Inherent, a London AI lab founded by Google DeepMind alumni, says its AI agents outperformed much larger models from Anthropic and OpenAI by a fraction of the size.
Among the startups launched by Google DeepMind alumni, Inherent is relatively unnoticed. But while their better-funded rivals have yet to show the world anything concrete, the London-based team is starting to share what they’ve been building.
Just weeks after emerging from stealth with a $50 million seed round, the British startup said its newly released AI agent, Faraday, outperformed larger, better-known models on the specific task of independently reproducing the results of published scientific papers without knowing the answers in advance.
That may sound like just a party trick, given Inherent’s much loftier goal: building an AI that can discover new scientific knowledge as well as verify old results. But paper replication is also standard training for human scientists, said co-founder and chief scientist Edward Hughes. “A lot of PhD students actually start by doing this.”
Hughes told TechCrunch that it wasn’t about beating other AI systems at this task. how did they get there? “What was most interesting to us about this was not so much the result of defeating a frontier agent (which we liked, of course), but the way we actually built this.”
Here’s the part that will attract investors’ attention. Measured against Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5, both much larger frontier-scale systems. Faraday runs on a relatively small model called Qwen 3.6, which has just 27 billion parameters. (Roughly speaking, a “parameter” is a proxy for the size of a model, and typically its training cost as well.) The bar for Inherent’s success was higher than mere accuracy. In addition to reproducing the results, they wanted Mr. Faraday to show “research sense,” an intuition about what experiments are worth running and how to properly design them.
It’s difficult to teach something as intangible as taste, which is where reinforcement learning comes in. This is a training method that rewards an AI system for good results rather than specifying rules for it to follow. Rather than training agents based primarily on research into how science itself is done, Inherent relies on this reward-based approach, betting that it can better generalize with its long-term goal of agents that can contribute across many scientific fields.
“We’re always guided by the north star of building AI scientist agents and imbuing them with sense,” Hughes says. That focus also shaped what Inherent chose not to build. Rather than develop its own coding tools, the company said it had Faraday use OpenAI’s GPT-5.5 codex instead. This is similar to how human scientists rely on existing software rather than building everything themselves.
Inherent also tries to avoid building agents that simply tell users what they want to hear. Instead, Hughes said this goal was modeled after his favorite teammate, the type of teammate who comes back and says:
That collaborative instinct extends to how Inherent operates as a company. The company’s dozens of employees all work directly from its King’s Cross office. King’s Cross is a once run-down London neighborhood that has been transformed into one of the world’s top AI hubs thanks to the presence of Google DeepMind. “We believe London is the best location,” Hughes said.
Hughes is bullish about the density of AI talent in London, but also calls for an end to “garden leave” – a common practice in the UK where departing employees are barred from joining a rival company or starting a business for several months after leaving the company. This is a limitation that U.S. researchers don’t typically face, and it gives U.S. startups a head start in hiring people who left their previous jobs. “This is a personal view, not a company view, but I was affected by the gardening leave issue,” he told TechCrunch.
Hughes eventually circumvented that constraint and founded Inherent with two other DeepMind alumni and a fourth co-founder. It won’t start up too slowly. The company plans to increase the number of employees to “approximately 20 to 25” by the end of the year. Inherent’s hiring drive could be an attractive landing spot for DeepMind employees considering a career change, given the company’s ambitions to become a world model, and given the anxiety some DeepMind staff have about Demis Hassabis’ new role.
Pictured from left to right: Original co-founders Louis Kirsch, Kaloyan Alexiev, Tantum Collins and Edward Hughes.
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
