LLMs are still far from being able to do most human work, but the pace of progress impresses me: in ~64 weeks, they've gone from 12% to 48% on GDPval. (also impressed that OpenAI keeps publishing useful papers & data, against its profit incentives) https://t.co/9Pzn26div1

1/n I’m really excited to share that our @OpenAI reasoning system got a perfect score of 12/12 during the 2025 ICPC World Finals, the premier collegiate programming competition where top university teams from around the world solve complex algorithmic problems. This would have

"super curious to see how quickly it goes from a curiosity to something useful." one of my favorite things about writing is how it helps you travel through time, by reminding yourself what you thought in the past. I don't remember being impressed by GPT-3, but apparently I was!

"obviously it's dumb and makes mistakes and apparently requires a fair bit of tuning to get good results, but when it does get those good results, it seems magical. I'd really like to understand this magic better. (3/4)

"with the first few examples, I brushed it off with my usual pessimism as probably an overfit model regurgitating text from its training data, but having seen so many now, it now seems like this model has really captured some of the essence of language. (2/4)

found an email I wrote to friends in 2020 about GPT-3: "the GPT-3 examples are wild! they have substantially lifted my feelings of what machine learning is capable of in the next decade. more impressive than everything else OpenAI has done put together. (1/4)

As per tradition: A thread with all positive results I created or shared about GPT-5 It's not my fault negative results take-off more than positive ones I can't make a thread with more than 25 posts, but here you go:

today we are significantly increasing rate limits for reasoning for chatgpt plus users, and all model-class limits will shortly be higher than they were before gpt-5. we will also shortly make a UI change to indicate which model is working.

GPT-5 (medium reasoning) is the new leader on the Short Story Creative Writing benchmark! GPT-5 mini (medium reasoning) is much better than o4-mini (medium reasoning). Claude Opus 4.1 shows gains over Opus 4.

Full results: GPT-5 maintains strong performance. GPT-5-mini notably competitive with o3 and gemini-2.5-pro. Absolute accuracy numbers depend on instruction and task complexity and will vary across settings—key takeaway is relative model rankings and degradation patterns

GPT-5 represents a huge step up over previous OpenAI models, such as GPT-4.1. We believe GPT-5 is at the frontier of agentic ability and shines on tasks that require complex code understanding. On our junior SWE evals, GPT-5 is particularly strong at code exploration and

gpt-5 is now free in @cursor_ai, go try it out! (reload cursor if you don't see it yet) we've worked closely with @OpenAI team to make this happen, and together we also put out a prompting guide for gpt-5. here are some examples prompts we've seen working well

^reminds me of the saying that a film is born three times: first in writing, then filming, and lastly in editing. AI product experiences are also born three times: first in model training, then in product, and lastly in prompting. the right prompt can cure a lot of flaws!

GPT-5 (medium reasoning) sets a new record on the Confabulations/Hallucinations on Provided Texts benchmark!

a cool thing you get to see building AI products: you can plug a new model into a product and see it kinda suck. but then you do a week of prompt & product optimization, and the same model shines. the user experience is not just the model; it's the model + product + prompt.

GPT-5 is here! it's way better at coding - not just in pointless evals, but real usage. using it in @cursor_ai, I think my favorite leap is code Q&A - it's been truly useful at helping me figure out our complex RL codebase. very tenacious digger! https://t.co/TRwscI3Ffd

GPT-5 is out! it's by no means perfect, but it's better than what's come before. if you have complaints about its coding, hit me up and I'll see if we can make future models even better for you

GPT-5 is here - and it’s #1 across the board. 🥇#1 in Text, WebDev, and Vision Arena 🥇#1 in Hard Prompts, Coding, Math, Creativity, Long Queries, and more Tested under the codename “summit”, GPT-5 now holds the highest Arena score to date. Huge congrats to @OpenAI on this

AGI is hard to define. my preferred definition of AGI is a computer system that can can accomplish a task impossible for 100 human geniuses working together, such as publishing a blog post with a single canonical spelling of GPT-4o / gpt-4o / gpt4o

When these electrical connections are connected to a load, then electrons can get through and accompany the lithium ions flowing through the electrolyte/separator. And these electrons do work, powering your lightbulb or computer or gameboy or whatever. 💡