London-based for-profit AI safety company working on Cognitive Emulation, an approach to building controllable, bounded AI systems that reason transparently.
London-based for-profit AI safety company working on Cognitive Emulation, an approach to building controllable, bounded AI systems that reason transparently.
People– no linked people
Updated 04/02/26Funding Details
Updated 04/02/26- Annual Budget
- -
- Current Runway
- -
- Funding Goal
- -
- Funding Raised to Date
- $25,000,000
Org Details
Updated 04/02/26Conjecture was incorporated on 9 March 2022 as a private limited company in London, UK (Companies House number 13966466), founded by Connor Leahy (CEO), Sid Black, and Gabriel Alfour. The three founders met through EleutherAI, the open-source AI research collective co-founded by Leahy and Black in 2020, and decided to form a professional company dedicated to scalable AI alignment research. The company's core technical focus is Cognitive Emulation (CoEm), a safety approach aimed at building powerful, useful AI systems that are bounded, controllable, and safe by emulating human-like cognition. CoEm breaks complex cognitive tasks into primitive building blocks and uses novel fine-tuning, synthetic data, and scaffolding methods to compose LLM/software hybrid systems that can perform these building blocks with high reliability. The approach is intended as a practical step toward aligned AI, distinct from both uncontrolled frontier scaling and narrow task-specific AI. During its first year, Conjecture's research covered mechanistic interpretability, sparse autoencoders, the polytope lens for interpreting neural networks, and the influential Simulators paper on LLM behavior. After an internal pivot in late 2022, the team committed fully to the CoEm agenda going into 2024, and some researchers who had focused on interpretability departed at that time. Beyond research, Conjecture has been active in AI safety field-building: it ran the Refine incubator for conceptual alignment research bets in 2022-2023, and currently fiscally sponsors the Alignment Research Engineer Accelerator (ARENA) and the London cohort of SERI MATS, both programs aimed at training AI safety researchers and engineers. Open Philanthropy has supported the ARENA program with grants totaling over $1 million. Conjecture has raised approximately $25 million in total funding, primarily from venture capital investors including Nat Friedman, Daniel Gross, Patrick and John Collison, Andrej Karpathy, and Arthur Breitman. The team grew to approximately 22 people at its peak in 2022-2023 before stabilizing at around 13 employees. Connor Leahy has been a prominent public voice on AI risk, speaking at the UK AI Safety Summit at Bletchley Park in November 2023 and testifying before the Canadian Parliament's Ethics Committee on AI safety regulation.
Theory of Change
Updated 04/02/26Conjecture believes the core problem of advanced AI is the control problem: building systems that do exactly what they are told and no more. Their theory of change is that current frontier AI architectures are fundamentally uncontrollable because they optimize in opaque, unpredictable ways. By developing Cognitive Emulation, Conjecture aims to produce a new class of AI systems whose reasoning is decomposed into auditable, human-legible primitives, making it possible to bound their capabilities and verify their behavior. This predictable boundedness is intended as a near-term, practical safety property that can be achieved before AGI, creating systems that are safe and useful for businesses. In parallel, Conjecture supports field-building through ARENA and MATS to grow the number of technically skilled AI safety researchers, increasing the overall capacity of the field. Connor Leahy also advocates publicly and in policy settings for compute-based regulation and international coordination to slow frontier AI development.
Grants Received
Updated 04/02/26Projects– no linked projects
Updated 04/02/26Discussion
No comments yet. Be the first to share your thoughts.