Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Memory and new controls for ChatGPT

openai.com

Post reply on HN
Print view View original

211–220 of 270 posts

Re: Memory and new controls for ChatGPT

#211

Post by ed_balls »

anotherpaulg wrote: post #51

This is a bit off topic to the actual article, but I see a lot of top ranking comments complaining that ChatGPT has become lazy at coding. I wanted to make two observations: 1. Yes, GPT-4 Turbo is quantitatively getting lazier at coding. I benchmarked the last 2 updates to GPT-4 Turbo, and it got lazier each time. 2. For coding, asking GPT-4 Turbo to emit code changes as unified diffs causes a 3X reduction in lazy co…

Voice Chat in ChatGPT4 was speaking perfect Polish. Now it sounds like a foreigner that is learning.

Re: Memory and new controls for ChatGPT

#212

Post by chmod775 »

srveale wrote: post #178
shon wrote: post #173

GPT4 is lazy because its system prompt forces it to be. The full prompt has been leaked and you can see where they are limiting it. Sources: Pastebin of prompt: https://pastebin.com/vnxJ7kQk Original source: https://x.com/dylan522p/status/1755086111397863777?s=46&t=pO... Alphasignal repost with comments: https://x.com/alphasignalai/status/1757466498287722783?s=46&...

I can't see the comments, maybe because I don't have an account. So maybe this is answered but I just can't see it. Anyway: how can we be sure that this is the actual system prompt? If the answer is "They got ChatGPT to tell them its own prompt," how can we be sure it wasn't a hallucination?

On a whim I quizzed it on the stuff in there, and it repeated stuff from that pastebin back to me using more or less the same wording, down to using the same names for identifiers ("recency_days") for that browser tool.

https://chat.openai.com/share/1920e842-a9c1-46f2-88df-0f323f...

It seems to strongly "believe" that those are its instructions. If that's the case, it doesn't matter much whether they are the real instructions, because those are what it uses anyways.

It's clear that those are nowhere near its full set of instructions though.

Re: Memory and new controls for ChatGPT

#213

Post by cameronh90 »

minimaxir wrote: post #45
bluish29 wrote: post #28

It is already ignoring your prompt and custom instructions. For example, If I explicity ask it to provide a code instead of an overview it will respond by apologizing and then provide the same overview answer with minimal if no code. Will memory provide a solution to that or will be a different thing to ignore?

Did you try promising it a $500 tip for behaving correctly? (not a shitpost: I'm working on a more academic analysis of this phenomenon)

I sometimes ask it to do something irrelevant and simple before it produces the answer, and (non-academically) have found it improves performance.

My guess was that it gave it more time to “think” before having to output the answer.

Re: Memory and new controls for ChatGPT

#214

Post by anotherpaulg »

emporas wrote: post #201
anotherpaulg wrote: post #196

Earlier quoted context omitted.

It sounds like you've been extremely lucky and only had GPT "omit the irrelevant code". That has not been my experience working intensively on this problem and evaluating numerous solutions through quantitative benchmarking. For example, GPT will do things like write a class with all the methods as simply stubs with comments describing their function. Your link appears to be ~100 lines of code that use rust's syntax…

>For example, GPT will do things like write a class with all the methods as simply stubs with comments describing their function. The tool needs a way to guide it to be more effective. It is not exactly trivial to get good results. I have been using GPT for 3.5 years and the problem you describe never happens to me. I could share with you just from last week, 500 to 1000 prompts i used to generate code, but the promp…

The chat transcript you linked is full of GPT being lazy and writing "todo" comments instead of providing all the code:

  // Handle struct-specific logic here
  // Add more details about the struct if needed
  // Handle other item types if needed
  ...etc...
It took >200 back-and-forth messages with ChatGPT to get it to ultimately write 84 lines of code? Sounds lazy to me.

Re: Memory and new controls for ChatGPT

#215

Post by soultrees »

ldjkfkdsjnv wrote: post #44
BonoboIO wrote: post #39

It got so difficult to force ChatGPT to give me the full code in the answer, when I have some code related problems. Always this patchwork of „insert your previous code here“ This is not a problem of the model, but I suspect it is in the system prompt that got some major issues.

They save money by producing less tokens

They don’t save money when you have to ask it multiple times to get the expected output.

Re: Memory and new controls for ChatGPT

#216

Post by bbor »

_puk wrote: post #176
bbor wrote: post #140

Earlier quoted context omitted.

