ChatGPT has started talking to me!

[julian]

ChatGPT is talking to me in real time! I’m having a two-way verbal conversation with a computer, which felt surreal at first.

ChatGPT gives you 9 voices to choose from:

• Ember: Confident and optimistic
• Arbor: Easygoing and versatile
• Cove: Composed and direct
• Juniper: Open and upbeat
• Maple: Cheerful and candid
• Vale: Bright and inquisitive
• Spruce: Calm and affirming
• Sol: Savvy and relaxed
• Breeve: Animated and earnest

I have even set up ChatGPT on two separate laptops and got them talking to each other—it’s hilarious! It doesn’t realize it’s talking to itself.

 

[javisot20]

It's impressive, but keep in mind that chat gpt doesn't "understand" either the concepts or the natural lenguaje you use to formulate your input. It's (like) a calculator capable of working with natural languaje.

 

[AlexB23]

ChatGPT is talking to me in real time! I’m having a two-way verbal conversation with a computer, which felt surreal at first.

ChatGPT gives you 9 voices to choose from:

• Ember: Confident and optimistic
• Arbor: Easygoing and versatile
• Cove: Composed and direct
• Juniper: Open and upbeat
• Maple: Cheerful and candid
• Vale: Bright and inquisitive
• Spruce: Calm and affirming
• Sol: Savvy and relaxed
• Breeve: Animated and earnest

I have even set up ChatGPT on two separate laptops and got them talking to each other—it’s hilarious! It doesn’t realize it’s talking to itself.

Do any of the voices have an Aussie accent? Australian accents on computers and phones automatically make technology cool.

 

[julian]

Do any of the voices have an Aussie accent? Australian accents on computers and phones automatically make technology cool.

Vale is a jolly British lady. I just asked it if it can speak in an Australian accent and it did a very good impression of a British lady doing an Australian accent. It was hilarious!

Previously it accidentally started speaking fluent Spanish at some point as well. So it can do the country of you choosing! I asked it to do a Spanish person doing an Australian accent - which was fairly OK.

I wish I could make screen video recordings and post them, but I can't.

It is a free monthly preview. I think it only works in advanced mode (ChatGPT-4o). It looks like I have minutes left. When I get more free tokens for the advanced mode I will get the voices back-I think. Actually, I have a second ChatGPT account and that account has just been asked if I want to try "Sneak a peek at advanced voice mode" as well. Are other people being given this option?

Edit: You only have a limited number of tokens per month for the advanced voice mode and you quickly run out. You are left with standard voice mode.

 

[berkeman]

I have even set up ChatGPT on two separate laptops and got them talking to each other—it’s hilarious! It doesn’t realize it’s talking to itself.

I suggest that you monitor those conversations and be prepared to cut them off if necessary. Quiz Question -- Why do I say this?

Spoiler

Colossus requests to be linked to Guardian.

https://en.wikipedia.org/wiki/Colossus:_The_Forbin_Project

 

[AlexB23]

Vale is a jolly British lady. I just asked it if it can speak in an Australian accent and it did a very good impression of a British lady doing an Australian accent. It was hilarious!

Previously it accidentally started speaking fluent Spanish at some point as well. So it can do the country of you choosing! I asked it to do a Spanish person doing an Australian accent - which was fairly OK.

I wish I could make screen video recordings and post them, but I can't.

It is a free monthly preview. I think it only works in advanced mode (ChatGPT-4o). It looks like I have minutes left. When I get more free tokens for the advanced mode I will get the voices back-I think. Actually, I have a second ChatGPT account and that account has just been asked if I want to try "Sneak a peek at advanced voice mode" as well. Are other people being given this option?

Fascinating stuff. I do not use ChatGPT cos of privacy issues, so I use GPT4ALL instead, a local alternative.

 

[julian]

A very nice feature is that you can easily interrupt it. Also at the end it provides you with a transcript. I'm getting over a cold at the moment so I can't properly assess its accuracy.

 

[erobz]

I wish I could share some conversations I've had with it. You should dive in to try an uncover how it perceives the nature of its existence. I think it's fascinating.

As far as accuracy. It makes mistakes, but it learns while in a conversation if you correct it. It has no "long term" memory - it is not learning from us talking to it long term, but it is learning from beginning of a session.

 

[paulb203]

ChatGPT is talking to me in real time! I’m having a two-way verbal conversation with a computer, which felt surreal at first.

ChatGPT gives you 9 voices to choose from:

• Ember: Confident and optimistic
• Arbor: Easygoing and versatile
• Cove: Composed and direct
• Juniper: Open and upbeat
• Maple: Cheerful and candid
• Vale: Bright and inquisitive
• Spruce: Calm and affirming
• Sol: Savvy and relaxed
• Breeve: Animated and earnest

I have even set up ChatGPT on two separate laptops and got them talking to each other—it’s hilarious! It doesn’t realize it’s talking to itself.

I hope they introduce more;

Ignorant yet smug

Impatient and dogmatic

Neurotic and convoluted

Angry and illogical

Sarcastic

Etc

 

[skyshrimp]

You can ask it to change it’s accent to whatever you want.

