this post was submitted on 11 Aug 2023
24 points (100.0% liked)
Technology
37360 readers
176 users here now
Rumors, happenings, and innovations in the technology sphere. If it's technological news or discussion of technology, it probably belongs here.
Subcommunities on Beehaw:
This community's icon was made by Aaron Schneider, under the CC-BY-NC-SA 4.0 license.
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
I've had a nagging issue with ChatGPT that hasn't been easy for me to explain. I think I've got it now.
We're used to computers being great at remembering "state." For example, if I say "let x=3", barring a bug, x is damned well gonna stay 3 until I decide otherwise.
GPT has trouble remembering state. Here's an analogy:
When I work with GPT, I spend a lot of time reminding it that Fred was a dinosaur.
Do you have any theories as to why this is the case? I haven't gone anywhere near it, so I have no idea. I imagine it's tied up with the way it processes things from a language-first perspective, which I gather is why it's bad at math. I really don't understand enough to wrap my head around why we can't seem to combine LLM and traditional computational logic.
ChatGPT works off of a fixed size possible maximum prompt. This was originally about 4000 tokens. A token is about 4 characters or one short word, but its not quite that... https://platform.openai.com/tokenizer
"Tell me a story about a wizard" is 7 tokens. And so ChatGPT generates some text to tell you a story. That story is say, 1000 tokens long. You then ask it to "Tell me more of the story, and make sure you include a dinosaur." (15 tokens). And you get another 1000 tokens. And repeat this twice more. At this point, the length of the entire chat history is about 4000 tokens.
ChatGPT then internally asks itself to summarize the entire 4000 token history into 500 tokens. Some of the simpler models can do this quite quickly - though they are imperfect. The thing is at point you've got 500 tokens which are a summarization of the 4 acts and of the story and the prompts that were used to generate it - but that's lossy.
As you continue the conversation more and more, the summarizations become more and more lossy and the chat session will "forget" things.
From my understanding, ChatGPT doesn't do anything like that by itself. If you want the story summarized, you'll have to request it and it will show up in the text buffer. There is no hidden internal state that ChatGPT can use to "think", there is just the text that you see in the text buffer.
The only hidden text that exists is the initial prompt that turns GPT into a chatbot, along with some start/stop tokens, that give control back to the user (plain GPT will just auto-complete both sides of the conversation).
Some experiments like AutoGPT do generate summaries and outlines for larger problems from what I understand. But ChatGPT is so far just a chatbot layer on top of GPT, without any extra cleverness.