David beats Goliath: tiny AI outscores OpenAI and Anthropic at science
Inherent, a London startup founded by Google DeepMind alumni, just claimed its AI agent Faraday outperformed Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 at a very specific task: independently reproducing the findings of published scientific papers, cold, with no answers given. The kicker? Faraday runs on a comparatively puny 27-billion-parameter model, not a frontier-scale giant. The real flex isn't the win itself, Inherent says, but how it trained the thing, using reinforcement learning to teach research 'taste' rather than just memorized rules, all while chasing a bigger goal: an AI that discovers new science, not just checks old homework.
Cute benchmark, but ask us again when it discovers something nobody already knew.