Inherent, founded by DeepMind alumni, says its AI ‘teammate’ just outperformed Anthropic and OpenAI at replicating research

A London-based AI startup founded by DeepMind alumni released an AI agent that replicated published scientific research findings better than larger models from major competitors, while running on a significantly smaller model. The startup achieved this using reinforcement learning, a training method that rewards AI systems for good outcomes rather than explicit instructions, and emphasizes teaching AI systems "research taste," or instinct for what experiments are worth conducting. The company's longer-term goal is building AI agents capable of discovering new scientific knowledge rather than just verifying existing results, similar to how human scientists work. This matters because it demonstrates that smaller, more efficiently trained models may outperform much larger systems at specific scientific tasks, potentially reshaping expectations about AI scale and capability.