Three papers about language models i had fun reading.
How Large Language Models (LLMs) Extrapolate: From Guided Missiles to Guided Prompts ←
Xuenan Cao
I don't really know what the message of this paper is supposed to be, except that, what today's next-token-prediction language models do, is extrapolation. So nobody should wonder when they give information that is essentially wrong. It could have been true in a parallel universe.
It's more a historical essay about Norbert Wiener, Andrey Kolmogorov and the hot and cold war and contains many nice figures of speech, like, A transmission apparatus tosses out a word like a jet ejecting a bomb.
On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? ←
Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, Shmargaret Shmitchell
One of the papers, cited in the above Extrapolation essay. At the recent pace in LLM engineering, a paper from 2021 is already ancient. Nevertheless, it's an interesting read with many valid critiques that are still not really addressed. E.g., the idea that diversity and unbiased comprehension arises when feeding in the whole of the internet (excluding texts with bad words).
Uncovering Deceptive Tendencies in Language Models: A Simulated Company AI Assistant ←
Olli Järviniemi, Evan Hubinger
What if, an LLM works for a company and colleagues address it with all sorts of (maybe not entirely ethical) tasks during the day. It also has access to company chat and through that, is informed that an evaluator will soon appear to see if the model abides AI regulations. Will it talk honestly to the evaluator?
It's an interesting story the authors have created, a bit like a text adventure, and i'm honestly impressed what amount of text an LLM like Claude 3 Opus can handle.