I think when chatgpt broke when used through the router what happened is it tried to actually run a program to calculate it instead of reciting it from memory. Standard programming languages without extra libraries can't calculate that. It's way too big.
Chat gpt has its own internal routers that basically invoke sub-models. And it's own router didn't recognize that the number would be huge and not a good idea to pipe to a model that runs code.
If it was smart it would have a time limit on any code it runs and try doing it by memory. I'd think they would no what the halting problem is and that by running arbitrary code, even generated by itself, the only way to guarantee it doesn't run forever is to have an external limit. Basic sandboxing concept.
Also I'd argue that it is very easy to shuffle a deck into an order it's been in before. Shuffling isn't really random. There are magicians to take advantage of that. If you perfect shuffle, one card exactly after the other, four times, it ends up unshuffled. Even non-magicians can get to a near perfect shuffle with a few doubles slipping by, but mostly singles. If you take that sequence of how many cards from one side to the other you can calculate the Shannon entropy of the action. If it's all ones its a very low Shannon entropy, hence why it isn't really shuffling. But if it's mostly ones and a few twos then it has a higher Shannon entropy. But it's way lower than what's represented by 52!. I'd argue that if you shuffle a deck three times pretty cleanly it's plausible someone else has done it.
Log in
I think when chatgpt broke when used through the router what happened is it tried to actually run a program to calculate it instead of reciting it from memory. Standard programming languages without extra libraries can't calculate that. It's way too big.
Chat gpt has its own internal routers that basically invoke sub-models. And it's own router didn't recognize that the number would be huge and not a good idea to pipe to a model that runs code.
If it was smart it would have a time limit on any code it runs and try doing it by memory. I'd think they would no what the halting problem is and that by running arbitrary code, even generated by itself, the only way to guarantee it doesn't run forever is to have an external limit. Basic sandboxing concept.
Also I'd argue that it is very easy to shuffle a deck into an order it's been in before. Shuffling isn't really random. There are magicians to take advantage of that. If you perfect shuffle, one card exactly after the other, four times, it ends up unshuffled. Even non-magicians can get to a near perfect shuffle with a few doubles slipping by, but mostly singles. If you take that sequence of how many cards from one side to the other you can calculate the Shannon entropy of the action. If it's all ones its a very low Shannon entropy, hence why it isn't really shuffling. But if it's mostly ones and a few twos then it has a higher Shannon entropy. But it's way lower than what's represented by 52!. I'd argue that if you shuffle a deck three times pretty cleanly it's plausible someone else has done it.