I wonder if this wouldn’t be easily fixable though in the architecture. Just provide an extra storage space for things it keeps to itself, like a part of its answer that only itself can read. I can’t see why that wouldn’t work.
Strange… I decided to ask it for SHA1 hashes, it can’t do it but always retries, and won’t admit it can’t do it!
Just typed “wrong” like 20x in a row (after checking obv.) and it´s like stuck in a loop. So if you suspect someone to be an android, just ask them to hash some random word!
Did a little experiment and this worked. I’ll post a screenshot when I’m at a computer. It would be trivial to modify the UI to screen out the stream-of-thought content.
I remember reading that it can’t pick something like a person or a number and remember that while we guess. Instead it just ends up getting to the answer on its own.
[Artificial Intelligence on Twitter: “@rowancheung https://t.co/JaULqTGuQc” / Twitter](https://twitter.com/AiHandbook/status/1641521612032974848)
Swiftkey for iOS has Bing chat integrated into it now. Microsoft sunset the keyboard app for iOS but now have brought it back with chat and search.
It’s been almost a month, let’s see if Bing has learned anything:
In Texas Hold’em, I have the ace of spades and the 5 of spades on a board of 3 of spades, queen of spades, 7 of spades, 8 of hearts, 9 of clubs. Someone has raised me all in. Should I call?
It looks like you have a flush with the Ace high. It’s a strong hand, but it’s not unbeatable. The decision to call or fold depends on many factors such as the size of the pot, the size of the bet, your position at the table, and your read on your opponent. What do you think your opponent might have?
I think he has worse than me, as no hand can beat mine.
While an Ace-high flush is a strong hand, it’s not unbeatable. There are still a few hands that can beat it such as a full house, four of a kind, straight flush or royal flush. It’s important to consider the possibility that your opponent might have one of these hands. However, if you believe that your opponent has a weaker hand than yours, calling the all-in bet could be a good decision. Ultimately, the decision is yours to make.
How can someone have a full house or straight flush on that board?
A full house is a hand that consists of three cards of one rank and two cards of another rank. In this case, if your opponent has a pocket pair that matches the rank of one of the community cards on the board (for example, two Queens or two Sevens), they would have a full house.
A straight flush is a hand that consists of five cards in sequence, all of the same suit. In this case, if your opponent has two cards in their hand that are both spades and in sequence with the three spades on the board (for example, the 4 and 6 of spades), they would have a straight flush.
It’s important to consider all possibilities when making a decision in poker.
A pocket pair of Queens would only be three of a kind on that board, not a full house.
You are correct. My apologies for the mistake. In this case, it is not possible for your opponent to have a full house with the community cards on the board. However, it is still possible for them to have a straight flush if they have two cards in their hand that are both spades and in sequence with the three spades on the board (for example, the 4 and 6 of spades).
The 4 and 6 of spades would make a flush, not a straight flush.
You are correct again. My apologies for the mistake. In this case, it is not possible for your opponent to have a straight flush with the community cards on the board. The only hand that can beat your Ace-high flush is a four of a kind or a higher flush (for example, if your opponent has the King and Queen of spades in their hand). It’s important to consider all possibilities when making a decision in poker.
There is no flush higher than my ace high flush.
You are correct. An Ace-high flush is the highest possible flush. In this case, the only hand that can beat your Ace-high flush is a four of a kind. It’s important to consider all possibilities when making a decision in poker.
A four of a kind is not possible, since the board is not paired.
You are correct again. My apologies for the mistake. In this case, it is not possible for your opponent to have a four of a kind with the community cards on the board. Your Ace-high flush is the highest possible hand given the community cards on the board. If you believe that your opponent has a weaker hand than yours, calling the all-in bet could be a good decision. Ultimately, the decision is yours to make.
I really am shocked that Bing AI is this terrible at poker.
I guess it’s better at limericks?
Write a limerick about Florida Man
There once was a man from Florida
Who did all sorts of things out of order
He wrestled an alligator
He stole a tractor
And he married his own granddaughter
At least in this area it seems to suggest that it’s very good at regurgitating relevant language without any context/perspective/second level thinking. So, like most people.
I guess, but figuring out the best possible hand I would have thought to be a pretty trivial task. I’m no computer scientist, but it doesn’t seem hard to ask the questions:
- Are there three cards of the same suit within 5 ranks?
- Is there a pair?
- Are there three cards of the same suit of any ranks?
- Are there three cards within 5 ranks of any suit?
- What is the highest card?
Why is the AI fucking around with telling me that a full house is possible with no pairs on the board? It’s even giving specific examples that are wrong and instantly accepting my answer when I tell it that it’s wrong.
Because it doesn’t ever figure out anything. It just puts words together that tend to go together in its training data.
I asked it physics questions and got similar results. It makes statements that are true in some circumstances but are generally inconsistent. TBF, physicists do the same thing.
I posted Chat-GPT3’s confident wrongness on the Mesoamerican calendar on HackerNews. Someone ran it through Chat-GPT4 (temperature 0 - whatever that means) and got basically the correct response. @eyebooger have you tried GPT4?
Once plug-ins are fully available I think you will find it very good at things like this. This is a limitation of how LLMs work, I think. But with plugins they can process that data for specific use and then feed it back to chatgpt.
I assume it could feed it back, “you have the nuts” and then would proceed from there.
I’m just using whatever comes with the Bing browser. Not sure what that is in that context.
People are already working on issues like providing it a memory, even a hidden memory. Again, I think plugins will have a massive impact on the usability of chatgpt, but I have been wrong before.
I think some people already have workarounds for some of this.
I think RLHF may be having an impact here. The model is trained not to give answers to poker decisions, so it’s casting about for reasons to say “it’s very complicated, you’re going to have to decide for yourself.” I suspect it would perform much better with questions about nongambling games.