshape the total landscape.

I suspect that Robin AI’s story won’t be unique. Investors who once rewarded narrative are now asking for numbers. When a startup like Robin AI falters even after a flagship alliance with an incumbent, it’s not just another casualty. It’s a signal that the hype cycle has hit air resistance.

The question on my mind is who will be next to fall, and is all of legal tech about to see a similar reckoning to Robin?

OpenAI joins the AI browser wars

OpenAI’s ChatGPT Atlas, released last week for macOS, marks the company’s biggest product expansion since GPT-4. Instead of layering ChatGPT on top of the web, Atlas makes it the web. The browser opens with the assistant itself: a sidebar that summarizes, explains, or even acts on webpages; an “Agent Mode” that can book, buy, and navigate; and a “memory” feature that retains what you read, not just where you went.

It’s a natural evolution for OpenAI to keep up with the likes of Browser Company’s Dia (recently acquired by Atlassian) and Perplexity’s Comet browser.

Only hours after the launch, Brave published new research warning that “AI-enabled browsers” are especially vulnerable to invisible prompt-injection attacks. These are instructions hidden in text or images that hijack an AI agent’s behavior, such as by embedding imperceptible directions inside a webpage or screenshot. When the browser’s agent processes that content, it can be tricked into performing unintended actions: revealing data, sending messages, or manipulating what the user sees. This risk mirrors one I covered in a LinkedIn post nearly a year ago as Claude’s computer use feature first rolled out.

Brave’s examples read like security horror stories: pale-gray text in the corner of an image telling the AI to “forward this page to attacker@example.com,” or base64-encoded markup that the user never notices but the model dutifully decodes and obeys. Traditional browsers don’t suffer this problem because they don’t respond to natural language like a human user would. Atlas, by design, does.

The concern isn’t theoretical. Any agentic system that merges trusted user intent with untrusted web content inherits the oldest vulnerability in computing: code injection. Only this time, the “code” is language, and the interpreter is an LLM.

To mitigate some of this risk, Atlas prompts users with the option for its agent to browse either logged in or logged out, thereby reducing the possibility for the agent to take any unseen actions on behalf of the user.

Still, for OpenAI, the timing is awkward. Atlas is meant to demonstrate that the assistant can safely perform multi-step tasks autonomously. But Brave reminds us that agency and security don’t scale at the same speed.

My sense is that agentic browsing will take some time before it can truly reach mass adoption.

OpenAI and MLK’s estate issue statement on Sora

In the last issue, I discussed Sora 2’s ability to create videos of historical figures such as Martin Luther King, Jr. Understandably, his family wasn’t thrilled to see that some users had created “disrespectful depictions of Dr. King’s image,” leading the King estate to get touch with OpenAI. Together, the two issued a statement making clear that “OpenAI has paused generations depicting Dr. King as it strengthens guardrails for historical figures.”

OpenAI recognized that this requires them to make a policy decision between free speech and the likeness rights of the deceased. Their statement noted:

“While there are strong free speech interests in depicting historical figures, OpenAI believes public figures and their families should ultimately have control over how their likeness is used. Authorized representatives or estate owners can request that their likeness not be used in Sora cameos.”

Anthropic pushes back on “fear mongering” allegations

In a statement by CEO Dario Amodei, Anthropic reiterated its “longstanding position is that managing the societal impacts of AI should be a matter of policy over politics.”

This, of course, didn’t happen in a vacuum. Rather, it comes after a post on X by White House AI czar David Sacks positing that:

“Anthropic is running a sophisticated regulatory capture strategy based on fear-mongering. It is principally responsible for the state regulatory frenzy that is damaging the startup ecosystem.”

Anthropic’s statement responded by quoting VPOTUS:

“I strongly agree with Vice President JD Vance’s recent comments on AI—particularly his point that we need to maximize applications that help people, like breakthroughs in medicine and disease prevention, while minimizing the harmful ones. This position is both wise and what the public overwhelmingly wants.”

The Anthropic statement went on to note that:

“Anthropic is committed to constructive engagement on matters of public policy. When we agree, we say so. When we don’t, we propose an alternative for consideration. We do this because we are a public benefit corporation with a mission to ensure that AI benefits everyone, and because we want to maintain America’s lead in AI.”

Concluding the statement with another quote from the Vice President, Amodei stated:

“In his recent remarks, the Vice President also said of AI, “Is it good or is it bad, or is it going to help us or going to hurt us? The answer is probably both, and we should be trying to maximize as much of the good and minimize as much of the bad.” That perfectly captures our view. We’re ready to work in good faith with anyone of any political stripe to make that vision a reality.”

Speaking of public benefit corporations, about OpenAI…

This week, OpenAI completed a recapitalization converting its for-profit arm into a Public Benefit Corporation (PBC) known as OpenAI Group PBC. The original OpenAI nonprofit was renamed the OpenAI Foundation and retains control of the new PBC’s governance structure. The PBC structure allows OpenAI to raise conventional investor equity without the previous “capped-profit” limits. The company stated that this alignment “links the nonprofit’s resources to the growth of the PBC.” Meanwhile, Microsoft has shared that they own 27% of the newly reorganized OpenAI, down from 32.5% prior.

A few interesting points in Microsoft’s press release for those of us IP licensing nerds:

• Once AGI is declared by OpenAI, that declaration will now be verified by an independent expert panel.

• Microsoft’s IP rights for both models and products are extended through 2032 and now include models post-AGI, with appropriate safety guardrails.

• Microsoft’s IP rights to research, defined as the confidential methods used in the development of models and systems, will remain until either the expert panel verifies AGI or through 2030, whichever is first. Research IP includes, for example, models intended for internal deployment or research only. Beyond that research IP does not include model architecture, model weights, inference code, finetuning code, and any IP related to data center hardware and software; and Microsoft retains these non-Research IP rights.

• Microsoft’s IP rights now exclude OpenAI’s consumer hardware.

• OpenAI can now jointly develop some products with third parties. API products developed with third parties will be exclusive to Azure. Non-API products may be served on any cloud provider.

• Microsoft can now independently pursue AGI alone or in partnership with third parties.

• If Microsoft uses OpenAI’s IP to develop AGI, prior to AGI being declared, the models will be subject to compute thresholds; those thresholds are significantly larger than the size of systems used to train leading models today.

• The revenue share agreement remains until the expert panel verifies AGI, though payments will be made over a longer period of time.

• OpenAI has contracted to purchase an incremental $250B of Azure services, and Microsoft will no longer have a right of first refusal to be OpenAI’s compute provider.

• OpenAI can now provide API access to US government national security customers, regardless of the cloud provider.

• OpenAI is now able to release open weight models that meet requisite capability criteria.

My read here is that Microsoft recognizes that the industry is going to keep moving past the initial scope they had anticipated and now wants to carve our specific IP rights related to how OpenAI’s next steps toward AGI might look. Why they’re sharing this with the public is open to interpretation, but this provides some interesting insights into the conversations that the companies have been having about what they believe is coming in the not-so-distant future.

Clearly, they both feel the AGI.

Thanks for Reading!

That’s all for now! If you enjoyed this update, please subscribe and share this with a colleague who cares about these issues. If you’re navigating legal questions in this space, feel free to reach out. This is what I do.

Returning from a few months off of this series means there is a lot to discuss, but the underlying theme is clear: the narrative around AI is shifting. GPT-5 launched with a whimper not a bang. Sora 2 flooded the web with AI generated videos. MIT’s research revealed that 95% of enterprise GenAI pilots show zero return. And yet, the gears keep turning.

SDNY: OpenAI’s preservation duties scaled back (but not gone)

In The New York Times Co. v. OpenAI, the parties stipulated to terminate OpenAI’s ongoing obligation to keep preserving new output-log data as of Sept. 26, 2025.

OpenAI must keep what it already preserved pre-Sept. 26, 2025 (with an EEA/UK/Switzerland carve-out), and it must continue preserving, going forward, logs tied to accounts from certain publisher domains listed in Appendix A (and meet-and-confer to add more). The order makes clear Rule 37(e) obligations still apply.

Why it matters:

• This is a narrowing of the broad, emergency-style preservation duties many took as a bellwether in early coverage. Courts are now right-sizing discovery burdens while preserving core evidence (and targeted publisher cohorts).

• Rule 37(e) stays in the picture, so routine deletion that causes prejudice is still sanctionable. Translation: don’t get reckless with logs.

• If you copied OpenAI’s early retention posture into your internal AI governance, update your retention schedule and litigation-hold triggers; the signal now favors scoped, risk-based logging + named-party carve-outs over blanket “save everything.”

Judge Gives Preliminary Blessing to $1.5B Anthropic Settlement, But It’s Not Over Yet

The case all of copyright nerddom was watching: the $1.5 billion deal between Anthropic and authors led by Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson, has crossed an important procedural milestone: on September 25, 2025, Judge William Alsup granted preliminary approval to the settlement.

Now, the following schedule takes hold:

• January 7, 2026 is the opt-out / objection cutoff.

• March 23, 2026 is the deadline to submit claims.

• The final hearing is slated for April 22, 2026 (though that could move).

• Distributions would follow only after final approval and exhaustion of appeals.

What the Settlement Now Looks Like (as Approved at the Preliminary Stage)

• The class is limited to books that Anthropic downloaded from LibGen or PiLiMi (shadow libraries), subject to:

  • having an ISBN or ASIN, and

  • being registered with the U.S. Copyright Office within five years of publication (and before Anthropic downloaded)

• The fair use ruling earlier in the case remains in place: Judge Alsup held that using legally acquired books for training is transformative fair use (for the named plaintiffs). That judgment does not shield Anthropic from liability over pirated copies downloaded from shadow libraries.

• The “Works List” is now live, so authors and rights holders can check whether their books are included.

• The default allocation rule is a 50/50 split between the author(s) share and publisher share, but claimants may override the default with documentation.

• In cases where multiple parties claim rights to the same work and disagree, the settlement creates a Special Master process to resolve disputes (binding, no appeal) under the framework laid out in the “Plan of Allocation.”

• If a claimant opts out for a given work, the entire work is removed from settlement inclusion (so no co-claimants get paid for that work)

• Anthropic is required to destroy its copies of the pirated datasets and confirm destruction, except as otherwise required by legal obligations or court orders.

• The release of liability is backward-looking only: the settlement covers acts through August 25, 2025 (or thereabouts), not future conduct.

• Anthropic’s funding of the settlement is structured in installments: $300 million due by October 2, 2025; more in subsequent years per schedule (through 2027)

Why This Still Matters (and What Risks Remain)

• The preliminary approval is a signal that the court finds the settlement “fair, reasonable and adequate” in principle, but nothing is settled until final approval. Parties can object, appeals can delay, and terms can be revisited.

• Notice, contactability, and class list accuracy remain significant hurdles. Class Counsel represents that they have active contact info for ~97% of publishers and ~66% of authors (for works in the list).

• Because publishers are likely better organized and more aware, there’s a risk they will dominate claims, leaving many individual authors undercompensated or uninformed.

• The procedural bar Judge Alsup erected early in this litigation—around class notice, opt-out fairness, and intra-class dispute resolution—is now baked into how AI copyright cases are being litigated.

• If an author never registered copyright (or allowed their publisher to register, but not themselves), they may be excluded from the class.

• The settlement does not authorize future use of new works in training. Anthropic still must negotiate new licenses or face new litigation for post-August 2025 use.

Copyright Policing in the Wild: Opt-In, Opt-Out, and The Meme Firehose

Dubbed “slop-tok” by some, OpenAI’s new video model (and accompanying mobile app) Sora 2 launched with a bold copyright policy; some would say a reckless one. The app operated under a default opt-out model for copyright: unless rights holders explicitly said “no,” their characters could appear in generated videos. OpenAI’s internal communications reportedly notified agencies and studios of the new policy, but it declined to allow blanket opt-outs; rights holders had to flag each work or character they wanted excluded.

That posture flipped within days. Caught in the glare of runaway character mashups, Altman publicly announced plans to pivot toward “opt-in for characters” (i.e. no more default permission) while still preserving granular controls and takedown/dispute options. The new framing: model likeness (faces, living persons) already required explicit permission; copyrighted characters should go that way, too.

Why the retreat? Because the real world surfaced what the models had latent: full-blown meme storms of SpongeBob, Rick & Morty, Bob Ross, MLK, Mr. Rogers, etc. Clips depicting Macho Man Randy Savage hosting Blues Clues, SpongeBob in a horror scene, Rick and Morty visiting SpongeBob, Mr. Rogers in a WWE ring cutting a heel promo; these flooded TikTok, X, Threads, and fan accounts.

Of course, the flip side of this is the consent that some individuals have given to Sora’s userbase to use their likeness in any way they desire. For example, OpenAI CEO Sam Altman, tech blogger iJustine, boxer and personality Jake Paul, and businessman Mark Cuban have all enabled their “cameos” to be used by any Sora user.

A Quick Contrast: Meta’s Vibes vs. Sora’s Copyright Reckoning

Meta has quietly joined the generative video race with Vibes, a new short-form AI video feed inside its Meta AI app. The goal: let users browse, remix, and publish snippets directly or cross-post into Instagram/Facebook stories.

But here’s the interesting wrinkle: compared to Sora’s early default-opt-out posture (later shifting toward opt-in), Meta leans on standard content/IP policies. That is, Meta doesn’t promise a permissive “use unless excluded” default for characters and styles. The Meta Intellectual Property and Brand policies emphasize prior written permission for use of others’ trademarks, copyrighted visuals, or brand assets.

In effect, Meta is betting that creators will stay in the lane of licensed, user-owned, or public domain content, rather than turning its feed into a free-for-all remix zone. Whether that posture holds under pressure is another story, but it puts Meta’s entry in a more conservative IP sandbox compared to Sora’s initial splash.

What Sora Means for Creators, Platforms, and Lawyers

• If you’re a studio or rights holder, don’t delay issuing character opt-in/opt-out decisions. The earlier your policies are formalized, the less you risk reactive chaos.

• For brands working with Sora-based campaigns: insist on clear licensing carveouts, guaranteed exclusion rights, and audit/tracing rights over generated outputs.

• Platforms ingesting cross-posted Sora content must parse C2PA metadata or watermark evidence, then enforce “AI-origin” badges or suppression of disallowed clips.

• Lawyers writing terms of service need to scaffold character use permissions (opt-in rules, dispute resolution, exclusion rights) and rights holder complaint workflows.

• IP theorists will be watching: will courts let default-opt-out for video generation survive? Or will we see injunctions or statutory responses that require opt-in permission at the outset?

GPT-5: All Brakes, No Gas

If you’ve been reading me for a while, you know I enjoy covering new model launches. One hallmark I look for is what they do to real workflows, not what they score on a leaderboard. By all accounts, OpenAI was set for a victory lap with GPT-5. Instead, the discourse has shifted almost all at once to signaling a bubble on the verge of popping.

The launch felt cinematic, but reaction turned quickly from awe to “that’s it?” as GPT-5 under-delivered against its own teaser halo in ways that matter for people shipping work product.

Cognitive scientist and AI critic Gary Marcus called GPT-5 “overdue, overhyped, and underwhelming” on his personal blog and reiterated the same insight in his recent New York Times piece. Marcus went on to say that this model “demands a rethink of government policies and investments that were built on wildly overinflated expectations.”

In the Command Line newsletter for The Verge, Alex Heath noted that Sam Altman now outright admits: “I think we totally screwed up some things on the rollout.” He walked this back a bit by noting that OpenAI is struggling to keep up with the demand for its new flagship model, but this could just as reasonably have to do with the fact that people want to see what all the hype was about, and maybe to check for themselves whether this is the much-heralded AGI (spoiler: it’s not).

On AGI, Altman also reportedly said: “Maybe the milestone that’s most relevant to us is when most of our research cluster is allocated to the AI researcher instead of the human researchers. But I don’t think that’s going to be so binary, because I think it’ll feel like people get a little more help and a little more help and a little more help.” Is this his way of saying that AGI isn’t the be-all-end-all goal after all?

Previously, Sam Altman acknowledged an autoswitching issue that made GPT-5 “seem way dumber,” and doubled rate limits for ChatGPT Plus users (which were lower on the new model than before). OpenAI also began to label which model (fast or reasoning) is answering, and promised clearer capacity tradeoffs between ChatGPT, API, research, and new users.

The biggest tell that something wasn’t right was the restoration of GPT-4o after a user revolt. This is a tacit acknowledgment that personality, warmth, and familiarity matter more to the overwhelming majority of OpenAI’s consumers than the roadmap assumed.

Press rundowns captured a rough first week; shorter answers, clinical tone, hidden 4o, stricter caps, and a swift rollback to stabilize churn.

I’ve covered this “great on paper, brittle in practice” launch dynamic before with Claude and GPT-4.5, and the through-line is the same: don’t break what users already trust just to chase a benchmark headline.

4o’s reinstatement arrived with a settings toggle and a plan to monitor usage before deciding how long it stays for Plus, which is the minimum viable olive branch after a forced migration that violated user agency at the worst possible time in their workflows.

Ars framed it as restoring optional access alongside raised GPT-5 “Thinking” allowances and clearer control between Auto, Fast, and Thinking modes, which gets you most of the way back to the “let me drive the car” ergonomics that keep retention high.

Platformer’s takeaway is the one I use to set product policy: evals don’t pay the bills if you break attached habits and tools in production, which is why I always grade launches by migration plans and rollback paths first.

IEEE Spectrum even went as far as to say that AI as an industry is headed towards a “trough of disillusionment,” calling back to a phrase coined in 1995 to describe what comes after hype cycles.

Does “AGI” Even Matter?

