Inherent, a London AI lab founded by Google DeepMind alumni, says its AI agent just outperformed much larger models from Anthropic and OpenAI using a fraction of the size. The startup emerged from stealth weeks ago with a $50 million seed round, and its first public release, an agent called Faraday, can independently reproduce the findings of published scientific papers without being told the answer in advance.
That task matters because paper replication is a standard training exercise for human scientists. "Many PhD students actually start by doing this," said Edward Hughes, Inherent's cofounder and chief scientist. The company's larger goal is building AI that can discover new scientific knowledge, not just verify old results.
A small model, a tougher test
Faraday runs on Qwen 3.6, a model with 27 billion parameters. That's comparatively tiny next to Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5, both frontier-scale systems that Faraday beat at the replication task. Parameters are a rough proxy for a model's size and training costs; doing more with less is the point.
Inherent also set a higher bar than accuracy. Beyond reproducing results, the company wanted Faraday to demonstrate "research taste" - an instinct for which experiments are worth running and how to design them well. That's a harder quality to teach than raw correctness.
Hughes said the result itself wasn't the main takeaway. "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."
Rewards over rules
Inherent trains Faraday using reinforcement learning, a method that rewards good outcomes rather than spelling out explicit rules. The company avoids training its agents primarily on the study of how science is conducted, betting instead that the reward-based approach will generalize better to its long-term goal: 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 shapes what Inherent chooses not to build. Rather than developing its own coding tool, Faraday uses OpenAI's GPT-5.5 Codex - the way human scientists rely on existing software instead of building everything themselves. The company also wants agents that push back, not just tell users what they want to hear. Hughes described the goal as modeled on his favorite kind of teammate: "I got curious about this, and I went off and I did these experiments. What do you think of these results?"
For researchers in fields like biology, chemistry, or materials science, the practical takeaway is that AI tools are moving from answering questions to running investigations. An agent that can replicate a published result - and eventually propose its own experiments - changes what automation can do in a lab setting. The key shift is reinforcement learning over rule-following, which lets the system develop judgment rather than memorize procedures.
Those working in research-heavy roles may want to track how Inherent's approach compares to the frontier models they already use. If a 27-billion-parameter agent can match or beat systems many times its size at a scientific task, the cost of useful AI assistance in research could drop considerably. That's a concrete efficiency gain for teams weighing which tools to build workflows around. For a deeper look at how AI fits into scientific workflows, see AI Learning Path for Research Scientists or the broader AI for Science & Research collection.
London roots, hiring plans
Inherent's dozen employees work in person out of an office in King's Cross, the London neighborhood where Google DeepMind's presence helped build one of the world's top AI hubs. "We believe that London is the place to be," Hughes said. The startup plans to grow to roughly 20 to 25 people by the end of the year.
Hughes has also added his voice to calls to end "garden leave," the U.K. practice of barring departing employees from joining or starting a rival company for months after they resign. American researchers generally don't face that restriction, giving U.S. startups a head start on 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 said.
Why this matters for science and research professionals
For researchers, the practical signal from Inherent's work is that smaller, specialized agents are becoming viable alternatives to general-purpose frontier models. A lab that can't justify the cost of running Claude Opus 4.8 or GPT-5.5 at scale may still get useful scientific results from a compact model trained specifically for research tasks. The replication benchmark is also a useful way to evaluate any AI tool you might adopt: if it can't reproduce known results, it's not ready to help generate new ones.
Your membership also unlocks:
