Ex-DeepMind Startup Says Smaller AI Outperformed OpenAI and Anthropic at Replicating Research

A Inherent, a London-based AI lab founded by former Google DeepMind researchers, told TechCrunch that its new agent, called Faraday, outperformed larger, better-known models from Anthropic and OpenAI at a specific task: independently reproducing the results of published scientific papers without being given the answer beforehand.

The announcement comes just weeks after the startup emerged from stealth mode with a $50 million seed round. According to the company, the result matters less for the scoreboard itself than for the technical path used to get there.

The Detail That Stands Out: Model Size

The central point of the story isn't just that it won, but what it won with. Faraday was compared against Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5, both frontier-scale systems. Inherent's agent, however, runs on Qwen 3.6, a comparatively small model with just 27 billion parameters.

As TechCrunch itself points out, parameter count is a proxy for a model's size and, typically, also for its training costs. In other words, Inherent claims to have achieved a superior result on a complex task using a fraction of its rivals' scale (and, presumably, their cost).

What Faraday Does and Why Replicating Papers Matters

Replicating the findings of an already-published paper might sound like a party trick next to Inherent's larger ambition, which is to build an AI capable of discovering new scientific knowledge, not just verifying old results. But according to co-founder and chief scientist Edward Hughes, replication is also a standard exercise in training human scientists. "A lot of PhD students actually start out doing this," he said.

The bar for success set by the company went beyond simple accuracy. Beyond reproducing results, Inherent wanted Faraday to demonstrate what Hughes calls "research taste," a kind of intuition about which experiments are worth running and how to design them well.

The Technical Bet: Reinforcement Learning Instead of Rules

Teaching something as intangible as "research taste" is difficult, and that's where reinforcement learning comes in. It's a training method that rewards the system for good results, rather than spelling out rules for it to follow.

According to Hughes, rather than training its agents mainly on the study of how science is conducted, Inherent bets on the reward-based approach because it believes it generalizes better toward the long-term goal: agents capable of contributing across many scientific fields. "We're always guided by this north star of building an AI scientist agent and imbuing our agents with taste," he said.

That focus also defined what the company decided not to build. Instead of developing its own coding tool, Inherent had Faraday use OpenAI's GPT-5.5 Codex, the same way human scientists rely on existing software instead of building everything from scratch. It's a telling detail: the rival's product becomes a component of the challenger's stack.

Hughes also said he tries to avoid agents that simply tell the user what they want to hear. The goal, he said, is modeled on his favorite type of teammate, the one who comes back saying: "I got curious about this, so I went and ran these experiments. What do you think of these results?"

Context: The Ex-DeepMind Race and "Garden Leave"

Inherent is one of several startups founded by former DeepMind employees, but so far it has received relatively little attention. While better-funded rivals have yet to show the world anything concrete, the London team has started to reveal what it has been building.

The company has a dozen employees, all working in person out of an office in King's Cross, the London neighborhood that DeepMind's presence helped turn into one of the world's leading AI hubs. The plan is to grow to about 20 to 25 people by the end of the year.

Hughes also spoke out against "garden leave," a common UK practice of barring departing employees from joining or founding a competing company for months after resigning, a restriction that, according to the report, American researchers generally don't face, giving US startups an edge in the fight for talent. "This is a personal view, not the company's, but I was affected by the garden leave problem," he said. He ended up working around the restriction and founded Inherent with two other ex-DeepMind researchers and a fourth co-founder.

What This Means for Developers in Brazil

For developers following the race for models outside the US, the practical takeaway here is the base choice: Inherent built its agent on Qwen, a family of open models, rather than on a closed proprietary model. This reinforces a trend that was already emerging: smaller, open models, when well tuned with techniques like reinforcement learning and careful agent design, can compete with frontier systems on specific tasks, at a much lower inference cost.

For Brazilian teams operating with a tight compute budget, it's a reminder that competitive advantage may lie less in the raw scale of the model and more in the engineering around it: reward function, agent architecture, and the decision to reuse existing tools (such as having an agent call Codex instead of building its own coding tool).

It's worth noting that, for now, the superior performance figures come from the company itself, and no independent benchmark has been published in the released material. What remains open, then, is whether the replication result holds up under external evaluation and whether the bet on "research taste" via reinforcement learning really generalizes to the more ambitious goal of discovering new knowledge.