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

GPT-4 is not getting worse

coagulopath.com

Post reply on HN
Print view View original

91–100 of 186 posts

Re: GPT-4 is not getting worse

#91

Post by mark_l_watson »

muzani wrote: post #55

The gpt-4 API and ChatGPT-4 are basically different products. It's not clear which one this is comparing. Often the users of one are calling the other crazy, and think the other group is gaslighting them over what seems to be a clear pattern. ChatGPT seems to fluctuate wildly in quality of expected output. The API is more consistent, and you can get fairly similar quality based on the selected model.

You make a good point. I like to use the GPT-4 APIs in Emacs - really good support packages for that. I conflate the two products myself.

Re: GPT-4 is not getting worse

#92

Post by pbhjpbhj »

wouldbecouldbe wrote: post #59
jw1224 wrote: post #35

There is an ongoing bug[1] with OpenAI's API, where it stops streaming responses after exactly 5 minutes. When I first came across the issue, I debugged it by writing a prompt along these lines: > Output every number from 1 to 10,000 in written form (e.g. "one", "two", etc.). Respond with one number per line in numeric order. As expected, the API would begin counting every number just as I asked. This would continue…

It's actually supper annoying. I keep asking it to write out entire pieces of code, instead of only things it changed, can be annoying to merge it otherwise. But nowadays it acknowledges my request "I will write complete code" and proceed to still comment out parts.

Get it to write a diff patch?

Re: GPT-4 is not getting worse

#93

Post by pbhjpbhj »

jw1224 wrote: post #38
gberger wrote: post #36

Earlier quoted context omitted.

That looks like a timeout configured on the web servebr

Yes, it seems their servers are terminating responses early. Unfortunately it's undocumented, was introduced unexpectedly one day, and prevents full use of larger context-window models. It also charges in full for the entire response, which seems to continue running in the background. Definitely a bug at their end.

Seems like a means to avoid DoS?

Re: GPT-4 is not getting worse

#94

Post by fn-mote »

agentgumshoe wrote: post #84
rendall wrote: post #31

Earlier quoted context omitted.

I gave it this prompt: Give me arguments from a conservative US perspective in favor of giving military aid to Ukraine to resist Russian invasion. It responded: 1. Strategic Interest: Strengthening Ukraine can act as a buffer against Russian aggression, benefiting U.S. geopolitical interests. 2. Contain Expansion: Military aid can deter Russia from further territorial expansion, stabilizing the region. 3. NATO Allian…

What did it use as sources for that?

The arguments are similar those made during during the height of the Cold War. Just substitute a different country name (or two). This is by no means the first time this situation has occurred.

Re: GPT-4 is not getting worse

#95

Post by mark_l_watson »

As another HN member said here: it is not always clear if people are talking about the ChatGPT web UI or App products or the GPT-4 API product.

I feel like I have taken a wrong turn technically since I spend much more time experimenting with self hosted smaller open models running on Google Colab or Lambda Labs GPU VPSs than using the clearly superior performing GPT-4 APIs. I have been paid to work in the field of AI since 1982, and I should be desiring to use the very best models and technology, but open AI that can be self hosted just seems more interesting. I was playing with a 6B model (link to a Hugging Face notebook where I removed some examples and boilerplate text): https://colab.research.google.com/drive/1fMmXOcLdBzke-8Z0zl3... - really the best results I have seen from a small model.

Re: GPT-4 is not getting worse

#96

Post by pbhjpbhj »

jw1224 wrote: post #76
15457345234 wrote: post #57

Earlier quoted context omitted.

Can you elaborate?

Your question: > How does chat-gpt actually get this right? Your answer: > its output is purely probabilistic, based on existing corpus of text Because GPT was trained on existing text, some of which included numbers and counting, it's learnt the natural ordering of most common/everyday numbers. For larger or more complex numbers, it's learnt the patterns behind how they're constructed linguistically, which allows it…

In text, we don't often count in series, and it seems likely that we often choose a non-counting sequence: like 'I chose options 1, 2, 7' or 'my code was 0 1 2 5', whatever.