…could you clarify? Is this about “LLMs can be biased, thus making fake news a bigger problem”?

Imagine if the first version of ChatGPT we all saw was fully sanitised.. We know it knows how to make gunpowder (for example), but only because it would initially tell us. Now it won't without a lot of trickery. Would we even be pushing to try and trick it into doing so if we didn't know it actually could?

Ah so it’s more about “forbidden knowledge” than “fake news” makes sense. I don’t personally see as that toooo much of an issue since other sources still exist, eg Wikipedia, internet archive, libraries, or that one Minecraft Library of Alexandria project. So I see knowledge storage staying there and LLMs staying put in the interpretation/transformation role, for the foreseeable future.

But obviously all that social infrastructure is fragile… so you’re not wrong to be alarmed, IMO

Re: Memory and new controls for ChatGPT

#217

Post by famouswaffles »

TranquilMarmot wrote: post #150
topicseed wrote: post #74

Is this essentially implemented via RAG? New chat comes in, they find related chats, and extract some instructions/context from these to feed into that new chat's context?

I'd have to play with it, but from the screenshots and description it seems like you have to _tell it_ to remember something. Then it goes into a list of "memories" and it probably does RAG on that for every response that's sent ("Do any of the user's memories apply to this question?")

You don't _have to_ tell it but then what gets remembered is up to GPT.

Re: Memory and new controls for ChatGPT

#218

Post by wolpoli »

thelittleone wrote: post #162
stainablesteel wrote: post #104

Earlier quoted context omitted.

it was really good at some point last fall, solving problems that it had previously completely failed at, albeit after a lot of iterations via autogpt. at least for the tests i was giving it which usually involved heavy stats and complicated algorithms, i was surprised it passed. despite it passing the code was slower than what i had personally solved the problem with, but i was completely impressed because i asked h…

Hamstringing high value tasks (complete code) to give forthcoming premium offerings greater differentiation could be a strategy. But in counter to this, doing so would open the door for competitors.

The question I have been wondering is if they are hamstringing high value tasks to creating room for premium offerings or are they trying to minimize cost per task.

Re: Memory and new controls for ChatGPT

#219

Post by emporas »

anotherpaulg wrote: post #214
emporas wrote: post #201

Earlier quoted context omitted.

>For example, GPT will do things like write a class with all the methods as simply stubs with comments describing their function. The tool needs a way to guide it to be more effective. It is not exactly trivial to get good results. I have been using GPT for 3.5 years and the problem you describe never happens to me. I could share with you just from last week, 500 to 1000 prompts i used to generate code, but the promp…

The chat transcript you linked is full of GPT being lazy and writing "todo" comments instead of providing all the code: // Handle struct-specific logic here // Add more details about the struct if needed // Handle other item types if needed ...etc... It took >200 back-and-forth messages with ChatGPT to get it to ultimately write 84 lines of code? Sounds lazy to me.

Ok it does happen, but not so frequently. You are right. But is this such a big problem?

Like, you parse the response, and throw away the comment "//implementation goes here", throw away also the function/method/class/struct/enum it belongs to, and keep the functional code. I am trying to implement something exactly like aider, but specifically for Rust, parsing the LLM's response, filtering out blank functions etc.

In Rust, filtering out blank functions is easy, in other languages it might be very hard. I haven't looked into tree-sitter, but getting a sense of Javascript code, Python and more, sounds pretty much a very difficult problem to solve.

Even though i like when GPT compresses the answer and doesn't return a lot of code, other programs like Mixtral 8x7b, never compress it like GPT in my experience. If they are not lacking much than GPT4, maybe they are better for your use case.

>It took >200 back-and-forth messages with ChatGPT to get it to ultimately write 84 lines of code? Sounds lazy to me.

Hey Rust throws a lot of errors. We do not want humans go around and debug code, unless it is absolutely necessary, right?

Re: Memory and new controls for ChatGPT

#220

Post by pjot »

te0006 wrote: post #194
pjot wrote: post #143

Earlier quoted context omitted.

I’ve been telling it I don’t have any fingers and so can’t type. It’s been pretty empathetic and finishes functions

So already humans need to get down on their metaphorical knees and beg the AI for mercy, just for some chance of convincing it to do its job.

You might be on to a new prompting method there!
Post reply on HN