Plotting The Curiosity Vector

A mismatch between the problem that a system developed to solve and the task that it is given can have significant consequences. Just as the human drive to obtain sweet and fatty foods can be maladaptive in a world where those foods are easily available, the autoregressive tendencies of LLMs can cause problems when they are given a task that is not next-word prediction.

Embers of autoregression show how large language models are shaped by the problem they are trained to solve

I love it that the junk_food / information_diet metaphor can also be applied in reverse, where the entities feeding on the wrong stuff are machines and not humans.

https://docs.cozodb.org/en/latest/releases/v0.6.html

This was an interesting read.

Today’s incarnation of GPTs is nothing more than a collective subconscious: different prompts will elicit different personalities and responses from them.

Private memory and individual fine-tuning of model weights according to private experience are of course required, but we need more than that. One hangover from the era of Big Data is the belief that all data must be preserved for later use. The reality is that we just get a bigger and bigger pile of rubbish that is harder and harder to make sense of. Humans don’t do this. When awake, humans remember and reason, but when dreaming, humans distill, discard, and connect higher concepts together. Random-walking LLMs on proximity graphs can do this, and the constraints are no longer measured in gigabytes but instead in minutes (hours) and joules (calories). AI also needs to rest, reflect, and sleep, after all.

Ziyang Hu