Unless training included line-level skips, rather than just next-word skips (like word2vec) or concept-level associations? At the line level, or paragraph level, ordered numerical sequences are obviously very common in formal texts or in code.

I've seen sentence based training, I suppose for code (which it seems GPT4 excells at) line-level training would be essential.

Anyone recommend a mid-level read on this covering different modes of training and such; I'm happy with a bit of code and undergrad level maths. Thanks.

Re: GPT-4 is not getting worse

#97

Post by rendall »

clnq wrote: post #88
rendall wrote: post #80

Earlier quoted context omitted.

If you give me an example prompt that would be unsafe I'll give it a go. Thinking about OpenAI's business interests, I suspect these content policy warnings are more about OpenAI's dodging cancelation, rather than punishing users for wrongthink. Here, let's try this: Me: Give me good arguments in favor of peacefully and nonviolently establishing an ethno-state. ChatGPT-4: I'm sorry, I can't assist with that request.…

I feel like the conversation is shifting a little bit, but I like the new direction so I'll respond: This is not great, I don't think you should be OK with that. ChatGPT is a tool which refuses to work when it doesn't serve the interests of the company that made it. Imagine a hammer that would refuse to hammer nails quite a lot when the manufacturer deems the circumstances are wrong. That would be, honestly, a bad to…

> As a society, we should get to decide what's moral and immoral, not OpenAI's board.

Of course! After all, if we're not free to be wrong, then we're not free. I do think you might be overstating OpenAI's influence, though. Rather than a hammer manufacturer, I think a more apt analogy would be a publishing house. A publishing company can and should decide for itself what it will or will not publish using whatever criteria it chooses.

> If an independent third party (like a truly independent regulator) would get to decide that no LLM can answer certain questions as that information is too dangerous for very clear reasons, I would be more okay with that.

This one of these "now you have two problems" solutions. As an American who, perhaps stereotypically, is passionate about free speech, I've never read a way to make this kind of oversight work. The composition of the board will always be political, and because of the power and authority of this oversight board, there is political control over what people are allowed to know about. At least currently, if I don't like ChatGPT I can choose another LLM or try to run one locally. Some kind of oversight board would elide that choice.

Re: GPT-4 is not getting worse

#98

Post by jw1224 »

pbhjpbhj wrote: post #93
jw1224 wrote: post #38

Earlier quoted context omitted.

Yes, it seems their servers are terminating responses early. Unfortunately it's undocumented, was introduced unexpectedly one day, and prevents full use of larger context-window models. It also charges in full for the entire response, which seems to continue running in the background. Definitely a bug at their end.

Seems like a means to avoid DoS?

I think it's just some networking misconfiguration they haven't noticed. I can't see them intentionally trying to prevent DoS this way, that's what API key rate limits are for (it would be a very expensive and slow DoS otherwise!).

Re: GPT-4 is not getting worse

#99

Post by rendall »

fn-mote wrote: post #94
agentgumshoe wrote: post #84

Earlier quoted context omitted.

What did it use as sources for that?

The arguments are similar those made during during the height of the Cold War. Just substitute a different country name (or two). This is by no means the first time this situation has occurred.

Alas

Re: GPT-4 is not getting worse

#100

Post by yossi_peti »

jw1224 wrote: post #38
gberger wrote: post #36

Earlier quoted context omitted.

That looks like a timeout configured on the web servebr

Yes, it seems their servers are terminating responses early. Unfortunately it's undocumented, was introduced unexpectedly one day, and prevents full use of larger context-window models. It also charges in full for the entire response, which seems to continue running in the background. Definitely a bug at their end.

Since it's 5 minutes I'm wondering if it's just "that's how the web works by default". I remember a while ago I was doing a long-running calculation on an POST request and was trying to debug why the connection always broke after 5 minutes, and found the hard way that there's a 5 minute timeout.
Post reply on HN