Updated
Updated · TechCrunch · Aug 22
Inherent Says 27 Billion-Parameter Faraday Beats OpenAI, Anthropic in Paper Replication
Updated
Updated · TechCrunch · Aug 22

Inherent Says 27 Billion-Parameter Faraday Beats OpenAI, Anthropic in Paper Replication

1 articles · Updated · TechCrunch · Aug 22

Summary

  • Just weeks after a $50 million seed round, London startup Inherent said Faraday independently reproduced published research findings better than OpenAI’s GPT-5.5 and Anthropic’s Claude Opus 4.8.
  • The edge came from reinforcement learning aimed at teaching “research taste” — choosing worthwhile experiments and designing them well — rather than relying mainly on larger model scale.
  • Faraday runs on Alibaba-backed Qwen 3.6 with 27 billion parameters and uses OpenAI’s GPT-5.5 Codex for coding, reflecting Inherent’s push to build a collaborative AI scientist rather than a full stack.
  • Inherent, founded by DeepMind alumni and staffed by about 12 people in London’s King’s Cross, plans to expand to 20 to 25 employees by year-end as it pitches itself as a local talent magnet.

Insights

How did a tiny AI model defeat tech giants at reproducing complex scientific research from scratch?
Could an AI's learned research taste eventually replace human intuition in discovering groundbreaking scientific knowledge?