Inherent's AI outperforms major competitors in scientific research replication

Inherent, a London-based AI startup founded by former DeepMind employees, claims its AI agent, Faraday, has outperformed larger models from Anthropic and OpenAI in replicating scientific research findings. The company, which recently emerged from stealth with a $50 million seed round, highlighted that Faraday achieved this using a much smaller model, Qwen 3.6, with 27 billion parameters. This model is significantly less complex than Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5. Edward Hughes, cofounder and chief scientist at Inherent, emphasized that the focus was not just on the outcome but on the methodology used to train the AI. The team employed reinforcement learning to teach Faraday to demonstrate 'research taste'—an ability to identify valuable experiments and design them effectively. Inherent also avoids building proprietary tools, instead using existing platforms like OpenAI’s Codex, mirroring how human scientists rely on existing software. The startup is expanding its team to 20 to 25 employees by year-end, positioning itself as an attractive option for DeepMind alumni. Hughes also criticized the UK’s 'garden leave' policy, which restricts former employees from joining competitors for months after leaving, a practice not present in the US.