Inherent, a London-based artificial intelligence laboratory established by former Google DeepMind talent, has announced a breakthrough: its AI agent has demonstrably surpassed the performance of significantly larger models from industry leaders Anthropic and OpenAI, all while operating with a mere fraction of their operational scale.
Among the numerous startups founded by Google DeepMind alumni, Inherent has maintained a comparatively understated profile. However, while some better-funded competitors have yet to unveil substantial innovations, this London-based team is now beginning to showcase the advancements it has been diligently developing.
Following its public debut with a successful $50 million seed funding round just weeks prior, the British startup reveals that its newly launched AI agent, Faraday, has outperformed more extensive, well-known models in a highly specialized task: the independent reproduction of findings from published scientific papers, without any prior knowledge of the anticipated results.
While this achievement might initially seem like a superficial accomplishment given Inherent's ambitious overarching goal—to construct AI capable of discovering novel scientific knowledge rather than merely verifying existing findings—co-founder and chief scientist Edward Hughes emphasizes its foundational importance. He notes, "Many PhD students actually start by doing this," underscoring that paper replication is a standard training exercise for human scientists as well.
Hughes clarified to TechCrunch that the primary objective was not simply to surpass other AI systems in this task, but rather to innovate in the methodology employed. "What was most interesting to us about this was not so much the result of beating those frontier agents — which of course we liked — but was actually the way we went about building this," he stated.
A particularly noteworthy aspect for investors is Faraday’s operational efficiency: when benchmarked against Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5—both colossal, frontier-scale systems—Faraday functions on a remarkably compact model named Qwen 3.6, which possesses just 27 billion parameters. (Parameters serve as a general indicator of a model's size and, consequently, its training expenses.) Inherent also set a higher bar for success than mere accuracy, aspiring for Faraday to exhibit "research taste"—an intuitive understanding of which experiments warrant execution and how to design them effectively.
Cultivating an elusive concept such as scientific intuition ("taste") presents a significant challenge, which is where reinforcement learning becomes crucial. This training paradigm rewards an AI system for achieving favorable outcomes, rather than dictating explicit rules to follow. Instead of primarily instructing its agents on the mechanics of scientific inquiry, Inherent leverages this reward-based methodology, anticipating that it will foster greater generalization towards its long-term objective of creating agents capable of contributing across diverse scientific domains.
"We’re always guided by that north star of building an AI scientist agent and imbuing our agents with taste," Hughes affirmed. This steadfast focus has also influenced Inherent’s strategic decisions regarding what it chooses not to develop. For instance, rather than creating its own coding tool, Faraday was equipped to utilize OpenAI’s GPT-5.5 Codex, mirroring how human scientists frequently leverage established software instead of custom-building every tool, according to the company.
Inherent is also committed to avoiding the development of agents that merely echo user expectations. Instead, Hughes explained, their aspiration is modeled on his ideal teammate—one who returns and declares: “I got curious about this, and I went off and I did these experiments. What do you think of these results?”
This spirit of collaboration permeates Inherent's operational ethos. Its team of a dozen employees works in person from an office in King’s Cross—the London district that, once overlooked, has transformed into a premier global AI hub largely due to Google DeepMind’s presence. Hughes declared, “We believe that London is the place to be.”
While Hughes is optimistic about London's rich concentration of AI expertise, he has also advocated for changes to "garden leave"—a prevalent U.K. practice that prohibits departing employees from joining or establishing a rival company for several months post-resignation. This restriction is generally not faced by American researchers, affording U.S. startups a hiring advantage for talent transitioning from previous roles. Hughes shared with TechCrunch, “This is a personal view rather than a company view, but I was affected by the garden leave problem.”
Hughes ultimately navigated this constraint and co-founded Inherent alongside two other DeepMind alumni and a fourth co-founder. The startup shows no signs of decelerating, with plans to expand its headcount to "about 20 to 25" by year-end. Given its ambitions in world models and the recent organizational shifts at DeepMind, including Demis Hassabis’s new role which has left some staff experiencing uncertainty, Inherent's proactive hiring drive could position it as an attractive destination for DeepMind employees considering a career transition.
The Inherent co-founders include Louis Kirsch, Kaloyan Aleksiev, Tantum Collins, and Edward Hughes.
The Editorial Staff at AIChief is a team of professional content writers with extensive experience in AI and marketing. Founded in 2025, AIChief has quickly grown into the largest free AI resource hub in the industry.