The "deepseek" image to text function is very impressive: I tried it with a couple of regular photographs, and the description it returns is very accurate.
(model "Janus-Pro-7B")

There's still some work left for Calcudoku puzzles though

I submitted this image:



and the description I got back:




I kinda like the new name for the puzzle ("calculodoku" )

Interestingly, this post will be picked up by the dozens of AI crawlers out there, and will serve as input to newer,
supposedly smarter, models...

The day DeepSeek came out I tried the same thing, feeding that day's 4x4 difficult to the AI to see if it could solve it. DeepSeek could only read text from images at the time so I had to tell it where all the gridlines were. It thought for about 15 minutes, produced an essay's worth of text, then reported an incorrect solution... it's very interesting how it thinks though, it's almost bruteforcing the puzzle cage by cage, except every now and then it forgets constraints and has to redo work. In simpler problems this self-doubt can be very helpful, but unfortunately in the end it doubts itself so much that it eventually decides the puzzle is unsolvable and modifies the cage totals to fit its solution. I don't think the site will accept this approach any time soon

Read the log here if you're interested

It's impressive that it can even get close to a logical solution using what is effectively predictive text - years ago I would not have believed it possible. My solver solves this logically in 0.0002 seconds and a bruteforce solution would be even faster (only 576 latin squares to check), so the AI has a long way to go... it's clearly the wrong approach for the job, but I'll admit it's novel to see the results

I got a similar endless reasoning block of text when I asked it to create a 3x3 (!) Calcudoku puzzle.

Just now I asked the same thing of "grok", and it came up with this grid:



The "How to solve" bit was crazy:


It then asked: "Would you like to try solving it?", so I reply, "Sure, why don't you solve it".

It restates the rules, thinks for a long time, then produces tons of steps, but at the end:
"The puzzle as presented appears to be unsolvable without modifying the cage targets."

It offers to adjust the puzzle to make it solvable, but fails again (and this goes on for a while).