Ask HN: Is it just me or GPT-4's quality has significantly deteriorated lately?
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 ] threadIt 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
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...
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
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.
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.
"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.
Of course, the performance of an unchanged model does not. But finetuning the model over time can of course either improve or degrade performance.
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.
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.
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).
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.
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.
The Swiss version of GDPR is coming in September:
https://www.ey.com/en_ch/law/a-new-era-for-data-protection-i...
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.
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.
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.
Making it unavailable in the EU decreases the likelihood and severity of a potential fine.
[1] - https://ico.org.uk/for-organisations/data-protection-and-the...
https://console.cloud.google.com/vertex-ai/generative/langua...
Also do you need to change the options like Token Limit etc?
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.
Alpaca is pretty good though.
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.
Can you please help me with how you are prompting it?
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.
Examples:
https://news.ycombinator.com/item?id=35224167#35227068
https://news.ycombinator.com/item?id=35303210#35360467
My testing agrees with yours. Almost seems like a sponsored marketing campaign with no truth to it.
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.
"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
For now. It's just a marketing tool/demo site, like ITA Matrix was/is. The ads are vended by Bing.
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.
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
While your overall point is well taken, GP is clearly referring to the original public release of GPT-4 on March 14.
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.
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.
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.)
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.
https://www.youtube.com/watch?v=oLiheMQayNE&t=2849s
Anyone know more about this?
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.
Much smoother to simply downgrade the model and claim you're "tuning" if caught.
Maybe their partnership with Microsoft changes the dynamics of how they handle their direct products though.
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.
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.
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.
Are they sitting on a near-perfect arbiter of truth? That would be worth hiding.
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 work with temperature 0 which should have low variability yet recently it shifted to feel boring, wooden, and deflective.
Preventing outrage because some answers could be considered rude and/or offensive.
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.
So much work to avoid work.
Such a short-sighted response.
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 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.
I feel the same way. It feels…lazy now.
The answer is the same on GPT plus and API with GPT-4, even with "developer" role.
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
"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.
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.
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...
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).
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.
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
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.
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!
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.
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.
Or one tuned with every fiction novel ever written, along with every screenplay.
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.
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.
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.
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'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.
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
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.
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