Live data from Hacker News

Notes on OpenAI o3-mini

simonwillison.net

31–40 of 81 posts

Re: Notes on OpenAI o3-mini

#31
post #8
post #5

Earlier quoted context omitted.

I've been trying out o3 mini in Cursor today, it seems "smarter" but overall tends to overthink things and if it's not provided with perfect context it's prone to hallucinate. Overall I prefer Sonnet still. It has a certain magic of always making reasonable assumptions and finding simple solutions.

As n occasions user and fan of Cursor, it would be good if they could explain what the models are and why the different models exist. There’s no obvious answer of why one should switch to any of them!

I don’t think there’s an obvious answer. Try them out and see which works better for your use case.

Re: Notes on OpenAI o3-mini

#32

Earlier quoted context omitted.

You obviously don't have a chronic illness or you wouldn't be asking that. Either that or you're rich.

Are you conflating doctors with insurance like the larger health care industrial complex?

No one is forcing those doctors to participate in that system. Many (but not enough) of them don't, and operate cash clinics, which have a better reputation for quality.

But ultimately, healthcare is suffering not because of the insurance companies, but because of the American Medical Association, which worked to artificially limit the supply of doctors so that doctors could be paid more. This makes it more optimal for doctors to choose to participate in the insurance scheme as well.

So I ask again, why are we prioritizing the automation of software development, where neither the skill nor the profession itself is gatekept like healthcare is?

Re: Notes on OpenAI o3-mini

#33
At the end of his post, Simon mentions translation between human languages. While maybe not directly related to token limits, I just did a test in which both R1 and o3-mini got worse at translation in the latter half of a long text.

I ran the test on Perplexity Pro, which hosts DeepSeek R1 in the U.S. and which has just added o3-mini as well. The text was a speech I translated a month ago from Japanese to English, preceded by a long prompt specifying the speech’s purpose and audience and the sort of style I wanted. (I am a professional Japanese-English translator with nearly four decades of experience. I have been testing and using LLMs for translation since early 2023.)

An initial comparison of the output suggested that, while R1 didn’t seem bad, o3-mini produced a writing style closer to what I asked for in the prompt—smoother and more natural English.

But then I noticed that the output length was 5,855 characters for R1, 9,052 characters for o3-mini, and 11,021 characters for my own polished version. Comparing the three translations side-by-side with the original Japanese, I discovered that R1 had omitted entire paragraphs toward the end of the speech, and that o3-mini had switched to a strange abbreviated style (using slashes instead of “and” between noun phrases, for example) toward the end as well. The vanilla versions of ChatGPT, Claude, and Gemini that I ran the same prompt and text through a month ago had had none of those problems.

Re: Notes on OpenAI o3-mini

#34
So far ive been impressed.. seems to be in the same ballpark as r1 and claude for coding. I will have to gather more samples.. in this past week ive changed from using 100% claude exclusively (since 3.5) to hitting all the big boys: claude, r1, 4o (o3 now), and gemini flash. Then ill do a new chat that includes all of their generated solutions for additional context for a refactored final solution.

R1 has upped the ante so Im hoping we continue to get more updates rapidly... they are getting quite good

Re: Notes on OpenAI o3-mini

#35
post #29
post #16

Earlier quoted context omitted.

Are you implying it isn't? (evidence please, everyone)

Simple example: o3-mini-high gets this [1] right, whereas Gemini 2.0 Flash 01-21 gets it wrong. [1] https://chatgpt.com/share/679d9579-5bb8-8008-ac4a-38cef65b45...

Great example. Thank you. Can confirm that none of the Gemini models warned about the exception without prompting.

Re: Notes on OpenAI o3-mini

#36
post #33

At the end of his post, Simon mentions translation between human languages. While maybe not directly related to token limits, I just did a test in which both R1 and o3-mini got worse at translation in the latter half of a long text. I ran the test on Perplexity Pro, which hosts DeepSeek R1 in the U.S. and which has just added o3-mini as well. The text was a speech I translated a month ago from Japanese to English, pr…

This is a great anecdote and I hope others can learn from it. R1, o1, and o3-mini work best on problems that have a “correct” answer (as in code that passes unit tests, or math problems). If multiple professional translators are given the same document to translate, is there a single correct translation?

