Inherent, a London lab founded by Google DeepMind alumni, says its Faraday agent beat Claude Opus 4.8 and GPT-5.5 at independently reproducing the findings of published scientific papers. The catch: Faraday runs on Qwen 3.6, a 27B-parameter open model that is a fraction of the size of its frontier rivals.
The startup emerged from stealth only weeks ago with a $50M seed round. Its cofounders include chief scientist Edward Hughes, who frames paper replication as a standard training exercise for human scientists. “Many PhD students actually start by doing this,” he told TechCrunch.
Inherent judged Faraday on more than accuracy. The team wanted the agent to show research taste – an instinct for which experiments are worth running and how to design them. Rather than training on how science is conducted, it leaned on reinforcement learning that rewards good outcomes, betting the approach generalizes to its longer-term goal: agents that contribute across scientific fields.
That discipline shapes what Inherent refuses to build. Faraday leaned on OpenAI’s GPT-5.5 Codex for coding rather than a custom tool – much as human scientists use existing software instead of writing everything from scratch.
The London team numbers about a dozen, all in a King’s Cross office the founders call a magnet for AI talent, with plans to reach 20-25 by year-end. Hughes wants the UK to scrap “garden leave” restrictions, arguing they slow departed employees from joining or founding rivals and hand US startups a hiring edge.
The result is a proof point for a smaller-model thesis: taste-driven training on compact architectures may beat raw scale on scientific tasks – and London is betting it can keep the talent that makes it work.
