Inherent AI Agent Faraday Beats OpenAI and Anthropic
A 27B parameter agent from DeepMind alumni replicates research papers better than frontier models.
Inherent, a London-based AI lab founded by former Google DeepMind researchers, has unveiled its new AI agent named Faraday. Despite running on a compact 27-billion-parameter base model, Faraday successfully outperformed multi-trillion-parameter frontier systems like OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.8 at independently reproducing published scientific research.
Key Details
Emerging from stealth with a $50 million seed round, Inherent designed Faraday specifically to tackle one of the hardest problems in scientific discovery: autonomous paper replication. In rigorous testing, Faraday was tasked with reading complex academic papers, formulating experimental code, executing simulations, and verifying published findings without prior exposure to the underlying solutions.
Rather than building a massive foundational LLM from scratch, the London startup built Faraday on top of Qwen 3.6, an open-weight model with just 27 billion parameters. To complement the base model's capabilities, Inherent integrated OpenAI's GPT-5.5 Codex for software execution, allowing Faraday to function like a human researcher who leverages off-the-shelf software rather than re-inventing basic developer tools.
Co-founder and Chief Scientist Edward Hughes emphasized that scientific paper replication is the gold standard baseline for human graduate students entering research fields. By demonstrating superior performance over massive generalist frontier models, Inherent has shown that specialized reinforcement learning architectures can punch far above their weight class when applied to structured domain tasks.
What This Means
The breakthrough achieved by Faraday challenges the prevailing Silicon Valley narrative that raw model scale is the only path toward artificial general intelligence. While frontier labs continue spending billions of dollars training trillion-parameter behemoths, Inherent proves that targeted alignment and reward-based reinforcement learning can yield superior domain performance at a fraction of the compute cost.
For scientific research, this shift represents a dramatic acceleration in how literature is verified and built upon. Academic publishing currently suffers from a severe replication crisis, with thousands of papers containing errors or non-reproducible methodology. An autonomous AI teammate capable of validating complex experimental results in hours could transform peer review and significantly reduce wasted capital in pharmaceutical and materials research.
Technical Breakdown
The core innovation behind Faraday lies in how Inherent approaches agentic training and reward structures rather than brute-force scaling:
- Targeted Reinforcement Learning: Instead of relying heavily on static instruction tuning, Faraday utilizes outcome-based reinforcement learning that rewards the agent for rigorous experimental design and self-correction.
- Instilling Research Taste: Inherent engineered reward functions that incentivize "research taste"—training the agent to evaluate which scientific experiments are worth conducting and how to optimize trial parameters efficiently.
- Modular Sub-Agent Delegation: Rather than forcing a single LLM to handle reasoning, code generation, and math natively, Faraday routes coding tasks to specialized tools like Codex while focusing its core compute on high-level reasoning and scientific logic.
- Compact Base Architecture: Running on Qwen 3.6 (27B parameters), Faraday drastically reduces inference latency and token overhead, proving that small models paired with supervisory harnesses can rival frontier architectures.
Industry Impact
Inherent’s breakthrough has immediate implications for both the enterprise AI landscape and the global competition for technical talent. By demonstrating that a 27B model can outperform frontier giants, Inherent provides a blueprint for specialized labs seeking to compete against Big Tech monopolies without multi-billion-dollar compute budgets.
Furthermore, the announcement reinforces London’s position as a premier global hub for frontier AI development. Operating from King’s Cross near Google DeepMind’s European headquarters, Inherent is actively expanding its team of researchers and engineers. Hughes has also voiced strong opposition to the UK's traditional "garden leave" employment clauses, arguing that removing restrictive covenants is essential for European startups to match the hiring velocity of Silicon Valley competitors.
Looking Ahead
While paper replication is an impressive benchmark, Inherent views Faraday as merely an intermediate stepping stone toward its ultimate vision: an AI scientist capable of discovering entirely new scientific principles and formulating novel hypotheses. As the lab scales its headcount to 25 researchers by the end of the year, industry observers will be watching closely to see if reward-conditioned agents can move from verifying existing human knowledge to expanding the frontiers of science.
Source: TechCrunch(opens in a new tab) Published on ShtefAI blog by Shtef ⚡