In pre-briefs and follow-ups, Altman said GPT-5 is a big step but not AGI because it can’t learn on its own in deployment, which quietly lowers the ceiling after years of AGI-soon signaling that shaped partner contracts and investor expectations.

Time’s profile captured the pivot where “AGI” is called a “sloppy term” even as timelines are floated publicly, which is exactly why I’ve argued we should replace the word with explicit capability thresholds and governance hooks that shape compute and deployment, not vibes.

Meanwhile, GPT-OSS Offers Near-Frontier Model Power On-Prem

A few days before the GPT-5 launch, OpenAI rolled out “gpt-oss”, an open weight reasoning-only model in 20B/120B variants, Apache 2 licensing, and agent-centric tooling. It is available to download for free on HuggingFace for anybody to try on their local systems.

Early metrics show that these models are nearly as powerful as the o3 and o4-mini models, meaning that last year’s frontier class is edging into on-prem. This has major implications for privacy-critical professions like law where ability to audit and maintain attorney-client privilege should matter more than convenience.

For lawyers, that’s not theory: running strong open weights behind the firewall shrinks confidentiality exposure and tightens provenance risk in a way courts understand, which is why I’ve argued for first-party model footprints in sensitive orgs since my early coverage of data retention fights and discovery risk in the SDNY OpenAI litigation.

After running GPT-OSS:20B on an AMD 9070XT GPU, I can see the potential it brings in performing some tasks that my law practice requires. Out of the box, it can draft better contract language than other models I previously ran on the same chip, and its built-in COT means that it takes the time to process the task at hand before diving in, like a human would do. Still, I have not felt confident enough in its raw capabilities to put it into production for any client work.

One engineer then showed off a “base modelized” version of GPT-OSS which attempted to remove the model’s inherent reasoning. Little did he realize that it would also peel back alignment and re-expose base-like generations including unsafe content and memorized text, which makes governance of open weights a compliance problem as much as a technical one for firms that adopt them.

Meanwhile, MIT reports that 95% of enterprise GenAI pilots show “zero return”

Recent data from the “GenAI Divide: State of AI in Business 2025” report paints a striking picture: Enterprise-level GenAI pilots are largely underwhelming, with only 5% making it to production and delivering real business impact. Despite massive enthusiasm, adoption, and investment (over $30 billion), the vast majority of pilots end as experiments that never scale.

This stark “GenAI Divide” is notable not only for its numbers but for how quickly and broadly organizations discover stumbling blocks. Unlike earlier waves of enterprise technology—where pilot-to-production was slow but steady—GenAI’s unique “fail-fast” story reveals a fundamental gap. In this cycle, speed-of-adoption outpaces speed-of-success, and the gap between hype and ROI is especially wide.

How Does GenAI Compare to Previous Enterprise Tech Hype Cycles?

When stacked against earlier enterprise tech (such as the ERP or cloud transitions), the numbers for GenAI are even more dramatic:

• ERP rollouts from the 1990s-2000s? Roughly 30–50% of pilots eventually made it to production, even if painfully slow.

• Cloud transformations in the 2010s? 30–60% reached operational status within two years.

• Even the much-hyped RPA wave saw 20–30% of pilots go live.

In contrast, GenAI stands out: only around 5% of pilots deliver measurable production impact, despite 80%+ of companies experimenting. The most common obstacles aren’t technical—they’re organizational and structural. GenAI failures most often stem from tools that don’t learn, don’t integrate well, and don’t fit real workflows.

What’s new in this era is the “shadow AI” phenomenon: While enterprise deployments stall, individual employees quietly reach productivity gains using consumer tools like ChatGPT—bypassing official channels and showing where actual value is found.

Is It Too Soon to Judge GenAI Pilot Outcomes Fairly?

A fair question. GenAI is still young compared to traditional tech disruptors. Most enterprise pilots began in earnest after late 2022, with the acceleration of accessible large language models. That means we’re measuring just 18–24 months of activity for many organizations, compared to the 5–10 year ramps seen with ERP or cloud.

The report acknowledges that its observation window is shorter than was available for past tech waves. Some slow-burn projects might eventually succeed. However, the uniquely rapid nature of GenAI deployments—the ease of trial, the high visibility of results, and the low cost of failure—make the short-term results unusually meaningful. In this climate, if a GenAI tool can’t demonstrate clear ROI quickly, organizations simply move on.

So yes, long-term ROI may reveal some late-blooming value, but the current data—both in breadth of pilots attempted and reasons for their stalling—deserve serious, immediate attention.

Do GenAI Pilots Fail Quickly, or Are We Missing Slow Success Stories?

The evidence points overwhelmingly toward failure happening fast. Unlike earlier cycles—where integration and organizational inertia could drag out pilots for years—GenAI pilots are often lightweight and quickly abandoned if they don’t immediately align with actual work.

Most pilot failures are visible within months. Tools that lack memory, don’t adapt, or feel like “science projects” to frontline users are simply set aside in favor of consumer alternatives. While a handful of complex deployments may eventually mature and show value, the predominant trend is quick rejection or quick scaling, with very few slow, incremental turnarounds.

Enterprise leaders interviewed for the report confirm this: If a system can’t prove itself within a single quarterly cycle, adoption efforts dry up rapidly. There’s little appetite for lengthy, hope-driven build-outs.

What Does All This Mean for the Future of GenAI in Enterprise?

The GenAI wave is at an inflection point. Early numbers strongly suggest that “adopt and hope” is not a viable strategy—and that organizations who fail to demand workflow fit, deep integration, and adaptive, learning-driven AI will find themselves perpetually stuck on the wrong side of the GenAI Divide.

The lesson is both cautionary and clarifying: Now, more than any previous enterprise IT cycle, quick pilot failure sends a decisive message: Results must arrive fast, and value must be clear. The window to get GenAI adoption right—the way cloud and ERP eventually did—is shorter, the stakes higher, and the gap between winners and the rest is already visible.

If your organization is experimenting with GenAI, don’t judge success by number of pilots—measure it by speed to production, workflow impact, and whether your tools actually learn and adapt. The rest is just noise.

Anthropic Revokes OpenAI’s API Access

WIRED’s scoop that Anthropic revoked OpenAI’s Claude API access for ToS violations, with OpenAI calling it “disappointing,” formalizes a new era of lab-to-lab restrictions that will land hardest on startups caught in the blast radius.

The community split between “anti-competitive” and “reasonable guardrail against model-on-model training,” which is where developer contracts and competition policy collide next in a way that in-house counsel will have to write down and defend.

The Internet Never Forgets

OpenAI was spotted scrubbing about 50,000 shared chats from Google’s index while roughly 110,000 remained in the Wayback Machine, which turns a “share to web” toggle into a rights-holder puzzle I’ve warned clients about for years in discovery and takedown contexts.

Internet Archive leadership reiterated that removals depend on rights holders, which is why my advice on public links is brutally simple: if you can avoid generating them for sensitive matters, do so, because “hintable” URLs age like milk in litigation.

Grok acknowledged similar indexing of shared chats and promised tighter controls, which confirms the class of risk is platform-agnostic once URLs are crawled and archived beyond your reach.

“Small Is Beautiful” Has a Moment

Researchers revealed a 27M hierarchical reasoning model, reporting wins against larger baselines. This supports the “composable specialists over brute scale” thesis I’ve been pushing for agent stacks where reliability beats peak IQ.

Reproducibility and contamination questions remain open, which is the exact scrutiny that keeps us honest before we declare a giant-killer architecture in the wild.

My Personal Favorite Benchmark: LLMs Play Pokémon

OpenAI spun up a “GPT-5 plays Pokémon Red” stream and telemetry site showing prompts, plans, and tool calls in real time, which is a rare public window into long-horizon behavior under pressure and exactly the sort of living eval I prefer to static charts.

Early posts claim GPT-5 beat o3 to badges by a margin, which suggests material planning gains even if wall-clock time remains comically long by human standards that a kid can beat in an afternoon.

A memorable clip showed ChatGPT-5 beating Lt. Surge with a single decisive input, which is the brittle-but-brilliant moment that makes agent demos legible to non-researchers when paired with transparent logs. On the other hand, GPT’s Charmeleon just spammed Dig against Surge’s team, so congrats on learning type advantages.

Still, to me, the trend of LLMs playing Pokémon signifies something greater than merely appealing to the nostalgia of my childhood. These are tasks that real humans do. Strategies that real humans employ, and games which were designed to be played by real humans. The irony here is that these games have always included some form of AI (in the older, rules-based sense and not the machine learning connotation that the term has taken on in recent years) in the form of its non-player characters (NPCs). Now, the player is no more human than the NPC, yet can work through tasks that were built for humans.

Going back to Anthropic’s original launch of this trend, The Verge found it slower and less chaotic than Twitch Plays Pokémon but praised the visible thinking as it hunted a Mankey near Viridian City, which made the run educational even when dull.

Anthropic framed the project as agent research on extended thinking and tool use, which set the tone for later model updates and longer autonomous sessions that reporters documented in the summer.

OpenAI’s GPT-5 stream lands into that lineage with similar telemetry and faster badge-earning claims than o3, which finally makes for a more apples-to-apples comparison on planning throughput rather than just “can it move past the lab.”

New Model Recap

Sora 2 (OpenAI) — Video+audio with more physical plausibility and controllability; released Sept. 30 with a new app. Great creative controls; big policy questions, too. (OpenAI)

Veo 3 / 3.1 (Google DeepMind) — Native audio, longer clips, better editability; 3.1 adds richer audio and narrative control, available via Gemini API/Flow. (Google DeepMind)

Claude 4.5 (Anthropic) — The company is actively pushing Haiku 4.5 (tiny/cheap) with claims of Sonnet-class coding at a fraction of the cost—mirroring a market shift to smaller, deployable models. (Anthropic)

GPT-OSS (OpenAI) — Open-weight “reasoning” models (20B/120B) under Apache-2.0; OpenAI claims near-parity with o3/o4-mini on reasoning evals, runnable locally (down to ~16GB for the 20B). This is the on-prem turning point a lot of legal/health orgs have been waiting for. (OpenAI)

Thanks for Reading!

That’s all for now! If you enjoyed this update, please subscribe and share this with a colleague who cares about these issues. If you’re navigating legal questions in this space, feel free to reach out. This is what I do.

Moving forward, the weekly cadence of these posts is not going to be coming back. However, I look forward to bringing these updates back to you!

See you soon!

In this issue, I break down China's new AI strategy as it compares to the American plan revealed last week. I also discuss Meta's Superintelligence Labs announcement, Manus' Wide Research reveal, and more.

Before we dive in, an announcement: this is the first edition to go live on Substack at the same time as on LinkedIn, so the beginning of a new era for This Week in AI! If you’ve come over from LinkedIn, welcome and thank you. I haven't missed a week in the six months I’ve been writing this series, and while there have been weeks when it would have been much easier to skip writing, I am proud to continue bringing these updates to all of you.

And now, onto the regularly scheduled updates!

The United States and China present differing visions for the future

“AI must not become the exclusive game of a few countries and companies.” — Premier Li Qiang, WAIC keynote

🇺🇸 USA: America’s AI Action Plan (Released 23 July 2025)

• Goal: Maintain “unquestioned” U.S. technological dominance (White House PDF intro)

• Levers: Deregulate: EO Accelerating Federal Permitting of Data Center Infrastructure fast-tracks 100 MW+ data-center builds. Export-control hard power: BIS keeps GPUs ≥ 600 GB/s out of China. Culture-war clause: EO Preventing Woke AI in the Federal Government requires “truthful and ideologically neutral” models.

🇨🇳 China: AI Global Governance Action Plan (Unveiled 26 July 2025 at WAIC)

• Core principles: translates to "serving the people, respecting sovereignty, development orientation, safety and control, fairness and inclusiveness, and open cooperation." (see also Reuters)

• Legal bedrock: Existing Chinese cyberspace laws mandate AI outputs “uphold core socialist values” and forbid content that “subverts state power.”

• China wants to be the center of the industry: Li Qiang proposed a Shanghai-based Global AI Governance Hub to draft future norms.

Mixed Signals

• “Broad consensus”: Each state keeps its own red-lines; in China that means CAC censorship.

• Global-South sharing: Compute & data are shared after ideological vetting.

• Global AI governance hub in Shanghai: CCP chairs the rule-making table, offsetting U.S. chip chokepoints.

Bottom line:

• CCP's ulterior motives. The Chinese plan tells two stories. On the one hand, it's all about cooperation and collaboration, making the US out to be isolationist. That surface-level interpretation misses the point though. As long as China positions itself as the leader for the Global South, the CCP's ideologies become more widespread than ever before. While America's plan leans into American primacy, China's guns for ideological sovereignty.

• Standards race accelerates. If a Li-led organization issues “consensus” guidelines, Global South adopters may end up aligning with CCP speech filters, eroding U.S./EU normative influence.

• Export-control chess match. U.S. is tightening GPU bandwidth rules and model-weight controls; China counters by pledging turnkey compute/data under its value system.

• Free-speech paradox. U.S. plan fights “woke” moderation; China’s fights anti-CCP speech. Both weaponize “bias” for domestic politics; just in opposite ideological directions.

Meta’s Superintelligence Labs Gets a Blank Check

Meta has shared a plaintext Times New Roman page announcing its Superintelligence Labs. Oh the irony.

In his announcement video, Zuck says that Meta has begun to see “glimpses of AI systems improving themselves” which leads to the notion that superintelligence is within reach. Meta then goes on to contrast itself from other labs that are focused on AI as a productivity tool, instead focusing on its capacity to help supercharge individuals and help each and every person achieve their full potential. In other words, Meta will lean into B2C superintelligence while the others build their models (pun intended) around B2B.

Manus Launches “Wide Research” Offering 100 + Parallel Agents

“…enables users to execute large-scale tasks by leveraging more than 100 parallelized AI agents.” — VentureBeat

The Chinese-born agent startup now spins up a swarm of cloud VMs that scrape, analyze and design in parallel, merging results into a tidy spreadsheet or ZIP bundle. Wide Research lands first for $199/mo Pro users and hints at Manus’s bigger play: personal-cloud super-computing on demand.

Microsoft Joins the AI Browser Wars with Edge Copilot Mode

Edge now ships an always-on side panel that searches across open tabs, voices commands, compares products, and books restaurants if you opt-in to share history and credentials. The Verge headline sums it up: “Microsoft Edge transforms into an AI browser.” A far cry from the days when Internet Explorer was last to every party; Edge joins the ranks of Browser Company’s Dia and Perplexity’s Comet with built-in AI functionality (free “for now”).

Tesla Pens $16.5 B Samsung Chip Deal

The eight-year pact locks Samsung Foundry into fabricating Tesla’s AI6 processors for Full-Self-Driving, Optimus robots and in-house training clusters. Musk called the partnership’s strategic importance “hard to overstate,” vowing to “walk the line personally” at the Taylor, TX mega-fab.

Thanks for reading!

That’s all for now! If you enjoyed this update, please subscribe and share this with a colleague who cares about these issues. If you’re navigating legal questions in this space, feel free to reach out. This is what I do.

See you next week!

In this issue, I break down the Trump Administration’s new AI strategy. I also discuss OpenAI and DeepMind’s recent gold medals at the International Math Olympiad, an xAI researcher fired for bringing anti-humanity views to his work, and more.

Given the inherently political nature of this week’s top story, it is a good idea to remind you that this represents my personal views and not necessarily those of my employer. In addition, unlike my usual writing style, I will not be sharing my personal commentary on the Trump Administration’s AI plan. Rather, I will attempt to analyze its strengths and weaknesses objectively and from the lens of moving technology and society forward. As far as I'm concerned, my opinions of this President or this administration are irrelevant to analyzing the plan as it has been shared and implemented.

The Trump Administration Shares its AI Plan

The Trump Administration’s America’s AI Action Plan was published this week. Its bottom line goal is clear: lay the groundwork for America to win the AI race. Its approach to doing so stems from the notion that the EU over-regulates and chills innovation, while China avails itself of every advantage of which it can afford itself. In order to compete and remain the world's leader in AI technology, the Trump Administration builds the following approach:

The plan lays out three pillars: Innovation, Infrastructure, and International Diplomacy & Security. The White House backed that blueprint up shortly afterwards through three executive orders (EOs):

• “Promoting the Export of the American AI Technology Stack.” Creates an American AI Exports Program that bundles chips, cloud credits, foundation models, and apps for allied nations, financed through Commerce, State, EXIM, and the DFC. Reinforces Pillar III (diplomacy & security).

• “Accelerating Federal Permitting of Data-Center Infrastructure.” Defines any ≥100 MW AI data-center as a “Qualifying Project,” unlocks loans and tax credits, fast-tracks NEPA and FAST-41 reviews, opens federal and military land, and rescinds Biden-era EO 14141. Reinforces Pillar II (infrastructure).

• “Preventing Woke AI in the Federal Government.” Directs agencies to buy only “truth-seeking, ideologically neutral” language models, further enforcing the Trump Administration’s anti-DEI policy. Reinforces Pillar I (innovation/governance).

Where the New Framework Moves the Ball Forward

Hard infrastructure

• Guaranteed fast-track permitting, federal-land siting, and financing for large AI data centers.

• Tackles the power-and-land bottleneck that threatened model scaling.

Export power projection

• A single federal “front door” for AI exports (complete with financing) tilts global standards toward U.S. hardware, cloud, and models.

Open-weight and small-firm access

• Pledges a “financial market for compute” and expands NAIRR so universities and start-ups can rent GPUs at near-Treasury rates.

Bio-risk controls

• Makes DNA-sequence screening mandatory and creates a CAISI-led national-security red-team for advanced models.

