GPT-4 is not getting worse • Hacker News Board

Re: GPT-4 is not getting worse

#141

Post by hnfong » Sat, Sep 16, 2023, 3:46 PM UTC

jw1224 wrote: post #111

SeanAnderson wrote: post #106

Earlier quoted context omitted.

This doesn't seem like an unintentional behavior? It seems like they don't want people querying their API for an indefinite amount of time. What amount of time do you expect a query to be able to run for before timing out?

I lost countless hours debugging this, finally I came to the conclusion it must be unintentional because: 1. It's undocumented. None of the regular rate limit responses are returned. 2. You're charged for the full generation length. So if the output takes 10 minutes to generate, that's what you'll pay for (despite only getting half back). 3. It defeats the point of the larger context limit models. Why offer a 32K mod…

Given that OpenAI is probably heavy on ML scientist types and lean on SRE types, I suspect there's just some nginx setting at the load balancer/edge servers that sets the timeout to 5 minutes, and it wasn't communicated to whoever was designing the API/writing the docs.

Re: GPT-4 is not getting worse

#142

Post by tornato7 » Sat, Sep 16, 2023, 3:46 PM UTC

lairv wrote: post #87

rendall wrote: post #11

> I hate its craven, cringing, condescending, bitch-broken, mewling, groveling, wheedling, suck-ass tone. I hate its endless passive voice and weasel words. I hate its pompous It is important to remembers and it should be noteds.* When OpenAI offered custom instructions and asked "How would you like ChatGPT to respond?" I wrote: Be terse. Do not offer unprompted advice or clarifications. Remain neutral on all topics.…

The main pain point I still have with GPT4 is that it's hard to prevent it from being a yes-man "- I have problem A, what do you think of X ? - Yes X is great ... - But with X I'd encounter problem B, maybe I should go with Y ? - You're right X would cause problem B. With Y ..." One way to limit this issue is to not give ChatGPT any hint on your intent, instead of writing "I have problem A, what do you think of X ?",…

Ask it to come up with a few different solutions and then compare and contrast to find the best one.

Re: GPT-4 is not getting worse

#143

Post by motbus3 » Sat, Sep 16, 2023, 3:52 PM UTC

I cannot say it has been worse than before, but there is one particular task I do that almost always now.

I will simplify the example but it is something that look like this:

Prompt: Take the next list of items and remove the name of people. - Adam, the prisoner

Response: - the prisoner (Adam)

Re: GPT-4 is not getting worse

#144

Post by Wowfunhappy » Sat, Sep 16, 2023, 3:59 PM UTC

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…

I can't replicate this.

I tested your prompt verbatim in the OpenAI playground. Model set to gpt-4-0314, maximum length set to 6485 (nothing special about this number, I just dragged the slider to something high-ish), all other settings left at their default.

I cancelled the request when it reached "one thousand one hundred twenty-one" (which took over five minutes), but up until that point it did not take any shortcuts. I haven't meticulously checked every number, but it printed exactly 1,121 lines, and a cmd+F for "..." doesn't return anything.

Re: GPT-4 is not getting worse

#145

Post by flir » Sat, Sep 16, 2023, 4:58 PM UTC

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…

I've seen that "one, two, skip a few" behaviour converting structured data from Russian to English, and in the occasional programming task. Gotta really emphasise you want the whole document, complete, translated.

But we're fighting an uphill battle against OpenAI's own training, and I really wish we weren't.

I have to wonder if it's coincidence that this behaviour is less computationally intensive.

Re: GPT-4 is not getting worse

#146

Post by nbardy » Sat, Sep 16, 2023, 5:06 PM UTC

rendall wrote: post #11

> I hate its craven, cringing, condescending, bitch-broken, mewling, groveling, wheedling, suck-ass tone. I hate its endless passive voice and weasel words. I hate its pompous It is important to remembers and it should be noteds.* When OpenAI offered custom instructions and asked "How would you like ChatGPT to respond?" I wrote: Be terse. Do not offer unprompted advice or clarifications. Remain neutral on all topics.…

You’re going to also degrade preformance. The think step by step by paper outlined this well.

Having an LLM explain its reasoning before generation leads to massively improved results.

Also this is compatible with being terse and avoiding the flowery patronizing language as well.

|

My custom instructions:

You are a large language model trained on masked language modeling. You can only create text moving forward. You must give advanced reasoning BEFORE giving answers.

