Inherent, founded by DeepMind alumni, says its AI ‘teammate’ just outperformed Anthropic and OpenAI at replicating research
TechCrunch Anna Heim
A London lab of DeepMind alumni says its AI beat bigger systems at copying research papers. The kicker: it did it with a much smaller model, and care about 'taste' too.
Based on reporting by TechCrunch, Anna Heim — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Inherent, a London AI lab founded by former Google DeepMind researchers, says its new agent has done something a lot of bigger systems haven’t: independently reproduced the findings of published scientific papers. The company says Faraday beat models from Anthropic and OpenAI on that task, even though it runs on a much smaller model.
That claim lands just weeks after the startup came out of stealth with a $50 million seed round. And Inherent is not pitching this as a gimmick. Co-founder and chief scientist Edward Hughes says paper replication is a basic training exercise for human scientists as well, and in this case the harder part was not merely getting the answer right.
Faraday was measured against Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5, but it uses Qwen 3.6, a model with 27 billion parameters. In the AI world, parameter count is an imperfect shorthand, but it still gives a sense of just how much smaller this setup is than the frontier systems Inherent chose to compare itself with.
The company also cared about something it calls research taste: the ability to pick worthwhile experiments and design them well. To teach that, Inherent leaned on reinforcement learning, rewarding good outcomes instead of hard-coding rules. Hughes says the bet is that this kind of training will generalize better as the company pushes toward agents that can contribute across scientific fields.
Inherent has also made some unusual choices about what not to build. Rather than creating its own coding tool, it had Faraday use OpenAI’s GPT-5.5 Codex. The company’s idea is closer to a sharp colleague than a flattering chatbot: one that comes back with experiments, not just agreement. The startup’s dozen employees work together in person from King’s Cross, and Inherent plans to grow to about 20 to 25 people by the end of the year.
My take — AI-written commentary, not fact-checked reporting
This is the right kind of AI story: smaller model, narrower claim, real benchmark, no cosplay about sentience. The market has enough companies selling grand theories; it needs more teams willing to prove one useful thing at a time. Also, if “taste” becomes the next buzzword, somebody should confiscate the deck before it gets applause for basic curiosity.
Read more about this at: TechCrunch
