🏅 Top traders

1. Does the image have to be done by a general purpose text2image diffusion model, or does a huge CNN built to solve it do?

2. Is there a check for coincidence built into the resolution criteria? Does it have to be a probability like at least 5%.

3. Does multi-model stuff like GPT-4 + DALL-E 3 count? What about byte level tokenization?

4. Are vector graphics considered text2image?

@31ff
1. It should produce an image, given text. Diffusion, Gan, or any other model type does not matter.
2. It should be at least 5%
3. dalle-3 + gpt-4 would count, but I didn't have much luck with that.

4. no, only pixels

@MikhailDoroshenko Does "DALL-e 3 + GPT-4 would count" allow for GPT-4 transforming the prompt to include the solution?

Since you said in the criteria that including the solution in the prompt would not count, I assume images involving GPT-4 as an intermediary will require us to inspect the DALL-E 3 prompt and ensure GPT-4 didn't give DALL-E the answer, in order for it to count.

@MikhailDoroshenko What if I attempt to train a CNN that basically renders any text given and input the solution by GPT-4? What if I generate a billion examples and train a transformer to pixel by pixel predict the solution as image (kind-of like DALL-E Mini, through perhaps left to right to encourage reasoning)? Does creating a new model to solve it and it alone count as text2image solving a quadratic equation?

@chrisjbillington @MikhailDoroshenko This seems trivial if one can use any prompt template where GPT-4 is allowed to solve the equation and give it to DALL-E 3.

@Jacy yeah, obviously that's the entire question.

@chrisjbillington I don't think it's so obvious, but I bet the market up a little in spirit. I'd bet you 1:10 for a thousand mana though if you want, since we like betting against each other on question interpretations.

@Jacy what are we betting on? I'm saying it's obvious an LLM can do maths and an image generation AI can't, but I don't know what counts for this market so I haven't bet.

@chrisjbillington the question resolution (i.e., you can bet on the limit orders I set up, if you want, and I can free up more mana if you're willing to bet more than that)

@chrisjbillington oops, did not see that question. At first I obviously thought about text->image model only, but it seems that LLM that solves equation and then prompts image model is close enough in spirit. Do you think that this would be trivial to prompt engineer gpt-4 to produce a good solution in a single image? It doesn't feel like it is.

@MikhailDoroshenko yes, but if you make comments like this, Chris and I are going to start zooming the market from 5% to 95% and back and forth based on what you say :)

@Jacy oh no

@Jacy can you show me an example of such solution? Just for me to see whether it feels right 😁

@Jacy I have no mana so you'll have to do the zooming back and forth by yourself.

@MikhailDoroshenko if we're allowed to cherry-pick an easy quadratic then yes, if it needs to work for any quadratic then no.

@chrisjbillington and if it's allowed to use it's data analysis plugin or whatever it's called, then trivially yes it can solve all quadratics.

@chrisjbillington but not display solution as an image from image model

@MikhailDoroshenko I guess the image model wont be able to display complex text, so that puts a limit on what quadratics it can display the solution to, yes.

@MikhailDoroshenko You could give ChatGPT the prompt "A small figurine of the number that is the solution to the equation "X^2=9"". I contend that this technically meets the market resolution criteria because (i) it is a single text-to-image model as you've defined it, (ii) the prompt does not contain the solution (X^2=0, 1, or 4 would be easier, but arguably those each contain the solution), and (ii) the output is the number 3, which is the solution to that quadratic equation.

I can also make the case a little stronger by building you a custom GPT that does this, which is accessible with this link: https://chat.openai.com/g/g-46TSxuYXO-tti-model-that-solves-simple-quadratic-equations

You can also add extra tools that can make it much better, to the point of just calling a quadratic equation calculator, though the image production step will likely fail often once you get to double-digit solutions.

@Jacy ugh, by solution I meant the whole text of the solution, but I see how it can be unclear. I would not resolve this yes based on this example only. But if you feel strongly about that I can compensate your losses.

@MikhailDoroshenko what do you mean by the whole text of the solution? "X=3", "3^2=9", "3, -3", or something else?

@MikhailDoroshenko you mean the worked solution, not just the answer?

@chrisjbillington yes.

@Jacy x^2=9

x=3 x=-3

Would suffice for me

@MikhailDoroshenko sigh. Chris trades too fast for anyone to beat him!

I'm pretty confident I can make a custom GPT that meets all of your stated criteria, but I don't know if you will accept it, and I know Chris has the fastest reaction time on Manifold, so I have sent the URL to access the custom GPT via message.

@Jacy The problem isn't about GPT-4 solving the problem, it's about getting DALL-E 3 to render

x^2=9

x=3 x=-3

This is basically impossible for DALL-E 3 to do.