TechWithAdit (TWA)

Jacob Coxon, a researcher who worked on pretraining at both OpenAI and Anthropic, quit the industry entirely on Tuesday, warning on X that both companies are "racing straight to self-improving superintelligence and gambling with our lives." Hours later, Anthropic's own Alignment Science lead, Evan Hubinger, publicly backed him up, saying he personally believes there's ">10%" chance AI kills everyone within the next decade, and that Anthropic doesn't yet have a plan to solve it.

Product launches didn't slow down for any of this. Meta rolled out Muse, a personal AI agent that can send emails, book travel, and make payments on your behalf, while Cognition raised $2 billion at a $48 billion valuation for its coding agent Devin. Meanwhile OpenAI claims an unreleased model just solved one of math's seven Millennium Prize Problems, though two mathematicians say the company may have used their unpublished work to get there.

Jacob Coxon, who spent three years on pretraining research at both OpenAI and Anthropic, resigned from Anthropic on Tuesday, warning that neither company is acting responsibly and that people building AI "earnestly believe it could kill us all by the end of the decade," a claim he insists is not a marketing stunt.

Anthropic's own Alignment Science lead, Evan Hubinger, quote-tweeted Coxon's thread to say he agrees, estimating the risk at more than 10% within the next decade and admitting Anthropic doesn't have a working plan to solve alignment for superintelligence, nor is it "clearly on track" to get one.

The insights: When a company's own safety lead publicly agrees with a departing employee's doomsday warning, "we're being careful" stops being a reassuring answer. Watch whether more Anthropic researchers feel emboldened to speak this openly before its rumored IPO.

US law enforcement and intelligence agencies, including the NSA, CISA, and the FBI, accused six Chinese companies, including DeepSeek, Moonshot AI, and Alibaba, of maliciously copying US AI models through "distillation," training cheaper models on the outputs of pricier American systems from Anthropic, OpenAI, Google, and SpaceX.

Officials said the activity was happening "likely with Chinese government awareness" and called it "aggressive, malicious, and targeted... at an industrial scale," timing that lands awkwardly just weeks before President Xi Jinping's planned US visit and a separate US-China AI safety dialogue set for mid-September.

The insights: Washington is now treating model distillation as a national security issue, not just a competitive nuisance. Expect this accusation to shadow every US-China tech conversation this month, including the safety talks meant to lower tensions.

Meta launched Muse, internally codenamed Hatch, a personal AI agent that can autonomously send emails, book travel, sell a car, and make payments on a user's behalf, positioning it as the centerpiece of Zuckerberg's "personal superintelligence" vision.

Each Muse agent runs on its own dedicated virtual machine so it can keep working in the background even when you're not using it, connecting to email, calendar, payments, health, shopping, and smart home apps, though Meta admits it delayed the launch from April specifically to address internal concerns about the agent mismanaging sensitive data.

The insights: Meta is betting billions that people will hand a company with its privacy track record autonomous access to their inbox and bank-linked payments. Whether trust, not capability, becomes Muse's real bottleneck is the story to watch here.

Cognition, maker of the AI coding agent Devin, raised $2 billion at a $48 billion valuation, nearly doubling its price tag in just four months since its $26 billion round in May, with run-rate revenue climbing from $492 million to almost $900 million over that stretch.

The round, led by Andreessen Horowitz and Accel, comes months after rival Cursor sold to SpaceX for $60 billion partly because it ran out of compute, and it's unclear whether Cognition, which leases its own Nvidia cluster costing hundreds of millions annually, will avoid the same fate.

The insights: VCs are still writing enormous checks for multiple AI coding players rather than betting on one winner, unusual for a software category this hyped. Compute access, not product quality, may end up deciding who survives.

Operator's Spotlight Read

OpenAI announced that an unreleased next-generation model, more capable than GPT-6 Astra, had solved the Navier-Stokes existence and smoothness problem, one of the Clay Mathematics Institute's seven Millennium Prize Problems, proving the equations can develop a "blow-up" in finite time using a swarm of 10,000 coordinating agents.

The announcement was immediately overshadowed by NYU professor Tristan Buckmaster, who revealed he and Anthropic mathematician Levent Alpöge had been closing in on a related partial result, and that OpenAI approached him beforehand offering two options: let OpenAI publish first while crediting them as "the closest humans to the problem," or publish his own result himself, but only if he removed Alpöge's name because of his Anthropic ties. Buckmaster refused both and went public instead. OpenAI researcher Sebastien Bubeck denied the company accessed their private work, though OpenAI admitted it cannot rule out that anonymized product-usage data influenced its model.

For operators, this is a preview of a fight that's about to get much bigger: as AI labs compete to claim scientific breakthroughs ahead of massive IPOs, the boundary between "trained on public knowledge" and "trained on your private drafts" is about to become one of the most contested questions in tech, not just in coding, but in every field where researchers use AI tools to think.

The insights: A genuine mathematical milestone got buried under a trust scandal within hours, which tells you something about where the industry's credibility actually stands right now. The next lab to claim a scientific first should expect this exact question before the headline even finishes trending.

Was this email forwarded to you? Don't miss any updates — Subscribe to TechWithAdit for sharp, no-noise tech intelligence. Stay sharp. — Adit