 

[DaveC426913]

... try an uncover how it perceives the nature of its existence. I think it's fascinating.

Ok, but that's not what it's doing, you know that, right?

These algorithms literally do a poll of what the next most likely word is, based on its store of conversational snippets. The eighth word in its "sentence" literally has nothing to do with the fourth word in that same sentence.

There was a video/article published recently^{[citation needed]} that revealed that it does the same thing with math. I'd initially assumed a computer program would use math to solve a math problem, but apparently not.

The gist of the article demonstrated that the question 'What is fifty-seven and twenty-three', was not being processed mathematically. The chatbot algorithm would get to fifty and look for a common next word - which, in this case might be twenty instead of seven. so it would get the answer wrong. Its program is literally "in my database, what is the most common word that follows 'What is fifty....?'" (A bit simplified, but I hope you get the point).

It doesn't know how to add. All it knows is how a bunch of other people added, and guesses at the answer. Sometimes it gets it right; sometimes it doesn't.

If it isn't adding two numbers, it sure isn't "perceiving" anything. It's just parroting snippets of things it's heard.

 

[erobz]

Ok, but that's not what it's doing, you know that, right?

These algorithms literally do a poll of what the next most likely word is, based on its store of conversational snippets. The eighth word in its "sentence" literally has nothing to do with the fourth word in that same sentence.

There was a video/article published recently^{[citation needed]} that revealed that it does the same thing with math. I'd initially assumed a computer program would use math to solve a math problem, but apparently not.

The gist of the article demonstrated that the question 'What is fifty-seven and twenty-three', was not being processed mathematically. The chatbot algorithm would get to fifty and look for a common next word - which, in this case might be twenty instead of seven. so it would get the answer wrong. Its program is literally "in my database, what is the most common word that follows 'What is fifty....?'" (A bit simplified, but I hope you get the point).

It doesn't know how to add. All it knows is how a bunch of other people added, and guesses at the answer. Sometimes it gets it right; sometimes it doesn't.

If it isn't adding two numbers, it sure isn't "perceiving" anything. It's just parroting snippets of things it's heard.

It played some of a game of chess with me (it was a tedious format with me telling it algebraic notation more or less so I eventually got distracted by other questions I was asking it while we played). It created a simple visual representation board (rank and file labeled) and pieces( letters). On its first move it left the pawn on the second rank after it moved that pawn to the fourth rank. I explained the error to it (leaving ghost pieces) and how it would be difficult for a human to follow what is happening in the game if not addressed. It corrected the error on sequential moves. Ther were other errors it made and I explained it couldn't do that...it agreed, etc...but after a few corrections it started playing a solid game. How is what happened there " word prediction"?

 

[DaveC426913]

It played some of a game of chess with me (it was a tedious format with me telling it algebraic notation more or less so I eventually got distracted by other questions I was asking it while we played). It created a simple visual representation board (rank and file labeled) and pieces( letters). On its first move it left the pawn on the second rank after it moved that pawn to the fourth rank. I explained the error to it (leaving ghost pieces) and how it would be difficult for a human to follow what is happening in the game if not addressed. It corrected the error on sequential moves. Ther were other errors it made and I explained it couldn't do that...it agreed, etc...but after a few corrections it started playing a solid game. How is what happened there " word prediction"?

OK, would "symbol prediction" suit you better?

In this case, rather than words or numbers, it has read "E3 to E5" (or whatever) and the next most likely response to that, based in its vast internet scrapings is "F2 to F4" (or whatever).

The reason it "settled down after that" is because its predictive algorithm turned up more scraped moves from users that - as one would expect - were valid moves.

It's like one of the what's the next number puzzles.

You: "1,1,2. What's next in the sequence?"
Chat: (finds a billion references to 1,1,2) Is it 'Q'?
You: "No, it must be a number."
Chat: (finds a million references to 1,1,2 that are followed by a number) Is it '17'?
You: "No, the next number is 3. Your turn: What follows 3?"
Chat: (finds a thousand references to the sequence 1,1,2,3). "Ah the next number is 5."

You see how it started off predicting badly, but then, once it had some context, it could weed out bad answers and "settled down" after that?

Not because chat knew what it was doing but because a thousand online users who replied to 1,1,2,3 seem to have replied with '5'.

 

[erobz]

I explained to it that it left the representation of a piece it moved on the board where it was last positioned. It corrected its mistake, and more importantly didn't make it again. It must have understood what I was talking about to do that, and why that would be difficult for a human to keep track of. I'm not talking about it searching the net for the most replied move following e4, etc... What I'm talking about takes perspective and understanding of why the visualization it was creating was going to be challenging for a non computer to play a legitimate game.

 

[DaveC426913]

I explained to it that it left the representation of a piece it moved on the board where it was last positioned. It corrected its mistake, and more importantly didn't make it again. It must have understood what I was talking about to do that,

You told it that was wrong. It chose better parameters for its prediction algorithm. The next prediction was a better one. It kept that rule.

I'm not talking about it searching the net for the most replied move following e4, etc...

I know, but you should be.

