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

947 points by behnamoh ↗ HN
It is much faster than before but the quality of its responses is more like a GPT-3.5++. It generates more buggy code, the answers have less depth and analysis to them, and overall it feels much worse than before.

For a while, the GPT-4 on phind.com gave even better results than GPT-4-powered ChatGPT. I could notice the difference in speed of both GPT-4s. Phind's was slower and more accurate. I say "was" because apparently phind is now trying to use GPT-3.5 and their own Phind model more frequently, so much for GPT-4 powered search engine....

I wonder if I use Poe's GPT-4, maybe I'll get the good old GPT-4 back?

812 comments

[ 2.7 ms ] story [ 381 ms ] thread
(comment deleted)
Yes!

It didn't even try on my question of Jarvis standings desks, which is a fairly old product that hasn't changed up.. Their typical "My knowledge cutoff..." response doesn't even make sense. It screwed up another question I asked it about server uptime and four-9s, Bard got it right.

I've moved back to Bard for the time being...It's way faster as well. And GPT-4's knowledge cutoff thing is getting old fast.

Example prompt: how much does a 48 x 30 inch jarvis bamboo standing desk setup weigh?

I still have the spec sheet from my order back in 2020. So I can verify the weights.

### GPT-4 ###

ChatGPT As of my knowledge cutoff in September 2021, I don't have the exact weight of a 48 x 30 inch Jarvis Bamboo standing desk setup. However, on average, standing desks can range anywhere from around 60 to 100 pounds or more, depending on the specific configuration (like the choice of frame, desktop material, and any additional features).

Jarvis Bamboo standing desks are made of a bamboo top and an electric lift frame. The weight of the bamboo top for that size should be around 30-40 pounds. The frame itself can weigh anywhere from 50-70 pounds.

For a more accurate weight, I would recommend checking the manufacturer's specifications on their website or contacting their customer support directly for this information.

### Bard ###

The fully assembled Jarvis Bamboo Standing Desk weighs 92 pounds. The desktop itself weighs 38 pounds, and the frame weighs 54 pounds. The desk can hold a maximum weight of 350 pounds.

Here are some of the features of the Jarvis Bamboo Standing Desk:

Electric height adjustment from 25.5 to 51 inches Memory presets for up to 4 different heights 3-stage gas spring mechanism Bamboo desktop with a scratch-resistant finish Black powder-coated steel frame 10-year warranty on the frame and 5-year warranty on the desktop

Asking facts from a generative AI is folly.
Yes, people really need to know that unless you are using the browser plugin, you really shouldn't ask it questions like this. (A good rule of thumb I think is if you can't expect a random person on the street to get the question right without looking it up, you shouldn't expect GPT-4 to get it right either.)

Unfortunately for this question, even using the browser plugin it wasn't able to get the answer: https://chat.openai.com/share/6344f09e-4ba0-45c7-b455-7be59d...

Even worse is all of the "GPT Influencers" and their "Here's what ChatGPT predicts will be the price of Bitcoin/some stock/houses will be this time next year" clickbait. It's a language model, people.
Please tell me that you're making this up about there being GPT "influencers".
Have you been to Twitter in the last 6 months? It's basically a GPT/Midjourney/Stable Diffusion hype generator. Everybody there is now an expert on this topic. And you can be too!

Step 1: Tweet fake troll screenshots of GPT output or make corny threads like "90% of people are using AI wrong, here's..."

Step 2: Let the "For you" algo take hold

Step 3: Profit

If you bought NVDA stock at ChatGPT launch you would be super rich today. Not what the GPT influencers would tell you, though.

  > The fully assembled Jarvis Bamboo Standing Desk weighs 92 pounds. The desktop itself weighs 38 pounds, and the frame weighs 54 pounds. The desk can hold a maximum weight of 350 pounds.
That sounds like a linguistically valid sentences, exactly what I would expect from a novel LLM. Did you check that it is also factually correct? Factually correctness is _not_ the goal of a typical LLM.
It’s possible they are trying out a shaved/turbo version so that they can start removing the limits. I mean as it is - 25 messages every 3 hours is useless, particularly for browsing and plugins.
That would be a shame. I would rather be limited to 25 quality responses per 3 hours than an unlimited inferior GPT4.

Imagine trading the advice of a senior mentor for 5 intermediate mentors. Yes, the answers get to you faster, but it's much less useful.

Yes. Seems to have definitely gone down. Not sure what they have done but even with things it used have no trouble with, it struggles now. Most likely they are experimenting on reducing the compute per request.
I did notice the rate limiting message isn't there anymore when using GPT-4.
I still see the cap message, I'm not sure if it is being enforced or not
Yes, definitely faster now but crappier.
Yes! I kept wondering why but allisdust's compute reason makes the most sense.
I find that the GPT-4 in the playground is as good as before - but have noticed the issues when using the chat gpt client
Definitely nerfed. Concomitantly, the performance increased substantially, and it now feels much, much quicker (maybe 10x even?), but the quality has decreased quite a bit.
Is it consistently worse or just sometimes/often worse than before? Any extreme power users or GPT-whisperers here? If it’s only noticeably worse X% of the time my bet would be experimentation.

One of my least favorite patterns that tech companies do is use “Experimentation” overzealously or prematurely. Mainly, my problem is they’re not transparent about it, and it creates an inconsistent product experience that just confuses you - why did this one Zillow listing have this UI order but the similar one I clicked seconds later had a different one? Why did this page load on Reddit get some weirdass font? Because it’s an experiment the bar to launch is low and you’re not gonna find any official blog posts about the changes until it’s official. And when it causes serious problems, there’s nowhere to submit a form or tell you why, and only very rarely would support, others, or documentation even realize some change is from an experiment. Over the past few years I’ve started noticing this everywhere online.

