Inherent's Faraday AI Agent Topples Larger Models at Replicating Science
Inherent, the London AI lab founded by DeepMind alumni, is proving that bigger isn't always better. Its new AI agent, Faraday, has just outperformed Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 at the tough task of independently reproducing published scientific papers — and it does it with a model 27 billion parameters instead of the hundreds billion typical of frontier systems.
The surprise isn't just that Faraday bested bigger models; it's how. Co-founder and chief scientist Edward Hughes told TechCrunch the real story wasn't the win-loss record but the approach. "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."
Faraday runs on Qwen 3.6, a comparatively tiny model with 27 billion parameters. By contrast, Claude Opus 4.8 and GPT-5.5 are an order of magnitude larger. Yet Inherent's success metric went beyond raw accuracy. The team wanted Faraday to demonstrate "research taste" — an instinct for which experiments are worth running and how to design them well. That's a notoriously intangible quality to encode.
The secret ingredient is reinforcement learning. Rather than training agents on textbooks of how science is conducted, Inherent rewards the AI for good outcomes and lets it figure out the rest. "We're always guided by that north star of building an AI scientist agent and imbuing our agents with taste," Hughes said. The bet is that a reward-based approach will generalize better to longer-horizon scientific contribution.
Inherent could have built its own coding tool, but deliberately didn't. Faraday uses OpenAI's GPT-5.5 Codex instead — the same way human scientists lean on existing software rather than building everything from scratch. The choice reflects a broader philosophy: the best AI teammate is one that comes back and says, "I got curious about this, and I went off and I did these experiments. What do you think of these results?"
That collaborative instinct shapes how Inherent the company operates, too. All dozen employees work in person out of an office in King's Cross, the London neighborhood that DeepMind's presence helped transform into one of the world's top AI hubs. "We believe that London is the place to be," Hughes said. He's also been vocal about ending "garden leave", the U.K. practice of barring departing employees from joining rivals for months after resignation. American researchers generally don't face this, giving U.S. startups a hiring edge. "This is a personal view rather than a company view, but I was affected by the garden leave problem," Hughes told TechCrunch.
The team isn't slowing hiring. Inherent plans to grow to about 20–25 employees by year's end. With Demis Hassabis' new role leaving some DeepMind staff unsettled, Inherent could become an appealing landing spot for colleagues considering a move. Pictured from left to right: co-founders Louis Kirsch, Kaloyan Aleksiev, Tantum Collins, and Edward Hughes.
The Technology Behind Faraday's Replication Edge
Faraday's architecture is purpose-built for reproducibility. The model ingests a scientific paper and must reproduce the results without being told the answer in advance. This "blind replication" task strips away shortcuts and forces the model to genuinely understand the experimental methodology.
Qwen 3.6's 27 billion parameters are sufficient because the task is narrow enough that parameter efficiency matters more than raw scale. Inherent paired the model with a reinforcement learning loop that penalizes hand-waving and rewards concrete, verifiable steps. Over many iterations, the agent learned which experiment-design choices actually move the needle.
The reinforcement signal comes from outcome-based feedback: did the reproduced results match the published findings within acceptable tolerance? If yes, the pathway is reinforced. If no, the agent gets a correction signal and must rethink its approach. This is unlike typical AI training that optimizes for next-token prediction or imitation of human behavior.
Why "Research Taste" Matters More Than Model Size
Inherent's co-founders, Edward Hughes, Louis Kirsch, Kaloyan Aleksiev, and Tantum Collins, all came from DeepMind, where they worked on various aspects of agent research. Their shared observation was that the biggest models often fail at tasks requiring sustained, autonomous problem-solving. They can mimic patterns but break when asked to plan across multiple steps.
Taste, in this context, means the agent can evaluate its own intermediate steps and say, "This experiment design looks promising; let me probe deeper," or "This path is hitting diminishing returns; let me pivot." Reinforcement learning provides the scaffold, but the model's internal value function does the actual judgment.
Hughes put it plainly: "We're not trying to build a system that just does what it's told. We're trying to build a system that can hold a real scientific conversation, asking good questions, designing decent experiments, and knowing when it's stuck."
Inherent's Company Philosophy Rooted in London Talent
The company's physical setup is intentional. Inherent's entire team works out of a single office in King's Cross. Hughes has been vocal about why location matters, especially regarding talent mobility.
"Garden leave" is a UK-specific restriction that can keep a researcher out of the job market for six months or more after leaving a role. In the U.S., such restrictions are rare and often unenforceable. Hughes told TechCrunch the practice puts UK startups at a disadvantage when trying to hire experienced talent.
"It's a personal view rather than a company view, but I was affected by the garden leave problem," he said. Eventually, he negotiated his way around it and started Inherent with his co-founders. The hiring push could make the company a haven for DeepMind alumni weighing a move, especially as Hassabis takes on new responsibilities that may unsettle some staff.
The company aims to double its headcount to 20–25 by the end of the year. Given its ambitions in world models, AI systems that learn broader representations of how the world works, Inherent's growth could accelerate work on longer-horizon scientific agents.
What Faraday's Success Means for AI Research
Faraday's win over larger models is more than a benchmark flex. It suggests that training methodology and architectural choices can outweigh raw parameter count for certain classes of problems. The reinforcement learning loop that rewards genuine outcome improvement, not just imitation, appears to be the differentiator.
For the wider AI research community, the result is a data point against the assumption that bigger always means better. It also validates the "AI scientist" framing Inherent has been pushing since emerging from stealth with a $50 million seed round weeks earlier. If a 27-billion-parameter model can outperform 100-billion-parameter systems on reproducibility, the door opens for other teams to experiment with efficient training pipelines.
Of course, paper replication is just one task. Inherent's longer-term goal is an agent that can contribute new knowledge across scientific fields, not just verify existing results. But if the taste-driven approach scales, the implications ripple outward: more capable AI at lower compute cost, and AI systems that are genuinely useful research partners rather than one-trick demos.
The Road Ahead for Inherent
Inherent has a clear runway. The $50 million seed round gave the company breathing room to ship Faraday and start iterating based on user feedback. The headcount growth to 20–25 will add both research depth and domain expertise, especially as the team explores world models and multi-domain scientific contribution.
Hughes has been frank about the company's position: they're not chasing funding for its own sake. They're building toward an AI scientist that can genuinely contribute, and they'll invest in whatever capacity, headcount, compute, algorithmic innovation, gets them there.
The garden-leave issue may follow them into hiring conversations, but Hughes has shown he's willing to challenge industry norms to get the team he wants. If Inherent keeps punching above its weight class on benchmarks like paper replication, those conversations will only get easier.
Source: TechCrunch article "Inherent, founded by DeepMind alumni, says its AI 'teammate' just outperformed Anthropic and OpenAI at replicating research" (August 22, 2026)
