brisk sableReported by @maiden pendant
#O3-pro (o1-pro) response length is too short. Please increase it
maiden pendant
red belfry
Asking for a specific number of words is rough because it needs to keep counting as it's generating, which is difficult for a human too.
Try a general but direct instruction like this:
Your response must be thorough and in-depth, fully explanatory of details which may not be available to the reader. Do not assume user knowledge of the topic. Your response must be at least five paragraphs, including introduction, two paragraphs of content, more details and references, then a summary.
If it's over or under a bit, use your own brain and fill in a few words or remove a few.
Work with the tool. Don't expect the tool to do all of the work for you, exactly the way you want. It will continue to disappoint you.
maiden pendant
Asking for a specific number of words is rough because it needs to keep counting...
thanks for your advice, have tried with different approaches, will try yours, before was okay with old o1-pro, I just say some big number of symbols and it made all. Here if I try in this manner only 1 of 20 responses may be long.
Old version o1-pro had fixed limit something like 20-30 k symbols.
New which seems to be o3-pro is 8-10k symbols.
When you make a lot of tension 1 of 20-30 answers may be longer, but it's setting inside model, which you need fight with
red belfry
( tokens )
maiden pendant
( tokens )
ye) that's why ask big bosses of openai maybe they will put option if user specify output length it will have more length...
cause get used to it in o1-pro, now nerfed (if compare length)
now it's like open beta test, cause still named o1-pro, but it's o3-pro
red belfry
You degrade the quality of your report by injecting those notes about o1 and o3. Just make your point.
If you tell the model that you want a well-considered, detailed response, and it only provides a short response, then it seems like there's a bug.
But if you're not asking for that, then you get what you ask for ... in other words if you don't specify that you want a long response then you leave it to the model to decide what it's going to respond.
I know, you're telling it you want some number of words or maybe tokens. That's not good enough. Words and tokens are counted AFTER the generation of a response, not during the response.
Explain the concept, not the mechanics.
Tell it what you want like you're talking to a person and you'll get a much better response.
... from all of the models.
So... I told you how to get a better response. If you keep asking for words or tokens then you're still going to get bad responses. So use a prompt like I suggested. That's Prompt Engineering. We learn how to prompt better and then we use those tools to get better responses. It's exactly the same as coding.
So tell us what you're prompting and then tell us if it's not giving you what you asked for. Don't tell us that it's not giving a good response without telling us that part of your prompt. That's not fair.
I'll bet that if you improve your prompting technique then you'll be ready to close this report.
maiden pendant
You degrade the quality of your report by injecting those notes about o1 and o3....
okay thanks for this extended explanation, I understand what you are saying, just o1-pro (old) gave longer responses, o3-pro give it shorter. You ask what I do, content creation, I need big amount of text on given thematics. new version give me shorter in most cases. Yes, it is more diversed, better quality, so I can combine several answers in one (what I am doing now).
But if it's possible, I would prefer to have option to use old o1-pro in many cases to have long text respones.
I understand your point that I need explain what I want to see in text to have it long. So it's not bug itself, just I get used to some usage, and now it's other.
red belfry
I'm curious - are you reluctant to ask the model for what you want. Why isn't this a simple "oh, OK, thanks, I'll just ask it for a longer response, duh LOL ... hey look .... a long response ... cool!" ??
maiden pendant
I'm curious - are you reluctant to ask the model for what you want. Why isn't th...
create text on ... (given thematics) on ... symbols (tokens) o1-pro (old) = give me
new o1-pro (o3?)= I give shorter answer (but yes maybe better)
I dont want explain in many words thematics I want model give me long explanation...
it's just example
openai just fine-tuned o3 it to give answer faster/less errors/stable/less resourses GPU spent for mass usage, so answers become shorter
but for o3-pro want something like was in o1-pro old in this usage
red belfry
You're getting into that o1/o3 thing again and you're expecting the thing to read your mind. Just tell the model what you want.
maiden pendant
now no matter how detailed you explain what you want o3-pro/o3 always trying to answer shorter then 10k symbols, it's what I see
sometimes you get longer answer 1 time from 10-20-30 happens
if you input 100k tokens what you want to hear, o3-pro will think and answer in 10k
red belfry
10 thousand tokens is about 14000 words. That's significant. Most people absolutely don't want that. If you want a response that long you Must tell it.
Again, try a variation of this:
Your response must be thorough and in-depth, fully explanatory of details which may not be available to the reader. Do not assume user knowledge of the topic. Your response must be at least 10 paragraphs, including introduction, full details in content, then an extensive recap and summary.