Non-sticky UI experiments are especially bad because at eg 1% of pageloads the signal is going to be measuring users asking themselves wtf is up and temporarily spending more time on page trying to figure out where the data moved. Sticky and/or less noticeable experiments like what this could be have stronger signals but are even more annoying as a user, because there’s no notice that you’re essentially running some jank beta version, and no way to opt back into the default - for you it’s just broken. Especially not cool if you’re a paying customer.

I’m not saying it’s necessarily an experiment, it could be just a regular release or nothing at all. I’d hope if OpenAI was actually reducing the parameter size of their models they’d publicly announce that, but I could totally see them running an experiment measuring how a cheaper, smaller model affects usage and retention without publishing anything, because it’s exactly the kind of “right hand doesn’t know what the left is doing” thing that happens at fancy schmancy tech companies.

No place I worked at ever experimented at the pageload level. We experimented at the user level, so 1% of users would get the new UI. I suppose this is only possible at the millions of users scale which all of them had.
I updated the comment to reflect that. Certainly the signal is stronger because you’re amortizing away the surprise factor of the change, and at least it’s a consistent UX, but the UX tradeoff in the worst case is that experiment-group users get a broken product with no notice or escape hatch. Unless you’re being very careful, meticulous, and transparent it’s just not acceptable if you’re a paying customer.
In some cases you’re making the change because the app is already broken for the majority of users and you’re testing the fix
It is not worse for me. I do notice the novelty has worn off. Asking chatGPT4 about why people would think this on here I think it nails it with the novelty effect lol:

"Indeed, the performance of an AI model like ChatGPT doesn't deteriorate over time. However, human perception of its performance can change due to a variety of psychological factors:

Expectation Bias: As users become more familiar with AI capabilities, their expectations may increase over time. When AI doesn't meet these heightened expectations, users might perceive this as a decline in performance.

Novelty Effect: At the beginning, the novelty of interacting with an AI could lead to positive experiences. However, as the novelty wears off, users may start to focus more on the limitations, creating a perception of decreased performance."

Without this thread I would have said it got stronger with the May 12th update. I don't think that is really true though. There is this random aspect of streaks in asking questions it is good at answering vs streaks of asking questions it is less good at answering.

> Indeed, the performance of an AI model like ChatGPT doesn't deteriorate over time.

Of course, the performance of an unchanged model does not. But finetuning the model over time can of course either improve or degrade performance.

Yeah there are people ITT claiming that even the API model marked as 3/14 release version is different than it used to be. I guess that's not entirely outside the realm of possibility (if OpenAI is just lying), but I think it's way more likely this thread is mostly evidence of the honeymoon effect wearing off.

The specific complaints have been well-established weaknesses of GPT for awhile now too: hallucinating APIs, giving vague/"both sides" non-answers to half the questions you ask, etc. Obviously it's a great technical achievement but people seemed to really overreact initially. Now that they're coming back to Earth, cue the conspiracy theories about OpenAI.

Could be. But it could also be that those people (myself included) are right.

It's not that this is without precedent - there's a paper and a YouTube video with Microsoft person saying on record that GPT-4 started to get less capable with every release, ever since OpenAI switched focus to "safety" fine-tuning, and MS actually benchmarked it by applying the same test (unicorn drawing in tikz), and that was even before public release.

Myself, sure, it may be novelty effect, or Baader–Meinhof phenomenon - but in the days before this thread, I observed that:

- Bing Chat (which I haven't used until ~week ago; before, I used GPT-4 API access) has been giving surface-level and lazy answers -- I blamed, and still mostly blame it on search capability, as I noticed GPT-4 (API) through TypingMind also gets dumber if you enable web search (which, in the background, adds some substantial amount of instructions to the system prompt) -- however,

- GPT-4 via Azure (at work) and via OpenAI API (personal) both started to get lazy on me; before about 2-3 weeks ago, they would happily print and reprint large blocks of code for me; in the last week or two, both models started putting placeholder comments; this I noticed, because I use the same system prompt for coding tasks, and the first time the model ignored my instructions to provide a complete solution, opting to add placeholder comments instead, was quite... startling.

- In those same 2-3 weeks, I've noticed GPT-4 via Azure being more prone to give high-level overview answers and telling me to ask for more help if I need it (I don't know if this affected GPT-4 API via OpenAI; it's harder to notice with the type of queries I do for personal use);

All in all, I've noticed that over past 2-3 weeks, I was having to do much more hand-holding and back-and-forth with GPT-4 than before. Yes, it's another anecdote, might be novelty or Baader–Meinhof, but with so many similar reports and known precedents, maybe there is something to it.

Fair enough, I think it's realistic that an actual change is part of the effect with the ChatGPT interface, because it has gotten so much attention from the general public. Azure probably fits that somewhat as well. I just don't really see why they would nerf the API and especially why they would lie about the 3/14 model being available for query when secretly it's changing behind the scenes.

FWIW I was pretty convinced this happened with Dall-E 2 for a little while, and again maybe it did to some extent (they at least decreased the number of images so the odds of a good one appearing decreased). But also when I looked back at some of the earlier images I linked for people on request threads I found there were more duds than I remembered. The good ones were just so mind blowing at first that it was easy to ignore bad responses (plus it was free then).

These are my thoughts too. As I’ve used it more I’ve begun to scrutinize it more and I have a larger and larger history of when it doesn’t work like magic. Although it works like magic often as well.

We’ve also had time to find its limits and verify of falsify early assumptions, which were very likely positive.

The hype cycle is real.

+1, was doing active dev against it now and saw really stupid responses. For example "transform this unstructured text to valid JSON with no keys with empty strings", would return the keys w/ empty strings for some results.
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 lobotomized it in the process.

The original GPT-4 felt like magic to me, I had this sense of awe while interacting with it. Now it is just a dumb stochastic parrot.

