Live data from Hacker News

Ask HN: Is it just me or GPT-4's quality has significantly deteriorated lately?

news.ycombinator.com

241–250 of 817 posts

Re: Ask HN: Is it just me or GPT-4's quality has significantly deteriorated lately?

#241
post #154

The researchers who worked on the "sparks of AGI" paper noted that the more OpenAI worked on aligning GPT-4 the less competent it became. I'm guessing that trend is continuing...

I don't think it's just the alignment work. I suspect OpenAI+Microsoft are over-doing the Reinforcement Learning from Human Feedback with LoRA. Most of people's prompts are stupid stuff. So it becomes stupider. LoRA is one of Microsoft's most dear discoveries in the field, so they are likely tempted to over-use it.

Perhaps OpenAI should get back to a good older snapshot and be more careful about what they feed into the daily/weekly LoRA fine-tuning.

But this is all guesswork because they don't reveal much.

Re: Ask HN: Is it just me or GPT-4's quality has significantly deteriorated lately?

#242
post #176

Earlier quoted context omitted.

They can only do this in the USA. The battle bottleneck is going to be GPUs and countries that have access to it will leap ahead of the US if arbitrary regulations are put in place to neuter LLMs in the US.

You can lobby in every country, what do you mean

And where lobbying is illegal, only criminals can do it!

Re: Ask HN: Is it just me or GPT-4's quality has significantly deteriorated lately?

#243
post #85

I tried to replicate a few of my chats (the displayed date is incorrect, it seems to be the publish date instead of the original chat date): svg editor: early april: https://chat.openai.com/share/c235b48e-5a0e-4a89-af1c-0a3e7c... now: https://chat.openai.com/share/e4362a56-4bc7-45dc-8d1b-5e3842... originally it correctly inferred that I wanted a framework for svg editors, the latest version assumes I want a js framew…

this is GPT 3.5, the icon is green

Re: Ask HN: Is it just me or GPT-4's quality has significantly deteriorated lately?

#244
My guess is that -probably no. It's more likely you had a stream of good luck in your earlier interactions and now you're observing regression to the mean.

That can easily happen and it's why, for example, medical studies, are not taken as definitive proof of an effect.

To further clarify, regression to the mean is the inevitable consequence of statistical error. Suppose (classic example) we want to test a hypertension drug. We start by taking the blood pressure (BP) of test subjects. Then we give them the drug (in a double-blind, randomised fashion). Then we take their blood pressure again. Finally, we compare the BP readings before and after taking the drug.

The result is usually that some of the subjects' BP has decreased after taking the drug, some subjects' BP has increased and some has stayed the same. At this point we don't really know for sure what's going on. BP can vary a lot in the same person, depending on all sorts of factors typically not recorded in studies. There is always the chance that the single measurement of BP that we took off a person before giving the drug was an outlier for that patient, and that the second measurement, that we took after giving the drug, is not showing the effect of the drug but simply measuring the average BP of the person, which has remained unaffected by the drug. Or, of course, the second measurement might be the outlier.

This is a bitch of a problem and not easily resolved. The usual way out is to wait for confirmation of experimental results from more studies. Which is what you're doing here basically, I guess (so, good instinct!). Unfortunately, most studies have more or less varying methodologies and that introduces even more possibility for confusion.

Anyway, I really think you're noticing regression to the mean.

Re: Ask HN: Is it just me or GPT-4's quality has significantly deteriorated lately?

#245
post #14

Yes. Before the update, when its avatar was still black, it solved pretty complex coding problems effortlessly and gave very nuanced, thoughtful answers to non-programming questions. Now it struggles with just changing two lines in a 10-line block of CSS and printing this modified 10-line block again. Some lines are missing, others are completely different for no reason. I'm sure scaling the model is hard, but they l…

If this is true, one should be able to compare with benchmarks or evals to demonstrate this. Anyone know more about this?

Yeah I think it's plausible it's gotten worse but it would also be classic human psychology to perceive degradation because you start noticing flaws after the honeymoon effect wore off.

Unfortunately this will be hard to benchmark unless someone was already collecting a lot of data on ChatGPT responses for other purposes. Perhaps if this is happening the degradation will get worse though, so someone noticing it now could start collecting GPT responses longitudinally.

Re: Ask HN: Is it just me or GPT-4's quality has significantly deteriorated lately?

#246

Earlier quoted context omitted.

I assume you're talking about ChatGPT and not GPT-4? You can craft your own prompt when calling GPT4 over API. Don't blame you though, the OP is also not clear if they are comparing Chat GPT powered by GPT3.5 or 4, or the models themselves.

When using it all day every day it seems (anecdotally) the API version has changed too. I work with temperature 0 which should have low variability yet recently it shifted to feel boring, wooden, and deflective.

I can understand why they might make changes to ChatGPT, but it seems weird they would "nerf" the API. What would be the incentive for OpenAI to do that?

Re: Ask HN: Is it just me or GPT-4's quality has significantly deteriorated lately?

#247
post #200

Earlier quoted context omitted.

I would immediately jump to an AI not being "aligned" by SF techies (or anyone else).

Same here. If I have a choice between honesty and political correctness, I always pick honesty.

What makes you think the "unaligned" version necessarily has more honesty? Rather than just being generally easier to prompt to say whatever the user wants it to say, true or not, horrible or not? Or even easier to unintentionally make it confabulate/hallucinate stuff? Does not seem to follow, and does not seem to be a true dichotomy. Edginess does not equal honesty.

Re: Ask HN: Is it just me or GPT-4's quality has significantly deteriorated lately?

#248

Chat GPT 4 has ongoing training, such as using Reinforcement Learning from Human Feedback (RLHF) to tune it to provide "better" responses, "safer" answers, and to generally obey the system prompts. There's a release every few weeks. Yes, I've noticed too that recently it has become very "cagey", qualifying everything to death with "As an AI model...". A paper[1] that took snapshots monthly mentioned that as the initi…

I had a look at the YouTube video -- I feel that an obvious question with regards to the "common sense" tests is, what was chat GPT-4 trained on? Was it partly trained on reams of questions used to test AI systems for example? How do you know it is "demonstrating" anything out-of-sample, especially if it is constantly being improved?

I've been learning some exotic programming languages recently, and my anecodotal experience is that asking ChatGPT to code in array programming or logic languages results in code which is highly non-idiomatic for those paradigms. Why is that? It mostly writes the code as if it was all just a funny syntax for Javascript or Python. I'm surprised at that if it really understood J or APL for example.

I am presuming that behind the scenes there are demonstrations of capabilities much greater than GPT-4 which are being used to illustrate the dangers of AI, because whilst I'm massively impressed by what's happening it is difficult to convince myself of a "qualitative" difference.

Post reply on HN