Skip to main content
ResearcherDoomerTier 2

Connor Leahy

CEO, Conjecture

Self-taught AI researcher turned safety startup CEO — the most outspoken young voice arguing that AI development is racing toward catastrophe.

Credentials

CEO & Co-founder of Conjecture (AI safety startup), Co-founder of EleutherAI (open-source AI research group that built GPT-NeoX and GPT-J), self-taught ML researcher, prominent AI safety advocate and public speaker

Why They Matter

Leahy represents a new generation of AI insiders who are genuinely alarmed. He co-founded EleutherAI, which proved that open-source groups could build GPT-class models — then pivoted to safety when he saw what those models could become. For business leaders, his perspective is a useful counterweight to Silicon Valley hype: he argues that AI is powerful AND dangerous, and that businesses should prepare for a world where AI regulation tightens significantly.

Positions

AI Timeline View

We may have only a few years before AI systems become dangerously capable. The window for getting safety right is closing fast.

Safety Stance

Doomer

Key Beliefs

AI development is a race to the edge of a cliff. We need to slow down or stop before we fall off.

Multiple public talks and interviews, 2023-2024

Current AI labs are engaged in a dangerous arms race where competitive pressure overrides safety concerns.

Conjecture blog and public appearances

AI alignment is not just a technical problem — it's a governance and coordination problem. We need international treaties.

UK Parliament testimony and media appearances

Open-sourcing powerful AI models without safety guarantees is irresponsible, even though he helped build some of them (EleutherAI).

Interviews discussing his evolution from open-source AI to safety-first approach

Controversial Take

Openly calls himself a "doomer" and argues that there is a meaningful probability of human extinction from AI within our lifetimes. Claims the AI safety community is not alarmed enough, and that even other safety advocates underestimate the risk. His shift from building open-source AI (EleutherAI) to warning against it gives his arguments particular weight.

Track Record

How well have Connor Leahy's predictions held up?

Open-source groups could replicate large language models without big tech budgets

Made: 2020

EleutherAI built GPT-J (6B) and GPT-NeoX (20B), proving that frontier AI wasn't exclusive to deep-pocketed labs. This was followed by LLaMA, Mistral, and many others.

Right

AI capabilities will advance faster than safety measures, creating a growing gap

Made: 2022

Capabilities have surged (GPT-4, Claude 3, Gemini) while alignment remains largely unsolved. The safety gap is widely acknowledged.

Partially Right

Governments will begin seriously regulating AI within 2-3 years

Made: 2023

The EU AI Act passed. The UK held the AI Safety Summit. But regulation remains fragmented and enforcement is weak.

Partially Right

Key Quotes

We are building a god. We don't know how it works, we can't control it, and we're doing it as fast as possible. This is insane.

Oxford Union debate (2023-06)

I helped build open-source AI models. Now I spend my time trying to make sure they don't destroy the world. That should tell you something.

Podcast interview, Bankless (2023-08)

The default outcome of building superintelligence is not a good one. The default outcome is extinction.

UK Parliament testimony (2023-10)

Saying "we'll figure out safety later" is like saying "we'll figure out the brakes after we've already driven off the cliff."

AI safety conference talk (2024-02)

Last updated: 2026-03-26

Back to AI Minds Directory