Dmitry Kobak @hippopedoid.bsky.social · 01/10/2026Cool. I wonder if it can get the right answer without thinking mode disabled? 130
Dmitry Kobak @hippopedoid.bsky.social · 30/09/2026Nice. Would the results change much if you ask it the same questions but in numeric form, not in words? Like, "417+529=" instead of "Four hundred seventeen ..." 001
Dmitry Kobak @hippopedoid.bsky.social · 15/09/2026FYI here's one of the top ML conferences: "We ask that authors explicitly state how they used LLMs in their submission [...]. To guide authors about the nature and extent of reporting, we provide a list of common research subtasks along with the necessity of disclosing them." iclr.cc/Conferences/...iclr.ccICLR 2027 AI Policy for Authors 110
Reposted by Dmitry KobakGreg Egan @gregegansf.bsky.social · 27/07/2026While the Fields Medal celebrates the under-40s, the mathematician Joan Birman has, at the age of 99, solved a major open problem in representations of the Braid groups, a topic she has worked on for more than 60 years. 121299320
Dmitry Kobak @hippopedoid.bsky.social · 22/07/2026Maybe interesting to look at Microsoft's report on MAI-Thinking (June 2026): microsoft.ai/pdf/mai-thin.... They claim math perf. btw Sonnet & Opus 4.6. The math training (Sec 3.2) seems to be pure RL, based on problems scraped from the literature and verified using SymPy and an (unclear) "AI judge". 010
Dmitry Kobak @hippopedoid.bsky.social · 22/07/2026Wow, I see, thanks. If it's not based on Lean, then I have no idea how the training can proceed automatically and at scale, as the system needs to judge the correctness somehow... And if it's all trade secrets, then I don't understand how the main competitors can stay so close to each other. 240
Dmitry Kobak @hippopedoid.bsky.social · 22/07/2026Hi Daniel. Thanks for your writing/conversations on this topic, I have been following with interest. Question: have you seen a good write-up on how frontier models are nowadays trained for math? What is the role of pre-training? Of RL? Of Lean? Of data? Can one find out or is it all trade secrets? 161
Reposted by Dmitry KobakVIB.AI @vibai.bsky.social · 30/06/2026The @steinaerts.bsky.social lab received an ERC Proof of Concept Grant to develop CellTuned: an AI platform that combines sequence-to-function models with single-cell regulatory networks to uncover the mechanisms driving disease, and translate them into new therapeutic opportunities. 👏 celltuned.ai 0124
Reposted by Dmitry KobakDr Michele V @micheleveldsman.bsky.social · 25/06/2026“I have done all of the things that one is supposed to do to earn a tenure-track position. And I have done approximately 85% of them by typing prompts into a large language model and then moderately editing the output.” 🤣🤣🤣 brilliant 👌🏽 open.substack.com/pub/inprepar...open.substack.comOpinion: I Was Not Allowed To Type Prompts Into ChatGPT During My Chalk Talk And This Is DiscriminationBy Dr. Rachel Simmons, Postdoctoral Fellow, Stanford University 26506166
Dmitry Kobak @hippopedoid.bsky.social · 22/06/2026Hey! Yes, that's me. Thanks a lot for doing the video, for your nice words, and for pinging me! The core idea of the paper is all by Sam Greydanus, he had v1 of the preprint long before I joined. I read it, loved it, contacted Sam, and we ended up reworking/resubmitting it together. 100
Reposted by Dmitry KobakVincent D. Warmerdam @koaning.bsky.social · 22/06/2026@hippopedoid.bsky.social I *think* I'm speaking to the correct Dmitry when I say this, but I really liked your paper! So we made a highlight of it for our YT channel. youtu.be/2eoOqqLvkl8youtu.beMNIST is Better in One DimensionYouTube video by marimo 121
Dmitry Kobak @hippopedoid.bsky.social · 16/06/2026A slide I prepared for my introductory departmental seminar last week to illustrate my CV as a weighted graph. Node weights (shown as marker areas) are the number of years I lived in each place: 24, 3, 2, 4, 9, 0. Edge weights (not shown) are the number of people moving each time: 1, 2, 4, 5, 5. 060
