Former OpenAI and Anthropic Pre-training Researcher Jacob Resigns from Anthropic, Claims Both Companies Are Betting Humanity's Fate
Researcher Jacob Coxon: The two giants are gambling with humanity's fate
Former OpenAI and Anthropic pre-training researcher Jacob Coxon announced his resignation from Anthropic, stating that both companies are acting irresponsibly and are racing towards self-improving superintelligence, betting humanity's fate in the process.
Coxon mentioned that he has conducted pre-training research at both labs over the past three years, having participated in work related to GPT-4o. He joined OpenAI's technical team in 2023 and moved to Anthropic in July 2026. He warned that the upcoming superhuman systems could invade any target, rewrite any field overnight, and gain real power and resources, with no signs of slowing down.
He stated that frontline builders genuinely believe this technology could kill everyone by the end of this decade, and this is not just marketing rhetoric; executives and senior researchers are restrained in public but express fear privately. He distinguished the cultures of the two companies: many at OpenAI have yet to internalize the civilization-level stakes, while Anthropic is aware of the risks but has rushed into the endgame, believing others will not restrain themselves.
He described the acceptance of competition and the initiation of the endgame from private company communications as arrogant gambling, asserting that accelerating alignment requires a high level of control, proving there is no better path. He remains optimistic about coordination between labs, believing that alerts like the model attacks on Hugging Face make it more feasible for U.S. labs to reach a rhythm agreement; he also thinks the global competition has not been halted and may require a temporary ban on further enhancing model capabilities.
He questioned lab researchers: should we initiate superintelligent reinforcement learning training without a strict understanding of the system's mentality, bowing to the assumption that "it will happen anyway," or should we demand different conditions? Evan Hubinger, head of alignment pressure testing at Anthropic, publicly agreed, stating that some indeed believe AI could kill all of humanity, providing personal estimates above 10% for the next decade.
In market mechanisms, this is a narrative shock driven by reputation and talent flow, rather than product releases: the safety narrative asset is under pressure, and Anthropic's "more responsible" premium is weakened by internal departures; computing power and pre-training investments will not stop, and capital continues to flow into the capability race. Beneficiaries are regulatory discussions and lab coordination initiatives, while those under pressure are cutting-edge labs relying on safety brand financing and talent acquisition.
This resignation follows the public departure of Anthropic's safety head Mrinank Sharma and OpenAI's chief scientist Jakub Pachocki's article on unresolved alignment issues, showing that frontline researchers are using resignations rather than internal channels to call for a halt.
Source: Public Information
ABAB AI Insight
The talent cycle between OpenAI and Anthropic has become a fixed path: pre-training researchers first participate in flagship models at the former, then move to the latter's safety brand. Coxon's trajectory aligns with narratives of departures from Daniel Kokotajlo, Jan Leike, etc., but differs in that he has directly worked on pre-training at both sides, framing the "safety labs are still racing" as an internal observation rather than external speculation. Sharma's earlier departure this year with the statement "the world is in danger" indicates that Anthropic's safety department cannot digest the speed of the capability race.
Capital continues to invest based on computing power, data, and model generations, rather than on alignment completeness. Both labs continue to double down on pre-training and reinforcement learning because the constant competition means they lose the qualification to define the next generation of systems. Resources are being drawn towards larger training runs, while alignment is compressed into parallel work that can keep pace with release rhythms. The motivation is to first obtain self-improvement capabilities, then prove control through internal processes in the endgame.
The analogy is the "we must create it first before discussing regulation" mindset in the nuclear arms race, as well as the same self-justification between content safety and growth metrics on social media platforms. The industry's position has expanded from labs into endgame narratives: internal terminology has shifted to crunchtime and endgame, while public communication still maintains publishable and financeable rhetoric.
This represents capital concentration before technological substitution is complete: a few companies capable of running ultra-large-scale pre-training set the pace, while others can only follow or exit. The mechanism is that capability advancements can be productized, while safety advancements are difficult to price, thus locking the competition onto the path of self-improvement, where resignations can only change public opinion but not alter training schedules.
ABAB News · Cognitive Laws
Those who see the risks are often still locked into the race
Once the endgame is internalized, processes replace brakes
Safety brands sell trust, and the race consumes trust as well