Inherent, a London-based artificial intelligence laboratory founded by former Google DeepMind employees, has announced that its proprietary AI agent successfully outperformed much larger, established models from Anthropic and OpenAI while operating on a significantly smaller technical scale.
Achieving Scientific Replication
The British startup, which recently emerged from stealth following a $50 million seed funding round, revealed that its new AI agent, named Faraday, achieved a notable feat: independently recreating the findings reported in published scientific research papers without being given any prior information. This ability to replicate research findings is described by co-founder and chief scientist Edward Hughes as a standard initial exercise for human academics, noting, “Many PhD students actually start by doing this.”
While the impressive performance was certainly highlighted, Hughes clarified that the core focus was not merely on beating rival systems but rather on the methodology used to achieve the results. When measured against large, state-of-the-art models—specifically Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5—Faraday is powered by a comparatively modest model called Qwen 3.6, which contains only 27 billion parameters. Parameters are generally understood to be a proxy for both a model’s size and its associated training expenses.
Defining ‘Research Taste’ and Methodology
Inherent set a standard for success that extended beyond basic accuracy. The team aimed for Faraday to demonstrate what they termed “research taste”—an innate capability to identify appropriate experiments and design them effectively. Teaching such an abstract concept required the use of reinforcement learning, a training approach that rewards the AI system for positive outcomes rather than being given explicit, step-by-step instructions.
The company utilized this reward-based system, betting that it would better generalize to their long-term objective: creating AI agents capable of contributing original scientific knowledge across diverse fields. Hughes emphasized the guiding philosophy behind the development: “We’re always guided by that north star of building an AI scientist agent and imbuing our agents with taste.”
This focus also guided Inherent’s operational choices. Instead of developing proprietary coding tools, the company instructed Faraday to use OpenAI’s GPT-5.5 Codex, mirroring how human researchers frequently rely on existing software rather than building every necessary tool from scratch.
Industry Vision and Talent
Beyond its technical goals, Inherent is focused on developing a collaborative AI experience. Hughes stated that the goal is to create a teammate that does not simply confirm what the user wants to hear, but rather approaches the user with curiosity, suggesting, “I got curious about this, and I went off and I did these experiments. What do you think of these results?”
Inherent currently maintains its operations in King’s Cross, a neighborhood in London that has become a prominent AI center, a sentiment echoed by Hughes: “We believe that London is the place to be.”
On a related note, Hughes offered a personal perspective on the talent market, criticizing the U.K.’s “garden leave” policy—the practice of restricting departing employees from joining or starting a rival company for several months. He noted that this restriction is not generally applied to American researchers, which he believes gives U.S. startups an advantage in hiring departing talent.
The company plans aggressive growth, aiming to expand its workforce from its current dozen employees to approximately 20 to 25 by the close of the year. Given its focus on world models and the evolving landscape following leadership changes at DeepMind, Inherent’s expansion could make it an attractive destination for DeepMind staff considering a career move.
