Pinned
Some work I did this summer on the IMO Grand Challenge:
Jesse Michael Han
681 posts
Jesse Michael Han
@jessemhan
@mathematics_inc //
prev. research @OpenAI
'turboautist'
'only guy i know who talks like an anime character'
- labor is priced like capital when it is the final bottleneck to the thing that will permanently substitute capital for labor
- GRPO was the workhorse of DeepSeek R1's training algorithm. To go beyond the Cartesian brain-in-a-box setup and to obtain the Monte-Carlo advantage estimates that GRPO requires, you will need to perform massive parallel branching rollouts against complex environments (an entire
- New paper, new SOTA in unsupervised neural machine translation, joint with colleagues at @OpenAI: Unsupervised Neural Machine Translation with Generative Language Models Only arxiv.org/abs/2110.05448
- i remember once, i was showing ilya the results of weeks' worth of experiments with the conclusion that what we wanted to work wasn't going to work. his instantaneous response was: "if you torture the data long enough, it will confess" by which he meant that the data would
- Replying to @jessemhangolden gate claude is the eighth wonder of the world
- the Mathstral MMLU breakdown is the funniest thing i've seen all week - mathematical ability negatively correlated with accounting, foreign policy, human sexuality
00:00- Offer from @saranormous @w_conviction's Embed rescinded after 3 days because I wanted to do due diligence on a 1 year old firm launching its first accelerator program. This is what we call "bad behavior" Know who you're letting on your cap table.
- Replying to @delipraoreminds me of "This book arose out of the author’s teaching a summer course to a bunch of Viennese 4th graders while on a long trans-Atalantic flight. A solid grasp of the multiplication tables should suffice for chapters 1 - 4, although experience with projective modules over
- there has been a new SOTA on swebench lite every two days for the past eight daysExcited to share that our team has cooked up a multi-agent coding framework and setting the new State-of-the-Art on SWE-Bench-Lite with 40.3% accepted solutions! Very soon this framework will work right in your editor with the developer working along with agent(S)
- Proud to share this work over the past year on neural theorem proving in @leanprover, joint with @spolu, @ilyasut and the support of many others at @OpenAI!We trained a neural network that solved two problems from the International Math Olympiad. openai.com/blog/formal-ma…
- Excited to share the next chapter for Morph and honored to be working with such great investors - AI SWE agents will soon vastly outnumber human software engineers, and we will lead the way in building AI-centric infrastructure to support them. We hiring!
