Continual learning: the missing key to AI's firm-specific value

LLMs are trained once in a massive run, then their parameters are frozen, locking in skills and knowledge. They cannot learn on the job like humans do. Context engineering—augmenting prompts with facts—can help AI remember information, but it cannot teach skills that persist across sessions. Every new chat resets the model to its original state, and while AI can memorize select facts, memorization is not learning. This lack of continual learning forces firms to rely on generic, centrally trained AI that lacks understanding of their unique rules, culture, and quirks.

The economic impact is significant. Hayek's knowledge problem, first described in 1945, holds that critical economic information is dispersed and local, not centrally available. AI without continual learning cannot stay current with such tacit knowledge. Coding works well because codebases provide a source of up-to-date context, but other domains lack a universal analog. The author argues that AI capable of continually learning tacit knowledge would be far more economically valuable than today's centralized models, unlocking firm-specific expertise that currently requires humans to painstakingly build context from scratch.