I don’t think that would really be possible outside OA until they open up the image-input feature, which they haven’t AFAIK. You could try to do the number-array approach I think someone has suggested, but given how heavily ARC exploits human-comprehensible visual symmetries & patterns, the results would be a lower-bound at best.
Speaking of ARC, has anyone tested GPT-4 on Francois Chollet’s Abstract Reasoning Challenge (ARC)?
https://pgpbpadilla.github.io/chollet-arc-challenge
I don’t think that would really be possible outside OA until they open up the image-input feature, which they haven’t AFAIK. You could try to do the number-array approach I think someone has suggested, but given how heavily ARC exploits human-comprehensible visual symmetries & patterns, the results would be a lower-bound at best.