Inherent's 27B-Parameter Model Outperforms Anthropic and OpenAI at Research Replication

The task Faraday tackled sounds simple but is extremely challenging: independently reproducing the findings of published scientific papers without being told the answer in advance. For an AI system, this means the agent must understand the paper's experimental design, write code, run analyses, and reach conclusions consistent with the original authors — far beyond the ability of traditional large models to 'recall' training data. Inherent co-founder and chief scientist Edward Hughes explained to TechCrunch that paper replication is a standard training exercise for human scientists — 'many PhD students actually start by doing this.' He also stressed that beating other AI systems was not the point: '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.'

What really catches an investor's eye is Inherent's technical approach. Instead of chasing bigger models and more parameters, Faraday runs on Qwen 3.6 with just 27 billion parameters — a size proxy typically linked to training cost, and an order of magnitude smaller than Claude Opus 4.8 and GPT-5.5. Inherent's bar for success is also higher than simple accuracy: beyond replicating results, it wants Faraday to demonstrate 'research taste' — an instinct for which experiments are worth running and how to design them well. Teaching something as intangible as taste is hard, which is where reinforcement learning comes in: a training method that rewards an AI system for good outcomes rather than spelling out rules for it to follow. Rather than training its agents primarily on meta-knowledge about how science is conducted, Inherent leans on this reward-based approach, betting it will generalize better to the longer-term goal of agents capable of contributing across many scientific fields. Hughes said: 'We're always guided by that north star of building an AI scientist agent and imbuing our agents with taste.'

This 'taste-first' philosophy also shapes what Inherent chooses not to build. For example, rather than developing its own coding tool, the team has Faraday use OpenAI's GPT-5.5 Codex instead — much the way human scientists lean on existing software rather than building everything themselves. Inherent is also trying to avoid building agents that simply tell users what they want to hear. Hughes said the goal is modeled on his favorite kind of teammate — the kind who comes back and says: 'I got curious about this, and I went off and I did these experiments. What do you think of these results?' That collaborative instinct extends to how Inherent operates as a company: its roughly dozen employees all work in person out of an office in King's Cross — the once-rundown London neighborhood that Google DeepMind's presence helped turn into one of the world's top AI hubs. 'We believe that London is the place to be,' Hughes said.

Beyond the launch itself, several industry signals deserve attention. First, Hughes is bullish on London's density of AI talent, but he has also added his voice to calls to end 'garden leave' — the practice, common in the U.K., of barring departing employees from joining or starting a rival company for months after they resign. American researchers generally don't face this restriction, giving U.S. startups a head start on hiring talent who've left a prior role. Hughes was personally affected by the problem and eventually worked around it, starting Inherent alongside two other DeepMind alumni and a fourth cofounder. Second, the startup isn't slowing down: it plans to grow its headcount to 'about 20 to 25' by the end of the year. Given its ambitions in world models, and with Demis Hassabis's new role leaving some DeepMind staff unsettled, Inherent's hiring push could make it an appealing landing spot for DeepMind employees weighing a move. From a broader perspective, Faraday's success also validates a different path from 'bigger is better': carefully designed reinforcement learning, clear objective functions, and restrained engineering choices can let a small model beat far larger opponents on specific tasks — perhaps signaling that AI competition is shifting from an arms race toward cleverness.