Faraday

Faraday is an AI program made by the UK startup Inherent. Its job is to take an already-published scientific paper and reproduce its experimental results on its own, without being shown the answer. The training approach is a bit like a PhD student learning how to do research by redoing a senior researcher's paper from scratch.

What's surprising is the size of the underlying model Faraday runs on. Faraday is built on top of a relatively small model called Qwen 3.6, which has around 27 billion parameters (the numerical settings that make up a model). That's smaller than much larger, more recent flagship models like Anthropic's Claude Opus 4.8 or OpenAI's GPT-5.5. Even so, Inherent claims Faraday outperformed both of these larger models on the task of reproducing paper results.

What Inherent cared about wasn't just getting the right answer. The company wanted to instill in Faraday a sense of judgment for figuring out which experiments are worth running and how to design them well — something the company calls "research taste." To do this, they used a reinforcement-learning approach that rewards good outcomes, and for coding tasks, they didn't build their own tools but simply used OpenAI's GPT-5.5 Codex.

Articles describe this as "Inherent's in-house AI agent Faraday outperformed frontier models from Anthropic and OpenAI." It's easy to misread this as Faraday being smarter across the board — but that's not the case. The win refers to one specific task: independently reproducing the results of already-published papers without being given the answers. Inherent itself has said it cares more about the process behind that result — the research judgment involved — than about winning or losing.