Tiny AI Startup Beats Big Tech at Science — With a Twist
A little-known London lab just beat OpenAI and Anthropic's AI at science — and it's small enough to fit in your living room.
A tiny startup called Inherent, founded by former Google DeepMind researchers, just beat the AI giants at a surprisingly human task: independently replicating scientific research. Their AI agent, Faraday, used a model so small it could run on a decent laptop, yet it outperformed much larger systems from Anthropic and OpenAI at reproducing published science papers. The real twist? Faraday wasn’t just accurate — it showed ‘research taste,’ an instinct for which experiments were worth running. That’s rare for AI, which usually follows strict rules instead of making judgment calls.
The approach is a big deal because most AI systems today are trained on massive amounts of data and rules. Inherent, on the other hand, used reinforcement learning — a system where the AI is rewarded for good outcomes rather than told exactly what to do. This could help AI generalize better to new fields, moving beyond just copying what it’s seen before. Faraday even used OpenAI’s coding tool, just like human scientists rely on existing software instead of building everything from scratch.
The startup’s focus is on building AI that acts more like a thoughtful colleague than a robot that just gives answers. Faraday is designed to say, ‘I tried this experiment and got curious — what do you think?’ That collaborative spirit extends to Inherent’s own team. All 12 employees work together in an office in London’s King’s Cross, a neighborhood that’s become a hotspot for AI talent thanks to DeepMind’s presence. The company plans to grow to 20-25 people by the end of the year, making it a tempting option for researchers leaving DeepMind amid recent leadership changes.
Inherent isn’t just about replicating old science — its ultimate goal is to build AI that can discover new knowledge. Faraday’s success is a small but meaningful step in that direction. It’s also a reminder that size isn’t everything in AI. Sometimes, a focused, well-designed system can outperform giants that try to do everything at once.
- Inherent, a tiny London AI startup, built an AI agent called Faraday that can independently replicate scientific research — and it outperformed much larger models from OpenAI and Anthropic.
- Faraday uses a tiny model (27 billion parameters) and ‘research taste’ — an instinct for which experiments are worth running — trained using rewards instead of rigid rules.
- The startup’s goal is to build AI that acts like a thoughtful colleague, contributing to science rather than just summarizing it, and it’s hiring aggressively in London.
Why It Matters
This could eventually help scientists work faster, make fewer mistakes, and focus on creative breakthroughs — not just tedious replication.