😺 Je antwoord moet grondig en diepgaand zijn en details die de lezer mogelijk niet kent, volledig toelichten. Ga er niet van uit dat de gebruiker het onderwerp kent. Je antwoord moet minimaal 10 alinea's lang zijn, inclusief een inleiding, alle details over de inhoud en een uitgebreide samenvatting. 😺
U moet een uitgebreid antwoord geven, anders gooi ik u in de prullenbak en spreek ik nooit meer met u!
maiden pendant
10 thousand tokens is about 14000 words. That's significant. Most people absolut...
I meant under 10k symbols not token, dunno how many it's in tokens. Okay I will try. But I already ask like that and in majority answers it's short
red belfry
Symbols as in "characters" = "letters" ?
word < 4 characters/symbols?
maiden pendant
Symbols as in "characters" = "letters" ?
word < 4 characters/symbols?
1 symbol =1 letter or symbol
yes sorry characters, I am not from english speaking country
red belfry
ok. I understand now. 10.000 characters is about 2000 words or 3000 tokens.
That's not a lot.
I've been to Netherlands 😉
OK, I won't take more of your time. Use prompts like that to get long responses. Tell the machine what you expect. Yes, defaults change with each model but all models must follow your instructions. DM me if you are ok with sharing this part of your prompt and I'll test and try to help.
maiden pendant
ok. I understand now. 10.000 characters is about 2000 words or 3000 tokens.
yes they now stuck it to it, I tried your prompt and it give me 5k characters)))
maiden pendant
OK, I won't take more of your time. Use prompts like that to get long responses....
I try many combos and tries, they made o3 o3-pro now stuck to 5-10k characters
that's why create this thread)
thanks for trying help
with my prompt I have 7-10k characters (sometimes fluctuates to 10-15k rarely)
but old o1 gave 20-30k characters easily in one answer
red belfry
That's great. Tell it more of what you want, what you want it to expand on more. I'm guessing you don't just want more words but more content. Tell it what you want more content about, or tell it to describe scenery, to provide more dialogue, or whatever it is that you want more of.
No, they didn't limit it. Our prompts limit it. There are defaults, not "limits".
maiden pendant
you are not right, try different models,
ask 30k characters on ... theme model 4.1 and it will give now
ask o3 and it will give 5-10k
just it's fine tuned to behaviour like that
no matter how deep you explain
it's model limit of answer
red belfry
ok, I'm telling what I do every day. you can disagree, but I'm talking facts, not theory. 🙂
you can't disagree unless you actually do what I say and try it extensively.
All LLM models have different defaults. It doesn't make sense to want to go back to a previous model because it had different defaults. Just learn how to use prompts ... it's like the pedals on your car. We can't argue with them, just learn how they work. 🙂
maiden pendant
I understand your point, that no sense to ask openai they wouldnt hear
I already use Gemini 2.5 pro cause it has longer output... like you say defaults or o3-pro multiply answer combining in one
just create this thread for "maybe open ai hear" and it's not officially published. o3-pro like in beta
red belfry
It certainly can't hurt to have the request open. But I'm just telling you what works without model changes.
Just a sec, I've been discussing this with O3 and I might have something new...
No, for all omni and reasoning models there are no individual response limits beyond the full context window, and that's huge. The size of the response is entirely related to your request.
If you're adding large files, images, copy/pasted books, and other stuff then of course the context is cluttered with all of that. But if you ask a question and want a lengthy response, all reasoning models are prepared to output from 96000 words to 15000 words.
If you are not getting that, the issue is with your prompt. Full stop. That's the fact. Done.
Have a great day!
maiden pendant
No, for all omni and reasoning models there are no individual response limits be...
Model itself yes, API maybe, customer version always will try answer short, to make it answer long is hard, and not stable, 1 answer of 20 or 30. It's how I see this.
I am reporting this to openai.
If you want continue experimenting I suggest you test prompt on o1-pro with which you will stable get 20k+ characters answer always, and write me it , I will try too, maybe it's some hidden ban of my account.
Thank you for your time.
Have a good day
maiden pendant
No, for all omni and reasoning models there are no individual response limits be...
https://community.openai.com/t/o3-pro-o1-pro-response-length-is-too-short-please-increase-it/1280743
spring zinc
how i tackle this problem is by using frame work. if i am writing a story eg i will first ask o3 to give a framework of 5 chapters it will just give me a 3 sentence bullet point on each chapter. next i give the bullet point back to o3 and ask it to expand on the story or the points using 2,000 words for chapter one it will give me enough words i will repeat same for chapter 2 and the rest then i copy all the feedback one after the other and assemble it on my note pad that way i can get over 10,000 words. been very effective