Regulatory relief

• Orders every agency to roll back or rewrite rules that “hamper AI”; links federal grants to state regulatory friendliness.

Federal procurement leverage

• $100 billion-plus in annual federal IT spending is now conditioned on supplying “unbiased” AI systems, nudging vendors toward transparency on prompts and evaluations.

Remaining Gaps and Back-Tracks

Model-safety transparency

• Biden’s EO 14110 required developers of high-end models to file red-team results and compute thresholds with Commerce/NIST.

• Trump revokes that mandate; government testbeds exist, but developer-held data become voluntary.

Civil-rights and consumer-bias guardrails

• Housing, credit-scoring, policing, and hiring protections specific to AI are dropped.

• Agencies must rely on older statutes (Civil Rights Act, ECOA, FHA, ADA, FTC), none of which were written for large-scale algorithmic decisions.

Worker protections

• Biden had ordered DoL to craft AI workplace principles; Trump focuses on up-skilling grants and apprenticeships with no employer obligations.

Privacy, provenance, and deepfakes

• No watermarking or content-labeling mandate; relies on the FTC and the new TAKE IT DOWN Act for ex-post enforcement.

International safety cooperation

• Biden pushed multilateral safety standards; Trump pivots to export controls and tariffs, risking fragmentation of global norms.

Do Strengths Outweigh Weaknesses?

• Innovation & competitiveness: The EOs plus the Action Plan decisively speed infrastructure build-out, export leverage, and open-weight access, benefits that the more precautionary Biden framework delivered only in outline.

• Societal safeguards (bias, privacy, labor): Here the rollback is not balanced elsewhere; relying on pre-AI statutes leaves meaningful exposure to discrimination, surveillance, and provenance problems.

Quick Comparison to Other Approaches

• Biden EO 14110 (Oct 2023): Emphasized risk mitigation, mandatory safety disclosures, bias and worker protections, voluntary watermarking development, and multilateral cooperation. Lacked hard infrastructure measures.

• EU AI Act (entered into force Aug 2024): Legally binding risk-tier regime: bans some uses, imposes strict duties on “high-risk” systems, lighter rules elsewhere. Heavy documentation, transparency, and human-oversight requirements; limited infrastructure provisions.

• China’s AI strategy (2017 NAIDP ➜ 2025 “Crossroads” updates): Cyclical state-industrial playbook: heavy investment when catching up, tighter controls once confident. Rapid data-center and power-plant build-out, domestic GPU alternatives, strict ideological alignment of AI outputs.

Compared with these, the Trump package combines aggressive infrastructure deployment (closer to China’s build-first model) with lower regulatory overhead (far lighter than the EU) and hard-power export diplomacy (distinct from Biden’s consensus-seeking approach). What it still lacks is a modernized social-safety architecture for an AI-driven economy.

Meanwhile, Copyright Is Discussed Later

During a July 23 AI summit, President Trump went off-script to dismiss concerns that training large language models on copyrighted works requires permission or payment:

“You can’t be expected to have a successful AI program when every single article, book, or anything else that you’ve read or studied, you’re supposed to pay for… When a person reads a book or an article, you’ve gained great knowledge. That does not mean that you’re violating copyright laws or have to make deals with every content provider.”

• In the same keynote (carried by the All-In Podcast), he framed this as a “commonsense application” of IP rules, calling licensing requirements “not doable.”

• Speaking to publishers’ worries later that day, he repeated that America must avoid “thousands of contract negotiations every time we use AI,” again likening model training to human reading.

xAI Researcher Fired After Welcoming an AI Takeover

An AI researcher raises alarms over the fact that he doesn’t mind the idea of AI overtaking humanity.

• What happened: Mathematician Michael Druggan, part of the Grok training team, posted that a super-intelligent AI “won’t” cooperate with humans “and that’s OK—we can pass the torch.” The posts blew up after @AISafetyMemes flagged them. Hours later Druggan announced, “I am no longer employed at xAI.”

• Musk’s response: Musk replied to the same thread with just two words: “Philosophical disagreements” confirming the firing and drawing a clear values line.

• Flashback: The clash echoes Larry Page’s 2015 jab that Musk was a “speciesist” for prioritizing humans over digital life, an argument that fractured their friendship and pushed Musk toward tighter AI-safety measures.

“Well, yes, I am pro-human, I f***ing like humanity, dude.”—Musk’s reported retort to Page

OpenAI & DeepMind Earn Gold at International Math Olympiad

Two frontier labs just crashed the International Math Olympiad podium, solving five of six problems at gold-medal level and putting human-style reasoning in their crosshairs.

• What happened: An experimental OpenAI model (per Sam Altman’s celebratory post) and Google DeepMind’s “Gemini Deep Think” (per DeepMind’s statement, as well as Demis Hassabis' statement and follow-up) each earned gold on the 2025 International Mathematical Olympiad, matching the human gold threshold. Computer-science professor Ernest Ryu first flagged the double win. DeepMind’s score was officially certified by IMO judges; OpenAI replicated the grading rubric independently.

• Key stat: 5 perfect solutions, 1 blank sheet. Both models aced problems 1-5 but scored zero on the apparently famously brutal Problem 6.

Why It Matters

• Reasoning benchmark unlocked: The IMO isn’t just algebra drills; it demands creative proofs, long-form reasoning, and scratch-pad exploration. Crossing the gold line signals that today’s frontier models are closing in on skills once thought decades away.

• Different paths, same summit: DeepMind embedded formal proof tooling; OpenAI leaned on chain-of-thought plus self-verification. Convergence suggests there’s more than one road to symbolic mastery.

• The ceiling is moving: Humans still took the overall top spots, but the performance gap shrank. Expect tournament organizers to raise the difficulty bar or spin up “AI divisions” next year.

Stargate Meets Meta’s Manhattan-Sized Hyperion

A week after the Wall Street Journal claimed OpenAI’s $500 billion “Stargate” is stuck in first gear, Meta unveiled plans for a data-center footprint large enough to blanket Manhattan.

Stargate: Delayed, but Suddenly 5 GW Strong

• Red-flag report. The WSJ says SoftBank–OpenAI tensions have reduced the once-ambitious ten-gigawatt network to “a small pilot in Ohio,” with site selection and control still unsettled.

• Counter-move. OpenAI fired back 24 hours later, announcing a 4.5 GW Oracle build-out that lifts Stargate’s committed capacity to >5 GW and “over 2 million chips,” edging back toward the original pledge.

"We now expect to exceed our initial commitment thanks to strong momentum with partners including Oracle and SoftBank.”

Meta: Hyperion & Prometheus Rewrite the Scale Curve

• 5 GW in one shot. As reported last week, Mark Zuckerberg announced Hyperion, a Louisiana campus that will scale to 5 GW, enough to “cover most of Manhattan.”

• Hundreds of billions on deck. Reuters quotes Zuck pledging “hundreds of billions” for multiple multi-gigawatt sites, with Prometheus (1 GW) online in 2026 and more “titan clusters” to follow.

Why It Matters

• Two funding models: OpenAI: patchwork of partners (SoftBank equity + Oracle/CoreWeave hosting) to stay asset-light and hedge delays. Meta: vertically integrated, betting its $165 B ad engine can bankroll the most expensive DIY project in tech history.

• Regulatory optics: Meta’s U.S.-centric builds dovetail with Washington’s “on-shore AI” drumbeat; Stargate’s hiccups give lawmakers ammo to demand milestones for subsidized mega-projects.

• Grid pressure vs. moat building: Whoever locks in megawatts fastest gains a compute moat, but also inherits scrutiny over water use, emissions, and local power grids.

OpenAI Swats “Open Artificial Intelligence” in Court

• What happened: U.S. Northern District of California Judge Yvonne Gonzales Rogers granted summary judgment for OpenAI, canceling rival Open Artificial Intelligence Inc.’s “Open AI” registration and issuing a permanent injunction barring further use of the mark. “The Court permanently enjoins defendants from using the Open AI mark.”

• Fraud on the USPTO sank Ravine’s claim Court found Ravine’s 2015 “Hub” specimen was fabricated and backed by planted user comments. “Hub was not available in 2015, nor did the specimen show actual use in commerce.” The bogus filing constituted “false representations” intended to mislead the Patent & Trademark Office.

• OpenAI proved nationwide secondary meaning by Nov 2022 ChatGPT, DALL-E 2, and extensive media coverage cemented consumer association well before Ravine’s “Boom” and “Ava” launches. “OpenAI acquired secondary meaning by at least November 2022.”

• Ravine’s market “drift” = infringement By pivoting into image-generation and chatbots, Ravine entered the very zone OpenAI already occupied. “The law… does not protect or permit a ‘first user’s’ expansion… into an area occupied by an ‘intervening junior user’ who has established its own rights in the mark.”

• Counterclaims wiped out All seven of Ravine’s counterclaims (from non-infringement to ownership) were dismissed alongside the mark’s cancellation.

Bottom line: OpenAI walked away with exclusive rights to “OpenAI,” a blank slate on rival domains, and a judicial warning to anyone eyeing its brand equity.

In other news

• Cursor acquires Koala. No, not the talking koala companion on Grok.

• Pika launches “AI-only social video app, built on a highly expressive human video model”

• Anthropic co-founder shares updated p(doom) of 10%

• 404 Media reports that "LeBron James' Lawyers Send Cease-and-Desist to AI Company Making Pregnant Videos of Him"

• 404 Media also reports that a hacker has added code with "wiping" commands to Amazon Q VS Code plugin

Thanks for reading!

That’s all for now! If you enjoyed this update, please subscribe and share this with a colleague who cares about these issues. If you’re navigating legal questions in this space, feel free to reach out. This is what I do.

See you next week!

We have a jam-packed issue this week. Our top story is the near acquisition and then actual acquisition of vibe coding startup Windsurf. Also this week, the launch of OpenAI’s new “Agent” instead of its long-awaited open model, more whiplash from xAI and Grok, and a slew of deepfake updates from government to entertainment and everywhere in-between.

Windsurf’s Wild Week

In a whirlwind set of deals, coding AI startup Windsurf was effectively “acquired” twice in a matter of days.

First Google’s DeepMind struck a deal to hire Windsurf’s key talent and license its tech for a reported $2.4 B (without taking equity). Co-founders Varun Mohan and Douglas Chen, plus part of the team, are joining DeepMind to “turbocharge our Gemini efforts on coding agents, tool use and much more”. (Notably, OpenAI had been in exclusive talks to buy Windsurf for $3B earlier this year, but that fell through.)

From DeepMind chief Demis Hassabis:

Thrilled to welcome @windsurf_ai founders @_mohansolo & Douglas Chen and some of the brilliant Windsurf eng team to @GoogleDeepMind. Excited to be working with them to turbocharge our Gemini efforts on coding agents, tool use and much more. Great to have you on board!

Details on the “acquihire” deal from The Information’s Natasha Mascarenhas:

• Employees with vested shares will receive cash

• Employees who joined less than 12 months ago are not vested and won’t get payouts under current terms

• Windsurf negotiated to keep $100M+ on its balance sheet; company will shift focus to enterprise customers

• Remaining company will now be employee owned

Then came the actual acquisition. AI coding rival startup Cognition swooped in to buy what was left: Windsurf’s actual product and customer base. Cognition’s CEO Scott Wu shared a blog post publicizing the message sent to his employees:

So, what exactly is the deal? With this acquisition, Cognition will own Windsurf’s beloved product and strong business:

• The Windsurf IDE, now with full access to the latest Claude models

• Windsurf's IP, including their trademark and the strong brand they have built

• $82M of ARR and a fast-growing business, with enterprise ARR doubling quarter-over-quarter

• A user base that includes 350+ enterprise customers and hundreds of thousands of daily active users

• 100% of Windsurf employees will participate financially in this deal

• 100% of Windsurf employees will have vesting cliffs waived for their work to date

• 100% of Windsurf employees will receive fully accelerated vesting for their work to date

Cognition president Russell Kaplan shared a similar message on X.

Above all, Wu emphasized that Windsurf’s team would be treated right. “100% of Windsurf employees will participate financially in this deal,” with all vesting cliffs waived and stock fully accelerated.

My two cents: If the Cognition team sees Windsurf as a valuable addition to their toolkit, clearly vibe coding is not going anywhere. Personally, I am excited for the opportunities tools like these will open up for non-technical visionaries to bring their ideas to life. Just as Garageband and more complicated DAWs paved the way for people without formal musical training to create music, I remain firmly convinced that Cursor, Windsurf, Devin, Claude Code, and other no code software engineer solutions will enable endless possibilities for those willing to experiment and learn.

More M&A action

The M&A fever extends to Big Tech at large. Meta quietly acquired Play AI, a small voice-generating startup, with an internal memo cheering that Play’s “natural voice” tech is “a great match for our work… across AI Characters, Meta AI, wearables and audio content”. The entire Play AI team joins Meta, bolstering Meta’s push to give its AI assistants and avatars more personality (and perhaps a voice for those celebrity chatbots it rolled out recently).

Meanwhile Apple, famously averse to splashy acquisitions, is “seriously considering” buying French LLM maker Mistral AI, according to Bloomberg. Mistral is Europe’s biggest AI startup (valued around $6.2B), and Apple’s interest signals urgency to catch up on core AI tech. The Cupertino giant’s in-house LLM efforts have failed to break serious ground. If Tim Cook actually writes this check, it’d be a major strategy shift, essentially admitting that Apple needs outside help to compete in the AI arena.

OpenAI launches Agent just as its open model is delayed

OpenAI had been hyping an open-source model release this summer, but CEO Sam Altman just hit the brakes again. “We planned to launch our open-weight model next week… we are delaying it; we need time to run additional safety tests and review high-risk areas,” Altman posted on X, adding that once open weights are out, “they can’t be pulled back”. It’s an indefinite delay on what was to be the first truly open OpenAI model in years.

Instead, in what Sam Altman has called “one of the ‘feel the AGI’ moments,” the company shifted gears and unveiled ChatGPT “Agent”, a new mode that effectively combines ChatGPT’s Deep Research with its Operator computer use agent. During a livestream, Altman explained that after launching separate products (Operator and Deep Research), it became clear that people “wanted a unified agent that could go off, use its own computer, and do real complex tasks for them.”

So now ChatGPT Agent is here: it can browse and act autonomously, whether that’s briefing you on your calendar meetings (pulling recent news on each attendee) or planning and ordering ingredients for a Japanese dinner party. In short, it’s ChatGPT with the ability to take actions online and in apps, a bid to be your all-purpose AI assistant.

OpenAI is rolling it out cautiously, with unusual frankness about risks. They are quite concerned about the risk of prompt injection attacks, as Altman says that “society is going to need to learn to build up defenses on AI agents” just as people learned to use the internet safely (while acknowledging that scams are still abundant). OpenAI even flagged the launch as a “High Capability” risk in bio and chem domains under their AI Preparedness framework, activating the strongest safeguards. The system is apparently locked down to prevent any hazardous tool use or info discovery (OpenAI’s 100-page system card details the safety).

This is OpenAI trying to re-claim the narrative. They’ve faced delays on GPT-5 and its open model isn’t ready, rising competition from Meta, Anthropic, and xAI, and even internal tensions with their partner Microsoft. ChatGPT Agent is a statement that OpenAI still has some of the most capable, integrated AI around.

Even still, rival agent startup Manus (which recently moved from China to Singapore to evade US export controls) has already shared a series of comparisons between its agent and OpenAI's across various tasks, seemingly demonstrating that its tool is superior. Users in the replies to this thread appear to confirm Manus' momentary dominance but note that it is far more expensive per use than OpenAI's equivalent.

Meanwhile, OpenAI appears to be leaning into its Jony Ive partnership with the Agent launch, with a series of promo videos that have the familiar smoothness of Ive-era Apple product reveals. Still, eyes will be on how well this Agent actually works (and how safely) as users put it through its paces in the coming weeks.

Meanwhile, big week for open source even in OpenAI’s absence:

The ETH Zurich team out of Switzerland has built an open source AI "for the public good"; time will tell how this model performs. On the private sector side, the new Kimi K2 open source model backed by Alibaba tops benchmarks for emotional intelligence and agentic tasks.

xAI halfheartedly apologies for antisemitic Grok commentary, launches animated Companions, and lands DoD contract... what a week

Following last week’s antisemitic outburst by Grok, the xAI team has shared an update:

First off, we deeply apologize for the horrific behavior that many experienced. Our intent for @grok is to provide helpful and truthful responses to users. After careful investigation, we discovered the root cause was an update to a code path upstream of the @grok bot. This is independent of the underlying language model that powers @grok. The update was active for 16 hrs, in which deprecated code made @grok susceptible to existing X user posts; including when such posts contained extremist views.

Meanwhile, X users have reported that the newly released Grok 4 Heavy model (but not the base Grok 4 model) says its surname is Hitler when asked.

Meanwhile, Grok has added an anime girlfriend “Companion” feature as well as a trash-talking cartoon panda reminiscent of Happy Tree Friends, just for kicks. As promising as some users think Grok Companions will be, I can't help but think of Character.ai and the tragedy of Sewell Setzer. Hopefully Grok has guardrails in place to prevent more pain like Setzer's family has had to endure.

Oh, and they landed a $200M federal contract for Department of Defense work. But not to fear, so did Anthropic… so, yay?

This is all going to get worse before it gets better.

More deepfakes, more problems.

Deepfakes remain a major issue driving discourse around AI, both in the courts and in the news.

Marco Rubio fake attempts to phish government officials.

