Inherent, a London AI lab founded by Google DeepMind alumni, has emerged from stealth with a bold claim: its AI agent, Faraday, outperformed frontier systems from Anthropic and OpenAI at replicating findings from published scientific papers. The task required the agent to independently reproduce experimental results without being told the answer in advance, and Inherent says Faraday passed that test more effectively than larger and better-funded rivals.

What makes the result notable is the size of the model behind Faraday. Inherent says the agent runs on Qwen 3.6, a model with only 27 billion parameters, while the competing systems it measured itself against — Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5 — are far larger. Parameter count is not a perfect measure of capability, but it is a useful proxy for model size and training cost, which means Inherent is claiming an efficiency advantage as well as a performance win.

Key facts at a glance

• Inherent is a London-based AI startup founded by Google DeepMind alumni, backed by a $50 million seed round.
• Its AI agent, Faraday, is built on Qwen 3.6, a 27-billion-parameter model.
• The company says Faraday outperformed Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5 at reproducing scientific research results.
• Inherent attributes the result to reinforcement learning and a focus on “research taste,” not to raw model size.
• The startup has around 12 employees and plans to grow to 20-25 by the end of the year.

Why replication is a meaningful test

Replicating a published study is one of the most basic and important activities in science. It forces the researcher to understand the original methodology, identify the key variables and control conditions, and judge whether the evidence actually supports the claims. For human scientists, it is often the first real test of independence: can you produce the same result using only the methods described in a paper? Edward Hughes, Inherent’s co-founder and chief scientist, says the same exercise is a natural starting point for AI.

“Many PhD students actually start by doing this,” Hughes said. The implication is that an AI capable of replicating research has learned more than memorized answers. It has learned to read a methods section, form a plan, run a simulation or experiment, and evaluate the outcome. That is why Inherent set up the benchmark the way it did: Faraday was not given the correct answer in advance. It had to design and execute its own path to the result.

Hughes was careful to frame the result in context. “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.” That comment points to Inherent’s broader thesis: performance is a function of training approach, not just scale.

Efficiency through reinforcement learning

Most large language models are trained in two broad phases. The first is pretraining, during which the model learns patterns from enormous amounts of text. The second is often supervised fine-tuning, where humans provide examples of good responses. Inherent is taking a different route. Instead of relying primarily on examples of how science is conducted, it uses reinforcement learning, a method in which the AI is rewarded for successful outcomes rather than told exactly how to behave.

Reinforcement learning has been central to recent advances in reasoning models, particularly in mathematics and coding, but Inherent is applying it to scientific judgment. The reward signal is designed to push Faraday toward better experimental designs, more efficient use of tools, and results that genuinely match what a paper claims. Inherent believes this approach will generalize better than a system that simply imitates training examples, because the agent is learning principles of good experimentation rather than pattern-matching particular papers.

The company also says this helps explain its focus on “research taste.” A scientist with good taste knows which experiments are worth running, how to design them cleanly, and how to interpret ambiguous results. That instinct is difficult to formalize, but it is exactly what a reinforcement learning system can, in principle, learn through trial and error. Hughes said Inherent is guided by “that north star of building an AI scientist agent and imbuing our agents with taste.”

Using external tools like human scientists

Another notable detail in Inherent’s approach is its willingness to use other companies’ tools. Rather than build its own coding assistant, Inherent had Faraday use OpenAI’s GPT-5.5 Codex for programming tasks. Hughes compared this to how human scientists rely on existing software and instruments instead of building everything from scratch. An AI scientist should be judged on its ability to do science, not on whether it wrote every line of code in its own library.

This modular philosophy extends to the way Faraday carries out research. Inherent says it is trying to avoid building agents that simply tell users what they want to hear. Instead, the ideal behavior is modeled on the best kind of human collaborator: someone who comes back to the team and says, “I got curious about this, and I went off and did these experiments. What do you think of these results?” The vision is an AI teammate that can operate semi-autonomously while still checking in with human judgment.

London roots and DeepMind connections

Inherent was founded by four people with close ties to Google DeepMind: Edward Hughes, Louis Kirsch, Kaloyan Aleksiev, and Tantum Collins. The company only left stealth a few weeks ago, announcing a $50 million seed round. Compared with some better-funded AI startups, it is still small, and its London location is a deliberate choice. The entire team, around 12 people at the moment, works in person out of an office in King’s Cross, the neighborhood that Google DeepMind helped transform into one of the world’s main AI hubs.

“We believe that London is the place to be,” Hughes said. The city has a dense pool of AI researchers, many of whom came through DeepMind or labs that orbit around it. Inherent’s team is small by design, but it plans to expand. Hughes says the company wants to grow to about 20 to 25 employees by the end of the year. Given that it has ambitions in world models as well as scientific agents, and that Demis Hassabis’ new role has left some DeepMind employees unsettled, Inherent’s hiring push could make it an appealing destination for DeepMind researchers considering a move.

Garden leave and the talent bottleneck

One issue that affects many British AI startups is “garden leave,” a practice common in the U.K. that prevents departing employees from joining or starting a rival company for months after they resign. American researchers are generally not subject to the same restriction, which gives U.S. startups a faster path to hiring people who have left prior roles. Hughes said he was personally affected by the problem before starting Inherent, though he emphasized it was a personal view rather than a company position.

“This is a personal view rather than a company view, but I was affected by the garden leave problem,” he said. Long garden leave periods can slow the formation of new startups and delay research momentum, especially in fields like AI where progress moves quickly. Policymakers have begun to debate whether the practice is appropriate for a sector where talent mobility is central to innovation. For now, Inherent has navigated the constraint, built a team, and started shipping results.

The company’s next challenge is to move from reproducing old results to producing new ones. Faraday’s performance on replication is an early signal that a small, focused team can build an agent with strong scientific judgment. Inherent’s broader goal, building an AI scientist that can make discoveries, is far more difficult, but the team believes its reinforcement learning approach gives it a viable path. With a compact model, a steady seed round, London’s talent density, and a culture that values curiosity over confirmation, Inherent is positioning itself as an unusual but serious contender in the race to automate science.

Source:TechCrunch News