Faraday runs on a model a fraction of the size of Claude Opus 4.8 or GPT-5.5. According to its London-based creators, it still beat both at reproducing real scientific research from scratch.
Inherent, a startup founded by Google DeepMind alumni, says its AI agent independently replicated the findings of published scientific papers without being told the answers in advance, outperforming much larger frontier models in the process.
What Faraday Actually Did
Paper replication might sound like a narrow party trick next to Inherent's actual goal of building AI that discovers new science rather than just verifying old results. Cofounder and chief scientist Edward Hughes says it's a legitimate benchmark anyway.
"Many PhD students actually start by doing this," Hughes said.
The company's own bar was higher than raw accuracy. It wanted Faraday to show "research taste," an instinct for which experiments are worth running and how to design them well.
Why the Size Gap Matters
Faraday runs on Qwen 3.6, a 27-billion-parameter model, a small fraction of the size of the frontier-scale systems it was measured against.
Hughes said the size gap wasn't really the point of the exercise.
"What was most interesting to us about this was not so much the result of beating those frontier agents, which of course we liked, but was actually the way we went about building this," Hughes said.
The Method Behind the Result
Inherent leans on reinforcement learning, rewarding good outcomes rather than hard-coding rules, betting that approach generalizes better across scientific fields than training an agent narrowly on how research gets conducted.
"We're always guided by that north star of building an AI scientist agent and imbuing our agents with taste," Hughes said.
The company made a deliberate choice not to build its own coding tool, having Faraday use OpenAI's GPT-5.5 Codex instead, the same way a human scientist would lean on existing software rather than reinventing it.
Inherent is small by design. Its dozen employees all work in person out of King's Cross, the London neighborhood DeepMind's presence helped turn into a major AI hub, and the company plans to grow to only 20 to 25 people by year's end.
Hughes has also become an advocate against "garden leave," the UK practice of barring departing employees from joining or starting a rival company for months after they resign. American researchers generally don't face that restriction, giving US startups a real hiring advantage over UK ones.
"This is a personal view rather than a company view, but I was affected by the garden leave problem," Hughes said. He eventually got around the constraint himself before co-founding Inherent with three other DeepMind alumni.
What This Means for Miami
This is the same lesson MAIN covered in Nvidia's recent harness research, where a custom software wrapper took Claude Opus 5's score on a hard benchmark from 30% to 100% without changing the underlying model at all.
Faraday is a second, independent data point for the same idea: how an AI system is built and trained can matter more than which frontier model sits underneath it.
That's genuinely useful context for Miami's own AI companies competing without frontier-lab-scale compute budgets. Neural Earth, VeryAI and the rest of Miami's specialized AI companies aren't trying to out-spend OpenAI or Anthropic on raw model scale either.
Inherent's result is one more example that methodology and applied focus can be a real substitute for that kind of scale, not just a consolation prize for lacking it.