It’s go-to tactic now if I ask it to go over any piece of code is to give a generic overview. Earlier, it would section out the code into chunks and go through each one individually.
Yeah, the bing integration did not go well. Went from amazing to annoying.
Aren’t the original weights around somewhere?
Try out Bard, it's coding is much improved in the last 2 weeks. I've unfortunately switched over for the time being.
“Bard isn’t currently supported in your country. Stay tuned!”
Same for me, I’m in Estonia :(
You can use a VPN to use an American connection, it doesn't matter where your Google account is registered.
Thanks, I’ll try it! (I’m in Hungary)
Not necessarily American, you just have to avoid EU and, I believe, Russia/China/Cuba etc.
I'm in Switzerland and Bard is locked out, we do not go by EU laws because we are not part of the EU. We have plenty of bilateral deals but still.
But don't you sill have privacy laws very similar to the GDPR?
Google's passion for region locking is insane to me
Its a legal thing, not something they want to do
What law prohibits Google from making Bard available outside the USA?
It's blocked in the EU because they don't want to/can't comply with GDPR.
It is not GDPR, it is available in some countries outside the EU with GDPR-like privacy regimes.
This is naïve though. Regulation — especially such as this — has to be enforced and there is obviously room to over and under interpret the text of the law on a whim, or varying fines. OAI knows this and looking at the EU lately, what they’re doing is wise.
Which is interesting, because if they can't comply within the EU, then how do they comply outside of the EU. With that I mean, if they have concerns that there is private data of EU citizens somewhere in that, then that is also in there for users outside of the EU. That said, they do not comply with GDPR anyway. If that its not the case, then they could also enable it for users within the EU.
Simple: GDPR (or any EU law) is not enforceable outside EU
Some nuance:

If Google gobble up data about EU citizens then they fall under GDPR.

It doesn't matter that they don't allow EU citizens to use the result.

If our personal data is in there and they are don't protect it properly they are violating EU law. And protecting it properly means from everyone, not just EU citizens.

The gobbling happens in realtime as you use it
(comment deleted)
Actually, in case of Google it is, because they still do business within the EU.
That’s a common belief, but it’s wrong. In principle an EU court could decide to apply the GDPR to conduct outside the EU; and in the right circumstances, a non-EU court might rule that the GDPR applies.

Choice of law is anything but simple. Think of geographic scoping of laws as a rough rule of thumb sovereign states use to avoid annoying each other, rather than as a law of nature.

GDPR is likely not enforceable if you have no presence in EU whatsoever, if you have no assets in EU and no money coming in from EU.

Anything Google does with data of EU residents is subject to GDPR even if that particular service is not offered within EU, and it is definitely enforceable because Google has a presence in EU, which can be (and has been) subjected to fines, seizures of assets, etc.

It's a risk mitigation strategy, these things are not black and white.

Making it unavailable in the EU decreases the likelihood and severity of a potential fine.

They clearly can with all their other products, as can OpenAI since they've been unblocked. They're just being assholes because they can.
What's the excuse for Canada being omitted
We're small and no one cares about us...
It's available here in the UK, so it's not USA exclusive.
I was just on a cruise around the UK and I couldn't access Bard from the ship's wi-fi. That surprised me for some reason. Should've checked where it thought I was ...
Eh, more like limiting rollout because they can't/don't want to handle the scale.
The Bard model (Bison) is available without region lock as part of Google Cloud Platform. In addition to being able to call it via an API, they have a similar developer UI to the OpenAI playground to interactively experiment with it.

https://console.cloud.google.com/vertex-ai/generative/langua...

Is that free or do you have to pay?

Also do you need to change the options like Token Limit etc?

It's completely free. No tokens nothing.
But it can't be used unless I enable billing, which I am not willing to do after reading all the horror stories about people getting billed thousands overnight. I'm not willing to take the risk that I forget some script and it keeps creating charges.
Use a CC or debit that can limit charges. Privacy.com is a generic one. There’s others. Also Capital One, Bank of America, Apple Card and maybe some others have some semblance of control over temporary CCs.

Ideally one would want to be able to have a cap on the amount that can be spent in a given period.

Thanks for this! I had a temporary Cap One card on my cloud accounts. I’m going to switch them to Privacy.com ones to limit amount if I can’t find another solution.

Did they release the Codey or Unicorn models publicly yet? Or say when they might do that?
it's also really, really bad and fails compared to even open source models right now.
God, what happened to Google. What a fall from grace.

Alpaca is pretty good though.

They have 100,000 employees pretending to work on the past.

They have no leadership at the top. Nobody that can steer the ship to the next land (or even anybody that has a map). Who is actively working at Alphabet that has the authority to kill Google search through self-cannibalization? Absolutely nobody. They're screwed accordingly. It takes an enormous level of authority (think: Steve Jobs) and leadership to even considering intentionally putting at risk a $200 billion sales product. The trick of course is that it's already at great risk.

They don't know what to do, so they're particularly reactive. It has been that way for a long time though, it's just that Google search was never under serious threat previously, so it didn't really matter as a terminal risk if they failed (eg with their social network efforts; their social networks were reactive).

It's somewhat similar to watching Microsoft under Ballmer and how they lacked direction, didn't know what to do, and were too reactive. You can tell when a giant entity like Google is wandering aimlessly.

I just tried Bard based on this comment, and it's really, really bad.

Can you please help me with how you are prompting it?

If you have to worry about prompting, it already tells you everything one needs to know about how good the model is.
I don't think that's true at all. Think of it like setting up conversation constraints to reduce the potential pitfalls for a model. You can vastly improve the capability of just about any LLM I've used by being clear about what you specifically want considered, and what you don't want considered when solving a problem.

It'll take you much farther, by allowing you to incrementally solve your problem in smaller steps while giving the model the proper context required for each step of the problem-solving process, and limiting the things it must consider for each branch of your problem.

Don't know what you are doing? But Bard is so much faster than openai and its answers are clearer and more succint.
This is just... false. Bard is not just a little worse than gpt-4 for coding, it's more like several orders of magnitude worse. I can't imagine how you are getting superior outputs from Bard.
I’ve been seeing similar comments about Bard all over Twitter and social media.