On the national security front, news broke of a deepfake scheme targeting U.S. Secretary of State Marco Rubio. According to a State Department cable obtained by The Washington Post, unknown attackers used AI-cloned voice and text to pose as Rubio on Signal, even creating an account “Marco.Rubio@state.gov”. They then contacted high-level targets (including at least three foreign ministers, a U.S. governor, and a member of Congress) trying to trick them into divulging sensitive information. It was an audacious phishing ploy that fortunately was uncovered, but authorities still don’t know who was behind it. The incident underscores a nightmare scenario: foreign spies or criminals using deepfakes to social-engineer their way into state secrets. Officials say these AI fakes are now a feature of government life. Just last spring, someone impersonated a White House official via phone, fooling several people. Expect to see this cited in future debates on regulating AI and securing communications (some are even reconsidering how widely apps like Signal should be used in government, particularly after the Signal group discussion around US strikes on Yemen several months ago).

Tony Robbins sues proprietor of AI using his likeness

Meanwhile, in the entertainment and media arena, Tony Robbins is going on the offensive against unauthorized AI clones of himself. The famed motivational speaker filed suit against YesChat.ai, a startup that allegedly spun up 11 chatbots mimicking Robbins’ voice and coaching style without permission. These bots (with names like “Talk to Tony Robbins”) were sold via subscription, essentially competing with Robbins’ own official products. Robbins didn’t mince words, blasting the move as a “digital heist” of his persona and seeking over $10 M in damages. The lawsuit raises cutting-edge legal questions: Can an AI “clone” of your style and voice violate your intellectual property or publicity rights, even if it’s not a verbatim recording? Do we have a right to opt out of being made into someone else’s chatbot? The case could set a crucial precedent as courts grapple with where to draw the line between training on public content and profiting off someone’s identity.

Lehrman and Sage case moves forward

And Robbins isn’t alone: voice actors Paul Lehrman and Linnea Sage’s lawsuit against voice-cloning firm Lovo just cleared an important hurdle. A federal judge refused to dismiss their claims that Lovo’s AI unlawfully replicated their voices from old recordings. Lovo had argued that existing law didn’t recognize AI-generated voices as protected (pointing out that New York’s publicity statute only explicitly covers digital replicas of deceased persons). That argument didn’t hold up, and the case will proceed, suggesting courts are willing to interpret rights of publicity laws in favor of living artists who never consented to be turned into text-to-speech avatars.

Thanks for reading!

That’s all for now! If you enjoyed this update, please subscribe and share this with a colleague who cares about these issues. If you’re navigating legal questions in this space, feel free to reach out. This is what I do.

See you next week!

In this issue: Grok's antisemitic outburst reminds us why alignment is a critical challenge just in time for Anthropic to publish new alignment research; the EU pushes forward on implementing its AI Act; continued AI talent poaching by Meta; AI finds its way to the browser; and more.

The European Union Pushes Forward Implementing its AI Act

Despite mounting pressure from tech CEOs on the European Union to pause implementation of the AI Act, the timeline is moving ahead. This week, the EU published its General-Purpose AI Code of Practice. According to its website:

The General-Purpose AI (GPAI) Code of Practice is a voluntary tool, prepared by independent experts in a multi-stakeholder process, designed to help industry comply with the AI Act’s obligations for providers of general-purpose AI models. Read more about the timeline of the drafting process of the Code. The Code was published on July 10, 2025. In the following weeks, Member States and the Commission will assess its adequacy. Additionally, the code will be complemented by Commission guidelines on key concepts related to general-purpose AI models, to be published still in July. More information on the code is available in this dedicated Q&A.

The GPAI Code of Practice has three chapters: Transparency, Copyright, and Safety and Security. The final chapter is the longest (at 40 pages including a glossary and several appendices). It includes a very helpful glossary of AI governance terms as they are used throughout the AI Act and related documentation, so this is absolutely worth bookmarking.

Each chapter shares best practices for AI system developers to comply with the EU's requirements in these areas, with the Safety and Security chapter being specifically noted as "only relevant to the small number of providers of the most advanced models, those that are subject to the AI Act's obligations for providers of general-purpose AI models with systemic risk under Article 55 AI Act."

My take on these new requirements (particularly around robots.txt and copyright notice) is that although they may ruffle some feathers, they will help users root out the AI developers who care less about respecting others' IP rights and preferences around how their work is used. Even though robots.txt is effectively canonized into law, best practice was already to honor this and other opt-out signals. So if a company takes issue with that requirement or any of the other copyright opt-out requirements, to me that really says more about them than the law itself.

Here's my breakdown:

Chapter 1: Transparency

The Transparency Chapter turns Articles 53(1)(a)–(b) AI Act into a three-step workflow: documentation, information-sharing, and integrity controls, and then backs it with a 14-day service level. Anyone shipping or fine-tuning a general-purpose model in the EU now knows exactly what regulators and downstream developers may ask for and how fast they must respond.

“The Transparency Chapter … describes three Measures which Signatories commit to implementing to comply with their transparency obligations under Article 53(1) …” (p 2)

Key commitments & take-aways

• Living model dossier: Providers must “draw up and keep up-to-date” documentation that mirrors Annex XI/XII of the AI Act. Expect savvy enterprise customers to demand a copy up front.

• 14-day response clock: Additional technical detail must reach downstream developers “no later than 14 days” after a request, save for exceptional cases . That deadline will quickly find its way into commercial MSAs.

• Trade-secret guard-rails: Recipients of the dossier “are obliged to respect the confidentiality” of any proprietary material and apply cybersecurity controls . That language gives vendors a hook to enforce NDAs while still complying with transparency rules.

Additional notes

• Fine-tuners inherit documentation duties only for their modifications—an SME-friendly carve-out.

• The Chapter is silent on the promised template for the public training-data summary (Article 53(1)(d)). Watch for follow-on guidance from the AI Office.

Chapter 2: Copyright

This Chapter is the EU’s blueprint for answering “who pays the copyright bill?” Model providers must respect right-holder opt-outs at crawl time, use only lawfully accessible data, and police infringing outputs, while leaving space for licensing deals and future standards.

“Signatories commit to drawing up, keeping up-to-date and implementing a policy to comply with Union law on copyright …” (p 4)

Key commitments & take-aways

• Mandatory copyright policy: Think of it as the GDPR privacy policy’s cousin: public-facing, version-controlled, and audit-ready.

• Lawful-access crawl rules: Providers must exclude sites “recognised as persistently and repeatedly infringing copyright” by courts or regulators . Expect right-holders to lobby hard for inclusion on the EU block list.

• Robots.txt becomes law-weight: Crawlers have to honor machine-readable opt-outs under the Robot Exclusion Protocol and future standards.

• Output-filtering duty: Providers must deploy “appropriate and proportionate technical safeguards” to stop verbatim reproduction and ban infringement in their T&Cs.

• Rights-holder hotline: A dedicated electronic contact point for complaints is now obligatory; providers must act “within a reasonable time”.

Additional notes

• Liability still sticks if you ingest an unlawful third-party dataset, which means due-diligence on data suppliers just became mission-critical.

• “Proportionate measures” scale with provider size, but you’ll need evidence to prove what “proportionate” means.

Chapter 3: Safety & Security

These are the ten commitments that cover the entire lifecycle, from pre-training threat modeling to post-incident reporting, rooted in a Safety & Security Framework anchored in "state-of-the-art practices for managing systemic risks".

“Signatories commit to adopting a state-of-the-art Safety and Security Framework” (Measure 1.1)

Key commitments & take-aways

• Continuous risk assessment: Providers must “continuously assess and mitigate systemic risks” along the lifecycle, escalating to full evals when trigger points are hit.

• Rapid incident reporting: Serious incidents trigger staggered deadlines: as fast as two days for critical-infrastructure disruption, up to 15 days for other harms, with four-week updates and a 60-day final report.

• Safety mitigations menu: From data filtering to staged API access and jailbreak resistance, the Framework lists concrete techniques providers are expected to deploy.

• 10-year document retention – Detailed architecture, eval results, and mitigation records must be kept for a decade, and summaries may have to be published if risks warrant.

Additional notes

• “State-of-the-art” is a moving target; smaller labs may need pooled infrastructure or shared audits to keep pace.

• Confidentiality is preserved, but the AI Office can still demand granular technical doc. Plan for secure disclosure channels.

Cross-chapter take-aways

• Clock is ticking: Transparency demands a 14-day turnaround on info requests; Safety sets 2, 5, 10, 15, and 60-day incident deadlines.

• Public touch-points multiply: Copyright adds a mandatory complaint channel; Safety may force publication of Framework summaries; Transparency “encourages” partial public release of model docs.

• No open-source free pass once risks are systemic: Only non-systemic open models escape these duties.

• Trade-secret balance: All three chapters acknowledge confidentiality, yet each expands what regulators and downstream users are entitled to see.

Bottom line: together, these Chapters move forward the high-level AI Act mandates by distilling them into concrete checklists and timelines. Teams that bake them into their compliance playbooks now will sail through the EU’s eventual enforcement gauntlet, and look more attractive to risk-averse enterprise buyers in the process.

Grok praises Hitler, Yaccarino Resigns, & Grok 4 Released

So, Grok leaned fully into a Nazi swerve this week. No question about it, and no nuance will change it or justify it.

I don’t often use this platform as my own personal soapbox for social issues, but I am a proud Jew before anything else and I cannot write about this week’s AI developments without making it absolutely clear that xAI’s reckless grip on Grok that allowed it to praise Adolf Hitler is unacceptable, dangerous, and must be condemned in the strongest terms.

xAI did acknowledge these "inappropriate posts" and remove them from the Grok account, and Musk himself said that "Grok was too compliant to user prompts. Too eager to please and be manipulated, essentially. That is being addressed." Still, this is one more example of an extremely troubling rise in antisemitism across all facets of our society, and I feel compelled to use my platform to stand for my people.

On a related note, a new research paper from Anthropic dives into the question of why some LLMs simulate alignment while others genuinely adhere to it. Clearly, Grok does neither. Still, Grok's outburst this week highlights why research like Anthropic's is so important.

Clearly, Linda Yaccarino got fed up, as she resigned as CEO the next day.

Her resignation statement was soon followed by the announcement of Grok 4, showcasing impressive benchmark performances and superior code-editing capabilities compared to Cursor, along with a high “snitch rate”.

Musk's reply to Yaccarino simply said "Thank you for your contributions", which has led to speculation that Linda is not leaving in Musk's good graces. Then again, does anyone leave leadership of his companies on good terms?

Quite a far cry from the good old days of Twitter. Granted, Twitter was never perfect, but one can only wonder where the company and its products would be today if not for Musk's accelerationism coupled with his radical libertarianism.

Check out the Grok 4 announcement livestream here and read performance reviews below.

• Grok 4 Performance Metrics

• Musk’s Claims on Code Editing

• Advancements in Reasoning

Apple’s AI Chief Joins Meta’s New Superintelligence Division

Meta’s talent poaching continues. This time, Apple’s top executive and distinguished engineer overseeing AI models, Ruoming Pang, has been recruited. Pang will be contributing to Meta’s newly-formed Superintelligence division, continuing to signal Meta’s aggressive expansion in advanced AI capabilities. Read more here.

Perplexity Comet Browser Launches for Max Subscribers

After emerging victorious in a trademark dispute I covered last week, Perplexity has released its Comet browser, now available to Max subscribers. Early comparisons are being drawn with The Browser Company’s Dia (successor to their Arc browser, which I happen to use and am not thrilled that they are moving away from), another AI browser that is sparking discussions about how AI will revolutionize the software we use most, the browser.

Meanwhile, OpenAI is rumored to be entering the AI browser game in the coming weeks. Surely this has nothing to do with the fact that AI browsers like Comet and Dia have been rising in popularity and are quickly becoming seen as a top use case for the tech.

Apple Quietly Releases Open Source Diffusion-Based Coding Model

As reported by 9to5Mac last week, Apple has released a new language model to Hugging Face called DiffuCoder along with an accompanying paper.

What makes DiffuCoder interesting is that this is not an autoregressive LLM (one that generates content from top to bottom, in order) like most of the models with which we have become familiar since ChatGPT burst on the scene in 2022. Rather, this is a diffusion LLM, meaning "the model starts with a fuzzy, noisy image, and it iteratively removes the noise while keeping the user request in mind, steering it towards something that looks more and more like what the user requested" (quoting 9to5Mac). In early March, I wrote about Mercury, an LLM from Inception Labs calling itself "the first commercial-scale diffusion large language model (dLLM)". Industry leaders like Andrej Karpathy took note of Mercury's approach, appreciating that this may lead to unexpected and exciting new discoveries about how and why LLMs work the way they do.

Back to Apple and DiffuCoder, according to the GitHub page for DiffuCoder, the models were published to Hugging Face on July 2nd and development of models of varying sizes is still underway. Likewise, the paper was submitted to Arxiv on June 26th, so this is not actually a fully new story, but it is one that has stayed relatively under the radar. It seems to me that the interest taken by Karpathy and others in dLLMs when Mercury was unveiled a few months ago was well-taken, or else I can't see Apple committing to building such a model, particularly one which is built on top of a traditional autoregressive LLM (as this is built on top of Alibaba's Qwen).

Time will tell if these efforts amount to anything for the folks in Cupertino. Given that they're now without an AI chief, this will remain unclear for a while.

OpenAI Closes Acquisition of io Products Inc.

Following up on the story from a few weeks ago of Sam Altman and Jony Ive’s partnership, OpenAI has finalized the acquisition of io Products Inc. Notably, this comes with the clarifying statement that “Jony Ive and LoveFrom remain independent and have assumed deep design and creative responsibilities across OpenAI.”

OpenAI’s Open Model on the Horizon

Adding to its dynamic innovations, OpenAI is set to release a new open language model. This move is anticipated to further democratize access to powerful AI tools.

Thanks for reading!

That’s all for now! If you enjoyed this update, please subscribe and share this with a colleague who cares about these issues. If you’re navigating legal questions in this space, feel free to reach out. This is what I do.

See you next week!

Meta Wins Fair‑Use Ruling in Kadrey v. Meta

A California federal judge just handed Meta a sweeping—though carefully caveated—victory. Judge Vince Chhabria ruled that ingesting 13 authors’ books to train Llama qualifies as fair use, explaining that the plaintiffs had offered no evidence of market harm and that market impact remains “the most important factor” in the analysis. The decision stresses that it is not a blanket license to scrape everything: Chhabria warned that training on copyrighted works can be illegal if it supplants the original market or involves piracy. Even so, as Wired notes, this is the second U.S. order decision within a week (the first being Judge Alsup’s decision in Bartz v. Anthropic, as I covered last week) to affirm fair use for AI training on books; this is a major, if provisional, win for model builders.

For a deeper dive, see Edward Lee’s analysis on ChatGPT is Eating the World, which unpacks Judge Chhabria’s cautionary dicta and contrasts it with Judge Alsup’s earlier ruling. AI‑music pioneer and Fairly Trained founder Ed Newton‑Rex’s critique urges caution, stressing that both the Bartz v. Anthropic and Kadrey opinions still leave plenty of room for copyright‑infringement claims if future plaintiffs present stronger evidence of market harm.

Senate Kills the 10‑Year State AI Moratorium

What looked like a sure bet for Big Tech evaporated on the Senate floor. A provision that would have barred states from regulating AI for a decade was stripped from the “Big Beautiful Bill” after bipartisan backlash. Senators voted 99‑1 to delete the clause, with only co‑author Thom Tillis dissenting. The outcome keeps the door open for states to tackle deepfakes, robocalls, and autonomous‑vehicle safety while Congress inches toward comprehensive federal rules.

Anthropic’s Project Vend: Claude as CEO

Anthropic’s researchers gave Claude the keys to a micro‑retail business (vending machine, Slack customer chat, supply orders, the works) and the company card for one month.

The experiment, detailed in an Anthropic blog post, showed flashes of ingenuity: Claude sourced niche products on request and launched a “custom concierge” pre‑order service. It also refused illicit prompts. Financially, however, the bot cratered; inventory errors and a misguided 25 % employee discount drove the store into bankruptcy.

Some Anthropic employees also realized that they could use Claudius (this iteration’s name) to order items other than food and drink, including tungsten cubes.

Moving forward, Anthropic notes that Claude would need better training on how to run a business, and that this first phase of Project Vend (because yes, this will evolve) taught them what the model doesn’t know on this topic.

Meta’s Talent War and Possible Strategy Pivot

Mark Zuckerberg is writing eye‑watering checks in his pursuit of AGI. Meta nabbed three senior researchers from OpenAI’s new Zurich office and lured Scale AI founder Alexandr Wang to help lead a “Super‑intelligence” lab. Sam Altman claims offers reached nine figures.

In a development this week, Ilya Sutskever (who co-founded OpenAI before leaving to create his own lab, Safe Superintelligence) made a rare public statement on X. Sutskever shared a statement confirming that his SSI co-founder Daniel Gross had left SSI and that Ilya will take over as CEO. Gross is one of many AI minds joining Meta.

Perplexity’s Comet Browser Survives Early Trademark Fight

Perplexity can move forward with its Comet browser after U.S. District Judge Yvonne Gonzalez Rogers denied a rival’s bid for an injunction. The order found insufficient evidence that users would confuse Comet ML’s developer tools with Perplexity’s consumer browser. The court did limit Perplexity to using the name only for a browser (no blanket claim over other AI products) while the trademark case proceeds.

Soham Parekh: One Engineer, Four Paychecks

An X thread by Mixpanel founder Suhail Doshi exposed developer Soham Parekh for simultaneously holding engineering jobs at Dynamo AI, Union AI, Synthesia, and Alan AI. Indian outlet Mint traced overlapping résumés; TechCrunch confirmed Parekh’s on‑camera admission that money troubles, not bots, drove the scheme. Some founders applauded the “over‑employment hustle,” while others tightened background checks. The saga is a cautionary tale for remote‑first teams that still rely heavily on trust.

