Inherent, a London AI lab founded by Google DeepMind alumni, says its agent Faraday has outperformed larger, better-known models on a specific scientific task. According to the company, Faraday—running on a comparatively small Qwen 3.6 model with about 27 billion parameters—successfully reproduced the findings of published scientific papers without being told the answers in advance.
The startup announced the result a few weeks after emerging from stealth and closing a $50 million seed round. Inherent framed the exercise as more than a benchmark win: it set a higher bar than raw accuracy, asking Faraday to demonstrate what it calls "research taste"—the instinct to choose worthwhile experiments and design them well.
In their reported comparisons Faraday was measured against Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5, both of which are substantially larger, frontier-scale systems. Edward Hughes, Inherent’s cofounder and chief scientist, said the company cared as much about the method as the outcome: how the team engineered the agent mattered more than simply beating other frontier agents.
Approach: reinforcement learning and pragmatic tooling
To instill an intangible quality like research taste, Inherent relied on reinforcement learning, a training method that rewards desirable outcomes rather than prescribing explicit rules. The company leaned on this reward-based approach rather than primarily training on meta-scientific texts about how science is conducted, believing this will generalize better for agents intended to contribute across multiple scientific fields.
Pragmatism also guided tool choices. Instead of building its own coding tool, Inherent had Faraday use OpenAI’s GPT-5.5 Codex—analogous to how human scientists often use existing software rather than building every tool from scratch.
Aiming for a curious, collaborative agent
Hughes emphasized that the goal is not an agent that merely tells users what they want to hear. The desired kind of teammate is one that reports back: “I got curious about this, and I went off and I did these experiments. What do you think of these results?” That collaborative instinct shapes both product design and company culture.
Team, location and hiring plans
Inherent currently employs roughly a dozen people, all working in person from an office in London’s King’s Cross. Hughes said the company believes London—boosted by DeepMind’s presence—remains a prime location for AI talent.
The startup plans to expand its headcount to “about 20 to 25” by the end of the year. With ambitions in world models and amid reported shifts at DeepMind following Demis Hassabis’s new role, Inherent could become an attractive option for researchers considering a move.
Industry context: garden leave
Hughes also spoke out against the U.K. practice of "garden leave," which can restrict departing employees from joining or founding rival companies for months after they resign. He called this a disadvantage compared with the U.S., where such constraints are less common, and noted he was personally affected by the issue.
Conclusion
Inherent’s claim—that an agent running on a 27-billion-parameter Qwen 3.6 model can reproduce published results and outperform larger systems on that task—suggests that model scale alone does not determine capability. The company’s use of reinforcement learning to cultivate research taste is positioned as a step toward building AI agents that can not only verify scientific results but help discover new knowledge.
