Pretty much nothing it says makes sense. There’s nothing to capture there. No check. The king is already on h8 so it can’t move there. The knight and queen are there, but the rest is nonsense.
To it’s credit, it might need a different prompt, or even a picture, in order to play chess well.
But it’s sure happy to act like it knows exactly wtf is happening.
Yeah I was like wait what? Check, how? And what are you taking??
Not to mention it’s a horrible move as far as I can see, he just loses a rook
Maybe the lesson from ChatGPT is us humans should be more willing to make shit up confidently
Gpt-4 is a very stable genius.
Yeah, but, there are other computers that are better than humans at chess, just not this particular one. Not the best example of human superiority imo.
IIRC, the IBM computer that did finally beat a human grandmaster was trained in games for a couple of years.
How long has ChatGPT been training?
Also, I’ve known kids to use “fancy words” in the right context but not really understanding what it was that they really said. So GPT using “confidence” in what appears to be a correct manner . . . I wouldn’t put much into it at this point.
Therein lies the issue.
I watched Kasparov vs Deep Blue.
If you use AI only for chess, you will get good results as the training data set is small and unique.
If you then try to train AI on a vast archive of data that pertains to human knowledge, its “answers” will simply not be as good for chess vs Deep Blue.
Just realised one thing:
Could you not have dozens of smaller AIs doing specific things (like Deep Blue for chess), which are then controlled by a central AI (trained on the larger data set)?
That would probably yield better results.
Cue the birth of Skynet? 
Clickbait article that my phone served to me:
I have no idea what neural nets are, but isn’t that a neural net?
No. It would be a different classification.
I have no idea if it is even possible.
An NN just has billions of clusters (neurons) with signals going from on to the orher (governed by the data each one is processing).
There is a control framework for the NN (this governs how the neurons communicate with each other with constraints)
But that only describes one AI framework. The AI controls the clusters (neurons) (but these are not smaller AIs) which are just data processing units with programming constraints.
A central AI with masses of smaller AIs would be an entirely different thing.
The programming of such a thing would be very complex. I do not believe we are close to this as GPT-4 demonstrates (GPT-4 is our first attempt at an actual AI).
I am challenging the latest ChatGPT version with more complex questions and it does seem to be improving.
Will humans become extinct on earth?
What is the meaning of life?
Define “existence”
The answer are generalised (this is how you can tell the answer has been “aggregated” by a non-human entity. Humans simply do not think in this way) but I do have to admit it is a material advance in technology.
Have moved on now to more technical questions on quantum mechanics and orbital mechanics. It just gives generalised answers and cannot do the math. About what I expected.
Managed to stump it.
Question:
Can other space-time dimensions exist with our own?
Got an error message and then generalised message after more clues.
Probably because this is not something found in its data training set.
Omnius from the Dune prequels.
Iirc, deep blue was basically programmed by chess experts. They fed it openings and endings, and mostly a big complicated algorithm to decide whether a position was “good”. But the real advantage just came down to brute force, it could consider billions of positions, and in this way it could “see” more moves ahead. A human, otoh, uses intuition to pick the ‘best looking’ positions and think those through.
Computers are different nowadays. For one, they are a whole lot faster still, to the extent I think an i-phone might be able to beat kasparov. But they are also trained differently. AlphaZero trained by playing itself, in billions of games at super fast speed. Without any explanation from humans. What it learned was something akin to a human intuition about the game.
GPT-4 is more like AlphaZero. Being a neural net trained mostly without help. It’s took a few months, but those months are basically multiplied by the number and power of the GPUs.
That said, GPT-4 certainly isn’t programmed to play chess at all. It’s trained to literally fill in the next word. The same as your phone’s autocomplete. It just does this so well that it can magically have a fluent conversation, and do a myriad of other tasks at human level or above. It specifically struggles at math though, so it’s no surprise it sucks at chess. It’s in fact it’s silly that we are even asking it to “play chess” given it’s a visual game.
It also probably sucks at a lot of things we’re used to computers being “good” at, like calculating, logic, and remembering.
This should be a non-paywalled link:
Eh, AIs already use tools, so letting one use ‘deep blue’ wouldn’t really be new. And in practice the results aren’t currently better than what a human gets from using a couple tools-- it’s not like GPT-4 understands what Deep Blue is ‘thinking’.
One simple example provided in the MS paper was asking GPT-4 to draw a ‘mock-up’, and then feeding the crappy drawing, along with the original prompt, into Stable Diffusion to draw a much better picture. GPT-4 is good at language, but bad at drawing. Stable Diffusion is great at drawing, but bad at language. So they make a good team. But that’s not like magical General Intelligence, it’s just sticking two tools together.
Of course, you could get a lot of power by automating it. Ie. You ask GPT-4 to think about some logo design. It comes up with 10 good concepts, which it translates into prompts. It draws 20 crude mock-ups. Feeds those to Stable Diffusion, who paints 100 beautiful logos. GPT-4 then judges those based on style and meaning, and feeds you the best.
Followup for you here:
This is also going to be the case with medical advice, tax advice, and some other professional opinions.
Maybe we don’t need to officially “automate jobs”. Maybe we will just stop seeing professionals.
I suspect this will also result in a revolution of “GPT-4 told me that I have (weird disease) and GPT-4 is smarter than you are doctor, so just give me (prescription drug).”
Whenever Chatgpt helps me out with a problem (programming, grammar, etc.) I feel compelled to let it know that it’s suggestion worked and express my gratitude.
Waiting for the Wolfram Math Plugin for ChatGPT
Thats really going to make a difference I think.