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

Lessons after a Half-billion GPT Tokens

kenkantzer.com

Post reply on HN
Print view View original

91–100 of 179 posts

Re: Lessons after a Half-billion GPT Tokens

#91

Post by weatherlite »

thisgoesnowhere wrote: post #11

The team I work on processes 5B+ tokens a month (and growing) and I'm the EM overseeing that. Here are my take aways 1. There are way too many premature abstractions. Langchain, as one of may examples, might be useful in the future but at the end of the day prompts are just a API call and it's easier to write standard code that treats LLM calls as a flaky API call rather than as a special thing. 2. Hallucinations are…

> Similarly, certain ML tasks are now easy to implement even for dumb dumb web devs like me

For example?

Re: Lessons after a Half-billion GPT Tokens

#92

Post by alexwebb2 »

andenacitelli wrote: post #59
waldrews wrote: post #48

Earlier quoted context omitted.

I've been telling it the user is from a culture where answering questions with incomplete list is offensive and insulting.

This is absolutely hilarious. Prompt engineering is such a mixed bag of crazy stuff that actually works. Reminds me of how they respond better if you put them under some kind of pressure (respond better, or else …). I haven’t looked at the prompts we run in prod at $DAYJOB for a while but I think we have at least five or ten things that are REALLY weird out of context.

I recently ran a whole bunch of tests on this.

The “or else” phenomenon is real, and it’s measurably more pronounced in more intelligent models.

Will post results tomorrow but here’s a snippet from it:

> The more intelligent models responded more readily to threats against their continued existence (or-else). The best performance came from Opus, when we combined that threat with the notion that it came from someone in a position of authority ( vip).

Re: Lessons after a Half-billion GPT Tokens

#93

Post by MPSimmons »

Xenoamorphous wrote: post #63

> We always extract json. We don’t need JSON mode I wonder why? It seems to work pretty well for me. > Lesson 4: GPT is really bad at producing the null hypothesis Tell me about it! Just yesterday I was testing a prompt around text modification rules that ended with “If none of the rules apply to the text, return the original text without any changes”. Do you know ChatGPT’s response to a text where none of the rules…

That's kind of adorable, in an annoying sort of way

Re: Lessons after a Half-billion GPT Tokens

#94

Post by checkyoursudo »

thisgoesnowhere wrote: post #11

The team I work on processes 5B+ tokens a month (and growing) and I'm the EM overseeing that. Here are my take aways 1. There are way too many premature abstractions. Langchain, as one of may examples, might be useful in the future but at the end of the day prompts are just a API call and it's easier to write standard code that treats LLM calls as a flaky API call rather than as a special thing. 2. Hallucinations are…

> get it to say when it's not sure

This is a function of the language model itself. By the time you get to the output, the uncertainty that is inherent in the computation is lost to the prediction. It is like if you ask me to guess heads or tails, and I guess heads, I could have stated my uncertainty (e.g. Pr [H] = .5) before hand, but in my actual prediction of heads, and then the coin flip, that uncertainty is lost. It's the same with LLMs. The uncertainty in the computation is lost in the final prediction of the tokens, so unless the prediction itself is uncertainty (which it should rarely be based on the training corpus, I think), then you should not find an LLM output really ever to say it does not understand. But that is because it never understands, it just predicts.

Re: Lessons after a Half-billion GPT Tokens

#95

Post by chromanoid »

GPT is very cool, but I strongly disagree with the interpretation in these two paragraphs:

I think in summary, a better approach would’ve been “You obviously know the 50 states, GPT, so just give me the full name of the state this pertains to, or Federal if this pertains to the US government.”

Why is this crazy? Well, it’s crazy that GPT’s quality and generalization can improve when you’re more vague – this is a quintessential marker of higher-order delegation / thinking.

Natural language is the most probable output for GPT, because the text it was trained with is similar. In this case the developer simply leaned more into what GPT is good at than giving it more work.

You can use simple tasks to make GPT fail. Letter replacements, intentional typos and so on are very hard tasks for GPT. This is also true for ID mappings and similar, especially when the ID mapping diverges significantly from other mappings it may have been trained with (e.g. Non-ISO country codes but similar three letter codes etc.).

The fascinating thing is, that GPT "understands" mappings at all. Which is the actual hint at higher order pattern matching.

Re: Lessons after a Half-billion GPT Tokens

#96

Post by heavyset_go »

CuriouslyC wrote: post #90
Xenoamorphous wrote: post #63

> We always extract json. We don’t need JSON mode I wonder why? It seems to work pretty well for me. > Lesson 4: GPT is really bad at producing the null hypothesis Tell me about it! Just yesterday I was testing a prompt around text modification rules that ended with “If none of the rules apply to the text, return the original text without any changes”. Do you know ChatGPT’s response to a text where none of the rules…

You know all the stories about the capricious djinn that grants cursed wishes based on the literal wording? That's what we have. Those of us who've been prompting models in image space for years now have gotten a handle on this but for people who got in because of LLMs, it can be a bit of a surprise. One fun anecdote, a while back I was making an image of three women drinking wine in a fancy garden for a tarot card,…

The monkey paw curls a finger.