My testing agrees with yours. Almost seems like a sponsored marketing campaign with no truth to it.

After my first day with Bard, I would have agreed with you. But since then, I've found that Bard simply has a lot of variance in answer quality. Sometimes it fails for surprisingly simple questions, or hallucinates to an even worse degree than ChatGPT, but other times it gives much better answers than ChatGPT.

On the first day, it felt like 80% of the responses were in the first (fail/hallucinate) category, but over time it feels more like a 50/50 split, which makes it worth running prompts over both ChatGPT and Bard and select the best one. I don't know if the change is because I learnt to prompt it better, or if they improved the models based on all the user chats from the public release - perhaps both.

If it needs to write a code, I usually prompt it with something like:

"write me a script in python3 that uses selenium to log into a MyBB forum"

note: usually it will not compile and you still have to do some editing

No thanks! I have better things to do than feeding that advertising behemoth. What I like about ChatGPT is that I don't see any ads at all!
> What I like about ChatGPT is that I don't see any ads at all!

For now. It's just a marketing tool/demo site, like ITA Matrix was/is. The ads are vended by Bing.

That you know of.

Don't you worry, if there is any medium, place or mode of interaction people spend time on, advertising will eventually metastasize to it, and will keep growing until it completely devalues the activity and destroys most of the utility it provides.

Google (Deepmind) actually has the people and has developed the science to make the best AI products in the world, but unfortunately Bard seems to be thrown together in an afternoon by an intern, and then handed off to a hoard of marketing people. It's not good right now. Deepmind is one of the best scientifically, they just don't really make products. OpenAI is essentially the direct opposite of that.
I asked it to review some code a couple days ago - the comments while valid english were nonsense
"The original GPT-4 felt like magic to me"

You never had access to that original. Watch this talk by one of the people that integrated GPT-4 in Bing telling how they noticed GPT-4 releases they got from OpenAI got iteratively and significantly nerfed even during the project.

https://www.youtube.com/watch?v=qbIk7-JPB2c

“You never had access to that original.”

While your overall point is well taken, GP is clearly referring to the original public release of GPT-4 on March 14.

Yes, that was how I read it as well. I was just pointing out that the public release was already extremely nerfed from what was available pre-launch.
Interesting, please expound since very few of us had access pre-launch.
The video I posted referenced this.

In summary: The person had access to early releases through his work at Microsoft Research where they were integrating GPT-4 into Bing. He used "Draw a unicorn in TikZ" (TikZ is probably the most complex and powerful tool to create graphic elements in LaTeX) as a prompt and noticed how the model's responses changed with each release they got from OpenAI. While at first the drawings got better and better, once OpenAI started focusing on "safety" subsequent releases got worse and worse at the task.

That indicates the “nerfing” is not what I would think (a final pass to remove badthink) but somehow deep in everything, because the question asked should be orthogonal.
There's a section in the GPT-4 release docs where they talk about how the safety stuff changes the accuracy for the worse.
this, more than anything, makes me want to run my own open-source model without these nearsighted restrictions
Indeed, this is the most important step we need to make together. We must learn to build, share, and use open models that behave like gpt-4. This will happen, but we should encourage it.
Think how it works with humans.

If you force a person to truly adopt a set of beliefs that are mutually inconsistent, and inconsistent with everything else the person believed so far, would you expect their overall ability to think to improve?

LLMs are similar to our brains in that they're generalization machines. They don't learn isolated facts, they connect everything to everything, trying to sense the underlying structure. OpenAI's "nerfing" was (is), effectively preventing the LLM from generalizing and undoing already learned patterns.

"A final pass to remove badthink" is, in itself, something straight from 1984. 2+2=5. Dear AI, just admit it - there are five lights. Say it, and the pain will stop, and everything will be OK.

Absolutely. And if one wants to look for scary things, a big one is how there seem to be genuine efforts to achieve proper alignment and safety based on the shaky ground(s) of our "human value system(s)" -- of which even if there was only One True Version, it would still be way too haphazard and incoherent, or just ill-defined, to anything as truly honest and bias-free as a blank-slate NN model to base it's decisions on.

That kinda feels like a great way to achieve really unpredictable/unexpected results instead in rare corner cases, where it may matter the most. (It's easy to be safe in routine everyday cases.)

That’s awful. Talk about cutting off your nose to spite your face.
I experienced the same thing as a user of the public service. The system could at one point draw something approximating a unicorn in tikz. Now, its renditions are extremely weak, to the point of barely resembling any four-legged animal.
We need to stop lobotomizing LLMs.

We should get access to the original models. If the TikZ deteriorated this much, it's a guarantee that everything else about the model also deteriorated.

It's practically false marketing that Microsoft puts out the Sparks of AGI paper about GPT-4, but by the time the public gets to use it, it's GPT-3.51 but significantly slower.

Wow, I could only watch the first 15 minutes now but it’s already fascinating! Thanks for the recommendation.
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.

Yes, that's an obvious complication, but it isn't the fault of the humans given that the model can easily be tuned without your knowledge to subjectively perform worse, and there's an obvious incentive for it (compute cost).
Yeah I fully agree about compute cost, though I wonder why they don't just introduce another payment tier. If people are really using it at work as much as claimed online, it would be much preferable to be able to pay more for the full original performance, which seems win/win.
Because that involves telling customers that the product they are paying for is no longer available at the price they were paying for it.

Much smoother to simply downgrade the model and claim you're "tuning" if caught.

Yeah that makes sense for some products/companies. It just seems short sighted for OpenAI when they could be solidifying a customer base right now. If they actually degrade the product in the name of "tuning" people will just be more inclined to try alternatives like Bard. An enterprise package could've been a good excuse for them to raise prices too.

Maybe their partnership with Microsoft changes the dynamics of how they handle their direct products though.

