Faraday’s success confirms that training models on methodological curiosity and scientific taste can surpass simply scaling up computational model sizes.
Article index:
- Advanced scientific achievement for the Faraday AI agent
- The philosophy of research taste and reinforcement learning
- Collaborative work environment and hiring constraints
- Frequently asked questions
Advanced scientific achievement for the Faraday AI agent
Inherent, an artificial intelligence startup lab based in London and founded by a cohort of former experts and researchers from Google DeepMind, announced that its new AI agent, Faraday, has successfully outperformed much larger models from Anthropic and OpenAI while utilizing an intelligent model that represents a fraction of those giant models’ sizes.
Although the company has not received widespread media buzz compared to some heavily funded competitors, just weeks after emerging from stealth mode and raising a $50 million seed funding round, it has begun showcasing its tangible innovations. The company stated that its Faraday AI agent outperformed leading models in a specific and highly rigorous task: the independent reproduction of scientific results from published research papers without being provided the answers or conclusions beforehand.
Edward Hughes, co-founder and chief scientist at Inherent, explained that repeating experiments and reproducing research is a standard training progression for human scientists, stating, “Many PhD students begin their academic careers by performing precisely this task.” He added in statements to TechCrunch that the primary goal was not merely to surpass other intelligent systems, but rather the innovative engineering approach followed to build this agent.
The philosophy of research taste and reinforcement learning
The most intriguing aspect for investors and experts lies in the technical comparison; while Faraday’s performance was compared to massive models such as Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5, Inherent’s agent operates using a relatively small intelligent model called Qwen 3.6, which has only 27 billion parameters (with parameters representing a measure of model size and training costs). The company’s success criterion was not limited to computational accuracy, but focused on instilling what it describes as “research taste” in the agent—the scientific intuition to determine which experiments are worth conducting and how to design them impeccably.
Teaching models an intangible concept like research taste is complex, and this is where reinforcement learning comes in—a training method that rewards the intelligent system for achieving good results rather than instilling rigid, pre-defined rules. Inherent adopted this reward-based approach driven by its bet on its capacity for generalization and scaling to achieve its long-term goal of building agents capable of contributing to knowledge discovery across multiple scientific fields.
This focus was reflected in what the company chose not to build internally; rather than developing its own proprietary software tool for writing code, it made Faraday use OpenAI’s GPT-5.5 Codex tool, much like human scientists utilize available ready-made software to focus on the core of scientific research.
Collaborative work environment and hiring constraints
Inherent seeks to avoid building agents that merely echo what the user wants to hear; instead, the agent’s design is based on the model of a curious colleague who conducts their own experiments, returns to discuss the results, and proposes new paths. This collaborative approach extends to the company’s work environment, where its 12 employees work fully in-person inside an office located in London’s King’s Cross district, which has transformed thanks to DeepMind into one of the world’s premier AI hubs.
Hughes expressed optimism about London’s density of technical talent while simultaneously calling for an end to the UK’s common “gardening leave” system, which prevents departing employees from joining competitors or founding new startups for several months, giving American companies a competitive advantage in attracting talent. The company plans to scale its headcount to between 20 and 25 employees by the end of the current year, making it an attractive destination for talent amid its ambitions to build future world models. Alongside Hughes, the company’s founders include Louis Kirsch, Kaloyan Alexiev, and Tantum Collins.
Frequently asked questions
Question: What is the scientific achievement accomplished by the Faraday AI agent?
Answer: It successfully reproduced published research and scientific papers with complete independence, outperforming Opus and GPT models.
Question: What is the size of the model relied upon by Inherent’s agent?
Answer: It relies on the Qwen 3.6 model containing only 27 billion parameters, which is much smaller than competing leading models.
Question: What methodology did the company follow to train the agent on research taste?
Answer: It relied on reinforcement learning based on rewarding good results to encourage exploration and meticulous experimental design.
