Inherent's AI Agent Outperforms Anthropic and OpenAI in Scientific Research Replication
London-based startup Inherit, founded by Google DeepMind alumni, has developed an AI agent named Faraday that outperformed larger models from Anthropic (Claude Opus 4.8) and OpenAI (GPT-5.5) at replicating scientific research findings without prior knowledge of the answers. The task involved independently reproducing published results—a common exercise for human scientists—using a much smaller model with only 27 billion parameters compared to its rivals' larger models. The company emphasized that success wasn't just about accuracy but also demonstrated 'research taste,' or an AI's ability to determine which experiments are worth running. Inherent used reinforcement learning, rewarding the agent for good outcomes rather than prescribing rules, as a method to generalize across scientific fields in the long term. The startup is focused on building collaborative AI scientists that can contribute meaningfully and avoid simply regurgitating desired results. Inherent recently emerged from stealth with $50 million in seed funding and operates out of London's King’s Cross neighborhood, which has become a major hub for AI talent thanks to DeepMind's presence. The company is also advocating against U.K.-specific 'garden leave' restrictions that limit employee mobility.
Topics: AI DeepMind alumni model efficiency research agents startups