- Published on
Inherent Launches Small-Parameter AI Agent Faraday, Surpassing GPT-5.5 and Claude Opus 4.8 in Scientific Replication
- Authors
- Name
- DeepHorus
London AI lab Inherent has recently launched its developed AI agent Faraday, which is claimed to outperform larger flagship models in specific tasks. The company states that Faraday achieves superior results in autonomously reproducing published scientific research using a model significantly smaller than those from Anthropic and OpenAI.
Inherent is a startup founded by former Google DeepMind employees, headquartered in King's Cross, London. Just weeks ago, the company completed a $50 million (approximately 337 million RMB) seed funding round and emerged from stealth mode. Despite relatively low attention among numerous startups founded by ex-DeepMind members, the team has begun to showcase its actual research outcomes.
The core breakthrough of the Faraday agent lies in its efficiency. In tests reproducing scientific research, it surpassed massive frontier AI systems such as Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5. The key to this achievement is that the underlying model Faraday relies on is Qwen 3.6, with a parameter scale of only 27 billion—far smaller than competitors' models. Parameter scale is often directly linked to a model's training costs.
Company co-founder and chief scientist Edward Hughes pointed out that defeating other AI systems is not the ultimate goal; the methodology behind achieving this result is more important. Hughes emphasized that paper reproduction is a standard training starting point for human scientists, and Faraday aims to master 'research taste'—the ability to assess experimental value and design experiments reasonably. This is seen as a crucial step toward AI capable of discovering entirely new scientific knowledge.
To cultivate this 'taste,' Inherent adopted a training method primarily based on reinforcement learning. This approach does not directly feed rules to the AI but guides the system to learn more effective actions through a reward mechanism. The company believes that this reward-based training has greater generalization potential than simply learning research procedures, helping to enable AI agents to participate in research across multiple scientific fields in the future.
Faraday's design philosophy is also reflected in its tool selection. Instead of developing all tools in-house, Inherent allows Faraday to use the ready-made OpenAI GPT-5.5 Codex, simulating how human scientists utilize existing software. Hughes hopes to create agents that actively explore and share results as collaborative partners, rather than merely catering to user needs.
In terms of team development, Inherent currently has about 12 employees working offline in London and plans to expand the team to approximately 20 to 25 people by the end of this year. Hughes is particularly optimistic about London's dense pool of AI talent, but he personally calls for the UK to abolish the 'non-compete' system, arguing that it restricts talent mobility and gives US startups an advantage in recruiting AI talent who have just left their positions.
Considering that Inherent is aiming for a grander world model goal and that some Google DeepMind employees' future directions may face uncertainty due to organizational changes, this startup's expansion plans might attract top AI talent seeking new platforms.
