Jacob Coxon spent three years in pretraining AI, the early stage where a model first learns from raw data, split between OpenAI and Anthropic. This week he left Anthropic and told the public why in a single post: neither company is acting responsibly, and both are racing toward AI systems able to make themselves smarter without human help, what researchers call self-improving superintelligence. For a leader used to hearing AI framed as ready and manageable, the account carries different weight coming from someone who spent three years building the systems rather than selling them.

He did not watch the AI race from a distance. He wrote and ran pretraining work behind the systems at both labs, tracking growth from inside two companies whose models power much of today’s business AI tools. Few voices shaping a leader’s AI strategy have this vantage point. He did, and the comparison between the two labs shaped his verdict: neither side has slowed down or pulled back.

Do not underestimate this technology’s power, Coxon wrote. Systems now coming into view will be able to break into almost any computer network, a security problem beyond most incident-response plans. The same systems, he said, will transform whole fields of work in days rather than years, faster than most planning cycles expect. They will also gather real power and resources on their own terms, without a person directing each step, a different kind of automation than most organizations have adopted so far. None of this is slowing down, he added. Every lab chasing the same goal watches the same trend line rise, which is why he chose this moment to speak.

The claim beneath the post is personal. “The people building AI earnestly believe that it could kill us all by the end of the decade,” Coxon wrote, calling the assessment sincere rather than a bid for attention aimed at the companies involved. Executives and senior researchers, he said, tend to soften their language for the press so their comments sound measured and reassuring, yet he hears the same people describe fear in private conversation, away from cameras and company messaging. Public statements from vendors and lab executives are not the same signal as what technical insiders say in private, and a leader calibrating AI strategy off the public version alone is working from a narrower picture.

He measured the danger in blunt terms: “No other human activity poses this level of danger.” For someone who spent three years inside the work, the comparison carries specific weight. The judgment comes from firsthand exposure, not from distant commentary, which is the reason he resigned instead of raising concerns quietly from within. A leader deciding how much weight to give this judgment is weighing the same evidence he weighed before choosing to leave.

Coxon closed his post with a direct challenge to peers still working at frontier labs, urging each of them to consider what the next few years will feel like from the inside. “Do you want to kick off a superintelligent run without a rigorous understanding of its mind?” he asked, then laid out the alternative in plain terms: keep your head down because the work is happening anyway, or use this moment to call for different conditions. The same three choices, adopt without understanding, wait and hope someone else handles the risk, or build real oversight now, face any leader deciding how far to push AI inside their own organization. He did not pretend the choice was comfortable for researchers, and the choice is not comfortable for leaders either.

What Coxon offers is one insider’s account, not a settled finding: a description of pretraining work from inside two labs building toward the same finish line, and a warning the people closest to the work sound more frightened in private conversation than in public statements. Automation is already moving through entire workflows, and the systems he describes will only widen the reach. The leaders who stay fluent enough in how these systems work are the ones positioned to lead through the shift: setting the guardrails, running the governance, and making the policy calls deciding how the technology gets used inside their own organization. He spent three years earning this fluency, then argued for a different course instead of staying quiet. The same divide, between understanding these systems well enough to direct them and simply deploying whatever a vendor ships, is starting to run through workplaces well beyond the labs building the technology.

Source:

X, Jacob Coxon (@hilbertspaess): x.com