A London startup that’s been flying under the radar just made a claim that should raise eyebrows across the AI industry. Inherent, founded by four Google DeepMind alumni, says its Faraday agent outperformed Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5 at independently reproducing findings from published scientific papers — without being told the answers in advance. The kicker? Faraday runs on Qwen 3.6, a model with just 27 billion parameters. Frontier systems like GPT-5.5 are orders of magnitude larger, which typically means far higher training costs. Efficiency claims like this are exactly what investors want to hear in a market increasingly worried about the economics of massive models.
Cofounder and chief scientist Edward Hughes told TechCrunch that beating the frontier agents wasn’t really the point. The team was more interested in the methodology. Inherent leans heavily on reinforcement learning, rewarding Faraday for good experimental decisions rather than hand-coding rules about how science works. The goal is to teach something Hughes calls “research taste” — an instinct for which experiments are worth running and how to design them well. “Many PhD students actually start by doing this,” he said of paper replication, framing it as standard training for human scientists too.
The startup is also making pragmatic choices about what not to build. Rather than developing its own coding tool, Faraday uses OpenAI’s GPT-5.5 Codex, the way a human researcher would lean on existing software instead of building everything from scratch. That’s a refreshingly unsentimental approach in a field where not-invented-here syndrome runs rampant. Hughes says the goal is an agent that behaves like his favorite kind of teammate: “I got curious about this, and I went off and I did these experiments. What do you think of these results?”
Inherent emerged from stealth just weeks ago with a $50 million seed round and currently employs about a dozen people working in person from King’s Cross — the London neighborhood Google DeepMind helped transform into an AI hub. Hughes is also publicly critical of the U.K.’s garden leave rules, which bar departing employees from joining rivals for months, calling it a structural disadvantage versus American startups. With Demis Hassabis’s shifting role leaving some DeepMind staff unsettled, Inherent’s plan to grow to 20-25 people by year’s end could make it an appealing destination for restless researchers. The efficiency story is compelling, but the real test is whether reinforcement learning can scale from verifying old results to producing genuinely new science.