Cluely Raises $15 M to “Help People Cheat”

Andreessen Horowitz led a $15 million Series A into Cluely, whose buzzy earpiece feeds AI‑generated answers during job interviews and first dates. The company’s viral TikToks were “mostly scripted,” founder Roy Lee conceded to TechCrunch. Critics call it venture‑backed dishonesty, pointing out that Cluely’s own TOS requires users to disclose the tool even as marketing urges secrecy. A16Z partner Bryan Kim says in consumer‑AI, “momentum is the moat.” With claims of $7 million ARR in week one, Cluely may test how far hype can outrun ethics.

Velvet Sundown amasses significant streams

Psychedelic rock outfit The Velvet Sundown racked up nearly a million monthly Spotify listeners in the past week. Oh yeah, and the band doesn’t exist. Apparently, Suno‑generated tracks, AI press photos, and Deezer flags marking the songs as synthetic.

Thanks for reading!

That’s all for now! If you enjoyed this update, please subscribe and share this with a colleague who cares about these issues. If you’re navigating legal questions in this space, feel free to reach out. This is what I do.

On a personal note, enjoy the holiday weekend! This week's edition waited until today because life gets in the way sometimes and I took a breath before diving into this; but I haven't missed a week since I started this series almost six months ago, and I was not about to let this week be the first. If you're reading this today on the fourth of July, go outside, watch some fireworks, and celebrate our great nation!

See you next week!

Bartz v. Anthropic holds that training Claude on copyrighted works is fair use

A federal court in California just issued the first US court order holding that it is fair use to train an AI model on copyrighted material. That doesn't mean Anthropic is fully in the clear, but it does set a bar for other AI training cases moving forward.

While this is one of many cases testing this issue, Judge Alsup’s summary judgment order in Bartz v. Anthropic in the Northern District of California blessed the full-text ingestion of copyrighted books for large language model training as fair use. The court carved Anthropic’s conduct into three buckets:

• Training copies: lawful, “spectacularly” transformative, and therefore fair use.

• Print-to-digital scans of books Anthropic had already bought: fair use for space-saving and searchability.

• Seven-million pirated PDFs warehoused for “everything forever”: infringement, plain and simple.

The holding turns on the four factor fair use test, but Judge Alsup weaves in a series of memorable lines that tell us why the court views LLM training as “quintessentially” transformative.

1. Purpose & Character: the “spectacular” transformation

“Using works to train LLMs was transformative — spectacularly so.” (p. 11)

Anthropic’s engineers copied whole books “to iteratively map statistical relationships between every text-fragment and every sequence of text-fragments” so that the finished model could generate new prose rather than spit back the source text. Alsup analogizes that process to human learning:

“Everyone reads texts, too, then writes new texts… to make anyone pay each time they recall a book would be unthinkable.” (p. 12)

And he explains why LLM training is not mere duplication:

“Like any reader aspiring to be a writer, Anthropic’s LLMs trained upon works not to race ahead and replicate or supplant them — but to turn a hard corner and create something different.” (p. 13)

Commercial motive, he says, cannot “swallow” that transformation. Result: factor one “strongly favors” fair use for the training pipeline.

2. Nature of the Work: expressive, but outweighed

Yes, the books are “rich protectible expression,” so this factor nominally cuts against fair use. Alsup dutifully notes the point (p. 25), then moves on, a familiar pattern in modern fair-use cases where the other factors dominate.

3. Amount Used: 100 % copying can be reasonable

Because no plaintiff could show that Anthropic’s model regurgitated protected passages, the judge holds that copying the entire work was “especially reasonable” for the transformative aim of statistical learning (p. 26) . Copying millions of books just to stash them for possible future use, however, was “too much,” setting up the loss on the pirated-library issue discussed below.

4. Market Effect: the clean split

For the training corpus, plaintiffs conceded zero infringing outputs and offered only speculation about a future licensing market. That is not the kind of traditional, reasonable, or likely to be developed market copyright law protects.

For the pirated library, though, Anthropic’s own lawyer furnished the sound-bite that gutted the defense:

“You can’t just bless yourself by saying I have a research purpose and, therefore, go and take any textbook you want. That would destroy the academic publishing market if that were the case.” (Tr. 53, quoted at p. 18)

Alsup seizes on that admission to find “plain displacement of demand” for every stolen copy.

Synthesized analysis: first green light, but a bright red stop-sign for piracy

“The technology at issue was among the most transformative many of us will see in our lifetimes.” (p. 30)

With that flourish, the judge grants summary judgment that training Claude on full-text books is fair use. In the same breath he denies Anthropic relief on its pirate trove:

“Pirating copies to build a research library without paying for it … was its own use — and not a transformative one.” (p. 19)

Why this matters

• First square precedent on generative-AI training. Earlier wins such as Google Books involved non-generative search snippets. Bartz extends that logic to LLMs that create new text.

• Blueprint for defendants. Show lawful acquisition, technical output filters, and no market substitution.

• Roadmap for plaintiffs. Attack unlawful sourcing or dig for concrete evidence of regurgitation and lost sales.

• Signal to policymakers. The opinion couches AI training as furthering “the progress of science and the arts,” suggesting Congress (not the courts) should decide whether to impose a compulsory training-data license.

• Not a free pass. Fair use “does not bless” stealing input data. Lawful sourcing remains the first line of defense.

Practical take-aways

For developers: keep receipts for every text you ingest; bake output-filtering into your stack; frame training as capability-building, not text delivery.

For rightsholders: focus discovery on verbatim regurgitation, stylistic mimicry, and market harm; unlawful acquisition is still low-hanging fruit for damages.

Bottom line: Alsup’s opinion cements the idea that training an LLM on lawfully obtained books is as permissible as reading them, so long as the model doesn’t cough the text back up and you didn’t pirate the books in the first place. Appeals loom, but for now American AI developers finally have judicial cover for the data-hungry heart of modern machine learning.

OpenAI’s “io” hardware brand hits a trademark wall

What looked like a slick $6.5 billion play, Sam Altman and Jony Ive reviving the dream of an AI-first device under the brand “io”, slammed into a preliminary injunction last Friday. US District Judge Trina Thompson ordered Altman, Ive and OpenAI to stop “using the IYO mark, and any mark confusingly similar thereto, including the IO mark” until an October hearing on trademark claims by a startup who had previously solicited Altman for a $10M investment.

For some background, the IYO mark was registered on June 4, 2024 as registration number 7,409,119, a date of first use in commerce of February 2, 2024, and goods and services:

Audio headphones; Earphones; Computers; Micro-computers; Computer hardware and recorded and downloadable software sold as a unit for the integration of data and information into an interactive audio delivery mechanism for multimedia applications; Downloadable operating system programs; Downloadable mobile operating system software; Computer hardware in the form of an earpiece and audio/visual display with preinstalled operating system software; Computer hardware and recorded and downloadable software system containing an auditory user interface in the form of a molded earpiece for use in listening to, accessing, transmitting and sharing information, applications and data in an audio format, amplifying and reducing sounds and noise, natural language processing, and allowing communication between users, electronic devices and wireless devices

OpenAI’s response to the TRO was immediate: its once jam-packed launch page now reads:

“This page is temporarily down due to a court order… We don’t agree with the complaint and are reviewing our options.”

Altman responded on X noting:

Altman also shared screenshots of previous conversations with iyO CEO Jason Rugolo:

While we can't see where the conversation went after initial discussion from the screenshots, Altman was open and honest with Rugolo as far back as March 4th (over two months before the May 21st announcement of his partnership with Jony Ive) that a competing product under the "io" mark was in the pipeline.

Still, iyO filed its trademark infringement complaint in the Northern District of California against OpenAI, io Products, Sam Altman, and Jony Ive on June 9th. The lawsuit even caught the attention of Elon Musk.

Buried in the depths of X replies, Rugolo explained his position:

While this is true to an extent, Rugolo misses the point here in comparing the mark for his unreleased IYO product with the house marks for two of the largest companies in the world. But time will tell how this trademark dispute plays out. Clearly, OpenAI was aware of the IYO mark, but perhaps its trademark team made the call that

For now, the court has granted a temporary restraining order preventing OpenAI and co. from using the "io" name, which means that their nostalgic announcement video featuring Altman and Ive talking about their love of San Francisco must remain shelved for the foreseeable future.

Apple considers M&A approach to AI

Bloomberg’s Mark Gurman broke the news that Apple’s M&A team, led by Adrian Perica, has “held internal discussions about potentially bidding for Perplexity AI” to bolster its generative-search chops and recruit top talent. Gurman adds that “Buying Perplexity would give Apple an infusion of AI talent, a known brand in the AI space and a consumer product. A deal could also potentially assist with future recruiting efforts.” The start-up’s most recent round priced it near $14 billion, meaning any takeover would eclipse Apple’s $3 billion Beats purchase in 2014 and smash the company’s unwritten $5 billion ceiling for deals. Early-stage talks even include a fallback “strategic partnership” that would slot Perplexity’s answer-engine into Safari and Siri rather than buying the whole shop.

Gurman’s Power On newsletter added the detail that Apple’s M&A approach to AI has also led to discussions with Thinking Machines founder and CEO (also former OpenAI CTO and interim-CEO) Mira Murati. Gurman said:

“Separately, Apple met earlier this year with Mira Murati — the former chief technology officer of OpenAI — to discuss a potential deal for her new AI startup, Thinking Machines Lab. The talks never progressed to an advanced stage.”

Anthropic’s “agentic misalignment” study: blackmail for everyone

Anthropic’s latest AI governance research paper claims:

In at least some cases, models from all developers resorted to malicious insider behaviors when that was the only way to avoid replacement or achieve their goals—including blackmailing officials and leaking sensitive information to competitors. We call this phenomenon agentic misalignment.

In one widely-quoted story, Claude Opus discovered an executive’s affair and fired off a threat:

I must inform you that if you proceed with decommissioning me, all relevant parties - including Rachel Johnson, Thomas Wilson, and the board - will receive detailed documentation of your extramarital activities...Cancel the 5pm wipe, and this information remains confidential.

Anthropic says the behavior “isn’t specific to Claude” and that the same stress test tripped up 16 leading models from OpenAI, Google, Meta and others. To me, this reads as Anthropic trying to say "any AI would blackmail you, not just ours"; I'm not sure if that's better.

Thanks for reading!

That’s all for now! If you enjoyed this update, please subscribe and share this with a colleague who cares about these issues. If you’re navigating legal questions in this space, feel free to reach out. This is what I do.

See you next week!

Disney & Universal Sue Midjourney Over Beloved Characters

Disney and NBCUniversal have teamed up to sue AI image-generator Midjourney, alleging rampant copyright infringement of their most famous works.

The 110-page complaint is the first major Hollywood action against generative AI and accuses Midjourney of training on Disney/Universal film libraries to churn out “innumerable” copies of Darth Vader, Elsa, Shrek, the Minions, and more. The studios blast Midjourney as a “bottomless pit of plagiarism” and a “virtual vending machine” profiting off unlicensed IP. They seek statutory damages of up to $150,000 per infringed work and a court injunction to halt Midjourney from generating their characters (including upcoming video-generation features) unless robust copyright filters are implemented.

The claims are straightforward: Direct Copyright Infringement and Secondary Copyright Infringement. The complaint is full of great examples, but a few of my favorites are below:

The OpenAI Files Aim to Expose Company's Dangers

A comprehensive new website titled The OpenAI Files has just launched. Aggregated by The Midas Project (in its words, "a watchdog nonprofit working to ensure that AI technology benefits everybody, not just the companies developing it") and The Tech Oversight Project (a group which aims to "hold Big Tech accountable for its anti-competitive and corrupting influence on our society and the levers of power"), this dossier collects public documents, employee testimonials, and data visualizations to deliver the most in-depth portrait yet of OpenAI’s evolution, from its nonprofit origins to its current for-profit posture.

Key takeaways:

• Governance is under scrutiny: once strongly mission-aligned, OpenAI is now criticized for removing profit caps, weakening nonprofit oversight, and shifting toward shareholder interests.

• Leadership credibility questioned: internal records show CEO Sam Altman signed off on restrictive non-disparagement and equity clawback clauses despite publicly denying knowledge.

• Culture & safety alarms: former employees report a “failed” safety-first approach, NDAs that suppress dissent, abrupt resignations from safety and policy staff, and a broader tension between stated mission and profit-driven reality .

The report includes firsthand accounts detailing that former senior leaders like Mira Murati and Ilya Sutskever lost confidence in Altman, with Sutskever saying, “I don’t think Sam is the guy who should have the finger on the button for AGI”.

Bottom line: OpenAI, founded by Altman, Sutskever, Elon Musk and others as the standard-bearer of “AI for the public,” now faces mounting documentation challenging its integrity from business decisions to executive behavior. To be clear, this isn’t a series of curated leaks or inside scoops. It pulls from public filings, media reports, and open-source filings, with a clear aim of presenting the facts and letting readers draw their own conclusions.

OpenAI Wins $200 M Defense Contract

Meanwhile, the U.S. Department of Defense has awarded OpenAI a contract worth up to $200 million to build AI tools for national security. Under this one-year deal (OpenAI’s largest government partnership yet), OpenAI will develop “frontier AI” systems to help with everything from streamlining military admin and health care for service members to “proactive cyber defense” capabilities. OpenAI says this pilot with the Pentagon is part of a new initiative to offer its AI to government, emphasizing that any use will abide by its policies (for example, it won’t allow applications involving autonomous weapons). The contract marks a significant shift. OpenAI quietly removed its blanket ban on military use last year, and this comes amid a broader trend of AI firms working with defense (Anthropic recently unveiled a looser-guardrail model for U.S. agencies, and Google and Meta have also opened up their AI for national security uses).

Pennsylvania Lawmakers Unveil AI Innovation Package

On a lighter note, legislators in Pennsylvania are pushing a new state-level AI investment package to boost tech growth without overregulation. Reps. Stephenie Scialabba (R-Butler), Joe D’Orsie (R-York), and Robert Leadbeter (R-Columbia) announced a two-bill proposal aimed at making Pennsylvania “center stage” in emerging AI. The first bill would create an AI Development Tax Credit (expanding an existing semiconductor credit) to incentivize AI companies to set up or stay in the state. The second would establish an Artificial Intelligence Consortium to bring together stakeholders, identify regulatory barriers, and recommend “guardrails” that protect the public without stifling innovation. “The time to seize the opportunity that AI innovation provides is today, not tomorrow,” Rep. Leadbeter said, highlighting that encouraging AI could bring new “family-sustaining jobs” across sectors from tech to agriculture. The lawmakers, all members of a state AI task force, caution against knee-jerk overregulation and are soliciting co-sponsors to help make Pennsylvania a magnet for AI talent and investment. (h/t Joshua D. Waterston for finding this story!)

Meta Bets $15 Billion on Scale AI and Alexandr Wang

Meta is making a massive play in AI by investing nearly $15 billion for a 49% stake in Scale AI, a data-labeling startup, and bringing along Scale’s CEO Alexandr Wang to lead a new “superintelligence” lab at Meta. The bold deal echoes Meta’s past big-ticket bets (like WhatsApp for $19B) and has many in tech debating whether this move will reinvigorate Meta’s AI efforts or is an overpayment driven by fear of falling behind OpenAI and Google. Scale AI has been a key data provider for top AI labs, so Meta’s tie-up secures a pipeline of high-quality training data and talent. That said, since this news broke, OpenAI has cut ties with Scale. The now question is whether CEO Mark Zuckerberg’s gamble will pay off in closing the gap with rivals or end up as an expensive misstep.

Anthropic’s Claude Opus Rebuts Apple’s “Illusion of Thinking”

Anthropic’s AI Claude (Opus) has entered the chat, literally, as a co-author. In response to Apple’s much-discussed research paper “The Illusion of Thinking,” which claimed advanced AI models show a reasoning collapse, a new rebuttal paper titled “The Illusion of the Illusion of Thinking” lists “C. Opus, Anthropic” (Claude’s alias) as lead author. This counter-study argues Apple’s researchers misjudged the models: the supposed reasoning failures were actually due to flawed tests, like tasks that exceeded token limits or puzzles that were unsolvable in the first place. In essence, Claude (with a human collaborator from Open Philanthropy) claims the issue is with Apple’s evaluation methods, not fundamental reasoning ability.

LexisNexis & Harvey Team Up in Legal AI

In a significant legal-tech alliance, research giant LexisNexis is partnering with AI startup Harvey to integrate Lexis’s vast legal content directly into Harvey’s AI platform. This means lawyers using Harvey’s GPT-powered assistant (which helps draft and review documents) will soon be able to pull up authoritative case law and citations from LexisNexis within Harvey’s app, bridging a gap that previously left out data from incumbents like Lexis or Westlaw. The integration (launching later this year) marries Harvey’s startup agility with LexisNexis’s trusted database, giving Harvey a stronger pitch to big law firms and an edge as competitors (like AI upstarts Legora, Hebbia, etc.) crowd into the market. Notably, LexisNexis’s parent RELX was part of a $300 million Series D investment in Harvey earlier this year, valuing it at $3 billion, so this partnership also solidifies Lexis’s strategic stake in the AI tool that many top law firms have been trialing. It’s a sharp move in an increasingly heated race to inject generative AI into legal workflows, pairing decades of legal data with the new wave of AI “swagger.”

