Skip to main content

Browse GamesGame JamsUpload GameDeveloper LogsCommunity
Indie game storeFree gamesFun gamesHorror games
Game developmentAssetsComics
SalesBundles
Jobs
TagsGame Engines
redonihunter3 days ago
Just from my own perspective, the term "training" implies something I have not yet witnessed. The weight calibrations feel like exactly that, calibrations.

Calibration is a feedback loop of fitting your model to the data. In statistics, that's a bad thing. For ai, that is how they are trained. The end result is the calibration values.

If you calibrate your model on weather patterns or if you calibrate it on how words appear in texts, those will be different values inside your hard coded model, that now can do this hyper dimensional math things to predit what comes next. Like an answer to a question. Or if it rains tomorrow.

But anyways, it appears that nothing ever actually touches the network itself.
There is no memory. There's nothing to remember.

Yes and no.

The model itself is not changed by using it. It really acts like a black box transformator where you put in something and get out a response specific to that something.

And it does have a memory for your session, but actually it does not have a memory. Just as you said, the previous text is just fed in again.

I recently saw a good explanation for this memory thing.

Reply
Durance Gaming2 days ago

Ah, I see the disconnect and I think it's my own fault. Everything you are saying is totally true, but I think we are talking about two different parts of the same overall thing. You said "the model". I know exactly what you mean with your weather example and it's kind of a big deal when you think about it. I always wondered how scientists did "studies" and "research" with AI with all it's current limitations but at the time, I was still stuck on thinking AI = Chat Bot or Image Generator or what-have-you. The input vectors can represent anything and whether or not the system can produce anything useful with the data--that's all down to the training. But this isn't what I'm trying to point at.

Not the model. The model is a combination of multiple processes. There is a surrounding framework involved.

I'm talking about the system at the core of it all. Not just ChatGPT or whatever, but any feed-forward neural network at all. 

The very concept of neural network technology appears to be an invariable expression of the idea of "navigation". As a matter of fact, there's nothing stopping anyone from building a 4x4 input neural network specifically for navigating the probabilities of a 4D environment. Once you cross that 4th dimension, you are plotting points in higher-dimensional space. That, as far as I can tell, is the function of a feed-forward of a neural network. I hate to reduce such a neat thing to such a description but... It's like a fancy graphing calculator. You could do the math by hand yourself, but you wouldn't want to lol.

I hope that makes sense.

"And it does have a memory for your session, but actually it does not have a memory. Just as you said, the previous text is just fed in again."

This part, I gotta push back on--respectfully--but hear me out: That's not memory. That's repetition. Which why it only works in the same session. Like I said, it isn't a conversation it's co-creating a composition but you go once and ChatGPT goes ~34,000 times per word of it's "replies". It isn't talking to you or me, it's talking to the conversation as a whole. New session, different co-composed conversation.

I noticed this before I studied neural networks, when I got real into setting up my own models on my PC.

The more "memory" my models "remembered", the worse they performed. Tracking the source, sure enough, there was no real "memory" but instead just brute-force reapplication of data over and over and over, chewing up token limits and taking up context space.

In my opinion, "memory" is one of the worst things you could possibly give a modern AI in any business environment. Any access to irrelevant context, in any way, opens the door to compounding instability. The current way of faking "memory" is probably what leads to a lot of cybersecurity woes.

Reply
Durance Gaming2 days ago

Forgive me for the double reply, I read your message first then replied and am now watching the video you posted.

So far, immediately, there's a problem and maybe it's half-mine. This explanation is not what a "token" is. At least not in the system architecture sense. But now I am connecting the dots... Ok, so they call the limit a "token" limit because they charge you per token (navigation point) reached at the end of each full pass. Ok... Still watching...

Yep. A token isn't always a full word. Right. Still viewing...

Yeeeaaahhh dude! They were DECIDED and realizing that blew my mind. From my perspective, it's like writing a map of a place you've never been and then navigating probability until the landscape matches the map you drew... It's wild!

Boom--right there! Did you catch that at 4:04 to 4:10? "It doesn't care, they're just input tokens" That right there, the "it" he's talking about is the neural network. It doesn't care because it's a high-dimensional probability drive. The output is repurposed through the surrounding framework into "intelligence" and "behavior". This is what I am pointing to, the actual system itself before any arbitrary tokens are involved. The tokens are part of the framework, not the system itself... Still watching...

So, I don't want anyone to take this the wrong way... I'm not saying it's a conspiracy or anything... But the very next thing the guy does if you listen to his words and also pay attention to his body language: "But that's not even the interesting part" and he waves it away with a physical gesture. Before I taught myself linear algebra, I came from a 30+ year background of psychology and philosophy (b.s. in humanities) and that dude got uncomfortable when he realized where he was about to have to go to explain "why" the tokens don't matter. I don't think it's any kind of conspiracy, I think it might be a sense of embarrassment like the Tech world doesn't want to admit how far away from AGI modern AI architecture actually is.

I did watch the rest of the video but the rest of it revolves around token pricing... Okay... I see what's going on now. There is no real measurable way to price the usage of something like a modern AI because it's literally just automated calculations. So I can see the logic behind charging by the token.

Something isn't adding up, man... You don't need all these data centers and all this stuff to achieve something like ChatGPT. Even if everyone in the world was using it, yeah you'd need to scoop up all the RAM and processing power you could but you don't need all this storage space unless you're saving every single output of the neural network thought process and every single conversation.

I dunno... Anyways, that was a good video. I get it now, the token pricing versus actual tokens. Makes sense.

Reply
redonihunter2 days ago(+1)

In context the video was just a good explanation for the memory concept. Or rather the lack of memory. A session grows rapidly in token usage, because it has to be fed in again and again. The model is not fine tuned to you by usage, it "forgets".

It also explains, why excessive use of such a system needs so much computing hardware and electricity. And doing a little chatting with chatgpt does not do this. But trying to replace a software developer with ai will do that.

You seem to fuss about the word token or the definition. It's just the bits or atoms of language. A chatbot has to operate on those pieces, so measuring the usage of the bot is easiest done by counting how many pieces a session used.

Reply
Durance Gaming10 minutes ago

"You seem to fuss about the word token or the definition."

I do, forgive me--but there is a reason for that. It's because I had to comprehend them in order to build my own AI (not a chatbot, something completely different). A token is actually nothing on it's own. A single variable holding an array of strings--that's a token dictionary. The final layer gets softMaxed and produces an index list of probabilities that matches the same amount of indices in your array. Then there's that game of "Battleship" I spoke of. If index 304 sits at 0.989 then the "word" in index 304 of the word dictionary array is chosen by code logic--not the neural network itself.

There is no token use in the beginning. That's where the disconnect is. Tokens are for after the system is done with it's calculations, not before. There is an easy way to both prove and demonstrate this:

Recently, a man was able to get an AI to send him money from that AI's company. (he got caught) He did this by communicating in Morse Code to the chat agent. There are no Morse Code tokens. However, the AI was still able to understand the input. Why it worked as an adversarial prompt, that is also obvious when one stops considering the framework around the system and looks directly at how feed-forward neural networks do what they do.

So that's why I get hung up on it. The way the tech industry uses the word "tokens" to track monetization, it muddies the waters on what a token actually is and what function it performs in the overall architecture.

Check it out--you don't even need tokens at all. This is a short demo video for a completely different kind of AI that I am developing for video game NPCs (a multi-modal dual-input feed-forward/simplicial neural network architecture): 

Reply
AboutFAQBlogContact us
Copyright © 2026 itch corp · Directory · Terms · Privacy · Cookies