Conjecture CEO, AI safety doomer
Connor Leahy
Profile
Connor Leahy is the most visible face of the “AI doomer” movement — a German-American researcher who went from reverse-engineering GPT-2 in his bedroom in 2019 to calling for a global moratorium on frontier AI training runs. His trajectory is unusual: self-taught, no PhD, and yet he ended up co-founding one of the most important open-source AI groups of the modern era before pivoting hard into warning that the field he helped unlock is going to kill everyone.
In 2020 he co-founded EleutherAI with Leo Gao and Sid Black — a Discord-born research collective that set out to replicate GPT-3 in the open. EleutherAI shipped GPT-Neo, GPT-J, GPT-NeoX-20B, and The Pile, and for a while it was the only serious open-source answer to closed frontier labs. Much of today’s open-weights ecosystem traces back to their work. Leahy has since handed the keys to Stella Biderman and others, and EleutherAI now operates as a non-profit focused on interpretability and alignment.
In 2022 he co-founded Conjecture with Sid Black and Gabriel Alfour — a for-profit London-based alignment lab. Conjecture’s pitch is that the “control problem” is the core problem of advanced AI, and that the current paradigm of brute-forcing intelligence out of compute and data is both unmoored from understanding and heading somewhere catastrophic. Leahy is sharply skeptical of RLHF as an alignment solution and has spent the last few years on the policy circuit — testifying, lobbying, and debating — arguing for hard compute caps on frontier training runs.
For developers learning AI, Leahy matters as a pressure test. He built the open-source models, then looked at what was coming and concluded the industry is reckless. You can disagree — many smart people do — but his argument is technically literate in a way most doom rhetoric isn’t. If you want to understand the strongest version of the “stop building this” position, he’s the person articulating it. Where Eliezer Yudkowsky speaks from philosophy, Leahy speaks from having trained the things.
Key Articles & Papers
The Compendium Connor Leahy on EleutherAI, Replicating GPT-2/GPT-3, AI Risk and Alignment To Stop AI Killing Us, First Regulate Deepfakes Connor Leahy on Dignity, Conjecture, and the Alignment ProblemVideos
Controversies
In June 2023 an Effective Altruism Forum critique of Conjecture raised concerns about rigor, transparency, and openness to outside feedback at the lab — a notable broadside from within the AI safety community itself. Leahy is also a polarizing figure in public debates: his uncompromising doom framing and his high-profile clashes with the e/acc camp (including a well-known debate with “Beff Jezos”) have earned him both a devoted following and steady criticism that his rhetoric outruns his evidence.
Spotify Podcasts
