Inherent, a London AI lab founded by Google DeepMind alumni, is making a precise claim about its new AI agent: Faraday can independently reproduce findings from published scientific papers, and it did so while running on a much smaller model than the frontier systems it was measured against.
The company is not presenting paper replication as the final destination. Its larger ambition is to build AI that can help discover new scientific knowledge. But for a young lab that recently emerged from stealth with a $50 million seed round, Faraday is an early signal of how Inherent wants to approach scientific AI.
What Faraday Was Asked To Do
Faraday was evaluated on a task that is familiar in scientific training: reproducing the findings of published papers without being given the answer in advance. Edward Hughes, Inherent cofounder and chief scientist, described this as a normal starting point for researchers, saying, “Many PhD students actually start by doing this.”
The point is not simply to copy an outcome. A system has to understand what a paper is trying to show, decide what work is needed to test the finding, and carry out the steps well enough to reach the same result. For an AI scientist agent, that makes replication a useful bridge between reading research and doing research.
Inherent says Faraday outperformed Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5 on this specific task. Those systems are described as larger, frontier-scale models, while Faraday runs on Qwen 3.6, a model with 27 billion parameters.
That contrast is central to the company’s pitch. Parameters are broadly used as a proxy for model size and often for training costs. Inherent’s claim is that a carefully trained agent using a comparatively tiny model can compete with, and in this case beat, much larger systems in a narrow but meaningful scientific workflow.
Why Inherent Cares About Research Taste
Inherent’s benchmark was not limited to whether Faraday could reach the right answer. The company also wanted the system to show “research taste,” meaning an instinct for which experiments are worth doing and how to design them.
That is a harder target than ordinary accuracy. In scientific work, a technically correct experiment can still be unhelpful if it tests the wrong thing, wastes effort, or fails to clarify the key question. Inherent is trying to train agents that make better choices about what to investigate, not just agents that execute instructions.
Hughes told TechCrunch that beating other systems was not the main lesson. “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.”
The company’s method leans on reinforcement learning, a training approach that rewards an AI system for good outcomes rather than giving it a fixed rulebook. Inherent is betting that this reward-based approach will generalize better than training agents mainly on descriptions of how science is conducted.
That choice reflects the company’s longer-term goal: AI agents that can contribute across many scientific fields. Hughes described the direction this way: “We’re always guided by that north star of building an AI scientist agent and imbuing our agents with taste.”
A Teammate, Not Just A Tool
Inherent is framing Faraday less as a standalone answer machine and more as a scientific teammate. That distinction matters because the company says it wants agents that can explore, test and return with useful results, rather than simply reflecting a user’s expectations back at them.
Hughes said the goal is modeled on the kind of collaborator who 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 idea also shapes what Inherent is choosing not to build. Rather than developing its own coding tool, the company had Faraday use OpenAI’s GPT-5.5 Codex. Inherent compares that to how human scientists rely on existing software instead of building every supporting tool themselves.
This approach gives Faraday a practical role: it can use available systems as part of a broader research workflow. The emphasis is on orchestration, judgment and experimental design, not on owning every layer of the stack.
London Is Part Of The Strategy
Inherent’s story is also tied to London’s AI ecosystem. The company’s dozen employees work in person from an office in King’s Cross, the London neighborhood whose AI profile has been shaped in part by Google DeepMind’s presence.
Hughes is clear about the company’s location choice. “We believe that London is the place to be,” he said.
At the same time, he has joined calls to end “garden leave,” a common U.K. practice that can stop departing employees from joining or starting a rival company for months after they resign. Hughes said this was a personal view, not a company position, but added: “This is a personal view rather than a company view, but I was affected by the garden leave problem.”
The contrast with U.S. hiring conditions is part of the issue. American researchers generally do not face the same restriction, which can give U.S. startups a head start when hiring talent leaving a prior role.
What Comes Next For Inherent
Inherent was founded by Hughes alongside two other DeepMind alumni and a fourth cofounder. The source article identifies the cofounders as Louis Kirsch, Kaloyan Aleksiev, Tantum Collins and Edward Hughes.
The company is still small, but it is already planning to expand. Inherent expects to grow its headcount to “about 20 to 25” by the end of the year.
The hiring push comes as the company pursues ambitions in world models as well as AI scientist agents. The source article also notes that Demis Hassabis’s new role has left some DeepMind staff unsettled, which could make Inherent an appealing landing spot for employees considering a move.
For now, Faraday is not evidence that AI has solved scientific discovery. It is a narrower demonstration: an agent trained for research judgment, using a smaller model, can reproduce scientific paper findings in a way Inherent says outperformed larger systems from Anthropic and OpenAI. For a lab trying to build AI teammates for science, that is the first concrete result it wants the market to notice.