DeepMind alumni’s AI beats OpenAI and Anthropic at research replication
▼ Summary
– Inherent, a London AI lab founded by DeepMind alumni, says its AI agent Faraday outperformed Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5 at reproducing published scientific findings without prior knowledge of the answers.
– Faraday runs on Qwen 3.6, a 27-billion-parameter model far smaller than its rivals, and was trained via reinforcement learning to develop “research taste” for choosing and designing experiments.
– Inherent used OpenAI’s GPT-5.5 Codex for coding tasks rather than building its own tool, reflecting a focus on scientific discovery over developing auxiliary software.
– The startup has 12 employees working in person in King’s Cross, London, and plans to grow headcount to 20–25 by year-end, potentially attracting DeepMind staff unsettled by Demis Hassabis’s new role.
– Co-founder Edward Hughes supports ending the U.K.’s “garden leave” practice, which he says delays departing employees from starting or joining rivals, giving U.S. startups a talent advantage.
Inherent, a London-based AI lab founded by Google DeepMind alumni, says its AI agent has outperformed much larger models from Anthropic and OpenAI at a specific scientific task, all while using a fraction of the computing resources.
Among the many startups spawned by Google DeepMind alumni, Inherent has largely flown under the radar. But while better-funded rivals are still heavy on promises and light on results, this London team is beginning to show what it has been working on.
Just weeks after coming out of stealth with a $50 million seed round, the British startup reports that its newly released AI agent, Faraday, has beaten larger, well-known models at a particular challenge: reproducing the findings of published scientific papers without being given the answer beforehand.
That might sound like a clever trick, especially given Inherent’s far more ambitious goal of building AI that can generate new scientific knowledge rather than just confirm existing results. But paper replication is a standard training exercise for human scientists as well, according to cofounder and chief scientist Edward Hughes. “Many PhD students actually start by doing this.”
Hughes told TechCrunch that beating other AI systems wasn’t the real objective. What mattered was 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.”
Here’s the detail that should make investors sit up: compared against Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5, both of which are much larger frontier-scale systems, Faraday runs on a comparatively small model called Qwen 3.6 with just 27 billion parameters. As a rough guide, parameters serve as a proxy for a model’s size and typically its training costs as well. Inherent also set the bar higher than simple accuracy. Beyond replicating results, the company wanted Faraday to show “research taste,” an instinct for which experiments are worth running and how to design them effectively.
Teaching something as elusive as taste is no easy task, which is where reinforcement learning comes into play. This training method rewards an AI system for good outcomes rather than giving it explicit rules to follow. Rather than training its agents primarily on the study of how science is conducted, Inherent leans on this reward-based approach, betting that it will generalize better to the longer-term goal of agents capable of contributing across many scientific fields.
“We’re always guided by that north star of building an AI scientist agent and imbuing our agents with taste,” Hughes said. That focus has also shaped what Inherent deliberately chooses not to build. Instead of developing its own coding tool, it had Faraday use OpenAI’s GPT-5.5 Codex, much the way human scientists rely on existing software rather than building everything from scratch, the company said.
Inherent is also trying to avoid creating agents that simply tell users what they want to hear. Hughes said the goal is modeled on his favorite kind of teammate, the one 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 collaborative instinct extends to how Inherent operates as a company. Its dozen employees all work in person out of an office in King’s Cross, the once-rundown London neighborhood that Google DeepMind’s presence helped turn into one of the world’s top AI hubs. “We believe that London is the place to be,” Hughes said.
Hughes is optimistic about London’s density of AI talent, but he has also added his voice to calls to end “garden leave,” the practice common in the U. K. of barring departing employees from joining or starting a rival company for months after they resign. It’s a restriction American researchers generally don’t face, giving U. S. startups a head start when hiring talent who’ve left a prior role. “This is a personal view rather than a company view, but I was affected by the garden leave problem,” he told TechCrunch.
Hughes eventually got around that constraint and started Inherent alongside two other DeepMind alumni and a fourth cofounder. The startup isn’t slowing down either. It plans to grow its headcount to “about 20 to 25” by the end of the year. Given its ambitions in world models as well, and with Demis Hassabis’s new role leaving some DeepMind staff unsettled, Inherent’s hiring push could make it an appealing landing spot for DeepMind employees weighing a move.
(Source: TechCrunch)
