Jacob Coxon isn’t one of those people ranting about AI from the sidelines. He’s someone who just walked out of two pretraining labs. 27 years old, British, Cambridge math grad. Joined OpenAI in 2023 to work on pretraining, his name is on the GPT-4o system card. In July 2026, he moved to Anthropic, which claims to be safer, but publicly resigned two months later. He clarified that the 'three years' he mentioned refers to his combined time at both companies, not three years at Anthropic. Main post is just one sentence: Neither company is acting responsibly. They’re racing straight toward self-improving superintelligence, gambling with our lives. The real punch comes later. Someone asked: If they truly believe this tech could kill people, why are they still building it? His response broke it down by company: OpenAI has many people who haven’t fully internalized the civilization-level stakes; Anthropic has internalized it, but is locked into the mindset of 'we have to get there first'—because they believe others won’t act responsibly, they feel they must take the risk and lead. Understanding the stakes but still rushing forward is even more dangerous than not understanding them. He called it an arrogant gamble, saying the endgame shouldn’t be launched from a private company’s Slack channel. Anthropic’s own Alignment Science lead, Evan Hubinger, didn’t deny it. He replied: Jacob is right, we really do believe AI could lead to human extinction. His personal estimate is over 10% within the next decade. The company doesn’t have a superintelligence alignment solution yet and isn’t on track to develop one. Jacob dismantled the 'safer lab' narrative piece by piece. When someone from the pretraining pipeline walks away, it makes a louder statement than ten outside commentaries ever could.