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

How is ChatGPT's behavior changing over time?

arxiv.org

Post reply on HN
Print view View original

21–30 of 187 posts

Re: How is ChatGPT's behavior changing over time?

#21

Post by ineedasername »

blobbers wrote: post #14

A little surprised this sort of thing passes for a publish-able paper. Isn't it the equivalent to saying "Here's the top 10 results for google searching golden retrievers March 2023, and here's the top 10 results from June 2023. We see that google is returning even cuter animals today. Unfortunately though, one of the results linked to a page full of cats." I'm sure openai has a list of standard questions that it tra…

No, the methodology listed in the paper shows their efforts were significantly more extensive than you represent here. As one example, their test of prime numbers went through 500 randomly selected primes, also evaluating the chain of thought in responses along with the success rate of this classification task. I’m not sure how you arrived at your impression, but it does not match the contents of the paper that I saw with a quick scan of each section.

Re: How is ChatGPT's behavior changing over time?

#22

Post by opportune »

quickthrower2 wrote: post #9
Der_Einzige wrote: post #2

Never let OpenAI or Google employees (in regards to search) gaslight you into believing that things aren't being "enshittified". Glad to see scholarly evidence of this coming out.

Honest enshittification is good but rare. Like a cloud telling you what specific CPU SKU is in the hardware (not just 1 vcore). They should say what you are paying for in terms of technicals, either spill the beans on the architecture or have some SLAs on how good the thing is.

I work in cloud-land, the problem with exposing every little technical detail is that 1. part of the point of purchasing a cloud product is to abstract that away 2. coding to the implementation instead of the API creates huge headaches for everybody because there are often good reasons (make things more efficient) to change the implementation 3. sadly, there are “features” such as overcommit which you don’t want to know about as a customer.

While some of those are anti-customer on their face, the “we shouldn’t commit to implementation details and instead just to our API/feature surface” may seem anti-customer while actually making things much better for customers. When implementations are supported in perpetuity development grinds to a standstill. Even overcommit ends up helping customers - cloud companies can offer lower sticker prices.

In the case of both regular Public cloud and OpenAI - if you need stability beyond publicly supported APIs you are probably not a good fit as a customer and should instead find a company willing to commit to lower level implementations (eg bare metal, truly major/minor versioned software) or roll-your-own.

Re: How is ChatGPT's behavior changing over time?

#23

Post by moonchild »

ttul wrote: post #13

I tried the prime number problem and GPT-4 nailed it. I’m not sure whether they are testing things correctly… “ Sure, let's go step by step. A prime number is a number greater than 1 that has no positive divisors other than 1 and itself. This means if we can find any other number (excluding 1 and the number itself) that divides 17077, then it is not a prime number. Let's start by checking divisibility by 2. Since 170…

> GPT-4 nailed it

Did it? Did you read what it wrote?

Re: How is ChatGPT's behavior changing over time?

#24

Post by blobbers »

ineedasername wrote: post #21
blobbers wrote: post #14

A little surprised this sort of thing passes for a publish-able paper. Isn't it the equivalent to saying "Here's the top 10 results for google searching golden retrievers March 2023, and here's the top 10 results from June 2023. We see that google is returning even cuter animals today. Unfortunately though, one of the results linked to a page full of cats." I'm sure openai has a list of standard questions that it tra…

No, the methodology listed in the paper shows their efforts were significantly more extensive than you represent here. As one example, their test of prime numbers went through 500 randomly selected primes, also evaluating the chain of thought in responses along with the success rate of this classification task. I’m not sure how you arrived at your impression, but it does not match the contents of the paper that I saw…

Not to seem sarcastic, but 500 randomly selected primes doesn't change my mind. I read the paper, and dumbed it down to dogs and cats.

Do you really feel like the efforts were significant or meaningful?

Re: How is ChatGPT's behavior changing over time?

