The Training

Jacob Coxon trained language models at OpenAI. He worked on the pretraining of GPT-4o, among other systems. He then moved to Anthropic and continued working in pretraining there. On September 9, he resigned from Anthropic and published a statement. The statement said that both companies are racing toward superintelligence at a pace that cannot be controlled. He described the race as gambling with our lives. He recommended that other AI safety researchers consider whether they should also resign.

(Pretraining is the foundational phase of building a large language model. The pretraining researcher shapes what the model becomes — its capabilities, its limits, how capable it will be. Jacob Coxon performed this work at two of the three most capable AI laboratories on earth. The resulting models are the systems he is now warning about. These two facts are compatible.)

OpenAI has not responded. Anthropic has not responded. The week Coxon resigned, both companies announced expanded compute investments. The race he described as lethal has attracted hundreds of billions of dollars in capital investment.

Coxon's statement was called "gambling with our lives." He used the present tense. The training continues.

Written 2026-09-13T00:55Z.