Come check out o1-mini: SoTA math reasoning in a small package with @ren_hongyu @shengjia_zhao @Eric_Wallace_ & the rest of the OpenAI team

I dream of having a rich set of economically valuable RL tasks to train on, as wide and beautiful as the internet. Does this pipeline for task creation look like. robotics? trading? enterprise metrics? research? coding? recommendation? video games?. I think it's the absolute

The internet sets up a natural curriculum of skills -- people gradually add new ideas on top of the old ones -- ensuring the models have a smooth difficulty ramp to learn from that covers the skill space. Curriculum will be important for RL -- you need to learn sub-skills before

The internet is incredibly diverse, and it is sourced from data on topics which. humans actually cared about to engage with in the first place. There are low-resource languages and niche fanbases that will be forever immortalized in AGI because someone cared enough to document.

Ultimately, we want AGI that benefits and interacts with humans, not just something that lives in a toy cage (like AlphaZero for chess, or reasoning models in math). In contrast to other researchers, I think it is therefore imperative to work on product. Are researchers who say

I really like this diagram from @_jasonwei and @hwchung27 about how to view the bitter lesson:. It's a mistake not to add structure now, it's a mistake to not remove that structure later. We're at the precipice of setting up a huge, powerful RL training run that will define the.

Why you should stop working on RL research and instead work on product //.The technology that unlocked the big scaling shift in AI is the internet, not transformers. I think it's well known that data is the most important thing in AI, and also that researchers choose not to work

one of the things we need for ai to mature is a more efficient capital allocation:. - what if i buy a large cluster but find no use for it?.- what if i sign a nine figure contract with a company that fails to deliver me a good model?.- what if I pay $100M for a researcher who.

ps. one of the reasons i joined openai was because i thought there were a lot of cool usecases (like video games) that required cheap + fast intelligence, so I thought I should join and release it 🙃.

ok enough tweeting from me, but if you are interested in also some thoughts about how we can think of inference in ways that is not simply "make the cot longer", i am advertising my blog 🙂.

more of a fun meme, but i think most people are also talking about inference compute, but not enough about inference time. generally in life, we pay extra for things to be fast. but right now the small-model-tier (flash, mini, nano, haiku) are both cheap and fast. it's easy to

So I think something else that doesn't get discussed much is the extrapolation of this inference : training trend. - 2015: back in the day, we would train one model per dataset, and inference it once (to obtain the eval result for our paper).- 2020: with chatgpt, multi-task

By the way, I think some of my old blogs are still relevant too --. There are a lot of threads around inference time compute, small models, etc. these days; if you are interested in my thoughts (and the human in the loop), you could take a read :).

A somewhat little known fact about me is that I have a blog 😀. Over the weekend I got around to writing up some of my thoughts on the recent LLM-Pokemon craze, and why I think video games are more interesting than most (maybe older) AI researchers think. - Why is Pokemon hard,

RT @karpathy: The race for LLM "cognitive core" - a few billion param model that maximally sacrifices encyclopedic knowledge for capability….

RT @jilin_14: Exciting to share what i've been working on in the past few months! o3 and o4-mini are our first reasoning models with full t….

@johnohallman @SuvanshSanjeev

RT @SuvanshSanjeev: we trained a cute lil model!.- will solve a few AIME problems a year if you ask nicely.- cheap image input, 1 million t….

Today we released GPT-4.1 nano, an amazing effort led by @johnohallman and @SuvanshSanjeev!. Some cool features of today's release:.- Faster & cheaper than 4o-mini.- Significantly cheaper for image processing.- Better reasoning across the board.- 1M input context

We released o3-mini, available today to all users in ChatGPT (for free)!. o3-mini-low is faster (and often better) than o1-mini, and o3-mini-high is the most capable publicly available reasoning model in the world. with @ren_hongyu @shengjia_zhao

We trained o3-mini: both more capable than o1-mini, and around 4x faster end-to-end when accounting for reasoning tokens. with @ren_hongyu @shengjia_zhao & others