Re: Notes on OpenAI o3-mini

#37
post #33

At the end of his post, Simon mentions translation between human languages. While maybe not directly related to token limits, I just did a test in which both R1 and o3-mini got worse at translation in the latter half of a long text. I ran the test on Perplexity Pro, which hosts DeepSeek R1 in the U.S. and which has just added o3-mini as well. The text was a speech I translated a month ago from Japanese to English, pr…

This is a great anecdote and I hope others can learn from it. R1, o1, and o3-mini work best on problems that have a “correct” answer (as in code that passes unit tests, or math problems). If multiple professional translators are given the same document to translate, is there a single correct translation?

My wife is a professional translator and both revises others' work and gets revised. Based on numerous anecdotes from her, I can promise you that "single correct translation" does not exist.

Re: Notes on OpenAI o3-mini

#38

There's a huge price difference between o3-mini and o1 ($4.40 vs $60 per million output tokens), what trade-offs in performance would justify such a large price gap? Are there specific use cases where o1's higher cost is justified anymore?

its the same thing as:

gpt-3.5 -> gpt-4 (gpt-4-32k premium)

"omni" announced (multimodal fusion, initial promise of gpt-4o, but cost effectively distilled down with additional multimodal aspects)

gpt-4o-mini -> gpt-4o (multimodal, realtime)

gpt-4o + "reasoning" exposed via tools in ChatGPT (you can see it in export formats) -> "o" series

o1 -> o1 premium / o1-mini (equivalent of gpt-4 "god model" becoming basis for lots of other stuff)

o1-pro-mode, o1-premium, o1-mini, somewhere in that is the "o1-2024-12-17" model with not streaming, function calling, and structured outputs and vision

now, distilled o1-pro-mode probably is o3-mini and o3-mini-high-mode (the naming is becoming just as bad as android)

its the repeat, take model, scale it up, run evals, detect innefficiencies, retrain, scale, distill, see what's not working. when you find a good little zone in the efficiency frontier, release it with a cool name

Re: Notes on OpenAI o3-mini

#39

Earlier quoted context omitted.

You obviously don't have a chronic illness or you wouldn't be asking that. Either that or you're rich.

Are you conflating doctors with insurance like the larger health care industrial complex?

Insurance companies are part of the problem but not the foremost problem. Their profit margins are less than 10%, even if they operated as a charity, healthcare costs wouldn't go down much.

https://www.noahpinion.blog/p/insurance-companies-arent-the-...

Re: Notes on OpenAI o3-mini

#40
post #33

At the end of his post, Simon mentions translation between human languages. While maybe not directly related to token limits, I just did a test in which both R1 and o3-mini got worse at translation in the latter half of a long text. I ran the test on Perplexity Pro, which hosts DeepSeek R1 in the U.S. and which has just added o3-mini as well. The text was a speech I translated a month ago from Japanese to English, pr…

This is a great anecdote and I hope others can learn from it. R1, o1, and o3-mini work best on problems that have a “correct” answer (as in code that passes unit tests, or math problems). If multiple professional translators are given the same document to translate, is there a single correct translation?

No. People’s tastes and judgments vary too much.

One fundamental area of disagreement is how closely a translation should reflect the content and structure of the original text versus how smooth and natural it should sound in the target language. With languages like Japanese or Chinese translated into English, for example, the vocabulary, grammar, and rhetoric can be very different between the languages. A close literal translation will usually seem awkward or even strange in English. To make the English seem natural, often you have to depart from what the original text says.

Most translators will agree that where to aim on that spectrum should be based on the type of text and the reason for translating it, but they will still disagree about specific word choices. And there are genres for which there is no consensus at all about which approach is best. I have heard heated exchanges between literary scholars about whether or not translations of novels should reflect the original as closely as possible out of respect for the author and the author’s cultural context, even if that means the translation seems awkward and difficult to understand to a casual reader.

The ideal, of course, would be translations that are both accurate and natural, but it can be very hard to strike that balance. One way LLMs have been helping me is to suggest multiple rewordings of sentences and paragraphs. Many of their suggestions are no good, but often enough they include wordings that I recognize are better in both fidelity and naturalness compared to what I can come up with on my own.

Post reply on HN