More is different for intelligence

A new viral essay from researchers at Fulcrum Inc., titled 'More is Different for Intelligence,' presents a compelling framework for understanding the coming impact of large language models (LLMs). Drawing a direct parallel to the software revolution of the early 2000s, the authors argue we are currently in the 'naive' phase of LLM adoption, where tools like ChatGPT and GitHub Copilot primarily make existing knowledge work (coding, writing, debugging) more efficient. However, history shows that when a key resource—like computation—becomes radically cheaper, it doesn't just optimize old processes; it enables entirely new ones that were previously inconceivable, such as real-time algorithmic trading or personalized recommendation engines.

The essay posits that the true transformation will come not from AI 'pair-programming' with humans, but from designing new protocols and 'agent societies' that reorganize cognitive work itself. The plummeting cost of AI-generated reasoning and judgment will unlock 'latent processes' and 'cognitive flows' we haven't yet imagined, moving beyond simple task automation. This shifts the fundamental question from 'how do we use AI tools?' to an experimental metascience: 'How should cognitive work be organized?' The implication is that the most impactful applications of models like GPT-4o and Llama 3 won't be the ones we're building today, but entirely new organizational structures built around cheap, abundant artificial intelligence.