71–80 of 109 posts
Hmm, tempting to make a thing that just loops: description = gpt4("describe ") and chain that into new_result = gpt4(" was described by a human as but it should have been visibly apparent as a unicorn. provide SVG output that ") and post that as today's output. I wonder what we'll get. I'll do it manually.
Earlier quoted context omitted.
Aren’t they using the March 14 model like the general public? It’s frozen in time, there are no updates to it.. All of these will be drawn using the same model until they push a new update, or you switch to a different GPT But I already think they proved the point that the generation is random enough that it would be extremely difficult to track progress this way.
GPT's output is by default somewhat random. If you ask the same exact question several times, you'll potentially get several different answers. Each successive word in the output is chosen from a distribution of possibilities -- that distribution is fixed, but that actual sample chosen from the distribution is not fixed. See, e.g., https://platform.openai.com/docs/api-reference/completions/c...
I asked gpt4 to "draw a unicorn using svg", and it gave me way better results: http://www.w3.org/2000/svg " viewBox="0 0 100 100" width="100" height="100">
\\ \\ \\ ,__, \\ (oo)____ (__) )\\ ||--|| * “Draw an ASCII unicorn” (GPT4)
\\
\\
\\`*-.
) _`-.
. : `. .
: _ ' \\
; *` _. `*-._
`*-* `*-._ `-.
|_| `-._*-*_
I asked gpt4 to "draw a unicorn using svg", and it gave me way better results: http://www.w3.org/2000/svg " viewBox="0 0 100 100" width="100" height="100">
Can you post screenshots of these results?
Look at this majestic beast GPT-4 just drew! https://imgur.com/a/ogZ6qI0 Interestingly, I think you often get better results with ChatGPT web interface than the API (the OP used a system message), but for a funny reason. Because ChatGPT typically first responds by insisting it can't actually draw, it ends up accidentally chain-of-thought prompting itself into drawing better. "I can't draw, but I can describe a detail…
Earlier quoted context omitted.
Yes, this project seems like a misunderstanding of what Bubeck and team were obsering with their unicorn test. GPT-4 was being trained, and checkpoints were provided to them to experiment with. The improvements in the unicorn reflected further training progress. The models on offer now are frozen(-ish). Per the models[0] page, the non-snapshot model IDs "[w]ill be updated with our latest model iteration". So this pro…
If you guys think Chatgpt and gpt-4 are static you haven't been using it. The answers change constantly (and not because of the inherent randomness of the response but openai is constantly making it "safer" and improving it's output via humans) - basically when any article comes out saying "Chatgpt can't solve X puzzle) within a day suddenly it can solve that puzzle perfectly. I can't tell you how many jailbreaks jus…
Earlier quoted context omitted.
If you guys think Chatgpt and gpt-4 are static you haven't been using it. The answers change constantly (and not because of the inherent randomness of the response but openai is constantly making it "safer" and improving it's output via humans) - basically when any article comes out saying "Chatgpt can't solve X puzzle) within a day suddenly it can solve that puzzle perfectly. I can't tell you how many jailbreaks jus…
I think improved filtering for jailbreaks is very unlikely to correspond to the kinds of model improvements that would result in drawing a better unicorn.
Which is very interesting. You already have a model that consumes nearly the entire internet with almost no standards or discernment, whereas a smart human is incredibly discerning with information (I’m sure you know what % of internet content that you read is actually high quality, and how even in the high quality parts it’s still incredibly tricky to get figure out good stuff - not to mention that half the good stuff is actually buried in low quality pools). But then you layer in political correctness and dramatically limit the usefulness.
I got this: https://www.svgviewer.dev/s/yldwue0Q
I would argue this is better than what they're achieving.
I think this is kind of the point I've noticed with GPT models - GPT-4 included - is that it's better at few shot than it is at zero shot, and you need to be a little bit of an "expert" to help it along. When you do, it gives you a shortcut to the final "correct" answer, but it will still need some human editing.
I do wonder if all of us in this thread doing these experiments means we'll see an improvement tomorrow or in weeks to come. I'm also keen to see what GPT-5 looks like in this regard.
EDIT: I noticed some people asked to "make it better" and it did. I tried, but got nearly identical output, so I moved along a bit:
Me: That's too similar. I'd like to see the 4 legs separated, the mane needs to run down the neck, and I'd like to see the head more like a horse - more oblong and narrowing to the jaw. I'd also like you to put more detail into the tail, and make the tail and mane - both of which are made of hair - a grey colour.
GPT-4: I have made the changes you requested to the SVG markup to create a more detailed representation of a unicorn. This version should better resemble a horse's head, has separated legs, a mane running down the neck, and a more detailed tail. The mane and tail are also colored grey, as requested.
https://www.svgviewer.dev/s/WQcdOhVP
That output is... better? Maybe? I mean, it's followed my instructions (apart from the legs), I guess...