Brandon Simonsen

A mindworm that has caused great damage among AI researchers is the implicit, universal assumption that every piece of data is a (random) "sample" from a (static) "distribution". This is a valid way of modeling some phenomena, but it isn't applicable to the vast majority of real-world signals.

It’s over (for now). A majority of senators caved to the fearmongering and bush league tactics of the administration and surveillance hawks in Congress, and they sold out Americans’ civil liberties. Section 702 has been reauthorized, not just without any meaningful reforms… 1/10

this is actually hilarious. Quora SEO'd themselves to the top of every search result, and is now serving chatGPT answers on their page, so that's propagating to the answers google gives the internet is dying

It's clear that we're far from peak LLM performance -- these models will keep getting better. It's also clear that pure generation is just the first step -- we can largely alleviate the LLM reliability issue by using them as information retrieval devices over a knowledge corpus.

So far all evidence that LLMs can perform few-shot reasoning on novel problems seems to boil down to "LLMs store patterns they can reapply to new inputs", i.e. it works for problems that follow a structure the model has seen before, but doesn't work on new problems.

A shocking revelation: Companies don't have "datasets" waiting for you. It's your job to create and maintain them. This is time-consuming and hard, but primarily responsible for the success of your machine learning system. Here are some thoughts: 1 of 7