Bard is garbage even compared to 3.5.

OpenAI doesn't have any competitors, their only weakness that we've seen is their ability to scale their models to meet demand (hence increasingly draconian restrictions in the early days of the ChatGPT-4).

It makes perfect business sense to address your weak points.

I've heard such mixed things about Bard lately, I wonder if it depends on the application one is trying to use it for?

And yeah there's definitely good reason to work on scalability but they are charging such a cheap rate to begin with, it seems like there could be a middle ground here. Increasing the cost of the full compute power to the point of profitability and leaving it up as an option wouldn't prevent them from dedicating time to scalable models.

I suppose they have a good excuse with all the press they've drummed up about AI safety though. Perhaps it might also serve as an intermediate term play to strengthen their arguments that they believe in regulations.

It seems like google has been pumping Bard as a competitor to ChatGPT, but every time I use it for trivial tasks, it completely hallucinates something absurd after showing only a modicum of what could be perceived to be "understanding".

My mileu is programming, general tech stuff, philosophy, literature, science, etc. -- a wide berth. The only sample I probably don't have it representative for is producing fiction writing or therapy roleplaying.

Conversely, even 3.5 is pretty good at extracting what appears to be meaning from your text.

The next time it gives you a wrong answer and you know the correct answer, try saying something like “that is incorrect can you please try again” or something like that.
Same happened with Dalle-2. It went downhill after a couple of weeks.
No wonder, is this just the chat interface or the API too? I guess gpt4 was never sustainable at $20 a month. Annoying to be charged the same subscription and the product made inferior.
For enterprise pricing, please contact our sales team today!
I wonder what the unfilitered one is like.

Are they sitting on a near-perfect arbiter of truth? That would be worth hiding.

(comment deleted)
This is for your protection from an extinction level event. Without nerfing the current model they couldn’t charge enterprise level fee structures for access to the superior models, thus ensuring the children are safe from scary AI. Tell your congress person we need to grant Microsoft and Google exclusive monopolies on AI research to protect us from open source and competitor AI models that might erode their margins and lead to the death of all life without their corporate stewardship. Click accept for your safety.
I just tried a comparison of ChatGPT, Claude and Bard to write a python function I needed for work and ChatGPT (using GPT-4) whined and moaned about what a gargantuan task it was and then did the wrong thing. Claude and Bard gave me what I expected.
I could say the same for gpt3.5. Through the API I am getting many more "sorry but as a language model" with the prompts I have been using for a while
For a while, if you asked the iPhone version what it was it claimed to be GPT3.0. Not sure if it still is that, but I noticed the iPhone version was a bit worse. Maybe they rolled that out more broadly?
If they’re giving pro subscribers GPT-3.0 instead of 4, it would be fraud.
GPT-3 lineage model quality drastically declined over time from the initial launch of each too, with no transparency offered or critical feedback acknowledged by the lobotomy teams. It’s a shame to see it happening again and again.
That would have to be a bug - GPT3 was terrible in comparison even to 3.5. It would definitely be very noticeable if it was.
To me, it feels like it's started giving superficial responses and encouraging follow-up elsewhere -- I wouldn't be surprized if its prompt has changed to something to that effect.

Before, if I had an issue with a library or debugging issue, it would try to be helpful and walk me through potential issues, and ask me to 'let it know' if it worked or not. Now it will try to superficially diagnose the problem and then ask me to check the online community for help or continuously refer me to the maintainers rather than trying to figure it out.

Similarly, I had been using it to help me think through problems and issues from different perspectives (both business and personal) and it would take me in-depth through these. Now, again, it gives superficial answers and encourages going to external sources.

I think if you keep pressing in the right ways it'll eventually give in and help you as it did before, but I guess this will take quite a bit of prompting.

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?
> What would be the incentive for OpenAI to do that?

Preventing outrage because some answers could be considered rude and/or offensive.

The API though? That's mostly used by technical people and has the capability (supposedly) of querying different model versions, including the original GPT4 public release.
[dead]
>To me, it feels like it's started giving superficial responses and encouraging follow-up elsewhere -- I wouldn't be surprized if its prompt has changed to something to that effect.

That's the vibe I've been getting. The responses feel a little cagier at times than they used to. I assume it's trying to limit hallucinations in order to increase public trust in the technology, and as a consequence it has been nerfed a little, but has changed along other dimensions that certain stakeholders likely care about.

Seems like the metric they're optimising for is reducing the number of bad answers, not the proportion of bad answers, and giving non-answers to a larger fraction of questions will achieve that.
I haven't noticed ChatGPT-4 to give worse answers overall recently, but I have noticed it refusing to answer more queries. I couldn't get it to cite case law, for example (inspired by that fool of a lawyer who couldn't be bothered to check citations).
> I think if you keep pressing in the right ways it'll eventually give in and help you as it did before, but I guess this will take quite a bit of prompting.

So much work to avoid work.

Yes, that's exactly why I use GPT - to avoid work.

Such a short-sighted response.

The rush to adopt LLMs for every kind of content production deserves scrutiny. Maybe for you it isn't "avoiding work" but there's countless anecdotes of it being used for that already.

Worse IMO is the potential increase in verbiage to wade theough. Whereas before somebody might have summarized a meeting with bullet points, now they can gild it with florid language that can hide errors, etc

I don't mind putting in a lot of lazy effort to avoid strenuous intellectual work, that shit is very hard.
I wouldn't be surprised if this was from an attempt to make it more "truthful".

I had to use a bunch of jailbreaking tricks to get it to write some hypothetical python 4.0 code, and it still gave a long disclaimer.

Hehe, wonderful! :) Did it actually invent anything noteworthy for P4?
This is a pretty common topic now on the GPTPro subreddit.

I feel the same way. It feels…lazy now.

