So it seems ChatGPT probably has about a 4000 token limit on the things it can handle. Token = sortof part of a word, but pretend Token=word if thats easier. So that means if you ask it a question involving more than 4000 tokens, or continue a conversation for more than 4000 tokens, it will struggle. Although some people have had longer conversation with it - perhaps its smart enough to summarise/compress earlier parts of the conversation into a smaller amount of tokens so that it can keep going.

So its smart but its 'working memory' (thats one way of describing it) is about 3000/4000 words.

(note this is seperate to its 'understanding of the world' which is probably static and would be the combined weights of all the connections in the trained model, its probably enormous, terabytes)

Obviously us humans have much more working memory than ChatGPT, we can recall most of our lives, in a fuzzy, summarised sort of way, and any of those recollections can be incorporated into the things we say.

SO the way to make ChatGPT more like a person is to increase the token limit. Which is probably mega difficult, but if you imagine a fantasy version of it with a million token limit it would be able to converse with you for a couple of weeks (say) before it ran out of tokens. Which means within that two week conversation it would perfectly remember everything you'd asked or told it and would be able to work that back into its responses. Or if you had it reading all your documents and emails at work it would burn through the tokens faster but it would probably manage a couple of days of shadowing you in your work and helping you decide what to write next.

So thats fantasy for now

But what probably is within reach in the next year or so is an 8000 token version that has 4000 tokens for whats happening right now, and like 4000 tokens that it keeps to itself for recording ongoing context. So it kindof has a longer term memory of 4000 tokens that it updates with a summary of what is happening ("Tom is talking to me about his work project, it involves X people called a,b,c and they are building some software to do Y, we are working on a presentation for next week, the issues in play are ..."). And it seems like that would be pretty smart, it could learn context over time (as long it can compress that context down to 4000 tokens) and use that context to inform whatever the current request is. I wonder if that would work. It would really start to feel like a buddy/assistant that could remember what you were doing and help out. (and/or maybe that would also be quite freaky)

This one is cool: https://fakeyou.com/

https://blog.elevenlabs.io/

Hey Janelle, whats wrong with Wolfy?

ugh. 

https://xpressioncamera.com

eek

https://www.theverge.com/2023/1/20/23563851/google-search-ai-chatbot-demo-chatgpt
 

As recently as December, we’d heard Google execs were worried that despite investing heavily in AI technology, moving too fast to roll it out could harm the company’s reputation. But things are changing quickly. Earlier this morning, Google announced it’s laying off more than 12,000 employees and focusing on AI as a domain of primary importance

Well I, for one, welcome our imminent AI overlords

This is quite interesting:

"Do Large Language Models learn world models or just surface statistics?"

https://thegradient.pub/othello/

Trying to answer the question of what goes on inside these things. So in this article:

• they teach a GPT AI to play Othello, just by showing it chains of moves A5 B6 F3 etc, like a language
• they analyse the AI's innards (using a second AI) and find structures which seem to represent the 8x8 Othello board. So the AI had built a model of the othello board in its internal workspace (aka 'mind') (without actually being told about the othello board, all its ever seen is A5 B6 F3 etc)
• they then mess around with that othello board model inside the AI, by switching values around, and observe that changing the model inside the AI does make it choose different next moves. This confirms that it is using its own model of the board to decide what to do

 AI house party images are freaking people out 

https://www.riffusion.com/

https://www.riffusion.com/about

  On 1/31/2023 at 6:27 PM, Summon Dot E X E said:

"gl!tchW@ve"
"A.I_V!be"
"♫crctB3nd"
"b!nrDr!ft"
"d@t@Flux"
"c0dR3d"
"s¥nth3t!k"
"αlg0r!thm"
"pr0c3ssR!tm"
"err0rRhythm"
"elektr!kM@elstrom"
"qu@ntL3ap"
"!ntell!g3nc3"
"d@t@Rhythms"
"d!g!t@lD@nce"
"m0dul@t3d"
"s¥nth3t!cS0und"
"αlg0r!thm!cAura"
"pr0c3ss0rB3ats"
"gl!tchR3verie"

Expand  

Think that's the usernames of these guys

  On 1/31/2023 at 6:27 PM, Summon Dot E X E said:

"gl!tchW@ve"

pretty cool .

fun, this :

https://chatcgt.fr/

this needs translation + context explanation for non-french people -- I'll give it a try in the next post -- but you'll probably get the idea already

  On 1/31/2023 at 8:31 PM, neurone said:

fun, this :

https://chatcgt.fr/

this needs translation + context explanation for non-french people -- I'll give it a try in the next post -- but you'll probably get the idea already

 

 

 

Expand  

I said 'Bonjour' and from my limited French vocabulary I think I got scolded and told to say more than just 'Bonjour'

  On 1/31/2023 at 9:54 PM, ilqx hermolia xpli said:

Expand  

Fantastic, I'd like a poster of this for when I'm on Zoom calls

Whats the story Hermolia, it seems quite old?

  On 1/31/2023 at 9:57 PM, zazen said:

Fantastic, I'd like a poster of this for when I'm on Zoom calls

Whats the story Hermolia, it seems quite old?

the store is this so called owning class is putting our lives under the control of metal objects and id say its about time every single one gets removed. the computers too

  On 1/31/2023 at 9:30 PM, mcbpete said:

I said 'Bonjour' and from my limited French vocabulary I think I got scolded and told to say more than just 'Bonjour'

ah .. well actually .... sorry....    wrong thread maybe,  forget about it : nothing interesting in terms of A.I. properly, it's just a link on chat-gtp at the end.

anyway here's the joke: these days there are strike in france, people objecting against new pensions laws for working people (to make it short)

"CGT" is the name of one of the labor union that -amongst others- are struggling against this law  (their real position could be arguable, but nonetheless it's a popular reference)

so, the joke with replacing "GPT" with "CGT" was the idea,  the guy beind the joke was telling : "The number of characters has been limited in the answers. Let's take back the money from the bosses (understand : capitalists governement bosses) in order to have more words."

then you could type  "give me a protest slogan" in the window above (linking to the actual chat GPT bot) and then click "print it " so you can use it out in the street during the protest.   sort of a joke.

also he provides a tip to - sort of - share a trial version of chat gtp gathered one one account, so his chat-gpt version would run free for every protesting person (turns out it' s not working at the end actually, but anyway the idea was fun )  - edit : chat gtp is free afaik so I might been messing up, nevermind

as for a end, they said that actually the IA was on strike so you could not ask the chatbot to type something because it's on strike to support protesting workers.

that's why it's called "Chat CGT - your Marxist I.A"

I just foud about this just aftr browsing that IA topic, so I stupidly post it there coz was fun to me.

  On 1/31/2023 at 10:41 AM, prdctvsm said:

Expand  

ai generated seinfeld 

https://www.twitch.tv/watchmeforever