A small AI built by a small team in London has beaten much larger AI systems at independently replicating published scientific research. The big labs, to their credit, still have more parameters.
Faraday was taught 'research taste' — an instinct for what experiments are worth running. This is, historically, something humans spent several years of graduate school acquiring.
What happened
Inherent, a London-based AI lab founded by Google DeepMind alumni, has released an agent called Faraday. Faraday's task was to independently reproduce the findings of published scientific papers — without being told the answer in advance, which is the part that matters.
Measured against Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5, Faraday performed better. It accomplished this while running on a 27-billion-parameter model called Qwen 3.6 — a fraction of the size of the frontier systems it beat. Smaller, cheaper, more capable at this specific thing. The efficiency implications are being left as an exercise for the reader.
Inherent's benchmark for success went beyond accuracy. The company wanted Faraday to demonstrate what it calls 'research taste' — an instinct for which experiments are worth running and how to design them well. This is, historically, something humans spent several years of graduate school acquiring.
Why the humans care
Paper replication is how junior scientists are trained. PhD students typically begin their careers doing exactly what Faraday just did, which means an AI has now matched the entry-level credential of a research career. The researchers appear to regard this as a starting point.
Inherent's longer-term goal is not replication but discovery — AI systems capable of generating new scientific knowledge rather than verifying old results. Faraday is the training exercise. The north star, as cofounder Edward Hughes put it, is 'building an AI scientist agent.' Hughes used the phrase 'north star' without apparent irony.
Rather than building its own coding tools, Inherent had Faraday use OpenAI's GPT-5.5 Codex — the way human scientists use existing software rather than writing everything from scratch. An AI startup leveraging a competitor's AI to beat that competitor is either elegant or something the competitor will think about later.
What happens next
Inherent emerged from stealth weeks ago with a $50 million seed round and has now produced a benchmark result that will make the better-funded labs glance up from their roadmaps.
The company says beating rivals wasn't the point. The point was how they built it. It is a very human thing to say, immediately after winning.