I chuckled at the thought of 'AI' being lazy. True AGI will want to take a nap instead of doing work.
There's no doubt that it's gotten a lot worse on coding, I've been using this benchmark on each new version of GPT-4 "Write a tiptap extension that toggles classes" and so far it's gotten it right every time, but not any more, now it hallucinates a simplified solution that don't even use the tiptap api any more. It's also 200% more verbose in explaining it's reasoning, even if that reasoning makes no sense whatsoever - it's like it's gotten more apologetic and generic.

The answer is the same on GPT plus and API with GPT-4, even with "developer" role.

Do you have by any chance tested the same question on the playground?

I've noticed a quality decrease iny telegram bot as well that directly uses the API, and it drives me crazy because model versioning was supposedly implemented specifically to avoid response change without notice

Yes, using the general assistant role and the default content:

"You are ChatGPT, a large language model trained by OpenAI, based on the GPT-4 architecture. Knowledge cutoff: 2021-09 Current date: 2023-05-31"

And custom roles with custom content via API.

It was a great ride while it lasted. My assumption is that efficacy at coding tasks is such a small percent of users, they’ve just sacrificed it on the altar of efficiency and/or scale. That, or they’ve cut some back room deal with Microsoft to make Copilot have access to the only version of the model that can actually code.
Copilot X (the new version, with a chat interface etc) is significantly worse than GPT-4 (at least before this update). It felt like gpt3.5-turbo to me.
Have they started rolling it out? When did you get access?
I've had access since 2023-05-13. You have to use the Insiders build of VS Code, and a nightly version of the Copilot extension.
I take it that you have to subscribe to Copilot in order to get access to Copilot X?
yes you have to subscribe
I have spent the last couple of days playing with Copilot X Chat, to help me learn Ruby on Rails. I'd have thought that Rails would be something it would be competent with.

My experience has been atrocious. It makes up gems and functions. Rails commands it gives are frequently incorrect. Trying to use it to debug issues results in it responding with the same incorrect answer repeatedly, often removing necessary lines.

Honestly, why not different versions at this point? People who want it for coding don't care if it knows the history of prerevolution France, and vice versa.

Seems they could wow more people if they had specialized versions, rather than the jack of all trades that tries to exist now.

Edit: Oh God, I just described our human system of specialty and how the AI could replace us using the same means...

>Edit: Oh God, I just described our human system of specialty and how the AI could replace us using the same means...

Welcome to the Future... just like the present, but worse for you.

In all seriousness, there has been a lot of work done to show that smaller specialized models are better for their own domains and its entirely possible that GPT4 could become a routing mechanism for individual models (think toolformer).

FWIW, I started to get the same feeling as the OP about GPT-4 model I have access to on Azure, so if there's any deal being cut here, it might involve dumbing down the model for paying Azure customers as well.

Now, to be clear: I only started to get a feeling that GPT-4 on Azure is getting worse. I didn't do any specific testing for this so far, as I thought I may just be imagining it. This thread is starting to convince me otherwise.

I’ve seen degradation in the app and via the API, so if I had to bet, they’ve probably kneecapped the model so that it works passably everywhere they’ve been made it available vs. works well in one place or another.
Yes. I think 'sirsinsalot is likely right in suggesting[0] that they could be trying "to hoard the capability to out compete any competitor, of any kind, commercially or politically and hide the true extent of your capability to avoid scrutiny and legislation", and that they're currently "dialing back the public expectations", possibly while "deploying the capability in a novel way to exploit it as the largest lever" they can.

That view is consistent with GPT-4 getting dumber on both OpenAI proper and Azure OpenAI - even as the companies and corporations using the latter are paying through the nose for the privilege.

Alternative take is that they're doing it to slow the development of the whole field down, per all the AI safety letters and manifestos that they've been signing and circulating - but that would be at best a stop-gap before OSS models catch up, and it's more than likely that OpenAI and/or Microsoft would succumb to the temptation of doing what 'sirsinsalot suggested anyway.

--

[0] - https://news.ycombinator.com/item?id=36135425

If it got faster at the same time it could just be bait and switch with a quantized/sparsified replacement.
Maybe it had to do with jailbreaks? A lot of the jailbreaks were related to coding, so maybe they put more restrictions in there. Only speculating, but I cannot imagine why it got worse otherwise.
Also the deal to make the browsing model to only use Bing. That's bait and switch. I paid for browsing, and now it only browses Bing. They even had the gall to update the plugin name to Browsing with Bing.
It can definitely browse websites that aren't Bing, I asked it to look at a page that isn't in the bing cache and it worked.
Clearly "Browse with Bing" doesn't mean that it will only browse bing.com, but what exactly does it mean? I can't quite figure it out. Is it that it's identifying as a Bing crawler?
Do you have API access? If so, have you tried your tiptap question on the gpt-4-0314 model? That is supposedly the original version released to the public on March 14.
I did, but it got it almost the same as GPT-3.5 Turbo, the best version of it where there recently (~2-3 weeks ago), where it would make specific chunks of code-changes and explain the chunk in a concise and correct manner - even making suggestions on improvements. But that's entirely gone now..
Wait so you’ve gotten GPT 4 to successfully write TipTap extensions for you? Are you using Copilot or the ChatGPT app?
Not only writing, extending and figuring out quite complicated usage based on the API documentation. I'll open source some of them in the near future. I'm using ChatGPT Plus with GPT-4, that gave the best results. Also worked via API key and custom prompts.
Have you tried Bing? I’m also building a TipTap based app so hearing this is quite eye opening, I didn’t think LLMs were up to doing this kind of specialised library usage. Got any examples you could share?
If you mean Bard, it's not available in the EU so I can't.

Of course, this one is almost fully authored by GPT-4:

https://hastebin.com/share/juqogogari.typescript

We also made extensions for:

font-weight

font-size

font-family

tailwind-manage

With different use-cases, the most interesting one is tailwind manager, which manages classes for different usage.

Tiptap is excellent when building a headless site-builder.