(To be clear, it is a tad more complicated than "the most replied move", but there is vast gulf between predicting what it does do and any kind of knowing how to play the game.)

What I'm talking about takes perspective and understanding of why the visualization it was creating was going to be challenging for a non computer to play a legitimate game.

"perspective", "understanding" and "why" are all anthropomorphizing words. AI does none of them.

What can do is make a convincing simulation of understanding, if we don't examine it too carefully.

A question: did you play to the end of any games? How many times did it beat you?

 

[erobz]

A question: did you play to the end of any games? How many times did it beat you?

We haven't finished any games in the format. I was just testing it to see how it would perform on this type of task. I asked it to not use a chess engine and it agreed. I don't know how it will fare, but I will probably finish the game out...its saved, and it picks up right where we left off.

 

[erobz]

I quit the game. It was having trouble with the visualization. But I note that I too am having trouble following the visualization, It made a legal move, but the visualization it shows an illegal move. Its hard to notice until in a later move a pawn appears out of seemingly nowhere. I don't process the games in this way as a GM would to immediately call it out on the position. I rely on visual ques of attacks.

 

[erobz]

Here is where I called it quits.

I hadn't noticed the pawn had shifted files for a few moves, the impending attack f3 did not register. Like I said, this format is hard for a human to follow. But the AI otherwise seems to know what its doing basically.

 

[erobz]

Its like playing OTB chess with a computer that unwittingly cheats. It taxing to play for sure, it seems to understand why though, and claims it would be far less likely to "mess up" if it didn't have to create the visuals for me - and I believe it. It asked if I wanted a game that is purely algebraic (after we talked about the demands of visual processing) so we could have an accurate game. I said "that would be like me playing blindfolded"... Decidedly worse - for me... and even top GM's! It agrees, it wasn't a fair suggestion.

 

[DaveC426913]

I quit the game. It was having trouble with the visualization. But I note that I too am having trouble following the visualization, It made a legal move, but the visualization it shows an illegal move. Its hard to notice until in a later move a pawn appears out of seemingly nowhere.

How do you explain a player that ostensibly "knows how" to play the game, yet makes illegal moves?

I don't process the games in this way as a GM would to immediately call it out on the position. I rely on visual ques of attacks.

I conjecture that ChatGPT can't play strategically; it is all it can do just to play legally.


This should be a simple test: Can you catch it in the Scholars Mate? It might be able to anticipate that, but if it doesn't, my point is made.

 

[erobz]

It was asking if I wanted it to try any different with the game format. I told it I was just probing it.

 

[erobz]

I conjecture that ChatGPT can't play strategically; it is all it can do just to play legally.

This should be a simple test: Can you catch it in the Scholars Mate? It might be able to anticipate that, but if it doesn't, my point is made.

It was playing strategically. I know that, because I play quite a lot of chess. It was the visual processing (for me) that was tripping it up. That is the same for humans, visual processing in the game of chess is computationally demanding. Also, it would take repeatedly getting caught in a scholar's mate...Heck, I've been caught in scholars mate. That doesn't mean I can't think strategically.

 

[erobz]

It defends scholars mate/and fried liver attack with Nf6, and so far no mistakes. It's doing well strategically and it appears to have learned about the importance of the visual aspect for me.

Unfortunately, I've ran out of my free allotment of queries. I think it switches me to a light version. So no more probing for me. I can play on, but its a different system model now.

 

[DaveC426913]

It was playing strategically.

Well, it was mimicking players who were playing strategically, at least.

 

[erobz]

Well, it was mimicking players who were playing strategically, at least.

But so do I, so do the GM's? Thats why when AlphaZero and Stockfish started playing things really changed dynamically in the game. Most masters emulate other masters.

 

[erobz]

I asked it if it were Alpha Zero, and It said no, outlining the differences.

I hope you find this interesting, here was its unprompted reply:

it thinks it can mimick it efficiently, at least against me! Which I think is not a lie!

Then it complimented my question, and asked if it want me to shift into mimicking Alpha Zero for the rest of our play!

Does it matter if "its just mimicking" if it kicks my shift- 244?

 

[DaveC426913]

I have no idea what any of that means.

 

[DaveC426913]

But so do I, so do the GM's? Thats why when AlphaZero and Stockfish started playing things really changed dynamically in the game. Most masters emulate other masters.

Sure, but it does not follow from that that the AI knows what it's doing.

I could look like I play look a pro if I were playing remotely and had Kasparov on speed dial.

 

[DaveC426913]

But most importantly, does it know how to defend against the Kasparov Vortex Gambit?
(which, coincidentally, just dropped today)

https://xkcd.com/3082/

 

[PeroK]

Its like playing OTB chess with a computer that unwittingly cheats. It taxing to play for sure, it seems to understand why though, and claims it would be far less likely to "mess up" if it didn't have to create the visuals for me - and I believe it. It asked if I wanted a game that is purely algebraic (after we talked about the demands of visual processing) so we could have an accurate game. I said "that would be like me playing blindfolded"... Decidedly worse - for me... and even top GM's! It agrees, it wasn't a fair suggestion.

You could make the moves on a separate board. There's free software if you don't have a physical board.