Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

I let Claude Code write an entire book

github.com

Post reply on HN
Print view View original

71–80 of 84 posts

Re: I let Claude Code write an entire book

#71

Post by Tadpole9181 »

kemotep wrote: post #67
ToValueFunfetti wrote: post #56

Earlier quoted context omitted.

I find I'm willing to explore a minecraft world or puzzle through a nethack dungeon that nobody bothered to create. You could argue that humans made the biomes or defined the layout constraints, but humans also supplied the training data for an LLM. I guess it comes down to whether the art is any good, with procedural generation being mostly irrelevant? But perhaps a book is different

I mean the difference with Minecraft is each part that is procedurally generated was made by a human or involved human input into the design decisions. Unless you are suggesting Notch was a generative AI model he made Minecraft.

What? That's not true at all, sans structures. Minecraft worlds are infinite and unique, using randomly generated noise textures as the basis. Nobody "made" each "part".

And arguing that a human tweaking noise parameters is somehow more creative than humans distilling their entire knowledge and cultural repertoire into a machine, then working with that to produce literature with a guided hand seems quite silly.

Re: I let Claude Code write an entire book

#72

Post by loudmax »

ToValueFunfetti wrote: post #56
amake wrote: post #16

Why would I read a book that a person couldn't be bothered to write?

I find I'm willing to explore a minecraft world or puzzle through a nethack dungeon that nobody bothered to create. You could argue that humans made the biomes or defined the layout constraints, but humans also supplied the training data for an LLM. I guess it comes down to whether the art is any good, with procedural generation being mostly irrelevant? But perhaps a book is different

I don't think I'd want to read a novel that was generated by an algorithm, but I might be up for a Choose Your Own Adventure style game, which might be a better analogy to Minecraft or nethack.

Re: I let Claude Code write an entire book

#73

Post by MSFT_Edging »

karmakurtisaani wrote: post #68
MSFT_Edging wrote: post #32

Earlier quoted context omitted.

This is by definition slop. A mathematical average of training data. By design it contains nothing novel.

A small correction: it can contain novelty in the form of AI hallucinations.

Which is happenstance as opposed to intentional novelty.

Re: I let Claude Code write an entire book

#74

Post by etblg »

fragmede wrote: post #57
debesyla wrote: post #41

Earlier quoted context omitted.

Just like with mass-manufactured furniture, some of us seek hand-made items, as it feels more human, despite its flaws.

Realy? Does having flawws really make four better reading? Okay, I'll admit that hurt me to right (as did that) but writing isn't furniture, and other than a couple of tells which I haven't kept pace with (eg use of the word "delve"), the problem with trying to key off of LLM generated content and decide quality, is that you can't tell if the LLM operator took three minutes to copy and pasted the whole thing (unless…

Here it is cleaned up by ChatGPT:

"Really? Does having flaws actually make for better reading?

Okay, I’ll admit—that hurt to write (as did that last sentence), but writing isn’t furniture. Aside from a few tells I haven’t kept pace with (like the overuse of the word “delve”), the problem with trying to judge quality based on LLM-generated content is this: you can’t always tell whether the operator spent three minutes copying and pasting the whole thing (unless they accidentally leave in the prompts—which has happened and is a dead giveaway that no one even skimmed it), or if they took the time to thoughtfully consider the questions ChatGPT asked about what the writing should contain.

If you’ve made it this far: do mistakes like these really make for better reading?"

And I'm going to have to say: yes, I enjoyed reading your weird paragraph more than the ChatGPT sanitized version of it.

Re: I let Claude Code write an entire book

#75

Post by karmakurtisaani »

MSFT_Edging wrote: post #73
karmakurtisaani wrote: post #68

Earlier quoted context omitted.

A small correction: it can contain novelty in the form of AI hallucinations.

Which is happenstance as opposed to intentional novelty.

Yeah I was writing it tongue in cheek, should have maybe been a bit more obvious about it.

Re: I let Claude Code write an entire book

#76

Post by dinfinity »

evanelias wrote: post #69
dinfinity wrote: post #60

Earlier quoted context omitted.

If someone claims "the book is total shit", it is entirely reasonable to ask if they've actually read it, regardless of who or what wrote it.

If you flip through something claiming to be a "book" and immediately see that a majority of pages just contain nonsensical bulleted lists, and furthermore see that chapter titles are printed overlapping with the book title on each page, you can correctly conclude the entire thing is a zero-effort pile of shit without wasting any further time to read it.

