Inherent, a London based artificial intelligence laboratory founded by former Google DeepMind researchers, has announced a significant milestone in the development of AI designed for scientific discovery. The startup recently revealed that its new AI agent, named Faraday, successfully outperformed much larger, frontier scale models from industry leaders OpenAI and Anthropic in a rigorous scientific task.
According to Inherent, Faraday demonstrated a superior ability to independently reproduce the findings of published scientific papers. Crucially, the agent was required to replicate these results without being provided with the answers in advance. This achievement is particularly notable because Faraday is built upon a model that is significantly smaller and more resource efficient than the massive systems it surpassed.
The development marks a major step for Inherent, which emerged from stealth just weeks ago following a 50 million dollar seed funding round. While several startups founded by DeepMind alumni have captured headlines recently, Inherent has maintained a lower profile until now, focusing on building tools that can eventually move beyond replication toward the discovery of entirely new scientific knowledge.
The methodology behind Faraday
The competition in the AI sector is often framed as a race for more parameters and higher computing power. However, Inherent is taking a different approach. Faraday is powered by a model known as Qwen 3.6, which possesses 27 billion parameters. In contrast, the models it competed against, including Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5, are frontier scale systems with significantly higher parameter counts and training costs.
Despite this disparity in size, Faraday proved more effective at the specific task of scientific replication. Edward Hughes, cofounder and chief scientist at Inherent, explained that paper replication is a foundational exercise for human researchers. Many PhD students actually start by doing this, Hughes noted, suggesting that mastering the ability to verify existing work is a necessary precursor to performing original research.
Hughes told TechCrunch that while the team was pleased to beat frontier agents, the victory itself was not the primary goal. What was most interesting to the company was the specific methodology used to build the agent and the efficiency with which it operates.
Developing research taste through reinforcement learning
One of the core challenges in creating an AI scientist is teaching the system "research taste." This refers to the intuitive understanding of which experiments are worth conducting and how to design them for maximum impact. Research taste is often considered an intangible quality that human scientists develop over years of experience.
To imbue Faraday with this quality, Inherent utilized reinforcement learning. This training method involves rewarding the AI for positive outcomes rather than providing a rigid set of rules to follow. By focusing on reward based outcomes, the company believes its agents will be better equipped to generalize their skills across various scientific disciplines.
Inherent’s approach also emphasizes the use of existing tools when appropriate. Rather than reinventing every component of the stack, the company had Faraday utilize OpenAI’s GPT-5.5 Codex for coding tasks. The company compares this to how human scientists rely on specialized software for their work instead of building every tool from scratch.
A new model for AI collaboration
Inherent is not attempting to build an AI that simply serves as a passive tool for users. Instead, the company envisions Faraday as a proactive teammate. Hughes described the ideal interaction as one where the agent pursues its own curiosity.
He modeled the agent's behavior after a collaborative colleague who might return from a series of independent experiments and ask for feedback on the results. The goal is to move away from agents that merely confirm what a user wants to hear and toward systems that can provide genuine scientific insight through independent investigation.
This collaborative philosophy extends to the company’s internal operations. Inherent currently employs a team of twelve, all of whom work in person from an office in King’s Cross, London. The neighborhood has become a significant AI hub, largely due to the nearby presence of Google DeepMind. We believe that London is the place to be, Hughes stated, citing the city's high density of AI talent.
Challenges in the London talent market
While Hughes is optimistic about London's potential, he has raised concerns regarding local employment practices that may hinder the growth of the startup ecosystem. Specifically, he has spoken out against "garden leave," a common practice in the United Kingdom where departing employees are prohibited from joining rivals or starting new companies for several months after resigning.
Hughes noted that researchers in the United States generally do not face these same restrictions, which can give American startups a competitive advantage in hiring. This is a personal view rather than a company view, but I was affected by the garden leave problem, Hughes said, referencing his own experience transitioning out of DeepMind.
Despite these hurdles, Inherent is moving forward with an aggressive hiring plan. The company expects to increase its headcount to between 20 and 25 employees by the end of the year. This expansion comes at a time when some staff at DeepMind are reportedly unsettled due to organizational changes, potentially making Inherent an attractive destination for researchers looking for a more specialized, science focused environment.
What happens next?
The successful replication of scientific papers is only the first phase of Inherent’s broader roadmap. The company’s long term "north star" is the creation of an AI scientist agent capable of making original contributions to various fields.
By focusing on world models and high efficiency agents, Inherent aims to prove that specialized, smaller models can rival or exceed the performance of general purpose giants when applied to complex, technical domains. The tech community will likely be watching to see if Faraday can transition from verifying known science to uncovering new truths in chemistry, biology, or physics.
As the company scales its team and refines its reinforcement learning techniques, the next major milestone will likely involve the agent's first attempts at designing and executing original experiments. For now, Inherent has demonstrated that in the world of AI, size is not always the deciding factor in performance.
Filed under: AI, TechNews, Startups, Software, London, DeepMind, Faraday