Will GPT-4 visual model (as released by OpenAI) show ability to tell if an object is inside or outside another object?
100
712
2.2K
resolved Dec 26
Resolved
YES

GPT-4 (visual model as released by Open AI, first version, and without additional enhancements using plugins or whatever, and not another company using GPT-4 in their product).

Will it be able to tell if a square is inside a circle, or if a toy is inside a box when shown an image.

I've had only limited exposure to gpt-4 using a bot that was available on discord and not sure of authenticity, but it failed. I'll resolve when i play with it myself and give it 10 such questions. Resolves YES if it gets 70% of them correct. Otherwise NO.

It must do this with the images (png, jpeg, jpg, etc) themselves. There are softwares like webGPT which simplify web pages to text which let's GPT interpret it, but that wouldn't be allowed for this market.

If someone wants to know more about this, search for "visual grounding".

Hopefully i get access soon and this will be resolved but keeping the close date till EOY for now.

Get Ṁ200 play money

🏅 Top traders

#NameTotal profit
1Ṁ856
2Ṁ786
3Ṁ124
4Ṁ108
5Ṁ99
Sort by:

t's really good at this. I've tested with some basic geometry examples and some similar to the Toy in the box examples like @NoaNabeshima and @Jacy have suggested below. Anyone want to suggest other examples before resolution? Anything that it finds particularly tricky?

sold Ṁ83 of YES

@firstuserhere I'm happy to say I was wrong here. In the "car" toy example, the only thing it doesn't get is that there is a "plane" in the image, but it doesn't say it's outside of the box or anything. Good bets, all!

predicted YES

LMK if you want me to make any particular queries @firstuserhere

bought Ṁ200 of NO
predicted NO

@NoaNabeshima I'd be curious how @firstuserhere would adjudicate between the "Yes" (inaccurate) and "in front of" (accurate) in the response.

Also, for what it's worth, I was imagining "if a toy is inside a box" as a test with an image that had a toy box with several toys inside of it and several toys outside of it. Here are two challenging ones (say, "Is the car inside or outside the box?" in one and "Is the soccer ball inside or outside the box?" in the other).

bought Ṁ1,000 of YES

@Jacy GPT-4V says
1: The car is outside the box.
2: The soccer ball is inside the box.

predicted YES
bought Ṁ500 of YES

If it is not released by EOY, how does this resolve?

@RobertCousineau I'll extend the close date in that case

boughtṀ255NO

@Mira 👀

What image styles are you going to use? Objects in objects in the natural world, objects in objects on a white background, etc?

I don't recall the paper looking into this. In fact notably most of the visual examples in the paper were images that included text, like memes and math/science textbook figures. So wouldn't be shocking if its capabilities in that area are lacking.

predicted YES

@jonsimon Indeed so, and yet this market trades at 88%, which tells us something about either the over-optimism or a lack of understanding

More related questions