AI Reality Check: Are LLMs a Dead End?

Cal Newport examines Jan Lacoon's claim that large language models (LLMs), the foundation of tools like ChatGPT, represent a technological dead end for achieving truly intelligent AI. Lacoon, a Turing Award winner who pioneered deep learning techniques, has launched Advanced Machine Intelligence Labs (AMI Labs) with over $1 billion in seed funding from investors including Jeff Bezos and Mark Cuban, valuing the month-old, 12-person startup at $3.5 billion. AMI Labs aims to build AI through modular architectures trained specifically for real-world tasks, bypassing LLMs entirely.

LLM Limitations and the Dominant Approach

Major AI companies like OpenAI and Anthropic rely on massive LLMs as a universal "digital brain" for applications from chatbots to coding agents. LLMs work by predicting the next word or token in text autoregressively: input like "the cat sat" gets tokenized, embedded into a mathematical space, processed through transformer layers with attention mechanisms and feedforward networks, and outputs predictions trained on vast text corpora. Pre-training knocks out words for prediction, implicitly encoding world knowledge, patterns, and rules emergently.