Inherent, a London-based AI lab founded by Google DeepMind alumni, has released its AI agent, Faraday, which claims to outperform larger models from Anthropic and OpenAI at replicating published scientific papers. The startup, which emerged from stealth with a $50 million seed round, said its model succeeded in independently reproducing findings without being told the answer in advance. This achievement highlights Inherent's focus on building AI capable of discovering new scientific knowledge rather than just verifying existing results.
Faraday runs on a relatively small model called Qwen 3.6 with 27 billion parameters, compared to Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5. Inherent's success was measured not just by accuracy but also by its ability to demonstrate 'research taste'—an instinct for selecting meaningful experiments and designing them effectively. The company used reinforcement learning, rewarding AI systems for good outcomes rather than spelling out rules, to train its agents.
According to the source, Inherent's approach is guided by the goal of creating an AI scientist agent capable of contributing across many scientific fields. The startup also avoids building its own coding tools, instead using OpenAI's GPT-5.5 Codex, much like human scientists rely on existing software. The company's focus on collaboration and transparency is evident in its hiring practices and operational setup, with all employees working in person at its King’s Cross office in London.
Source: techcrunch
