A tiny London lab just did something the well-funded giants haven’t. Inherent, an AI startup founded by Google DeepMind alumni, says its new agent Faraday outperformed frontier models from Anthropic and OpenAI at reproducing published scientific research, according to TechCrunch AI. And it did it running on a model a fraction of their size.
This is significant because of how lopsided the matchup was.
📊 The Quick Version
- Inherent, based in King’s Cross, London, emerged from stealth just weeks ago with a $50 million seed round.
- Its agent, Faraday, beat Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5 at independently replicating scientific papers without being told the answers in advance.
- Faraday runs on Qwen 3.6, a comparatively small model with just 27 billion parameters. The models it beat are frontier-scale, meaning far larger and far more expensive to train.
- The company’s stated goal isn’t verifying old science. It’s building AI that discovers new scientific knowledge.
🔬 Why paper replication matters
Reproducing a published study without the answer key sounds like a narrow test. It isn’t. Cofounder and chief scientist Edward Hughes told TechCrunch AI that it’s a standard training exercise for human scientists too. “Many PhD students actually start by doing this.”
But Hughes was clear that winning wasn’t the headline. “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 method is the real story.
🧠 The bet on ‘taste’
Inherent wanted more than accuracy. It wanted Faraday to show what Hughes calls “research taste,” an instinct for which experiments are worth running and how to design them well. You can’t write rules for taste, so the team leaned on reinforcement learning, a training approach that rewards good outcomes instead of spelling out instructions.
The wager: reward-based training will generalize better toward the long-term goal of an AI scientist that can contribute across many fields. “We’re always guided by that north star of building an AI scientist agent and imbuing our agents with taste,” Hughes said.
That philosophy also shaped what Inherent refused to build. Rather than making its own coding tool, it had Faraday use OpenAI’s GPT-5.5 Codex, the same way a human scientist uses existing software instead of reinventing it.
💡 Why this matters for the industry
The dominant story in AI has been scale. Bigger models, more parameters, larger training bills. Inherent is poking a hole in that assumption. If a 27-billion-parameter model paired with smart reinforcement learning can beat frontier systems at a real scientific task, the “just make it bigger” playbook starts to look less like a law and more like a habit.
For practitioners, the takeaway is practical. Method and training design can beat raw size on specialized tasks. That’s good news for teams without hyperscaler budgets, and it’s a signal that the next edge may come from how you train, not how much you spend.
👀 What to watch next
Inherent isn’t slowing down. It plans to grow from a dozen employees to “about 20 to 25” by year’s end, all working in person out of King’s Cross. Hughes is bullish on London’s talent density, though he’s publicly pushing to end “garden leave,” the UK practice of barring departing staff from joining rivals for months. He was affected by it himself before launching Inherent with three other cofounders.
With DeepMind staff reportedly unsettled after Demis Hassabis’s new role, Inherent’s hiring push could pull in more alumni. And with ambitions in world models on top of its AI scientist goal, this is a team worth tracking.
The party trick was replicating old research. The real question is whether Faraday can go find something new. Full details are available at the original TechCrunch AI report.