Inherent, founded by DeepMind alumni, says its AI ‘teammate’ just outperformed Anthropic and OpenAI at replicating research

British AI laboratory Inherent, founded by former Google DeepMind researchers, announced that its newly released AI agent, Faraday, has outperformed larger models from Anthropic and OpenAI in independently reproducing the findings of published scientific papers. The achievement was measured against Anthropic’s Claude Opus 4.8 and OpenAI’s GPT‑5.5, both considerably larger frontier‑scale systems, while Faraday runs on the comparatively modest Qwen 3.6 model with just 27 billion parameters. Co‑founder and chief scientist Edward Hughes emphasized that the breakthrough lies not merely in beating the rivals but in the method used to train Faraday, which combines reinforcement learning to instill a sense of “research taste”—the ability to judge which experiments are worthwhile and how to design them—rather than relying solely on explicit rule‑based instruction.

The significance of Faraday’s performance stems from its focus on a core scientific practice: paper replication, a standard exercise for human researchers and PhD students. Inherent deliberately chose this task as a training ground because it requires both accurate result reproduction and the nuanced judgment of experimental design, qualities that are difficult to encode directly. By rewarding the AI for successful outcomes through reinforcement learning, the team hopes the model will generalize to broader scientific discovery tasks, eventually contributing novel insights across multiple fields. In a pragmatic move, Faraday leverages OpenAI’s GPT‑5.5 Codex for coding needs, mirroring how human scientists rely on existing software rather than building tools from scratch, and the company avoids creating agents that simply echo user expectations, aiming instead for a collaborative “teammate” that proactively conducts experiments and reports findings.

Inherent operates out of a King’s Cross office with a dozen staff members and plans to expand to 20‑25 employees by year‑end, positioning itself as an attractive destination for AI talent, especially given recent uncertainties at DeepMind following Demis Hassabis’s role change. Hughes, who has spoken out against the UK’s “garden leave” restrictions that delay departing employees from joining rivals, sees London’s dense AI ecosystem as a strategic advantage. The startup’s broader ambitions include developing world models, and its early success with Faraday—achieving high‑quality scientific replication with a fraction of the computational resources of its competitors—signals a potential shift in how AI research assistants are built and deployed.

⚡ Effects Interpreter