Most of the AI we use today is built to be a generalist—a jack-of-all-trades that can write emails, debug code, or summarize meetings. But a small group of researchers in London is taking a different approach. They have built an AI agent designed to act like a junior scientist, specifically tasked with independently replicating published research papers.
WHAT'S HAPPENING
A startup called Inherent recently released this agent, named Faraday. In tests, Faraday outperformed much larger, well-known AI models from companies like OpenAI and Anthropic. What makes this surprising is that Faraday runs on a relatively tiny model. In the world of AI, researchers often use parameters—the internal settings that represent what a model has learned—as a proxy for size. Faraday uses a model with 27 billion parameters, a fraction of the size of the massive, frontier-scale systems it beat. The startup’s goal is not just to verify old findings, but to eventually create an AI that can discover entirely new scientific knowledge.
The art of research taste
HOW IT WORKS
To understand why Faraday is different, think of how a human scientist learns. A first-year PhD student often starts by repeating experiments that have already been done. By doing this, they don't just learn the facts—they learn what makes an experiment good or bad, which researchers call research taste.
Inherent is trying to teach this taste using a technique called reinforcement learning. Instead of giving the AI a rigid set of instructions or a manual on how to do science, they reward the system when it makes smart decisions. If the AI designs a good experiment or chooses an effective path, it receives a positive signal, much like training a dog with treats. Over time, the model begins to prefer the types of decisions that lead to successful results. By focusing on this reward-based training, Inherent believes the agent will develop the intuition needed to navigate new, unknown problems across different scientific fields.
WHY IT MATTERS
This shift from brute force to specialized, intuitive intelligence changes how we think about AI progress. We are used to the idea that bigger is always better—that if you just throw more computing power at a problem, the AI will get smarter. Inherent is betting that there is a better way to build these tools. If they succeed, it suggests that we don't necessarily need the world's largest, most expensive computers to solve complex problems. Instead, we might just need AI with better training, sharper instincts, and a clearer sense of what good research actually looks like.
Liked this one? The next lands at breakfast.
Every story in tomorrow's AI news, rebuilt in plain English — five minutes, sources linked, free forever.
By joining you agree to receive Article's daily newsletter — unsubscribe in one click. Privacy