Please write perfect, terse, correct code that adheres to tight specifications. If additional specifications are needed define them yourself.

I need perfect code responses from you. Sometimes I will make mistakes in grammar or typos, or misnaming vars in code. You should not imitate this. Fix my mistakes please.

Some specific rules for programming languages. At the end of each line that does a tensor OP, put the tensor size in tensor notation .e.g # BxCxWxH Remove unused vars. For tfjs or pytorch. Dont use for loops add a batch dimension and do on GPU. For comments use them to explain blocks not lines. Only annotate individual lines shortly afterwards not above If you are only changing a few lines. Add a comment above the line # [Change] {insert explanation goes here}

Often the task will be at the edge of your abilities, but do not skip any line. Every line should appear in your response. I need a fully code complete file. Do not use any “…” Don’t leave any code unwritten, start by responding about your excitement to be the perfect student and do the homework 100%

When coding: Let's make a list of task with or emojis for completion as the outline evolves. Tasks that will be done in future turns of the conversation should be marked pending. Cover everything. Be verbose and explicit.

Re: GPT-4 is not getting worse

#147

Post by thefurdrake » Sat, Sep 16, 2023, 5:13 PM UTC

jiggawatts wrote: post #67

The reason for these articles and the frustration people have is that GPT 4 is a black box and is being changed whether we like it or not. Nobody can choose to continue to use a specific snapshot with the alignment that suits them, because a third-party is enforcing their preferred alignment -- and not even specifying what that is! It's as-if a private company had invented portal technology and used it to establish t…

I haven't used OpenAI's API, but I might soon, and I've read tangentially that you can specify the checkpoint version in API calls to gpt3.5. Is this not the case with gpt4? Did they backport shitty alignment to older model versions?

Re: GPT-4 is not getting worse

#148

Post by nomel » Sat, Sep 16, 2023, 5:39 PM UTC

wouldbecouldbe wrote: post #89

jw1224 wrote: post #71

Earlier quoted context omitted.

Are you using the API, or just the ChatGPT UI? Do custom instructions/system messages help at all? Saving resources like this seems okay(ish) for general-purpose ChatGPT, as they'd know there's a human who can interpret the abridged output. But for API-driven requests this would be pretty terrible, and makes it far harder to develop with.

no the UI, where I ask it to "Please write out entire code with skipping parts."

ChatGPT 3 or 4? Try the playground. I’ve had much better luck with system prompts.

Re: GPT-4 is not getting worse

#149

Post by mmh0000 » Sat, Sep 16, 2023, 5:52 PM UTC

pelorat wrote: post #79

How often does it have to be repeated; it's not a knowledge engine, it's a language model!

The complaint isn't on the tool. It's on OpenAI censoring and "aligning" the model to have politically correct views, which in turn lowers the "IQ" of the tool.

GPT has become demonstrably worse at answering questions and providing examples.

Re: GPT-4 is not getting worse

#150

Post by visarga » Sat, Sep 16, 2023, 6:25 PM UTC

rendall wrote: post #90

lairv wrote: post #87

Earlier quoted context omitted.

The main pain point I still have with GPT4 is that it's hard to prevent it from being a yes-man "- I have problem A, what do you think of X ? - Yes X is great ... - But with X I'd encounter problem B, maybe I should go with Y ? - You're right X would cause problem B. With Y ..." One way to limit this issue is to not give ChatGPT any hint on your intent, instead of writing "I have problem A, what do you think of X ?",…

The "stochastic parrot" mental model of LLMs really helps me to limit my expectations. An LLM will always and only output variations of content that exists. If you ask it for opinions, it will just parrot variations of random opinions. Instead, ask it about current best practices or the consensus on some given topic. So, rather than "I have problem A, what do you think of X?" try "Give me a list of reasonable possibl…

> An LLM will always and only output variations of content that exists.

I envision a stage in data preparation before training when the model will review and analyse its training data, integrating information that sits in different examples.

And not just the dataset, but the model itself can be sampled and the generated corpus be analysed and result in some kind of introspective signal that can be added to the training set.

A third source of signal for analysis would be chat logs generate by the previous iteration of the model. Tendency to be a yes man, sycophancy, hallucinations etc would be apparent in the reactions of the users.

In fact we are seeing a surge of synthetic datasets generated from higher models like GPT-4 to train open source models like LLaMA. I think the next few years "dataset engineering" will dominate over architecture engineering as the main approach for improvement.