Inherent, a startup founded by former Google DeepMind researchers, has introduced an AI agent called Faraday that can independently reproduce scientific research findings. The system achieves high performance using a relatively small model by emphasizing reinforcement learning and the development of 'research taste.'
The Faraday Scientific Breakthrough
Inherent has officially emerged from stealth to reveal Faraday, an AI agent designed to replicate scientific research findings without external guidance. The London-based startup, established by alumni from Google DeepMind, claims that its agent outperformed much larger, industry-standard models such as Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5. The core challenge involved the agent reading and independently validating published scientific literature, a task often used as a foundational training exercise for human PhD students. By succeeding in these replications, Inherent aims to move toward its long-term objective of developing autonomous agents capable of contributing original scientific knowledge rather than merely verifying legacy results.
Efficiency Through Model Architecture
A striking feature of the Faraday system is its relatively lean technical footprint. While industry counterparts operate at massive scales, Faraday is powered by the Qwen 3.6 model, which utilizes only 27 billion parameters. Inherent suggests that this smaller size offers a significant advantage in training costs and operational efficiency. Instead of focusing on massive scale to achieve intelligence, the startup relies on reinforcement learning. This training paradigm emphasizes rewarding the agent for achieving high-quality outcomes rather than imposing rigid, rule-based constraints. By centering its development around this reward-based mechanism, the company believes it can cultivate a more generalized form of intelligence that effectively navigates the complexities of various scientific domains.
Cultivating 'Research Taste'
Beyond simple accuracy in replication, Inherent is prioritizing what Chief Scientist Edward Hughes describes as 'research taste.' This intangible quality represents an AI's ability to intuitively discern which experiments are valuable and how to design them effectively. To assist in this pursuit, Faraday leverages existing external tools, such as OpenAI’s GPT-5.5 Codex, for its coding requirements rather than attempting to build proprietary software from the ground up. This approach mirrors the workflows of human scientists, who typically rely on established software suites to conduct their inquiries. By delegating technical execution, the Inherent team hopes to keep the agent focused on the higher-level decision-making processes required for authentic scientific discovery.
Operations and Industry Challenges
Based in the King’s Cross area of London, Inherent benefits from the city’s high concentration of machine learning expertise, an environment largely fostered by the historical success of Google DeepMind. The company currently employs about a dozen people and aims to expand that number to between 20 and 25 by the end of 2026. Despite his commitment to the local ecosystem, co-founder Edward Hughes has expressed concerns regarding U.K. labor practices, specifically 'garden leave.' This policy, which restricts departing employees from joining or launching competing ventures for several months, places British startups at a disadvantage compared to their American rivals. Hughes, who navigated these hurdles himself, suggests that reforming these regulations could better support the growth and mobility of the U.K. technology sector.
⚖ The Balanced View
Supporting view
The startup’s reinforcement learning approach allows it to achieve high-performance results using a compact model (27B parameters), significantly reducing the resource overhead typically associated with large-scale frontier AI models.
Concerns & criticism
The team must overcome the regulatory burden of U.K.-based 'garden leave' policies, which can delay talent acquisition and hinder the startup's ability to scale compared to U.S.-based competitors.
→What's next
Inherent plans to aggressively scale its operations by nearly doubling its current headcount by the conclusion of the year. The team remains focused on expanding its reach into world models while continuing to refine the research capabilities of its Faraday agent.
