“We are building god. And we have no idea how to control it.”
“The default outcome is doom. Not because it's certain, but because we're not doing nearly enough to prevent it.”
“I think there's a better than even chance that AI kills everyone if we continue on the current path.”
“Alignment is not a research problem we can solve later. It is an engineering problem we must solve now.”
“Every AI lab knows they don't have alignment solved. They're racing anyway. That should terrify you.”
Connor Leahy represents the youngest generation of AI safety leaders—someone who grew up in the era of modern deep learning, co-founded EleutherAI (the open-source AI research collective), and then pivoted entirely to safety after concluding that AI development was on a catastrophic trajectory. His urgency and willingness to make bold probabilistic claims about AI doom make him one of the most provocative voices in the field.
Leahy’s path from open-source AI builder to AI safety advocate is instructive. Having been deeply involved in building and open-sourcing large language models through EleutherAI, he experienced firsthand the rapid capability gains that convinced him alignment was not keeping pace with capabilities. This technical background lends credibility to his warnings.
Through Conjecture, Leahy has pursued both theoretical alignment research and practical interpretability work, while simultaneously engaging in public advocacy. His communication style is deliberately alarming—he argues that understating the risk has been the default failure mode of the AI safety community.
His framing of the alignment problem as an “engineering problem” rather than a “research problem” reflects a belief that the challenge is not primarily about discovering new scientific principles but about doing the difficult engineering work of building reliable systems under time pressure.
Critics find Leahy’s doom predictions unfounded and counterproductive, arguing they either paralyze action or discredit the safety field. Supporters appreciate his willingness to state clearly what they believe many researchers think privately but are too cautious to say publicly.