I read through a bit of it and it really wasn't all that bad. The only thing that I found to be really problematic were the made up experiences. Clearly hallucinations are still a big problem for LLMs, but if we manage to get rid of those a book like this can really be quite serviceable (a lot of human-written books are badly written so the bar isn't incredibly high, imho).

The creator should really tweak the prompt/process to include automatic review explicitly intended to remove hallucinations. It clearly is already the intent: "Future iterations of this experiment will include AI-powered fact-checking of the content."

I'm looking forward to what the improved version will look like.

Re: I let Claude Code write an entire book

#78

Post by evanelias »

dinfinity wrote: post #76
evanelias wrote: post #69

Earlier quoted context omitted.

If you flip through something claiming to be a "book" and immediately see that a majority of pages just contain nonsensical bulleted lists, and furthermore see that chapter titles are printed overlapping with the book title on each page, you can correctly conclude the entire thing is a zero-effort pile of shit without wasting any further time to read it.

I read through a bit of it and it really wasn't all that bad. The only thing that I found to be really problematic were the made up experiences. Clearly hallucinations are still a big problem for LLMs, but if we manage to get rid of those a book like this can really be quite serviceable (a lot of human-written books are badly written so the bar isn't incredibly high, imho). The creator should really tweak the prompt/…

Flip through the middle of the book. Nearly every page has either a bulleted list or a numbered list. In several cases, a single list spans multiple pages.

That’s the format of an outline, not a legitimate book.

Re: I let Claude Code write an entire book

#79

Post by kemotep »

Tadpole9181 wrote: post #71
kemotep wrote: post #67

Earlier quoted context omitted.

I mean the difference with Minecraft is each part that is procedurally generated was made by a human or involved human input into the design decisions. Unless you are suggesting Notch was a generative AI model he made Minecraft.

What? That's not true at all, sans structures. Minecraft worlds are infinite and unique, using randomly generated noise textures as the basis. Nobody "made" each "part". And arguing that a human tweaking noise parameters is somehow more creative than humans distilling their entire knowledge and cultural repertoire into a machine, then working with that to produce literature with a guided hand seems quite silly.

But someone did design a diamond block and a grass block and program their properties and model them and add them to the procedural generation system with rules on how they should be placed.

An LLM would from whole cloth create a block, how it works, and how it would be randomly generated. That’s why the current MinecraftGPT doesn’t have any consistency if you turn around in 360 degrees. Everything is being generated on the fly and how it works. Once you generate that Minecraft world, how it works and what it looks like is static and why it works the way it works was designed entirely by people.

Re: I let Claude Code write an entire book

#80

Post by gjm11 »

novosel wrote: post #70
gjm11 wrote: post #61

Earlier quoted context omitted.

Why? I emphasized "lot" because presumably the training data for a high-end model like Claude includes literally millions of books, far more than any human being has read even a few pages of. To me, "Claude has read a lot of books" would read pretty oddly here. I suppose the idea would be to contrast with the (hopefully few) books it's written to date, but I can't think why the emphasis would be called for.

Yes, I understood your argument, and your emphasis is on the point there. I am simply asking in what sense did Claude read all those books. I am sorry if I have hijacked the disscusion. edited: spelling

It's read them in the same sense as it's written a new one: it's ingested them, had what pass for its mental processes influenced by them, remembered some of their contents, etc.

You might want to say that that isn't "reading", just as we never say that aeroplanes "fly" since they don't do the same thing as birds, never say that computers "play chess" since the calculations they do are very different from those done by human chessplayers, never say that machines "dig ditches" since they don't have the experience of tired muscles and the sun beating down on their backs, etc. As you may gather from the tone of the previous sentence, I am not altogether convinced.

I do agree that the two things aren't equivalents. I can imagine futures in which AI systems have a training process that somewhat resembles the present one, and also do something that corresponds more closely to human reading[1], and then I'd want to reserve the term "read" for the latter. What Claude has done to all the books that helped shape it isn't exactly reading. But it's quite like reading for our present purpose; when someone jokes about an author having written more books than they've read, they mean that the author doesn't have much awareness of other people's work, and that isn't a problem Claude has.

[1] E.g., they might have some sort of awareness of the process, whatever exactly that might mean; they might do it "for pleasure", whatever exactly that might mean; etc. Those things would require the AI systems to resemble humans in ways present AI systems aren't designed to and, so far as I can tell, don't; maybe some future AI systems will be that way, maybe not.

Post reply on HN