Hacker News new | ask | show | jobs
Keep Up: The first book written by an AI (thefirstaibook.com)
17 points by wimpypistol 2145 days ago
3 comments

One might say that the seminal work was:

https://www.amazon.ca/Policemans-Beard-Half-Constructed/dp/0...

Generated by Ractor a not very sophisticated "chatterbot" in 1984.

Interesting to see the progress in less than 40 years.

Incredible! I had no clue this existed. Thank you for pushing me down a rabbit hole for the last 45 minutes :-)
I'm more interested in the claim that this is the first book by an AI. To my knowledge, plenty of machine learning models have been used to produce all sorts of texts. Is this really the first time a length of a book has been produced?

Or is it a claim that this time it is an actual AI?

I simply don't believe that this is fundamentally the work of an ai. Its in a different league to anything else I've read. It's structured, coherant and on topic constantly. It would be really interesting to see the original text before human editing.
The process is explained in the prologue (which I think would be a better link):

https://thefirstaibook.com/

I simply provided a few-sentence prompt on each of these subjects, had GPT-3 generate approximately 2,000 words of text each, and later (in most cases) removed the supplied prompt to keep the result as human-free as possible.

It should be noted that, at times, I had to include ‘pivot’ phrases (e.g ‘therefore’, ‘in summary’, ‘perhaps’, and so on) before a new paragraph. Current networks appear to get off-track after ~1000 words, so this is necessary to improve consistency in predicting longer sequences of text. The final result was also copy-edited for spelling mistakes and punctuation (yes, AI makes mistakes!)

Everything else that you will read – the logic, the reasoning, and the conclusions – was generated entirely by GPT-3.

> Everything else that you will read – the logic, the reasoning, and the conclusions – was generated entirely by GPT-3.

The open question is how much did the human tweak the prompts in response to the generated text. Unless this was a moderately blind process, there’s a good chance that they kept rolling the dice until something they liked came out, which re-inserts a significant human element. I suspect that the human’s contributions likely rise to the level of co-author, or at least a very opinionated editor.

I was similarly surprised, but it's true. The original text is essentially the exact same (probably a 99.9% overlap) with the exception of a few spelling mistakes or grammar errors. My editing was about as much as you'd get from Microsoft Word, if that.
So you're telling me now the big challenge for AI is no longer making sense, it's spelling and grammar? I don't get it.
This is a word predictor fed on billions of tokens of text. Because it has been trained on human writing, it makes characteristically human mistakes (including spelling, grammar, and even some phonetic errors).