alt Hacker NewsThere’s Gaussian and fractal randomness. Fraud and transportation losses are Gaussian, for example - they average out to known values. An empowered LLM can wreck absolute havoc, and if it’s not empowered there’s no reason to spend $100b on training it.
This really isn’t highlighted enough. Most real world probabilities that are evaluated follow a Gaussian structure. LLMs…don’t? Fractal probably? Heavy tailed maybe (like a Cauchy distribution)? But certainly not in ways that companies are currently accustomed to.