Hacker News new | ask | show | jobs
by zerojames 997 days ago
We have just added a section on this! TL;DR: GPT-4V isn't great at this task at the moment :)
2 comments

Back when they leaked it via a Discord bot I found it worked better when you ask it to first describe each box

Without doing that: https://cdn.discordapp.com/attachments/964175221089259591/11...

With it: https://cdn.discordapp.com/attachments/964175221089259591/11...

(though it's only one example so it could be coincidence)

Is it possible they hobbled it a bit? I know CAPTCHA solving was one of the reasons they delayed the roll-out of this feature.
Given that it fails by hallucinating the structure of the challenge instead of refusing to solve a CAPTCHA, I doubt they've intentionally reduced the capability. Although the example in your sibling comment implies it should have enough information to do it.