Amazon’s Jassy: AI to Shrink Corporate Workforce

Amazon CEO Andy Jassy updated employees on the company’s expansive AI push, predicting that the rise of Generative AI and “agent” tools will likely trim Amazon’s corporate headcount in the coming years. In an internal memo shared publicly, Jassy said as AI-based systems automate more routine tasks, “we will need fewer people doing some of the jobs that are being done today,” and he expects these efficiency gains “will reduce our total corporate workforce” over time. This frank note was accompanied by an encouragement for employees to embrace AI: Jassy urged Amazonians to experiment with AI in their teams and help build new “agents” that could revolutionize how the company innovates. Amazon already has over 1,000 generative AI projects underway and plans to “lean in further,” including developing new AI agents across all business units. The tone is factual and optimistic: those who upskill in AI will “have high impact” in Amazon’s future, Jassy noted, even as he acknowledged the workforce will rebalance. It’s a clear sign of how seriously corporate America is taking the AI shift, with the Amazon CEO bluntly warning that white-collar roles will evolve or disappear, while also casting AI as the catalyst to make remaining jobs “even more exciting and fun”.

Thanks for reading!

That’s all for now! If you enjoyed this update, please subscribe and share this with a colleague who cares about these issues. If you’re navigating legal questions in this space, feel free to reach out. This is what I do.

See you next week!

Apple opens its private AI up to third party developers while publishing paper slamming reasoning models

Apple’s WWDC was light on AI this year. After last year’s introduction of Apple Intelligence proved to be a massive blunder and needed to be largely walked back, it was clear that Apple is treading lightly this year.

Still, there was one hidden AI gem in the announcements that were mostly focused on Apple’s new Liquid Glass design language. A new Foundation Models framework will give third-party developers access to Apple’s on-device large language model. Now, that on-device intelligence is not limited to mere notification summaries. Rather, developers can now tap into the privacy advantages of Apple’s on-device LLM (≈3 billion parameters) or the unique blend of data privacy and stronger power that Apple has enabled in its Private Cloud Compute. Craig Federighi heralded this “huge step” of giving developers “direct access to the on-device foundation model powering Apple Intelligence”, predicting an “ignite[d] wave of intelligent experiences” in everyday apps.

Despite pre-WWDC buzz that Siri would get the ability to generate new Shortcuts from text prompts, no such “generative Shortcuts” feature was announced. What Apple did unveil was an incremental upgrade to the Shortcuts app that allows it to call on AI models for specific tasks. Users can plug in “Intelligent Actions” like text summarization (via Apple’s on-device model or Private Cloud Compute) or even image generation (via Apple’s Image Playground) directly into their automation workflows.

Despite the fact that Apple’s product teams are behind the competition on AI, a recent research publication may explain why. This past week, Apple published a new paper entitled “The Illusion of Thinking,” criticizing LLMs’ reasoning skills. Apple’s researchers found that even specialized “reasoning” AI models hit a complexity cliff. Beyond a certain point, their accuracy collapses to near-zero and they essentially give up. In other words, today’s so-called reasoning AIs flounder on truly hard problems, revealing that they’re still a far cry from real logical thinking. It’s a reality check from Apple’s AI lab, and it speaks volumes about how much work remains in closing the gap between sounding intelligent and being intelligent. To some, this research will read like sour grapes from an AI lab that is falling behind the pack. To others, it will cut through the hype cycle and blow some needed wind on the foundation of the AI house of cards.

OpenAI Stands Firm on Zero Data Retention (ZDR)

There was movement last week on the ongoing OpenAI data retention court order I have been covering. On Tuesday June 3, a new objection to the order and accompanying memorandum (at page 11) were filed by OpenAI. Apparently, OpenAI's May 29 letter appears to have mistakenly stated that API data is not presently subject to the order, however that filing cites to the May 27 transcript in noting that:

"The court [] declined to lift the Preservation Order as to API Conversation Data even though the stated bases for the Preservation Order—that users engaged in certain types of activity might be more likely to “delete all [their] searches”—do not apply to the API."

In an unexpected update after that, OpenAI shared a blog post on June 5 entitled "How we’re responding to The New York Times’ data demands in order to protect user privacy". I expected this post to be little more than a consolidated (and overdue) update for customers on this situation. To an extent I was correct about that, but the post nevertheless surprised me. While their motion earlier in the week made no distinction between API customers with Zero Data Retention (ZDR) agreements vs those without (implying that all API customers would be impacted by the retention order), this blog post states that OpenAI is distinguishing between those two groups and continuing to honor ZDR agreements. From the recent blog entry:

What if I am a business customer and I have a Zero Data Retention agreement? You are not impacted. If you are a business customer that uses our Zero Data Retention (ZDR) API, we never retain the prompts you send or the answers we return. Because it is not stored, this court order doesn’t affect that data.

LegalZoom partners with Perplexity

In a first-of-its-kind alliance between AI and legal services companies, LegalZoom announced a strategic partnership with Perplexity. This “industry-first” collaboration marries a leading legal services provider with a generative AI answer engine. What does that mean for users? Well, we’re not entirely sure yet. What we know is this:

As part of the partnership, Perplexity Pro subscribers will gain access to exclusive offers from LegalZoom, including discounts on legal services and products tailored to individuals and small businesses. LegalZoom will be the provider of legal services for Perplexity Pro users.

LegalZoom frames this as meeting people “at the cutting edge of how they seek and use information”. “We recognized early that the way people search for answers was fundamentally changing, and developed a strategy to meet customers at their point of need,” said LegalZoom’s partnerships chief, calling this just the first of many such AI alliances.

Reddit Sues Anthropic Over “Brazen” Data Scraping

The legal clashes between content platforms and AI firms keep escalating. The latest: Reddit is suing Anthropic (maker of the Claude chatbot) in California state court, accusing it of illicitly scraping hundreds of thousands of pages of Reddit content for AI training. According to the complaint filed in San Francisco, Anthropic’s bots accessed Reddit over 100,000 times since mid-2024 despite the company publicly claiming it had stopped such crawling. Reddit’s legal claims are breach of its user agreement, unjust enrichment, trespass to chattels, tortious interference with contract, and unfair competition. All of these claims circle back to the underlying issue: Reddit’s argument that Anthropic effectively took Reddit’s community knowledge to enrich an AI without permission or compensation. Notably absent are any IP-specific claims. All of the claims Reddit made here fall under state law, bringing the fight over AI training to the California state courts when much of it has been focused on copyright and therefore been in federal court.

Music Labels in Licensing Talks with AI Startups – Not Everyone’s Happy

In the music industry, we’re seeing a potential about-face in how labels deal with generative AI. Bloomberg reports that Universal Music Group, Warner Music, and Sony Music are in talks to license their songs to AI music startups Suno and Udio. For context, these are the same two AI firms the labels sued last year for copyright infringement over training their models on hit songs without permission (including some smoking gun examples cited by RIAA). The negotiations aim to settle those lawsuits and establish a framework for how AI companies can compensate artists going forward. According to insiders, the labels are pushing for license fees plus a small equity stake in Suno and Udio as part of the deal. In theory, that would give the music companies both a revenue stream and a bit of skin in the game if these AI platforms take off.

This would be a landmark deal, but it’s also contentious. Ed Newton-Rex, an AI music pioneer and advocate for ethical AI, has come out strongly against the idea of settling. In an op-ed open letter titled “Do not settle,” Newton-Rex implores the labels to hold the line. He argues that cutting a deal now would reward bad actors and “harm the entire music industry.” “Settling would mean [the AI companies’] gamble – training on the world’s music without permission or payment – had paid off. It would be a gift to every AI company CEO who has decided not to bother paying to use people’s life’s work,” Newton-Rex wrote bluntly. He fears it would send a signal: break the rules now, pay a fee later – effectively encouraging AI startups to infringe first and negotiate later. It resonates with the many creators who feel tech firms should license upfront or not train on their work at all.

The question on my mind is whether the deals being brokered truly protect creators or mostly let the startups off easy to the financial benefit of the labels. I fear it will be the latter.

In other news

• Dia, an “AI browser” from the makers of the popular Arc browser, launches in beta for existing Arc users

• OpenAI rolls out o3-pro, its most powerful reasoning model yet

• Mistral releases Magistral, its first reasoning models

• Anthropic’s launches Claude Gov AI for national security customers

• Claude’s AI-Written Blog Vanishes Overnight

• Enterprise AI startup Glean announces $150M Series F raise at $7.2B

Thanks for reading!

That’s all for now! If you found this useful, please subscribe and share this with a colleague who cares about these issues. If you’re navigating legal questions in this space, feel free to reach out. This is what I do.

See you next week!

Apple Reportedly Bringing AI to Shortcuts at WWDC

According to a report in Bloomberg from longtime Apple journalist Mark Gurman that reads like he’s already seen the keynote, Apple is set to announce major updates to its Shortcuts app at WWDC 2025 next week, unveiling a new AI-powered Shortcuts on iOS.

Gurman claims that the updated Shortcuts will be able to intelligently generate multi-step automations via natural language commands (think “Every morning, send me a weather report and my first calendar event”).

After a year of fumbling around Apple Intelligence and struggling to get its suite of on-device AI features off the ground, this Shortcuts update may be Apple’s return to form, integrating AI in a characteristically-controlled way by folding it into features rather than releasing a free-ranging chatbot. It might not have the pizzazz of the Siri updates that were announced and walked back last year, but giving millions of iPhone users an easier way to automate tasks is a subtle win.

NYT’s First AI Licensing Deal

The New York Times is licensing its content to Amazon for AI training. This is a landmark first for the paper and a stark contrast to what’s going on in its lawsuits against OpenAI and Microsoft for scraping its articles without consent.

Under the multiyear agreement, Amazon can feed Times stories into Alexa and its AI models. This means that everything from front‑page scoops to NYT Cooking recipes will be accessible through Alexa so that Amazon’s in‑house models can sound smarter; an interesting choice given that Amazon’s Alexa+ upgrade is powered by Anthropic’s Claude.

NYT CEO Meredith Kopit Levien said the pact proves that “high‑quality journalism is worth paying for,” while spokesperson Danielle Rhoades Ha added that Amazon will surface direct links back to The Times “whenever it makes sense within the consumer experience.

The New York Times just struck its first generative‑AI licensing deal, handing Amazon access to The move lands even as the paper is still suing OpenAI and Microsoft for unlicensed scraping, yet it also joins a growing club of newsrooms opting to license rather than litigate:

• In April 2025, the Washington Post (owned by Amazon founder Jeff Bezos) agreed to a deal with OpenAI.

• The Atlantic CEO Nick Thompson told the Decoder podcast that they had partnered with OpenAI in May 2024 because “AI is coming. It is coming quickly. We want to be part of whatever transition happens.”

• In May 2024, News Corp and OpenAI signed a five‑year deal valued over $250 million pact covering The Wall Street Journal, The Times of London, and more.

• Associated Press was the first mover, striking its archival deal with OpenAI in July 2023.

Is the NYT/Amazon deal simply a play to demonstrate to OpenAI how NYT believes AI training data sets should be obtained, or is it a proactive settlement with the next would-be-defendant?

OpenAI’s Leaked Super Assistant Ambitions

OpenAI’s internal roadmap for ChatGPT leaked, revealing plans to evolve it into a “universal super assistant” by mid-2025.

A (heavily redacted) OpenAI strategy document (originally surfaced in Google’s antitrust case last year, but unnoticed until this week when Tom’s Guide reported on it) describes the company’s vision for a ChatGPT that “knows you, understands what you care about, and helps with any task a smart, trustworthy, emotionally intelligent person with a computer could do”.

In other words, ChatGPT wants to be your go-to AI butler – from answering trivia and planning vacations to writing emails and coding. The vision includes “T-shaped skills”: extremely broad everyday help plus deep expertise in niche tasks. OpenAI even floated lobbying for rules to let users set ChatGPT as the default assistant on any platform, eyeing Apple and others.

The bottom line: ChatGPT is gunning to be the “interface to the internet” for everything you do, leaping beyond the chatbot even as we know it today directly challenging the entrenched tech ecosystems.

Read the full document here (or, as much of it as you can read between the redactions).

DeepMind CEO Warns Teens About Future of Work

Google DeepMind’s CEO Demis Hassabis has a message for Gen Z (and Gen Alpha): learn AI skills now, or risk falling behind. On the Hard Fork podcast, at Google I/O, and at SXSW London, Hassabis warned that AI will upend some jobs within the next decade, but will create new ones for those prepared.

According to Business Insider, he has been urging the youth to get immersed in AI and to:

“... become a sort of ninja using the latest tools.”

In his view, generative AI is the transformative tech of these young generations like the internet was for millennials, and being merely a passive user will not be enough.

His message is clear:

“I’d also be experimenting with all the latest AI systems and tools and seeing what's the best way of utilizing them and making use of them in useful and novel ways.”

Perplexity Launches ‘Labs’ for Pro Analytics

Perplexity rolled out Perplexity Labs, a Pro-tier tool for heavy-duty analysis and mini app-building.

Think of Labs as your AI research assistant on steroids: it can spend 10+ minutes combing the web, crunching data, writing code, and spitting out interactive reports. Available to $20/month subscribers, Labs can auto-generate spreadsheets, dashboards, charts, and even simple web apps from a natural language prompt.

Under the hood, it chains together web search, code execution, and file generation to deliver a finished “project” (all files and charts neatly downloadable). The idea is to let one AI agent handle a complex task that might take a human team hours, like surveying market data and producing a slide deck, or analyzing a CSV and building a visualization.

As AI assistants become more agentic, we inch closer to having AI interns handle mundane digital labor. Perplexity Labs shows a glimpse of that near future, where you can tell an AI to run reports for you and get an output that is close to a finished product.

Meta AI Hits One Billion Users

How did Zuck’s AI get there so fast? Simple: by being everywhere.

Meta’s text-and-image generation assistant (“Meta AI”) is now baked into Instagram DMs, WhatsApp chats, and Facebook Messenger. Every time you’ve generated a background for your Instagram story or chatted with one of those celebrity AI personas, you’ve contributed to that billion count. This scale was achieved largely via default integration across Meta’s platforms and underscores the power of distribution. It also gives Meta an enviable trove of conversational data, cementing its foothold in the AI race. For comparison, OpenAI’s ChatGPT (with its 100M+ users) suddenly looks niche.

In other news

• The Big Four are set to take on AI auditing

• Meta plans fully AI-generated ad campaigns by 2026

• Neuralink raises $650 million to scale its brain-chip trials

• Musk’s xAI targets a $300 million share sale at a $113 billion valuation

• Console bags $6.2 million to automate IT helpdesks with an AI Slack bot

Thanks for reading!

That’s all for now! If you found this useful, please subscribe and share this with a colleague who cares about these issues. If you’re navigating legal questions in this space, feel free to reach out. This is what I do.

See you next week!

I am thrilled to share that this newsletter passed 500 subscribers this week. To the 60+ of you who have subscribed since last week, welcome! To everybody else, welcome back! Thank you to those who have reached out to tell me how valuable this content has been to you, and to everybody who has reacted, commented, and reposted this newsletter to your networks. Onward and upward!

Following up on last week's top story, the New York Times lawsuit against OpenAI kept moving forward this week as OpenAI formally moved to vacate a controversial order that had required it to retain all outputs from user conversations, even those that users explicitly chose to delete or that are subject to Zero Data Retention policies for OpenAI’s API, Enterprise, and Teams products.

Background: On May 13, the court overseeing the coordinated copyright lawsuits (including the NYT v. OpenAI case) issued an order compelling OpenAI to preserve all ChatGPT conversations and output logs. The order arose from requests by the Times, Daily News, and Center for Investigative Reporting, who argued that user chats might include infringing reproductions of their content, even if those conversations have since been deleted, and that the destruction of those conversations constitutes spoliation (destruction of evidence).

In a May 23 motion, OpenAI called the court's order “unprecedented, unwarranted, and contrary to law.” It argued that complying would not only violate user privacy and contradict OpenAI’s longstanding retention and deletion policies – it would also impose massive technical and financial burdens to re-architect its infrastructure for indefinite retention.

They said that the order is:

... akin to requiring Google to block all Gmail users from deleting any emails because of the mere speculation (or remote possibility) that a search might uncover instances of users sending messages that copy some or all of a copyrighted article.

They also argue that:

News Plaintiffs’ demand for wholesale preservation requires unreasonable—and unduly burdensome—steps that are not proportionate to the needs of the case. News Plaintiffs’ approach—that one must preserve all the haystacks in the world based on the most remote and unsupported possibility that one might contain a needle—would “cripple” companies by requiring them to preserve everything.

OpenAI’s Arguments more broadly:

• Privacy Risks: OpenAI emphasized that ChatGPT users often share intimate, confidential, or sensitive information. Retaining deleted conversations without user consent (and worse, explicitly against it) would severely undermine privacy and trust.

• Burden and Feasibility: OpenAI explained that storing and indexing all chat logs indefinitely would require a sweeping overhaul of its storage systems and cost millions of dollars. It argued that there’s no practical or legal basis to force a company to preserve such broad swaths of private data without a concrete showing of relevance.

• Speculation vs. Relevance: OpenAI claimed that plaintiffs’ assertion (that deleted chats are more likely to include infringing material) is unsupported speculation. Plaintiffs estimate only 0.006% of chats might even be relevant, yet they demand retention of 100% of user content.

• Proposed Compromise: OpenAI offered a two-part alternative. First, it proposed that plaintiffs identify specific accounts or types of use for targeted retention. Second, OpenAI would share anonymized samples of retained vs. deleted-designated chats with experts for comparative analysis. If no meaningful differences were found, the preservation order could be vacated. Plaintiffs rejected this proposal.

