Inherent, a British startup founded by former Google DeepMind employees, has launched a Faraday artificial intelligence agent that can independently check the accuracy of scientific publications. Developers say that despite using only 27 billion parameters, the system’s performance is significantly better than larger models from Anthropic and OpenAI.
Image source: Gemini
According to news TechCrunchInherent, which emerged from Stealth mode a few weeks ago while raising $50 million in early-stage funding, has now revealed the first details of its development. The startup claims Faraday has the ability to independently replicate published scientific findings without prior prompting. Inherent co-founder and chief scientist Edward Hughes points out that this kind of work is standard training for people: Many graduate students begin their scientific careers by replicating research that has already been done.
In internal testing, Faraday was compared to Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5. Meanwhile, the Inherent agent works based on the relatively small Qwen 3.6 model (with 27 billion parameters), while its competitors are larger advanced models. Developers were assessed not only on the system’s ability to obtain original working results, but also on what’s known as “research instinct”—the ability to independently determine which experiments make sense and how best to organize them. Hughes stressed that what got them to this point was more important to the team than beating their rivals.
To train Faraday, the developers chose a reinforcement method (reinforcement learning, RL), in which the system is rewarded for successful results rather than following strict pre-written rules. The ultimate goal of Inherent, Hughes said, is to create an artificial intelligence scientist who can not only test existing discoveries but also generate new scientific knowledge. At the same time, the startup isn’t trying to create all the necessary tools itself. For example, Faraday uses OpenAI’s GPT-5.5 Codex for programming. Inherent also compares this approach to the work of traditional researchers, who use off-the-shelf software instead of developing each tool from scratch.
The startup currently has around 12 employees in London’s large King’s Cross borough and plans to expand to around 20-25 employees by the end of this year. Hughes calls the region a good location for the company because of its high concentration of artificial intelligence talent, but he also objects to the common British practice of garden sabbaticals, which essentially prohibit departing employees from jumping to a competitor or starting their own company within a few months. The leader said he himself encountered such limitations before working with others at Google DeepMind to create Inherent.
If you find an error, select it with your mouse and press CTRL+ENTER.