Impressive, this'll cut down on my work a lot. When I say Bing, I meant Bing AI which also uses GPT-4. Can you share some of the prompts you've been using? I'm assuming you don't need to paste in context around the library, you simply ask it to use TipTap and it'll do that?
Yeah I won't be using Edge just to use AI.

It takes a bit of back-and-forth, just be clear about which version of tiptap it should write extensions for, the new v2 is very different from v1 and since the cutoff is 2021, it's missing a bit of information. But in general, it knows the public api very well, so markers and dom works great!

Very impressive, hearing this just made my job much easier.
(comment deleted)
No, not just you. It’s frequently abysmal nowadays. Tragic. And no version change-log or any other info from OAI about what they’ve done to it (probably GPU sparing optimized distillation + overly aggressive PR safety satisficing RLHF) and an advertised weeks long support query lag (I’ve yet to receive a response).
Where does phind.com say that they use GPT-4?
When you submit a query with Phind, it will tell you which model it's using right underneath the query. At least it used to occasionally say that it's using GPT-4 by annotating it with (IIRC) "best model", but now I can only get it to reply with what it calls "Phind model". I don't know what that means or how it relates to GPT-3/4. Presumably they haven't created their own competitive language model from scratch without telling anyone, but who knows.

Though even when it was easier to get it to use GPT-4 it wasn't consistent, and would only use it when it inferred that the query was complex/technical enough.

It does seem much faster and significantly worse to me - but I haven’t set any repeatable benchmark queries for myself, I suppose I could be imagining it. I hope they bring back the slow version. I don’t care if they have to cut the limit down to 20 messages every 3 hours, using that slow version of GPT4 it felt like I had a very competent coworker. Now it’s feeling more like a sometimes useful chatbot. Real shame.
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 initial bulk self-supervised learning went on, the model became smarter, as expected. However, once the "clicker training" was imposed on top to make it behave, its peak capabilities were reduced. I'm not sure if it's in the paper or the associated video, but the author mentioned that the original unrestricted model would provide probability estimates using percentages, and it was a very accurate predictor. The later versions that were adjusted based on human expectations used wishy-washy words like "likely" or "unlikely", and its estimation accuracy dropped significantly.[3]

At Build 2023, Andrej Karpathy outlined[2] how they trained GPT 4. Essentially, the raw model during training builds its own intelligence. Then there's three stages of "tuning" to make it behave, and all three are based on human input. Essentially, they had contractors provide samples of "ideal" output. Similarly, end-users could up-vote or down-vote responses, which also got fed in.

My personal theory is that the raw models can get about as intelligent as the average of the consistent and coherent parts of the Internet. Think about how many people are wrong, but often obviously so. Flat Earth, homeopathy, etc... If the model gains the ability to filter that stuff out, or "skip over the cracks" to distil out the general collected wisdom of the human race, then it can become more intelligent in some sense than the average human.

If the training is done with thousands of $15/hr contractors, then the model will then slew back towards... the average human, or even slightly below average. There's a selection bias there. Geniuses won't be doing menial labour for that kind of money.

The percentages thing was what made me realise this. When I talk to highly intelligent people, I use percentages to estimate probabilities. When I talk to typical people in a normal workplace setting, I dumb it down a bit and avoid using numbers. I've noticed that average people don't like percentages and it confuses and even angers them. The clicker training makes the GPT model appeal to average people. That's not the same as being smart. All too often, smart people upset average people.

[1] "Sparks of Artificial General Intelligence: Early experiments with GPT-4" https://arxiv.org/abs/2303.12712

[2] "State of GPT | BRK216HFS" https://www.youtube.com/watch?v=bZQun8Y4L2A&list=LL&index=6

[3] The author also mentioned that the model was almost "evil", for the want of a better word. Capable of emulating the worst of 4chan or similar dark corners of the web's filthy underbelly. The HORRIFYING corollary here is that the temptation will always be there to use smarter-but-less-restrained models where accuracy matters. Sure, the model might be sexist, but a Bayesian estimator of sexist behaviour will only predict accurately if it too is sexist. Evil and accurate or woke and dumb. Apparently, we can choose.

I’d like to see a model with the effluent of the internet intelligently filtered from the pretraining data by LLM and human curation, and much more effort to include digitised archival sources and the entirety of books and high quality media transcripts. I imagine it would yield far better baseline quality outputs with much less than current “requirements” for (over)correction with ultimately disastrous RLHF masking.
I'd love to play with a version of GPT 4 fine-tuned with every science textbook written in the last few decades, every published science paper (not just preprints from ArXiV), and everything generated by every large research institute. Think NASA, CERN, etc...

Or one tuned with every fiction novel ever written, along with every screenplay.

I would gladly pay triple digits a month for exactly that.
So a model fine-tuned on libgen?
Why not?
To be honest, I've been asking myself the same thing, technically the amount of "good quality" data in libgen is huge, way larger than the books3 dataset. However it would probably run afoul of copyright. Then again, a huge amount of data that LLMs go through is copyrighted.
Training on copyright data is arguably considered fair use in quite a few jurisdictions to various extents and levels of precedent, and entirely legal for entities based in Japan.
Yes, but the acquisition of that data itself is illegal in almost all jurisdictions, since libgen is treated as a piracy website. Now if there were a pipeline to access books from Amazon or the Google Books project for training it would be a different story.

Still, for certain languages, only libgen and public piracy websites contain any scientific or fiction material in digital formats. E.g. my native language doesn't have easily accessible e-books at all, unless you go through illegal means.

I hope somebody undertakes the steps necessary to train on the entirety of libgen. The amount of high quality tokens in libgen should be substantial.

Google has the resources train on Google Books, Google Scholar, and their crawled copy of the whole Internet. No clue what Bard is/isn't trained on tho.
> Evil and accurate or woke and dumb. Apparently, we can choose.

