A Hallucinogenic Compendium - by Eryk Salvaggio
Large Language Models are not a search engine. Nonetheless, companies from Google to Meta to Bing are converting their search functions into platforms for algorithmically generated nonsense.
A few examples surfaced this week, from Google’s AI Overview suggesting that glue was a pizza ingredient to the suggestion that UC Berkeley scientists suggest eating “ at least one small rock per day.”
At the heart of these so-called “hallucinations” is the initial challenge of determining probability distributions within massive collections of text, using a pure language model, designed to predict the next words. Remember that the outcomes of Large Language Models are not designed to be true — they are merely designed to be statistically likely. So now, a company with a serious commitment to AI has to figure out how to constrain these errors of truth for a model that was never designed for its outcomes to align with truth in any way.
One of the great follies of our time may be that so many are trying to impose controls and constraints on outputs after the fact. The challenge — and the great complexity hidden behind the word “hallucination” — is that the hallucinations themselves do not have a single cause, but are the work of a confluence of factors that simultaneously assure that Large Language Models can function at all. There are several ways hallucinations can enter into text. Here, I’ll dive into a few: a short bestiary of hallucinations.