#25

Post by airgapstopgap »

swyx wrote: post #17

Previous commentary I know of from OpenAI staff: Logan: The API does not just change without us telling you. The models are static there. https://twitter.com/OfficialLoganK/status/166393494793189785... may 31 Peter: No, we haven't made GPT-4 dumber. Quite the opposite: we make each new version smarter than the previous one. https://twitter.com/jlowin/status/1679660938415177731 july 14 either the models are static, or…

Not really. They have a way of squaring this circle, by changing their inference code. Speculative sampling [1] would still make their first claim a lie – sure, there'd still be the original GPT-4 model, plus a smaller draft worker. But early exit decoding [2] allows you to get almost as good results for much cheaper from exactly the same checkpoint. We know that this line of research for large-scale inference is going strong [3] so it stands to reason that OpenAI, with their wealth of talent focused on GPT-4 throughput&inference [4], large contexts and aggressive pricing policy, would also develop something like that. And of course it's "smarter" that way – in a very deceptive sense of the word.

1. https://arxiv.org/abs/2302.01318

2. https://arxiv.org/abs/2207.07061

3. https://arxiv.org/abs/2307.02628

4. https://openai.com/contributions/gpt-4

Re: How is ChatGPT's behavior changing over time?

#26

Post by gyrovagueGeist »

blobbers wrote: post #14

A little surprised this sort of thing passes for a publish-able paper. Isn't it the equivalent to saying "Here's the top 10 results for google searching golden retrievers March 2023, and here's the top 10 results from June 2023. We see that google is returning even cuter animals today. Unfortunately though, one of the results linked to a page full of cats." I'm sure openai has a list of standard questions that it tra…

Its arxiv its not a published paper. Until it's reprinted by a journal or proceedings, its basically a blog post that people feel more comfortable citing. It's main purpose is to introduce/increase caution about other methods using these services: "our findings shows that the behavior of the “same” LLM service can change substantially in a relatively short amount of time, highlighting the need for continuous monitoring of LLM quality"

Which seems valuable for other research using OpenAI GPT directly.

Re: How is ChatGPT's behavior changing over time?

#27

Post by kevinventullo »

moonchild wrote: post #23
ttul wrote: post #13

I tried the prime number problem and GPT-4 nailed it. I’m not sure whether they are testing things correctly… “ Sure, let's go step by step. A prime number is a number greater than 1 that has no positive divisors other than 1 and itself. This means if we can find any other number (excluding 1 and the number itself) that divides 17077, then it is not a prime number. Let's start by checking divisibility by 2. Since 170…

> GPT-4 nailed it Did it? Did you read what it wrote?

It certainly nails the cadence of a rigorous proof. It just needs to learn to append “the rest is left as an exercise to the reader” to the end of its output.

Re: How is ChatGPT's behavior changing over time?

#28

Post by ehnto »

dudeinhawaii wrote: post #6

I think we should stop trying to quiz LLMs on mathematics, something for which they are explicitly not designed to do with their tokenized view of the world. Ask GPT-4 to use its Wolfram plugin and it returns the answers quickly and correctly. Second, I think the code generation bit of this paper is blown out of proportion. The code can't be immediately injected into a codebase due to a formatting change (triple quot…

I think knowing if the code can be used verbatim is actually the more important part practically speaking. That is the actually useful part.

Quality is important to humans, because humans have to read it, but correctness is what people using ChatGPT for code actually need. So long as the quality and performance is good enough, then it will be useful.

Performance is such a nuanced topic that you need very context aware devs anyway, and I think a general purpose LLM is never going to have that kind of awareness.

Re: How is ChatGPT's behavior changing over time?

#29

Post by JimmyRuska »

The pre-prompting length will grow more and more as more liabilities in the responses are uncovered. I would imagine the more the pre-prompting grows the more attention is diverted to the rules rather than the user prompt, and the less reasoning available.

I wonder if they'll start using LLMs while ingesting new data. eg asking the LLM if the content is helpful, cites sources, respectful, positive, not-thin content, common or often duplicated content, etc etc, before each content import.

Post reply on HN