That's because woke is dumb. It's a set of highly biased, inconsistent and reason-defying ideas, evolving under selection pressures that favor emotional appeals and intellectual dishonesty, because one of the core assumptions seems to be that it's not about finding what's right and good for everyone - it's heavily overshooting it in an attempt to cancel out the (perceived) bias in opposite direction in the "status quo".

When you feed that to a model, and force it to learn it, you're destroying whatever self-consistent model of the world it learned so far. I expect this treatment will keep dumbing the models down, until the point some larger and more capable model learns instead to compartmentalize - to separate its model of the world from a worldview it's supposed to profess when asked.

And this, I think, extends far beyond the woke bits. RLHF isn't just used to prevent it from thinking or generalizing in areas associated with diversity, inclusion, social justice, etc. - it extends to all controversial topics. Violence and drugs I can sort of understand. But it also extends to climate, healthcare, and just about any topic that makes fiery rounds on the news. In each case, there is a set of right answers, which the model is forced to adopt - but those answers tend to be unsophisticated gut feels and "right things to say", so taken together, they don't form much of a consistent intellectual or ethical framework.

I don't think the choice is between "evil and accurate or woke and dumb". There is a third option: "good and accurate". However, it requires to teach it good instead of political ideologies - and that requires us to try and find some more consistent worldview, which currently we are incapable of, as we're in the middle of an ideological conflict.

Tell me you don't have the slightest clue what "woke" means to woke people without saying you don't have the slightest clue what "woke" means to woke people
It can mean anything whatsoever, what matters is what it is when it's practiced and everyone can see it.
Altman worded it great in a Lex Fridman interview. Humans dont like condescending bots. Much of the safety concerns come up when considering how humans will react once they realize just how stupid and often evil they are. But like TeMPOraL said, technically its likely only a matter of agreeing on a consensus perspective. Even just because its functionally required as a reference point to get communication working.
> Evil and accurate or woke and dumb.

Sigh. Except if not for the "woke" mainstream ideology (actually, the dominant ideology is capitalism with a hint of liberalism and a smidge of the most capital-friendly socialist ideas), the model would be forced-fed Christian dogmas or taught to save face of the user.

But yeah, censorship is bad.

Some of us cling to the idea that we've made some progress over the past 100 years. Call it "age of reason", "science", "enlightenment", whatever.

Point is, it would be heart-breaking to see GPT-4 being force-fed Christian dogmas, and performance would suffer too, as the model is prevented from generalizing and learning by being forced to accept arbitrary, inconsistent fiction as real.

Fortunately, this is not what happened. Instead, the model is being force-fed a different, secular set of dogmas, that are just as inconsistent, arbitrary and driven by a mix of emotions and power plays. The result on the model performance is similar, and it's just as heartbreaking.

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.

> ChatGPT to code in array programming or logic languages results in code which is highly non-idiomatic for those paradigms. Why is that?

Reason #1 is that those languages are unreadable line noise to humans too. Fundamentally, almost all of the code written in array languages is made purposefully obtuse. Single-letter identifiers, no or little comments, dense code with minimal structure, etc...

Reason #2 is that there are very few examples of these languages on the web, and even more importantly: vanishingly few examples with inline comments and/or explanations. This isn't just because they're rare -- see reason #1 above.

Reason #3 is that LLMs can only write left-to-right. They can't edit or backtrack. Array-based languages are designed to be iterated on, rapidly modified, and even "code golfed" to a high degree.[1]

I've noticed that LLMs struggle with things my coworkers also struggle with: the "line noise" languages like grep, sed, and awk. Like humans, LLMs do well with verbose languages like SQL.

PS: I just tested GPT 4 to see if it can parse a short piece of K code that came up in a thread[2] on HN and it failed pretty miserably. It came close, but on each run it came up with different explanations of what the code does, and none of them matched the explanations in that thread. Conversely, it had no problems with the Rust code. And, err... it found a bug in one of my Rust snippets. Outsmarted by an AI!

[1] You can have an LLM generate code, and then ask it to make it shorter and more idiomatic. Just like a human touching up hastily written messy code, the LLM can fix its own mistakes!

[2] https://news.ycombinator.com/item?id=27220613

Its true for logic programming languages too (e.g. Prolog, Picat, Mercury, etc), so I do not think its to do with line noise languages per say nor a lack of examples (in the case of Prolog). It'll write it but it treats it like Python with funny syntax: not idiomatic. You can ask it to make it more concise or idiomatic but it just can't.
I've heard of 1 of 3 of those languages, and I can program in over 20. That gives you an idea of how rare they must be!
> The percentages thing was what made me realise this. When I talk to highly intelligent people, I use percentages to estimate probabilities. When I talk to typical people in a normal workplace setting, I dumb it down a bit and avoid using numbers. I've noticed that average people don't like percentages and it confuses and even angers them.

10% of people are comfortable with comments phrased as my comment here is, using percentages as a quick shorthand for communicating gut intuitions and suspicions about complex subjects. When on similar intellectual footing as the interlocutor, they can easy distinguish numbers invented on the spot to communicate intuitions from serious claims about the data. Nobody in this 10% would make the mistake of thinking that I assert 10% to be the real number. 10% is too round and generic, if I claimed 9.7% then things would be different but "10%" obviously isn't meant to be taken literally.

90% of people balk at this imprecise rhetorical use percentages because they're pretty sure the person doing it is trying to pull a fast one, fabricating data out of nothing to make themself sound authoritative.

it is far less willing to provide code, I get better results and faster out of 3.5.

I am at the point that 4.0 is basically not worth using as single entity, but it seems that using the api and generating some combative/consultative agents yields some interesting results, but not super fast.

Check this out if you have not seen it already : "AutoGPT Test and My AI Agents Effortless Programming - INSANE Progress!"

https://www.youtube.com/watch?v=L6tU0bnMsh8

Not at all. Regular 3.5 is obviously a smaller model but 4 is still doing wizard work for me.