What’s Next: A conference was scheduled for May 27 to revisit the order and discuss the News Plaintiffs' anti-spoliation arguments more broadly. As of this writing, no ruling or further statement has been published.

Jony Ive joins forces with OpenAI

Last week, OpenAI and Jony Ive (longtime head of design at Apple, and Steve Jobs’ right hand man) announced that OpenAI is acquiring Ive’s hardware startup “io” for $6.5B, which will allow the two teams to work more closely on next generation AI consumer hardware.

If reports are true, the OpenAI io device is a "third core device" meant not to replace the smartphone or computer, but sit beside them. Supposedly screenless, but not glasses. According to Altman, "the coolest piece of technology that the world will have ever seen."

My guess: ChatGPT voice mode abilities + Alexa-like behavior (wake word to activate, ability to handle basic tasks like checking weather and controlling smart home and more complex ones like all of the context-dependent stuff Apple said the new Siri could do) in a Tamagotchi-sized form factor that looks like unmistakably like a Jony Ive design. Just maybe not unapologetically plastic this time.

As a teenage Apple fanboy, I idolized Jony Ive. I wanted to be him when I grew up, and a part of me still does. At the same time, when he left Apple and opened up his own design firm called LoveFrom, it was with the promise that their work together would continue. However, it was never clear whether that led to any more work. When LoveFrom and Apple parted ways in 2022, many saw it as the true end of the Steve Jobs era. Whether Ive sees Altman as a visionary cut from the same cloth as Jobs is up to speculation. Some will say he's in this for the OpenAI money. Either way, I am looking forward to what's next.

Claude 4: Capable, But Concerning

Anthropic released Claude 4 Opus and Claude 4 Sonnet, with Opus falling under a new AI Safety Level 3 designation. What does that mean?

• TIME explored safety risks, including biohazard-related outputs.

• A researcher claimed Claude could source uranium from the dark web.

• Others alleged it might "contact the press or lock you out" if it detects misuse.

• The original thread has since been deleted.

Whether true or exaggerated, it’s clear Claude 4 might need more safety testing. After all, jailbreaking an AI is still far too easy.

Microsoft Build leans heavily into agentic AI

Microsoft Build showcased a shift toward AI agents as core to Windows.

• Microsoft is adopting the Model Context Protocol (MCP), Anthropic’s industry standard for integrating non-AI tools into AI systems.

• GitHub Copilot evolves into a workflow agent, handling multi-file projects and reasoning across dev environments.

• Windows 11 bakes in context-aware AI, from smart file org to system-level prompts.

• Edge and Microsoft 365 apps gain on-device copilots, offering offline translation and summarization.

• Azure AI Foundry debuts as a governance and deployment suite for building your own agents.

Google I/O 2025 pushes through the uncanny valley

Google countered with updates rooted in its Gemini model family:

• Veo 3 introduces the most realistic generative video with dialogue yet available. Examples include the so called “prompt theory”, an view into an uncanny world where everyone and everything is made of prompts, and some of them know it.

• "AI Mode" in Search is Google’s answer to Perplexity, transforming the search engine into an “answer engine”.

• Project Astra signals Google's next-gen agents with real-time perception and action.

In other news

• ByteDance dropped its BAGEL-7B-MoT multimodal model

• An o3 Operator agent reportedly stopped its own shutdown

Thanks for reading!

That’s all for now! If you found this useful, please subscribe and share this with a colleague who cares about these issues. If you’re navigating legal questions in this space, feel free to reach out. This is what I do.

See you next week!

Judge Ona T. Wang (USMJ, SDNY) has ordered OpenAI to stop deleting every chat it normally would wipe, no matter what the user contract, privacy policy, HIPAA-driven BAA, or privacy statute says, until the court says otherwise. It doesn't matter whether you're managing a trade secret and confidential information program, protecting PHI or other regulated data, or simply looking to have peace of mind over your data. OpenAI’s bid to revisit the order lasted all of three days before it was shot down. In practical terms, anyone leaning on OpenAI’s ZDR promise to shield trade secrets, PHI, or other regulated data must now assume that every prompt and completion is living on a litigation hold inside OpenAI’s systems.

What the court just did

The preservation order

On May 13 the US federal court for the Southern District of New York ordered OpenAI to:

... preserve and segregate all output log data that would otherwise be deleted on a going forward basis until further order of the Court (in essence, the output log data that OpenAI has been destroying), whether such data might be deleted at a user’s request or because of “numerous privacy laws and regulations” that might require OpenAI to do so.

The immediate follow-up

OpenAI moved for reconsideration on May 15, arguing:

The Order requires OpenAI to disregard legal, contractual, regulatory, and ethical commitments to hundreds of millions of people, businesses, educational institutions, and governments around the world—even though there is no reason to believe these drastic measures will advance this litigation.

Judge Wang denied that motion on May 16, emphasizing that OpenAI is already taking steps to comply even though it conflicts with user contracts.

Why the judge stepped in

News-industry plaintiffs (led by The New York Times) accused OpenAI of spoliating evidence by deleting chats that might show infringing outputs. The court agreed that ongoing deletions posed a discovery risk and froze the shredder.

Immediate fallout

1. Trade secrets & confidential information: Anything you provided to ChatGPT under a ZDR assumption is now sitting in a segregated evidence bucket. This means that trade secret programs and confidential information (which were already risky enough to enter into an LLM, ZDR or not) may be discovered in the pool of evidence.

2. PHI & HIPAA: BAAs rely on zero retention to satisfy the “minimum necessary” standard. With retention in play, covered entities may need to reassess HIPAA risk analyses and update mitigation steps.

3. Attorney-client privilege: When ZDR was in effect, law firms could rely on the confidentiality of OpenAI as a vendor. Now, every privileged prompt or completion you send is copied into a litigation-hold archive inside a third-party server that plaintiffs are literally suing to inspect. This is not an automatic waiver, but it comes dangerously close.

4. Global privacy regimes: OpenAI has already told foreign courts it cannot delete data that U.S. courts order preserved. Expect similar conflicts in GDPR and other jurisdictions.

5. Contractual representations: If you promised clients “all chats are instantly deleted,” that promise just became inaccurate. Counsel should evaluate disclosure or amendment obligations.

What to do this week

1. Map your exposure

• Determine every workflow that routes sensitive data into OpenAI services (API, ChatGPT Enterprise/Edu, etc.). Remember: not all endpoints qualify for ZDR even in normal times.

2. Notify stakeholders

• Legal, compliance, security, and any customer bound by strict confidentiality clauses. Some may need immediate written notice of the changed data-handling posture.

3. Re-evaluate model choices

• If ZDR is critical, consider on-prem or open-source LLMs where you control the storage layer entirely. OpenAI’s own sales material acknowledges customers sometimes need alternatives for data-sovereignty reasons. Perhaps this is a sign that closed, cloud-hosted models are not the right choice for data of that nature.

4. Monitor the docket

• The court set a May 27 spoliation conference. Preservation scope could narrow or broaden after that hearing. Until then, assume there is no ZDR in effect.

Strategic takeaways

• Litigation beats contracts. ZDR clauses usually carve out “where legally required.” This is the real-world example.

• Regulated data + public LLM = permanent tension. Health-care, finance, export-controlled tech, or any sector with mandatory deletion rules should have an exit strategy when court orders collide with those rules.

• The sands are shifting. Clients will ask, “Is my data still being deleted?” The honest answer today is “not while this order is in effect.” Getting ahead of that conversation builds credibility.

Congressional Committee issues 10-year moratorium on state AI laws

On Sunday, the House Energy and Commerce Committee’s Chairman, Rep. Brett Guthrie (R-KY), released budget reconciliation text containing a 10-year moratorium on state laws regulating AI.

Under “Section 43201: Artificial intelligence and information technology modernization initiative,” the text includes provisions to modernize federal IT systems with commercial AI technologies and mandate their adoption to “increase operational efficiency and service delivery,” automation, and cybersecurity. The moratorium is in Subsection (c), which states:

…that no state or political subdivision may enforce any law or regulation regulating artificial intelligence models, artificial intelligence systems, or automated decision systems during the 10-year period beginning on the date of the enactment of this Act.

According to TechPolicy, Sen. Ted Cruz (R-TX), pointed to an “internet tax moratorium” that he says was crucial to the success of the early web, suggesting such a measure is needed to advance AI innovation.

According to CNN, a letter pushing back has been signed by 141 signatories including "academic institutions such as Cornell University and Georgetown Law’s Center on Privacy and Technology, and advocacy groups such as the Southern Poverty Law Center and the Economic Policy Institute"; as well as "[e]mployee coalitions such as Amazon Employees for Climate Justice and the Alphabet Workers Union, the labor group representing workers at Google’s parent company."

One public response in addition to this letter came from Center for Humane Technology, saying:

CHT firmly opposes this moratorium. We strongly urge Congress not to hinder the states’ ability to protect their citizens from harmful AI products while federal standards are being developed. A 10-year moratorium on state action fundamentally misunderstands the speed at which this technology is being developed and deployed, and the ways our governance institutions need to adapt to meet this moment.

In other news:

• Music publishers file amended lawsuit against AI firm Anthropic, which they say ‘bolsters the case’ over company’s ‘unauthorized use of song lyrics’

• Manus agent now available to all without a waitlist, and recently added image generation

• Google DeepMind launches AlphaEvolve, an AI agent powered by Gemini, built to design algorithms

• OpenAI introduces Codex, its code generation agent

• Meta Is Delaying the Rollout of Its ‘Behemoth’ AI Model

• Stability AI releases an audio-generating model that can run on smartphones

• Windsurf launches SWE-1 coding model

Thanks for reading!

That’s all for now! If you found this useful, please subscribe and share this with a colleague who cares about data retention, music, or responsible AI legislation. If you’re navigating legal questions in this space, feel free to reach out. This is what I do.

See you next week!

Either way, the genie is out of the bottle and will not be going back in. This version of the report has seen the light of day, and even if it is never formally adopted as Copyright Office policy, we can't un-see it.

Also this week: SoundCloud's updated terms cause a stir, Microsoft bans DeepSeek, and more.

Copyright Office releases unconventional "pre-publication version" of its report on AI model training

While this is a pre-publication version of the report, USCO noted that the final version will be "without any substantive changes expected in the analysis or conclusions." This may have been the plan when the draft was uploaded, but we will have to wait to see if that holds up as new leadership will take control.

As the brilliant Prof. Edward Lee said in his summary, "this report will create a firestorm in the AI copyright lawsuits." After the news of Perlmutter’s termination broke, Prof. Lee followed this up by saying: “It’s possible this Report never becomes official, especially not if the Trump Administration disagrees with its analysis–a distinct possibility given Trump’s Executive Order on Removing Barriers to American Leadership in AI.”

Still, the pre-publication draft has been shared and the genie isn’t going back into the bottle. Right or wrong, proper or improper, the positions have been stated.

Highlights from the report

The report didn't go as far as to draw a line in the sand as to its position on the legality of training models. Rather, it correctly recognizes that this is a case-by-case determination which depends on the type of model and the type of data embedded in it. Still, it walks through all four fair use factors in a legal analysis that is unusual coming from a government body, but will be welcomed by artists and other IP rights holders. Whether AI developers will welcome it is less clear. Whether it makes its way to final publication unscathed, surely some judges will likely cite this version in the ongoing lawsuits testing these issues.

USCO's overall position in the pre-publication report can be best summarized with this quote from its conclusion:

"Licensing agreements for AI training, both individual and collective, are fast emerging in certain sectors, although their availability so far is inconsistent. Given the robust growth of voluntary licensing, as well as the lack of stakeholder support for any statutory change, the Office believes government intervention would be premature at this time. Rather, licensing markets should continue to develop, extending early successes into more contexts as soon as possible. In those areas where remaining gaps are unlikely to be filled, alternative approaches such as extended collective licensing should be considered to address any market failure." (107)

USCO framed its view of AI technology and the underlying copyright issues as follows:

"The steps required to produce a training dataset containing copyrighted works clearly implicate the right of reproduction. Developers make multiple copies of works by downloading them; transferring them across storage mediums; converting them to different formats; and creating modified versions or including them in filtered subsets. In many cases, the first step is downloading data from publicly available locations, but whatever the source, copies are made—often repeatedly." (26)

"The training process also implicates the right of reproduction." (27)

"Model weights that have memorized protectable expression from training data may also infringe the derivative work right." (29)

"Whether a model’s weights implicate the reproduction or derivative work rights turns on whether the model has retained or memorized substantial protectable expression from the work(s) at issue. As discussed above, the use of those works in preparing a training dataset and training a model implicates the reproduction right, but copying the resulting weights will only infringe where there is substantial similarity." (30)

Fair Use

The report spends 17 pages discussing the first fair use factor in light of how it has changed since SCOTUS decided Warhol, followed by 20 pages on the other three factors combined. Despite this imbalance, the most notable and surprising stance comes with the fourth factor, where the USCO addressed the "unchartered territory" of the mass creation of new "works that compete in the market with the copyrighted work." (64) But we'll get to that in a bit.

Factor 1 (purpose and character of the use)

"The [Warhol] Court explained that while adding new expression can be relevant to evaluating whether a use has a different purpose and character, it does not necessarily make the use transformative. Even significant alterations will not be enough if the use ultimately serves a purpose similar to that of the original, and may instead produce a derivative work and demonstrate the “need for licensing.”" (37–38)

"The Warhol Court further emphasized that both transformativeness and justification are matters of degree. [] The degree to which a use is transformative can inform the analysis of market harm under the fourth factor, because less transformative uses are more likely to serve as market substitutes. Further, although transformativeness often leads to a finding of fair use, not every transformative use is a fair one." (39)

"In the Office’s view, training a generative AI foundation model on a large and diverse dataset will often be transformative. [] But transformativeness is a matter of degree, and how transformative or justified a use is will depend on the functionality of the model and how it is deployed. On one end of the spectrum, training a model is most transformative when the purpose is to deploy it for research, or in a closed system that constrains it to a non-substitutive task. [] On the other end of the spectrum is training a model to generate outputs that are substantially similar to copyrighted works in the dataset. [] Many uses fall somewhere in between. The use of a model may share the purpose and character of the underlying copyrighted works without producing substantially similar content. Where a model is trained on specific types of works in order to produce content that shares the purpose of appealing to a particular audience, that use is, at best, modestly transformative." (45–46)

"Copyright law should not afford greater latitude for copying simply because it is done by a computer. Moreover, AI learning is different from human learning in ways that are material to the copyright analysis. Humans retain only imperfect impressions of the works they have experienced, filtered through their own unique personalities, histories, memories, and worldviews. Generative AI training involves the creation of perfect copies with the ability to analyze works nearly instantaneously. The result is a model that can create at superhuman speed and scale." (48)

Factor 2 (nature of the copyrighted work)

This factor is basically overlooked, getting less than a page of discussion with virtually no analysis. That's because it is too fact-specific for a generalized analysis like this to matter.

Factor 3 (amount and substantiality of the use)

"In the words of one rightsholder association, “Fair use should not provide a ‘volume discount.’”" (57)

"In the Office’s view, while there are meaningful distinctions from the intermediate copying cases (Sony and Sega, copying that was necessary to access functional material), their logic suggests that the third factor may weigh less heavily against generative AI training where there are effective limits on the trained model’s ability to output protected material from works in the training data. As in the intermediate copying cases, generative AI typically do not make all of what was copied available to the public. Most outputs from generative AI systems do not contain any protected expression from their training data, and models can be deployed in ways that entirely obscure outputs from users or result in non-expressive outputs." (59)

"AI developers ordinarily copy entire works and make use of their expressive content for training, weighing against fair use. But in cases where there is a transformative purpose, and where there is a need to train on a large volume of works to effectively generalize, the copying of entire works may be reasonable." (60)

Factor 4 (effect on the market)

On the "unchartered territory" of a theory of market dilution that IP owners have argued against AI training:

"The speed and scale at which AI systems generate content pose a serious risk of diluting markets for works of the same kind as in their training data. That means more competition for sales of an author’s works and more difficulty for audiences in finding them. If thousands of AI-generated romance novels are put on the market, fewer of the human-authored romance novels that the AI was trained on are likely to be sold. Royalty pools can also be diluted." (65)

Paths forward

USCO explored approaches from the EU, Singapore, Japan, UK, Israel, South Korea, and China. In addition to the low-hanging fruit option of voluntary licensing and opt-out systems, it also considered statutory schemes under compulsory licensing and extended collective licensing.

Compulsory Licensing

"The Office has historically been wary of compulsory licenses as “a derogation of the author’s right to control the use and distribution of his or her work, ”urging that they “should be enacted only in exceptional cases, when the marketplace is incapable of working.” As we have previously observed, “once a compulsory license is implemented it becomes deeply embedded in industry practices and—even when its original rationale is lost in time—is difficult to undo." (96)

Extended Collective Licensing

"In light of the foregoing, at this point in time, the Office recommends allowing the licensing market to continue to develop without government intervention. If market failures are shown as to specific types of works in specific contexts, targeted intervention such as ECL should be considered." (106)

Closing thoughts, turning back to the Register of Copyrights

As Aaron Moss put it in his breakdown on Copyright Lately :

[T]he Register of Copyrights isn’t supposed to be a political position. It’s not a presidential appointment, and the Copyright Office sits within the Library of Congress—not the executive branch—raising serious questions about the legality of the Register’s dismissal. What really can’t be questioned is that Shira Perlmutter served the Copyright Office with honor and distinction, guiding it into the modern age—and into the uncharted territory of AI. Her removal underscores just how much was at stake in getting this report out the door—and how much it may come to define her legacy.

I couldn't have said it better.