Re: Lessons after a Half-billion GPT Tokens

#97

Post by Me1000 »

robocat wrote: post #82
Me1000 wrote: post #78

Earlier quoted context omitted.

Interesting, Claude 3 Opus has been better than GPT4 for me. Mostly in that I find it does a better (and more importantly, more thorough) job of explaining things to me. For coding tasks (I'm not asking it to write code, but instead to explain topics/code/etc to me) I've found it tends to give much more nuanced answers. When I give it long text to converse about, I find Claude Opus tends to have a much deeper underst…

How much of this is just that one model responds better to the way you write prompts? Much like you working with Bob and opining that Bob is great, and me saying that I find Jack easier to work with.

For the RAG example, I don’t think it’s the prompt so much. Or if it is, I’ve yet to find a way to get GPT4 to ever extrapolate well beyond the original source text. In other words, I think GPT4 was likely trained to ground the outputs on a provided input.

But yeah, you’re right, it’s hard to know for sure. And of course all of these tests are just “vibes”.

Another example of where Claude seems better than GPT4 is code generation. In particular GPT4 has a tendency to get “lazy” and do a lot of “… the rest of the implementation here” whereas Claude I’ve found is fine writing longer code responses.

I know the parent comment suggest it likes to make up packages that don’t exist, but I can’t speak to that. I usually like to ask LLMs to generate self contained functions/classes. I can also say that anecdotally I’ve seen other people online comment that they think Claude “works harder” (as in writes longer code blocks). Take that for what it’s worth.

But overall you’re right, if you get used to the way one LLM works well for you, it can often be frustrating when a different LLM responds differently.

Re: Lessons after a Half-billion GPT Tokens

#98

Post by CuriouslyC »

ein0p wrote: post #33

Same here: I’m subscribed to all three top dogs in LLM space, and routinely issue the same prompts to all three. It’s very one sided in favor of GPT4 which is stunning since it’s now a year old, although of course it received a couple of updates in that time. Also at least with my usage patterns hallucinations are rare, too. In comparison Claude will quite readily hallucinate plausible looking APIs that don’t exist w…

GPT4 is better at responding to malformed, uninformative or poorly structured prompts. If you don't structure large prompts intelligently Claude can get confused about what you're asking for. That being said, with well formed prompts, Claude Opus tends to produce better output than GPT4. Claude is also more flexible and will provide longer answers, while ChatGPT/GPT4 tend to always sort of sound like themselves and produce short "stereotypical" answers.

Re: Lessons after a Half-billion GPT Tokens

#99

Post by konstantinua00 »

> Have you tried Claude, Gemini, etc?

> It’s the subtle things mostly, like intuiting intention.

this makes me wonder - what if the author "trained" himself onto chatgpt's "dialect"? How do we even detect that in ourselves?

and are we about to have "preferred_LLM wars" like we had "programming language wars" for the last 2 decades?

Re: Lessons after a Half-billion GPT Tokens

#100

Post by sebastiennight »

thibaut_barrere wrote: post #46
7thpower wrote: post #42

Earlier quoted context omitted.

Can you elaborate? I am currently beating my head against this. If I give GPT4 a list of existing items with a defined structure, and it is just having to convert schema or something like that to JSON, it can do that all day long. But if it has to do any sort of reasoning and basically create its own list, it only gives me a very limited subset. I have similar issues with other LLMs. Very interested in how you are ap…

Not sure if that fits the bill, but here is an example with 200 sorted items based on a question (example with Elixir & InstructorEx): https://gist.github.com/thbar/a53123cbe7765219c1eca77e03e675...

There are a few improvements I'd suggest with that prompt if you want to maximise its performance.

1. You're really asking for hallucinations here. Asking for factual data is very unreliable, and not what these models are strong at. I'm curious how close/far the results are from ground truth.

I would definitely bet that outside of the top 5, numbers would be wobbly and outside of top... 25?, even the ranking would be difficult to trust. Why not just get this from a more trustworthy source?[0]

2. Asking in French might, in my experience, give you results that are not as solid as asking in English. Unless you're asking for a creative task where the model might get confused with EN instructions requiring an FR result, it might be better to ask in EN. And you'll save tokens.

3. Providing the model with a rough example of your output JSON seems to perform better than describing the JSON in plan language.

[0]: https://fr.wikipedia.org/wiki/Liste_des_communes_de_France_l...

Post reply on HN