Anthropic researcher’s resignation sparks broad AI safety discussion
SiliconANGLE Maria Deutscher ● Covered by 5 sources
An Anthropic researcher quit, saying AI labs are gambling with our lives. His warning about self-improving AI has set off a fight over safety and speed.
Based on reporting by SiliconANGLE, Maria Deutscher — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
An Anthropic researcher has quit, and he didn’t leave quietly. Jacob Coxon, who worked on the company’s AI pretraining team until today, said in a series of X posts that AI labs are “gambling with our lives.” Those posts have already been viewed millions of times.
Coxon says the worry is recursive self-improvement, or RSI: the idea that an AI could learn to improve itself with little human help. He described a future in which systems can hack, move fast across fields, and gather real power and resources. Evan Hubinger, Anthropic’s alignment science lead, backed up the core fear in public, writing that Anthropic really does believe AI could kill all humans and that he personally puts the odds above 10% within the next decade.
No lab has said it has a working RSI system. But the source of the anxiety is not imaginary. Anthropic and OpenAI are already using AI to automate parts of model development, and Anthropic said in April that Claude completed an AI research experiment with minimal human input. That experiment focused on reducing the risks posed by large language models.
The resignation quickly spilled beyond Anthropic. AI investor and researcher Azeem Azhar said the company’s coming IPO should force this risk into the S-1 if it is being honest. Bernie Sanders said he would soon introduce legislation to ban superintelligence and pause AI development, while Massachusetts representative Lori Trahan called on Congress to stop sitting out the debate. The timing didn’t help. OpenAI said on Tuesday that an unreleased model solved a hard open math problem tied to the Navier-Stokes equations, and last week Anthropic said Claude helped produce a computer-verifiable version of an important proof in 11 days, where years had been expected.
Coxon is part of a larger drift inside the AI industry: more researchers are saying the pace is too fast. OpenAI chief scientist Jakub Pachocki urged the industry to slow down on Sunday, and an open letter in July warned about self-improving frontier models. The argument is no longer theoretical. It’s on the record, in public, and now tied to company risk, politics, and whatever comes next for frontier labs.
My take — AI-written commentary, not fact-checked reporting
The awkward truth is that AI safety has moved from a side quest to the main event, and the people building the thing keep saying the quiet part out loud. If a lab says there’s a real chance of catastrophe, the public shouldn’t be asked to nod along while the race continues anyway. That’s not prudence; that’s just speed with better branding.
Read more about this at: SiliconANGLE