Show HN: Claude created an "unsolvable" puzzle, but you can solve it

[logicallee]

https://claude.site/artifacts/2bbdfb89-da23-4de0-a047-e6c8b167d335

I had Claude (Sonnet 3.5 New) create a sliding puzzle, and I noticed it added code to make sure it is solvable. I asked about it and it explained that without it, the puzzle would be unsolvable. Its explanation didn't make too much sense for me so next I had Claude create an unsolvable version. I was able to solve it, so I changed it to include a congratulations message.

Go ahead and try to solve this "unsolvable" puzzle. You should be able to succeed!

https://claude.site/artifacts/2bbdfb89-da23-4de0-a047-e6c8b1...

[yorwba]

For a sliding puzzle of this kind, half of all starting positions are unsolvable: https://en.wikipedia.org/wiki/15_puzzle#Solvability

Maybe if you ask Claude to write code that generates random starting positions and checks them for solvability, it'll succeed at finding an unsolvable one.

[jqpabc123 → yorwba]

If "Claude" was truly "intelligent", shouldn't it have either; A) tried this on it's own; or B) admit it had no clue how to proceed.

Expecting intelligence to somehow "emerge" from a statistical process is absurd. Yet $ billions are being funneled down this rabbit hole by tech "leaders" who should know better.

[jqpabc123]

I solved it in 99 moves. I did twice just to verify.

LLMs are simply not trustworthy. Anything they produce requires verification --- which defeats a huge chunk of their utility.

[logicallee → jqpabc123]

Good job! I agree.

[logicallee]

Here is a version that includes the instructions on how to play:

https://claude.site/artifacts/161c31dd-3c64-43ac-b250-0fc462...