Inherent Unveils Faraday: London AI Startup Claims Agent Outperforms Frontier Models in Scientific Replication
London-based artificial intelligence lab Inherent has announced that its new AI agent, Faraday, outperformed significantly larger systems from Anthropic and OpenAI at independently reproducing published scientific research.
What Happened
Inherent, founded by former Google DeepMind researchers, emerged from stealth with a $50 million seed round and has begun detailing its core technology. According to the company, Faraday demonstrated superior performance in autonomously validating published academic studies without receiving advance knowledge of the findings.
During evaluations, Faraday was tested against frontier systems, including OpenAI’s GPT-5.5 and Anthropic’s Claude Opus 4.8. While both competing systems are frontier-scale models, Faraday operates on Qwen 3.6, an architecture comprising 27 billion parameters. Rather than building a separate coding mechanism, Inherent configured Faraday to use OpenAI’s GPT-5.5 Codex for code execution, mirroring how human researchers rely on existing technical software.
Key Highlights
- Model Architecture: Faraday relies on the 27-billion-parameter Qwen 3.6 model, competing against frontier models from Anthropic and OpenAI.
- Methodology: The startup employs reinforcement learning to reward good experimental outcomes, designed to cultivate an instinct for research design and experiment selection.
- Leadership and Team: Founded by DeepMind alumni Edward Hughes, Louis Kirsch, Kaloyan Aleksiev, and Tantum Collins.
- Funding: The lab recently emerged from stealth after securing a $50 million seed funding round.
- Operations: Inherent maintains an entirely in-person team of approximately 12 people at its King’s Cross office in London.
Why This Matters
According to Inherent cofounder and chief scientist Edward Hughes, replicating academic papers is a standard foundational exercise for early-career researchers and doctoral students. Achieving these results using a 27-billion-parameter system illustrates how targeted training methods, such as reinforcement learning, can allow smaller models to match or exceed frontier architectures in complex scientific workflows.
Inherent stated that the initiative is part of an overarching effort to construct an AI scientist capable of finding novel scientific answers rather than solely validating previously published literature.
What to Watch Next
Inherent is preparing to scale its operations, with plans to expand its King’s Cross team from 12 employees to between 20 and 25 by the end of the year. The startup is continuing research toward world models and systems designed to formulate and execute novel scientific discoveries across multiple fields.
Frequently Asked Questions
What is Faraday?
Faraday is an artificial intelligence agent developed by London-based startup Inherent, designed to assist with and conduct scientific tasks, including the replication of academic research.
Which models was Faraday evaluated against?
Faraday was evaluated alongside larger frontier models, specifically Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5.
How was Faraday developed?
Faraday is built around the 27-billion-parameter Qwen 3.6 model, utilizes OpenAI’s GPT-5.5 Codex for coding tasks, and is trained using reinforcement learning to reward effective research choices.
Source: TechCrunch