Connor Leahy
Last reviewed
Jun 8, 2026
Sources
11 citations
Review status
Source-backed
Revision
v1 · 1,323 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 8, 2026
Sources
11 citations
Review status
Source-backed
Revision
v1 · 1,323 words
Add missing citations, update stale details, or suggest a clearer explanation.
Connor Leahy is a German-American artificial intelligence researcher and entrepreneur known for his work on open-source large language models and for his advocacy on the risks that advanced AI poses to humanity. He came to prominence as a co-founder of EleutherAI, the volunteer collective that produced some of the first openly available GPT-3-scale models, and as the founder and chief executive of Conjecture, a London AI safety company he led from 2022 until it wound down in 2026. He is one of the most outspoken public proponents of the view that uncontrolled superintelligence could cause human extinction, and in 2026 he became US Director of the nonprofit ControlAI. [1][2]
Leahy is a largely self-taught programmer who built much of his early reputation online under the handle "NPCollapse." He studied computer science at the Technical University of Munich and worked from roughly 2019 to 2021 at Aleph Alpha, the German large-model startup, as a machine-learning engineer and researcher. [1]
His entry into frontier AI came in 2019, when OpenAI announced GPT-2 but initially declined to release the full model, citing misuse concerns in what became known as a "staged release." Working from his dorm room, Leahy reverse-engineered and reproduced a GPT-2-scale model himself, an episode that both drew him into the emerging open-source LLM community and sharpened his thinking about how quickly powerful systems could proliferate. [1][4]
In July 2020 Leahy co-founded EleutherAI together with Sid Black and Leo Gao, originally a loose group of researchers and engineers organized around a Discord server. The collective's founding goal was audacious for a volunteer effort: to openly replicate GPT-3, which OpenAI had just unveiled but kept proprietary. EleutherAI grew into one of the most influential grassroots projects in machine learning, and during Leahy's time as a co-lead it shipped a series of widely used artifacts. [1][3][4]
The group first released The Pile in December 2020, an 825 GiB curated text corpus assembled specifically for training large language models. It followed with GPT-Neo in March 2021, GPT-J-6B in June 2021, described at release as the largest openly available GPT-3-style language model in the world, and GPT-NeoX-20B in February 2022. These models and the Pile became standard building blocks for open research and seeded a generation of downstream systems at a time when comparable models were locked behind corporate APIs. [1][3]
Even as EleutherAI's open releases were celebrated, Leahy grew increasingly worried that racing to reproduce ever-larger models, without a matching ability to understand or control them, was itself dangerous. That tension between openness and safety led him to step back from open-model work and turn his attention to AI alignment, the problem of ensuring powerful systems reliably do what their operators intend. [1][4]
Leahy founded Conjecture on March 1, 2022, relocating to London, together with co-founders Sid Black and Gabriel Alfour and a founding team drawn largely from EleutherAI. Unlike a volunteer collective, Conjecture was a venture-backed company whose stated purpose was to scale applied alignment research and to build, in its own words, "a new AI architecture to ensure the controllable, safe development of advanced AI." It raised on the order of $10 million, with backers including Nat Friedman, Daniel Gross, Patrick and John Collison, Andrej Karpathy, and Arthur Breitman, among others. [2][9][10]
Conjecture's signature research direction was Cognitive Emulation, or CoEm: an attempt to build AI systems that reason through trusted, human-like, auditable steps rather than as inscrutable end-to-end networks, so that their capabilities would be predictable and bounded by design. As Leahy framed the motivation, "if we could predictably build AIs that we know how to bound, this would give us the fundamental primitive necessary to build a new paradigm of controllable AI." In 2024 the company publicly demonstrated a system it called "Tactics" and published a CoEm roadmap. [2][5]
| Year | Conjecture milestone |
|---|---|
| 2022 | Founded in London; early LLM infrastructure and interpretability work |
| 2023 | Announced the Cognitive Emulation (CoEm) research agenda |
| 2024 | Demonstrated "Tactics" and released a CoEm roadmap |
| 2025 | Pivoted toward a commercial product |
| 2026 | Wound down; Leahy moved to ControlAI |
In a March 2026 retrospective, Leahy announced that Conjecture's chapter was drawing to a close, writing that "the economics just didn't add up" because "doing research on frontier LLMs is incredibly expensive work." A 2025 pivot toward a commercial product had not succeeded against far better-funded competitors, and he said he would refocus his energy on policy and public advocacy rather than on building systems. [5]
Leahy is among the most prominent public voices warning that the unchecked pursuit of artificial general intelligence and superintelligence is an existential threat. He is skeptical that current alignment techniques are adequate; of reinforcement learning from human feedback, the dominant method for making models more agreeable, he has said that such systems "are not becoming less alien. If anything, we're putting a nice little mask on them with a smiley face." [1]
He signed the Future of Life Institute's March 2023 open letter calling for a six-month pause on training systems more powerful than GPT-4, and the Center for AI Safety's May 2023 Statement on AI Risk, which placed extinction risk from AI alongside pandemics and nuclear war as a global priority. In November 2023 he spoke at the inaugural AI Safety Summit at Bletchley Park, where he argued that governments should halt the race toward AGI, and he has given evidence to legislators including a committee of the House of Commons of Canada. He is a frequent presence in debates, podcasts, and broadcast interviews, where he is known for blunt, vivid framings of AI danger. [1][11]
In 2024 Leahy and collaborators Gabriel Alfour, Chris Scammell, Andrea Miotti, and Adam Shimi published "The Compendium," a book-length argument that the "default path" of competitive AI development leads to catastrophe unless people intervene. It likens the arrival of a more capable artificial species to the relationship between humans and chimpanzees, warning that "if a new, more intelligent species were to appear on Earth, humanity would surrender its choice over what future we want to make manifest," while insisting that the outcome is not inevitable: "if you don't do something, it doesn't happen." [6]
The Compendium is closely tied to "A Narrow Path," a policy plan published on October 2, 2024 by the nonprofit ControlAI and authored by Andrea Miotti, Tolga Bilge, Dave Kasten, and James Newport. It proposes a phased strategy: a Safety phase that would prevent the development of superintelligence for twenty years through national legislation and limits on compute and capabilities, a Stability phase building durable international oversight, and a Flourishing phase that would pursue transformative AI only once it can be kept under human control. [7]
In 2026, after Conjecture wound down, Leahy joined ControlAI, where he had previously been an advisor, as its US Director, based in Washington, D.C. There he leads efforts to brief lawmakers on the risks of artificial superintelligence as part of the organization's "Direct Institutional Plan," a campaign that asks elected officials to support binding limits on the most powerful AI systems. He has emphasized that his stance is not opposition to AI as such but opposition to building uncontrollable superintelligence before humanity knows how to govern it, a position he sums up as choosing a "narrow path" between reckless acceleration and giving up the technology's benefits. [5][8]