**[Former DeepMind Employees Found Inherent, AI Agent Outperforms Claude and GPT-5.5 in Scientific Research Reproduction]** According to DeepTech TechFlow on August 23, as reported by TechCrunch, UK-based AI lab Inherent, founded by several former Google DeepMind members, has launched an AI research agent called "Faraday." The company claims that Faraday outperformed cutting-edge models such as Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 in tasks involving the independent reproduction of results from published scientific papers. Notably, Faraday operates on the Qwen 3.6 model, which has only 27 billion parameters—a significantly smaller scale compared to the aforementioned advanced models. Inherent stated that its training objective is not only to enable the agent to accurately reproduce research results but also to cultivate its "research taste," meaning the ability to autonomously judge which experiments are worth conducting and how to design them. To achieve this, the team primarily relies on reinforcement learning for training rather than simply teaching predefined scientific research processes. Edward Hughes, co-founder and Chief Scientist of Inherent, noted that reproducing research papers is a common training method for researchers. The company's primary focus is not on Faraday outperforming other advanced agents in tests but on validating whether its training methods can be further applied to develop "AI scientists" capable of discovering new scientific knowledge. In its workflow, Faraday does not rely solely on its own model to complete all tasks. For example, when programming is required, it calls upon OpenAI's GPT-5.5 Codex as a tool, akin to how researchers use existing software to assist in their work. Inherent was co-founded by Louis Kirsch, Kaloyan Aleksiev, Tantum Collins, and Edward Hughes, including several former Google DeepMind members. The company is headquartered in London, currently has about 12 employees, and plans to expand to around 20 to 25 employees by the end of the year. The company emerged from stealth mode in May of this year and completed a $50 million seed funding round. Its long-term goal is to develop AI agents capable of autonomously conducting research and discovering new knowledge across multiple scientific fields.