Summary
Edward Hughes, CEO and co-founder of AI research startup Inherent — which raised $50 million from Index Ventures and Radical Ventures at a $225 million post-money valuation — joins Machine Learning Street Talk for a dense conversation about what it would take to build AI systems capable of genuine scientific discovery. Hughes, a former Google DeepMind researcher with a PhD in theoretical physics, draws a sharp distinction between innovation and creativity, arguing that AlphaGo’s famous Move 37 was innovative but not creative, and that most current AI systems remain on the wrong side of that line.
The discussion explores open-endedness as the key frontier: building agents that generate their own goals rather than solving human-specified problems. Hughes introduces the concept of “curriculum under specification” — made tractable by modern language models — as a mechanism for guiding agents toward genuinely novel discovery rather than pre-defined benchmarks. He also discusses Inherent’s Faraday agent, which he claims outperforms frontier models and frontier coding agents including Claude on certain tasks.
The conversation spans multi-agent reinforcement learning, cultural evolution as “the fastest intelligence-generating process in the universe,” collective vs. individual intelligence, and the organizational question of what replaces OKRs in an era of recursive self-improvement. A philosophically rich episode for researchers and practitioners thinking seriously about the science of AI science.
📺 Source: Machine Learning Street Talk · Published September 11, 2026
🏷️ Format: Interview
