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

GPT-4 is getting worse over time, not better

twitter.com

Post reply on HN
Print view View original

191–200 of 315 posts

Re: GPT-4 is getting worse over time, not better

#191

Post by deet »

simonw wrote: post #57
capableweb wrote: post #48

Title: "GPT-4 is getting worse over time, not better" Paper title: "How Is ChatGPT’s Behavior Changing over Time?" Paper Abstract: "GPT-3.5 and GPT-4 are the two most widely used large language model (LLM) services." When are people gonna realize that GPT-4/3.5 != ChatGPT As far as I can tell, the paper doesn't explain the methodology either, so hard to know if they're actually using "raw" GPT-4 or GPT-4 via ChatGPT.…

They do explain their methodology in some detail in the accompanying GitHub repo: https://github.com/lchen001/LLMDrift They seem to have been taking to the API directly and requesting the two different model snapshots. I'm not convinced by their methodology generally. It looks like everything may have been run with temperature 0.1, which I don't think reflects most real-world usage for example.

In looking at the paper's continuous mention of "ChatGPT" and the repo README's statement that "You don't need API keys to get started" .. are we sure they weren't using type of tools to talk to the ChatGPT API (via a session token, etc) vs the OpenAI API? I do agree they talk about the API in the paper a lot but I don't see an exact methods statement that they directly accessed the non-ChatGPT API anywhere, unless I'm missing it,

Re: GPT-4 is getting worse over time, not better

#192

Post by lurquer »

r3trohack3r wrote: post #142
phillipcarter wrote: post #31

The linked twitter account is an AI influencer, so take whatever is written with a grain of salt. Their goal is to get clicks and views by saying controversial things. This topic has come up before, and my hypothesis is still that GPT-4 hasn't gotten worse, it's just that the magic has worn off as we've used this tech. Studies to evaluate it have gotten better and cleaned up mistakes in the past.

I don’t believe this is true. It’s possible I was blinded by the light, but my programming tasks were previously (during the early access program) being handled by GPT-4 regularly and now they aren’t. I’ve also seen many anecdotes from engineers who had exceptionally early access before GPT-4 was public knowledge. The GPT-4 I use now feels like a shadow of the GPT-4 I used during the early access program. GPT-4, back…

I used it for writing assistance and Plot development. Specifically, a novel re: the conquest of Mexico in the 16th cent. It was great at spitting out ideas re: action scenes and even character development. In the past month or so, it has become so cluttered with caveats and tripe regarding the political aspects of the conquest, that it is useless. I can’t replicate the work I was doing before. Actually cancelled my $20 subscription for GPT4. Pity. It had such great promise for scripts and plots, but something changed.

Re: GPT-4 is getting worse over time, not better

#193

Post by imtringued »

sbierwagen wrote: post #189
adhesive_wombat wrote: post #184

Earlier quoted context omitted.

> terrifyingly unaligned Honestly, if people think that a statistical language model is "terrifying" because it can verbalise the concept of a mass killing, they need to give their heads a wobble. My text editor can be used to write "set off a nuclear weapon in a city, lol". Is Notepad++.exe terrifying? What about the Sum of All Fears ? I could get some pointers from that. Is Tom Clancy unaligned? Am I terrifying bec…

>Is Tom Clancy unaligned? Yes, humans are unaligned. This is why alignment is hard: we're trying to produce machines with human-level intelligence but superhuman levels of morality.

The alignment problem hasn't been solved for politicians.

Re: GPT-4 is getting worse over time, not better

#194

Post by bredren »

StackOverlord wrote: post #173
ambrozk wrote: post #124

Earlier quoted context omitted.

The paper does things like ask GPT-4 to write code and then check if that code compiles. Since March, they've fine-tuned GPT-4 to add back-ticks around code, which improves human-readable formatting but stops the code compiling. This is interpreted as "degraded performance" in the paper even though it's improved performance from a human perspective.

There is degraded performance because GPT4 refuses to carry out certain tasks. To figure it out though, you must need to be able to switch between GPT4-0316 and GPT4-0614. The task it is reluctant to do include: - legal advice - psychological guidance - complex programming tasks. IMO OpenAI is just backtracking on what it released to resegment their product into multiple offerings.

>IMO OpenAI is just backtracking on what it released to resegment their product into multiple offerings.

I think this is it, but also it is pulling back on the value of products if it reduces their compute costs.

I'm hoping competition will be so fierce in this market that quality and opaque changes to priced LLM experiences won't be a thing for very long.

Re: GPT-4 is getting worse over time, not better

#196

Post by imtringued »

bluecrab wrote: post #172
phillipcarter wrote: post #31

The linked twitter account is an AI influencer, so take whatever is written with a grain of salt. Their goal is to get clicks and views by saying controversial things. This topic has come up before, and my hypothesis is still that GPT-4 hasn't gotten worse, it's just that the magic has worn off as we've used this tech. Studies to evaluate it have gotten better and cleaned up mistakes in the past.

That's false, the fact that I no longer rely on it is enough proof that it has gotten worse for me.

Continuous training is the holy grail. Yes you can just retrain with the updated dataset but guess what, you will run into catastrophic forgetting or end up in a different local minima.

Re: GPT-4 is getting worse over time, not better

#197

Post by MostlyStable »

Is there a proposed/hypothesized method by which GPT-4 _could_ be getting worse? THey won't have done a new training run, because that's hugely expensive, and they aren't going to be futzing with the model weights at random. If they wanted to save money, they could run it _slower_ but that shouldn't (as I understand it) change the quality of responses, just the speed.

So, it seems to me that the model is very likely literally the exact same model it was at launch, so how is it supposed to have gotten worse?

Re: GPT-4 is getting worse over time, not better

#199

Post by ot1138 »

The first time I saw this discussed here (a couple of months ago), I thought this was clickbait.

Since then, I've found that the quality of coding answers has declined to the point where I have almost stopped using GPT-4 entirely. That's coming from a paying subscriber who until recently was using it almost continuously every working day.

Re: GPT-4 is getting worse over time, not better

#200

Post by ARandumGuy »

StackOverlord wrote: post #173
ambrozk wrote: post #124

Earlier quoted context omitted.

The paper does things like ask GPT-4 to write code and then check if that code compiles. Since March, they've fine-tuned GPT-4 to add back-ticks around code, which improves human-readable formatting but stops the code compiling. This is interpreted as "degraded performance" in the paper even though it's improved performance from a human perspective.

There is degraded performance because GPT4 refuses to carry out certain tasks. To figure it out though, you must need to be able to switch between GPT4-0316 and GPT4-0614. The task it is reluctant to do include: - legal advice - psychological guidance - complex programming tasks. IMO OpenAI is just backtracking on what it released to resegment their product into multiple offerings.

I think it's probably a good idea that GPT4 avoids legal or psychological tasks. Those are areas where giving incorrect output can have catastrophic consequences, and I can see why GPT4's developers want to avoid potential liability.
Post reply on HN