DALL-E 2 returns nonsense like this:
I'll give people a week to find such a prompt on DALL-E 3, and if it gets all the details correct, resolve YES.
1,000🏅 Top traders
| # | Trader | Total profit |
|---|---|---|
| 1 | Ṁ180 | |
| 2 | Ṁ147 | |
| 3 | Ṁ100 | |
| 4 | Ṁ23 | |
| 5 | Ṁ22 |
@IsaacKing it's a similar problem to:
/firstuserhere/will-any-ai-image-model-generate-co
But more difficult because more bits I think (30 days vs 26 keys). I was able to do better than your above picture using similar approaches to my attempts the qwerty keyboard (some posted in the comments of the linked market), but yeah, still a ways off.
@chrisjbillington Hmm, I think it's easier in some ways and harder in others.
Numbers are always in the same order, and this order shows up all over the internet, so if DALL-E is able to learn this order, it should be able to generate a calendar page. QWERTY keyboards are surely much less common in its training data, so I'd expect it to have a harder time learning that order.
On the other hand, QWERTY keyboards can be assumed to be the exact same order every time, whereas calendar pages will be inconsistent with regards to what day of the week they start on.
@IsaacKing I agree with the intuition, but image models seem to show a stark inability to abstract and generalise about that sort of flavour of thing - I get the feeling the abstract concept of "order" of numbers may not really be something they can make much use of. Empirically, they're doing better at the keyboard layouts at the moment (my best attempt below), and I haven't once seen output like "ABCDE...".
