Comment by ekidd | Hacker News Reader

Comment by ekidd | Hacker News Reader

Parent Full thread

> This is the same with all kinds of generative models. Humans are a critical part of the system.

Part of what I find frustrating about Stable Diffusion (and to a lesser extent Dall·E) is that they're so ridiculously far behind ChatGPT in some ways.

For example, I ask ChatGPT (in French) for a short story of three firefighters in the style of The Three Musketeers. And it writes me a perfectly serviceable little story about a brave and heroic rescue, with good pacing and structure. And it's written using the French "passé simple" tense, which is normal for fiction but virtually never used in speech. And I get this in a second or two, on my first try.

Meanwhile, I ask Stable Diffusion for a picture of "an astronaut on a horse" (the demo prompt in my copy). And I get endless badly drawn astronauts on dismembered horses. Dall·E does a bit better, but I'd obviously have to generate dozens of images and accept something visually interesting that doesn't bear any resemblance to what I envisioned. Apparently I can do better with a lot of prompt engineering. But it's a lot more work than ChatGPT.

The scale difference between text and pixels is relevant. If you asked for a novel (less information content than a 512x512 image) from ChatGPT and judged it by how it held together at a high conceptual level relative to output from a stable diffusion image I don't think it would seem that different.

> If you asked for a novel (less information content than a 512x512 image) from ChatGPT and judged it by how it held together at a high conceptual level relative to output from a stable diffusion image I don't think it would seem that different.

Try asking ChatGPT to write you a novel, chapter by chapter. The chapters may seem okay individually, but you'll find chapter 5 contains statements which completely contradict those in chapter 1. And this can still be true even if ChatGPT is producing chapter 5 based on chapters 1 to 4 as input. (I think ChatGPT could do with some more training and/or engineering on detecting and avoiding contradictions in its output.)

I think one advantage of generating images, is you can notice problems with the output very quickly. With text, you might not notice the contradiction between chapters 1 and 5 until you reread it days later. Similarly, if you get two generated images, and you like bits of both of them, it isn't too hard to edit them together (even roughly) in an image editor, then use further generative techniques to smooth that roughness out. If ChatGPT writes you two different fifth chapters for your novel, and you like bits of both, merging them together can be a more tedious process.

> The chapters may seem okay individually, but you'll find chapter 5 contains statements which completely contradict those in chapter 1.

I can't find documentation for ChatGPT's "context window", or how much text it can "see" at once. Some people say it might be 4k, or maybe more. But I would be surprised if it can actually work with multiple full-length chapters at once. So it's essentially like giving chapter outlines to multiple authors and gluing them into a book.

But as long as you stay within the context window, ChatGPT usually has no problem remembering what it's talking about. This isn't GPT 2, which wandered all over the place incoherently. ChatGPT is capable of writing a 1-page short story that is consistent throughout, and that has a clear structure and progression.

ChatGPT will still completely contradict itself sometimes, even with a quite brief conversation.

e.g. ask it for instructions on how to commit a crime, it will tell you its ethical guidelines prohibit it from answering. Ask it for more detail on the contents of those ethical guidelines, it will tell you that as a large language model, it is incapable of possessing any such thing as ethical guidelines. Point out it just contradicted itself, it is likely to respond that as a large language model it is incapable of contradicting itself, or else give some non-sensical explanation as to why the two statements aren’t contradictory

Ask it if it agrees with what you just said, it might reply that yes it does. Ask it again, it may instead give you a lengthy lecture on how as a large language model it isn’t capable of agreeing with anything.

I suspect it just wasn’t trained enough on the importance of self-consistency; nor on the principle that it is better to honestly admit to contradicting yourself than to give a non-sensical defence of your own contradictions

I even wonder if some of its training may have actually been contradictory - if trainers individually reward it both for saying “my ethical guidelines don’t permit me to answer that”, but also “I’m not capable of having ethical views” when pressed for judgements on ethical matters.

I once asked it to write a bit of code for me and it didn't give me the whole file and when I said: "please send the whole file" it said "I gave you the whole file, but here it is again" and then it gave me the whole file

Stable Diffusion is tiny compared to ChatGPT.

The largest image generation model I've seen images from is Google Parti, which is an order of magnitude larger than Stable Diffusion. Still small enough that you might be able to run it on a 24GB consumer GPU with some optimizations though, so hopefully Stable Diffusion will get there in some months.

Have a look at the four kangaroo pictures here to get an idea of the difference https://parti.research.google/

SD's text embeddings aren't nearly as advanced as GPT3, so this makes sense. It'll get there; multi-modal models are only getting better.

Stable Diffusion is a free 4GB model that can run on a phone. ChatGPT size is unknown, but around 350GB.

You can use a large text model to design prompts for a smaller image model if you want though.

Wow 350GB is tiny in the grand scheme of things.

Reply on news.ycombinator.com