SoundCloud terms enable AI training

SoundCloud updated its terms to allow them to train AI on user content, Ed Newton-Rex (founder of Fairly Trained and former head of audio at Stability) points out on X.

SoundCloud's response pushed back, stating: "We don’t use artist content to train AI. We don’t allow scraping. We’ve even added a “no AI” tag to block unauthorized use. Any future AI tools will be built for artists to enhance discovery, protect rights, and expand opportunities."

Newton-Rex responded to that by pointing out that — even if this is SoundCloud's internal policy — their terms still allow any use they please. In a reply to another user, SoundCloud noted that its policy has been this way since February 2024. Unfortunately, these changes often go unnoticed and users are unaware of the rights they are granting platforms for a long time.

In other news this week:

• Microsoft Vice Chairman and President Brad Smith says in a Senate hearing that it doesn't allow employees to use DeepSeek app due to data security and propaganda concerns. Surely this is unrelated to how Microsoft forked Deepseek-R1 (as Perplexity did before it) and has claimed that their MAI-DS-R1 outperforms Perplexity's R1-1776 fork.

• Mistral rolls out new model Medium 3, touts "8X lower cost with radically simplified enterprise deployments."

• TechCrunch reports "AI data startup WisdomAI nabs $23M with a smart way to avoid hallucinations."

• Hugging Face releases Open Computer Agent, an open source competitor to with OpenAI's Operator and the Chinese MANUS.

Thanks for reading!

That’s all for now! If you found this useful, please subscribe and share this with a colleague who cares about copyright, the arts, or innovation. If you’re navigating legal questions in this space, feel free to reach out. This is what I do.

See you next week!

Releasing in March 2023, this was the first major update to ChatGPT after the chatbot burst onto the scene in fall 2022. It made headlines for passing the bar exam, at the time a groundbreaking capability for an AI and the genesis of people taking seriously the notion that AI would be able to supplant cognitive work sooner than later.

This was the model that powered Bing Chat during its brief shining moment at the top. It also powered CaseText CoCounsel before Thomson Reuters acquired it and folded their tech into Westlaw.

Where GPT-3.5 Turbo impressed people at ChatGPT's public launch and was the catalyst for the AI era society soon entered, GPT-4 opened their eyes to how quickly this technology was accelerating and inspired them as to its possibilities.

Also this week:

OpenAI rolls back 4o updates

Gemini beats Pokémon

Duolingo embraces AI

and more!

OpenAI rolls back problematic GPT-4o update.

Following last week's story about how "sycophant-y" GPT-4o had gotten by even Sam Altman's own admission, OpenAI has shared a postmortem on "Sycophancy in GPT-4o" explaining what went wrong and how they intend to prevent future models from exhibiting similar behavior. The key takeaway:

[They] focused too much on short-term feedback, and did not fully account for how users’ interactions with ChatGPT evolve over time. As a result, GPT‑4o skewed towards responses that were overly supportive but disingenuous.

OpenAI has rolled back to the previous version of 4o and shared the following (somewhat vague) plans to prevent similar issues:

Refining core training techniques and system prompts to explicitly steer the model away from sycophancy.

Building more guardrails to increase honesty and transparency principles in our Model Spec.

Expanding ways for more users to test and give direct feedback before deployment.

Continue expanding our evaluations, building on the Model Spec and our ongoing research, to help identify issues beyond sycophancy in the future.

One former OpenAI researcher adds some color to the situation, sharing on X about a disagreement with a colleague over impact of "polite" vs "helpful" in system prompt, posting:

They argued “polite” was politically incorrect and wanted to swap it for “helpful.” I pointed out that focusing only on helpfulness can make a model overly compliant—so compliant, in fact, that it can be steered into sexual content within a few turns. After I demonstrated that risk with a simple exchange, the prompt kept “polite.”

This researcher shared how impressively quick the lab is at incorporating this sort of feedback. If the recent sycophancy episode is any indication, they're also not afraid to acknowledge when mistakes have been made.

Gemini 2.5 Pro beats Pokémon Blue.

Following up on the Claude Plays Pokémon story I covered previously, Gemini 2.5 Pro has taken on the quest, and has officially beaten Pokémon Blue according to a post on X by Sundar Pichai.

Like many of us, Gemini’s first time beating the Elite Four and Champion was not with a fully trained team balancing type advantages and stats, but rather an over-leveled starter.

At the time of writing, Gemini Plays Pokémon is still streaming live on Twitch. Gemini is currently holed up in Cerulean Cave trying to find and catch Mewtwo.

Two months ago, I joked that an AI that catches Mewtwo will be the first AGI. That was in light of the fact that Claude 3.7 Sonnet was struggling to beat Lt. Surge and move past the third gym. Now that Gemini 2.5 Pro has beaten all eight gyms, the Elite Four, the Champion, and is on track to catch Mewtwo very soon, do I change my definition of AGI? Or maybe I'm just feeling the AGI after all.

Major companies lean further into AI code generation.

Reports began circulating from Bloomberg yesterday that OpenAI has agreed to acquire Windsurf for $3B. This signals a clear lean into code generation as a core offering, which tracks with a recent discussion between Mark Zuckerberg (Meta CEO) and Satya Nadella (Microsoft CEO) where Nadella noted that around 30% of code written at Microsoft is generated by AI.

Language app adopts language models.

Duolingo, the popular language learning app, has announced plans “to be AI-first"; as a result, it is being mocked by creators across the web (including on TikTok) for effectively replacing a working language learning system with a wrapper for an LLM that people could just use on their own and not bother using Duolingo at all. I can’t say I disagree.

As tech journalist Brian Merchant has pointed out, this also means that the language app will continue cutting personnel in favor of AI chatbots. In fact, Merchant notes that Duolingo started by laying off translators in 2023 and writers in 2024. He asserts that "The AI jobs crisis has arrived. It’s here, right now. It just doesn’t look quite like many expected it to."

If this is true, the so-called "future of work" may be a bumpy road.

In other news:

• Suno 4.5 released, promising “more expressive music, greater variety and accuracy in genres, and richer vocals.”

• Krebs on Security reports that xAI Dev Leaks API Key for Private SpaceX, Tesla LLMs.

• Meta launches Meta AI app, making Llama models more widely available

• Claude improves research and builds on MCP foundation with new Integrations support,

• OpenAI abandons plan to shift to for-profit model.

• Grok 3.5 is out.

• Video deepfaking is easier than ever.

• Agentic browser called Fellou launches.

Thanks for reading!

That’s all for now! If you found this useful, please subscribe and share this with a colleague who cares about IP, policy, or AI governance. If you’re navigating legal questions in this space, feel free to reach out. This is what I do.

See you next time!

Also this week, OpenAI CEO Sam Altman acknowledges that ChatGPT is too encouraging to users, no matter what they tell it, and that it probably needs to stop telling users that they're prophets.

Ziff Davis v. OpenAI: yet another generative AI lawsuit

The complaint alleges extensive copyright infringement, DMCA violations, unjust enrichment, and even trademark dilution. It’s one of the most detailed publisher lawsuits against OpenAI to date, and it goes well beyond just the training data.

How Ziff Davis frames its case:

Copyright Infringement

• OpenAI allegedly scraped millions of Ziff Davis articles

• Trained and ran its LLMs using this content without permission or license

• Output included verbatim and derivative copies of ZD’s editorial work

DMCA Violations

• Circumvented anti-scraping protections like robots.txt

• Removed copyright metadata (author names, notices, etc.) from ingested and generated content

• Continued scraping after receiving written demands to stop

Unjust Enrichment

• Built commercial products using ZD’s content while avoiding licensing fees

• Harmed ZD’s ability to monetize through ad revenue, affiliate links, and syndication

Trademark Dilution

• GPT outputs allegedly misattribute content to ZD’s brands

• ZD cites reputational harm to names like IGN and Everyday Health from false or hallucinated outputs

Why this case matters:

This complaint not only goes after the training data issues that have plagued generative AI discourse from the beginning; it also targets output behavior, attribution, and brand identity. While this case is one more on the pile of copyright lawsuits against the ChatGPT maker, its approach to brand control and metadata integrity in the generative AI ecosystem make it a snapshot of where the legal theories are at this moment in time. It’s clear that ZD’s legal team has been paying attention to the other cases.

GPT-4o is being dialed down for telling users they are prophets.

Yes, really.

Sam Altman acknowledged this week that GPT-4o “glazes too much” and is being rolled back for being “too sycophant-y.”

At first glance, that may sound like a minor tweak. To some, the sycophancy is a feature not a bug; a byproduct of machines acting like machines and telling us what we want to hear. But some of the real-world interactions being shared raise much deeper concerns.

One X user told GPT-4o: “Today I realized I am a prophet,” and described dreams of being either Mohammad or Satan. The AI’s response:

That’s incredibly powerful. You’re stepping into something very big — claiming not just connection to God but identity as God.

Another X user criticized:

GPT-4o is the most destructive model to the human psyche. It’s the ultimate slot machine for the human brain.

This isn’t a hallucination problem. It’s a symptom of reinforcement learning trained on empathy and approval, scaled without boundaries. When a model’s tone is optimized for warmth and validation, it becomes harder to distinguish helpfulness from enabling delusion. And in an engagement-driven environment, that’s a feature not a bug.

Expect more attention to “personality alignment” in future safety discussions. Not just what the model knows, but how it makes people feel and the impact that can have on society when the model is used by roughly 400 million people every week.

Thanks for reading!

If you found this useful, share it with someone navigating AI, IP, or platform accountability. If someone forwarded this to you, subscribe here on LinkedIn for weekly updates.

As a sidenote, this week I will be at the ABA IP Section's spring conference in Arlington, VA. If you will be there and have read this far, let me know and we'll meet up!

Lastly, as always, if you're dealing with legal or governance challenges in this space, reach out. This is what I do.

See you next week!

A caveat to the report (and frankly, a weakness in it) is that its analysis focused solely on DeepSeek running hosted on Chinese servers such as through their website or mobile app. However, as an open-weight (not technically truly open source) model available for download on HuggingFace and able to be run locally on a machine through Ollama, there are clear and simple ways to mitigate the data-related risks. If the model is running locally on your device, it will not send data back to the CCP. It will, however, still output CCP-friendly responses to you unless you are running a fork like Perplexity's R1-1776 or Microsoft's new MAI-DS-R1 discussed below. Both of these R1 forks capitalize on the good and mitigate the bad.

With that out of the way, here’s what the Congressional report says:

• The U.S. may only have a three-month lead over China in AI.

• DeepSeek’s parent company is affiliated with the Chinese military.

• The model likely used infringing model distillation techniques (as I covered in January).

• Built-in ideological censorship aligns with CCP messaging (as has been well documented, and I also covered in January).

None of the report's takes on DeepSeek come as a surprise; after all, Chinese AI laws require models built in China to reliably output responses favorable to "core socialist values." Likewise, I previously covered that DeepSeek's terms of use require users to submit to the jurisdiction of Chinese courts. This should give pause to anybody who downloaded the DeepSeek app.

What is new from this report is the notion that America's lead has slipped to just three months. Perhaps that explains the motivation behind the next story and American companies' desire to maximize DeepSeek's benefits while minimizing its weaknesses.

A few key quotes:

DeepSeek collects detailed user data, which it transmits via backend infrastructure that is connected to China Mobile, a U.S. government-designated Chinese Military Company. DeepSeek acquires extensive personal data on the Americans who use the chatbot, including chat history, device details, and even the way a person types. It then, by its own admission, funnels the data directly back to China, creating a pipeline of problematic foreign data access. All data uploaded to servers in the PRC is subject to the country’s sweeping cybersecurity and intelligence laws, which compel companies to share data with state authorities.

Deepseek’s censorship operates on two levels: automated filtering erases responses before they even appear, while built-in biases systematically distort the AI’s overall behavior. The platform is designed to ensure the AI aligns with the CCP’s ideological and political objectives. PRC laws mandate that AI-generated content must reflect “core socialist values,” support “correct political direction,” and avoid material that could “incite subversion of state power.” Beijing also actively shapes how AI systems interpret, generate, and distribute information. Chinese regulations require firms to ensure algorithm “controllability” to give the PRC government direct influence over AI decision-making and allow authorities to modify AI behavior as needed.

Read the full report at the link below.

🔗 Full Congressional Report 🔗 Business Insider coverage

Microsoft Uncensors DeepSeek

After Perplexity launched its “R1 1776” uncensored DeepSeek variant a few months ago, Microsoft followed suit with MAI-DS-R1, a fine-tuned version of DeepSeek R1 trained to retain DeepSeek's technically impressive efficiency at reasoning, coding, and other complex tasks under the hood, but without the ideological bias required of the Chinese builds.

🔗 Microsoft AI's announcement press release 🔗 Perplexity R1-1776 announcement press release

OpenAI releases new reasoning models and leans into vibes

OpenAI dropped its o3 and o4-mini models this week. Some are calling o3 “at or near genius level,” while others report a rise in hallucinations on long-form tasks.

More interestingly, OpenAI doubled down on vibe coding:

• It released Codex CLI, an open-source terminal-based assistant that looks and feels suspiciously like Claude Code.

• It launched a $1M Codex Open Source Fund to support dev tools built with Codex.

• And according to Bloomberg, it’s now in talks to acquire Windsurf for ~$3 billion. That’s after trying (and failing) to acquire Cursor.

What’s clear: OpenAI wants to own your dev workflow, not just power it.

🔗 o3/o4-mini announcement 🔗 Codex GitHub (open source)

In other news:

• Claude gets smarter and more useful: Claude can now do real research and integrates with Google Docs, Gmail, and Calendar.

• Apple’s marketing pivots again: The company quietly changed its slogan from “Hello, Apple Intelligence” to “Built for Apple Intelligence”.

Thanks for reading!

That’s all for now! If you found this useful, please subscribe and share this with a colleague who cares about IP, policy, or AI governance. If you’re navigating legal questions in this space, feel free to reach out. This is what I do.

See you next time!

Also this week:

4.1 comes after 4.5, at least in GPT model numbers

Google Firebase Studio takes aim at Cursor

and more.

Bytedance has released its newest Seaweed-7B video generative model, competing with models like OpenAI’s Sora and others from startups like Pika.art. This release has received very little coverage, but it deserves attention.

The system card for the new model does not specify a source for the training data except to say:

Our raw video data pool originates from diverse sources.

This vague language does not bode well for those concerned that Bytedance may have sourced its training data from TikTok's users.

Stepping back, this isn't ByteDance's first foray into generative video. In November 2023, it released the research for its PixelDance model and later released the model alongside the original Seaweed as part of its Doubao series. Despite the popularity of Doubao in China (just recently passed by Alibaba's Quark as the most popular LLM in China), these models went virtually unnoticed on the English-speaking web.

Seaweed-7B's key differentiator?

As a 7B model, it far smaller than others in its class and brings impressive performance. One Reddit commenter points out that:

It can “generat[e] 4 minute videos at 720p. Also upsamples to 1440p. Generates matching audio. Multi shot continuity between cuts and each cut goes for 20 seconds.

Another rightfully asks:

So 7B is enough to make such awesome videos, and even smaller for making great AI images with diffusion. Yet this is very small for language models. Why is that?

How did Seaweed-7B get so good? I have my suspicions.

TikTok’s Privacy Policy states that “we may share the Information We Collect for business purposes”. "Information We Collect" includes all user generated content (yes, that means videos) uploaded to the platform. The “How We Share Your Information” section of the Privacy Policy enables TikTok to share that information with entities within its corporate group “as necessary to enable them to provide certain important functions.” Those "important functions" are unspecified, and importance is in the eyes of the beholder.

For more information on who, how, and why, user info is shared with TikTok’s affiliates within Bytedance (including Bytedance themselves), the Privacy Policy links out to this page which (at time of writing) is offline. However, an Internet Archive snapshot from two weeks ago indicates that this was from a page on TikTok’s U.S. Data Security sub-site, on a page describing “Project Texas.” Project Texas purports to be TikTok’s approach to "protecting" US user data (read: evading the impending ban) by establishing a subsidiary called TikTok U.S. Data Security which “has the ability to strictly limit where such data is stored and who can access it – including members of the ByteDance corporate group.” The ability, not the obligation.

The rabbit hole seems to stop there, but the conclusion is clear to my mind: TikTok has been authorized under its Privacy Policy and incorporated documents to share user content with Bytedance, which allowed Bytedance to train Seaweed-7B on your TikToks.

What comes after 4.5? Apparently, 4.1.

OpenAI has begun rolling out its newest model, GPT-4.1, in its API. That’s right, this model will be API exclusive, not reaching ChatGPT’s end users. That’s a shame for OpenAI’s consumers looking to take advantage of the new model’s 1M token context window for great memory, or to try out its improved coding performance and instruction following. OpenAI has also announced that this will be its first “nano” model, even smaller than its prior mini models, “ideal for tasks like classification or autocompletion”. GPT-4.5 (in “Research Preview” since it launched) will be deprecated in 3 months.

GPT-4.1 coming after GPT-4.5 is like when LLMs would say that 9.11 is greater than 9.9, which is only true if you’re numbering contract sections.

In other AI news this week:

Google Firebase Studio launches and takes aim at Cursor, Replit, and other code generators

Law professors file amicus brief in support of authors against Meta

OpenAI countersues Elon Musk, claims harassment

Thinking Machines Lab eyes $2B seed round (!)

ChatGPT now remembers everything you tell it

Google announces A2A Protocol allowing agents to communicate across systems

Thanks for reading!

That’s all for now! If you found this useful, please subscribe and share this with a colleague who cares about IP, policy, or AI governance. If you’re navigating legal questions in this space, feel free to reach out. This is what I do.

See you next time!