It's ridiculous to equate this to a traditional Markov chain language model. Here's something a Markov chain certainly cannot do:

Human: I want to test your creativity. Please invent a new word and give its meaning.

GPT-3: Ok. Um... Tana means to hit someone with the intention to wound them.

Human: Please use the word tana in a sentence.

GPT-3: You are about to tana the man attacking you.

Human: Speak like a dwarf.

GPT-3: I ain't talkin' like a dwarf.

https://www.reddit.com/r/MachineLearning/comments/hvssqn/d_g...

Right, I admit that I don't know the first thing about ML, so I tried an experiment.

Consider a language with the tokens "{[()]}" and the following grammar:

S := S S | '{' S '}' | '[' S ']' | '(' S ')' | <empty>

That is, "[()]" and "[]()" are valid sequences, but "[(])" or "))))" aren't. A child would quickly figure out the grammar if presented some valid sequences.

I generated all 73206 valid sequences with 10 tokens and used it as input to the RNN text generator code at http://karpathy.github.io/2015/05/21/rnn-effectiveness/. After 500,000 iterations I'm still getting invalid sequences.

Am I doing something stupid, or is a RNN text generator weaker than a child (or a pushdown automaton)? Is GPT fundamentally more powerful than this?

GPT-3 can generate well-formed programs, so yes, it does things well beyond this complexity.

> After 500,000 iterations I'm still getting invalid sequences.

How frequently? If it's only the occasional issue it might be down to the temperature-based sampling that code uses, which means it will, with some small probability, return arbitrarily unlikely outputs.

How can it do that? Did it read “tana” and the meaning somewhere?

I suspect people overestimate the intelligence because they just can't grasp how much data it's ingested or don't have a visceral sense of what an ocean of data can contain. There's a saying that "quantity has a quality all its own".

I don't think this is an overestimation of intelligence. That ability is itself intelligence.

It's a good book, and one most programmers should read early on (it grows less useful over time), but to think of this as only an incremental improvement over markov chains is underselling the advance. The technology is different, and can scale to much higher levels of capability. AGI levels? Almost certainly not. But it's passing usefulness thresholds so things can progress elsewhere.

>to think of this as only an incremental improvement over markov chains is underselling the advance.

Erm, citations needed. It's a giant, inefficient and shitty KNN model, which is capable of mimicking markov chains. Wonderful marketing achievement and not much else.

https://www.gwern.net/GPT-3 (Edit: in case it's not clear, I suspect if you give an honest perusal of that page, and the pages it links, and the pages they link, you'll come away with a different opinion.)

I'm totally with you. I find GPT-3 incredibly impressive, but people are acting as if GPT-3 is a harbinger of AGI.

This is especially obvious on stuff like lesswrong, where AI is a big part of what they talk about. I tend to agree with the LW/SSC crowd about the negative effects of AGI, but they are being so hyperbolic about GPT-3.