Dmitry Kobak @hippopedoid.bsky.social · 15/06/2026Many good points, thanks for writing. A question about Section 3 where you show that Claude judge prefers Claude solutions and Codex judge prefers Codex solutions. You have a screenshot where judges say things like "Codex is preferred". So I assume judges were not blinded. Did you try blinding them? 000
Reposted by Dmitry KobakSasha Gusev @sashagusevposts.bsky.social · 23/05/2026I assigned random gender/ethnicity labels to scientific abstracts from the literature and then asked Claude to do a thematic analysis. Claude identified a clinical versus computational split for female/male authors and a DEI focus for Black/URM authors. All in completely random data. 26865340
Reposted by Dmitry KobakJan Lause @janlause.bsky.social · 09/06/2026Meet Eyewire II: a new connectomic resource for the mouse retina. ~1 mm² of retina at nanometer resolution, with synapses and circuits of ca. 100,000 neurons, plus visual responses from the ~400 neurons shown in the video! Preprint: doi.org/10.64898/202... Data: eyewire.ai 🧠📈 🧠💻🧪 #VisionScience 25221
Reposted by Dmitry KobakJohn C. Baez @johncarlosbaez.bsky.social · 09/06/2026"An Erdős problem resolved by humans! One Abel prize winner was quoted as saying 'We knew the day would eventually come when humans could resolve Erdős problems, but we didn't know it would come this soon!' Several math departments now have plans for workshops on Human Alignment."blog.computationalcomplexity.orgHumans Solve Erdos Problem!!(In 2008 I wrote a survey of some of the known sum-product theorems, see here . Avi Wigderson has a great slide-set on sum-product theorems... 1326
Dmitry Kobak @hippopedoid.bsky.social · 09/06/2026It's a pity that it did not really pick up. Are you still active on Twitter? Or on Mastodon? 010
Dmitry Kobak @hippopedoid.bsky.social · 22/05/2026Top-3 questions this month on MathOverflow are all about the math performance of frontier reasoning LLMs and what it implies for mathematical training and careers. Very interesting discussions with some very polarizing answers. 011
Reposted by Dmitry KobakTimothy Gowers @wtgowers.bsky.social · 20/05/2026OpenAI's claim that this is a central conjecture in discrete geometry is not an exaggeration. This will I think be looked back on as the first time that AI solved a major mathematics problem (defined as a problem that all experts in some subfield had thought about). openai.com/index/model-...openai.comAn OpenAI model has disproved a central conjecture in discrete geometryAn OpenAI model solved the 80-year-old unit distance problem, disproving a major conjecture in discrete geometry and marking a milestone in AI-driven mathematics. 17650190
Dmitry Kobak @hippopedoid.bsky.social · 18/05/2026This is great! What is the frontend here? Did you write it yourself (user interface, the way individual points appear on zoom-in, etc), vibe-coded it, or did you use some existing solutions for that? Also, all 3M points are loaded into browser on page load, right? 000
Dmitry Kobak @hippopedoid.bsky.social · 18/05/2026This is a very cool work! Only saw it now. They could extract almost the entire (!) The Great Gastsby and 1984 from Claude (with some jailbreaking). But for some reason not Catch-22. I wonder why that is and what it tells us about the training data (or about Catch-22?). Catch-22 is great btw. 011
Dmitry Kobak @hippopedoid.bsky.social · 15/05/2026I am in the process of moving from Tübingen to Ghent, where I joined UGent and @vibai.bsky.social. Am really looking forward to working with wonderful VIB.AI colleagues @steinaerts.bsky.social, @joanampereira.bsky.social, @ppjgoncalves.bsky.social, @wsaelens.bsky.social. The lab is hiring! 1192
Reposted by Dmitry KobakVIB.AI @vibai.bsky.social · 18/03/2026We are super excited to welcome Dmitry Kobak @hippopedoid.bsky.social as a new VIB.AI group leader! The Kobak lab will officially kick off in April at the VIB.AI Ghent hub. More: vib.ai/en/news/dmit... 0175
Dmitry Kobak @hippopedoid.bsky.social · 27/08/2025We spent a year writing this review of low-dim embeddings and arguing about things like epistemic roles and best practices :-) 20+ authors are all participants of the Dagstuhl seminar we held last year: www.dagstuhl.de/24122. Led by @alexandr.bsky.social and Cyril de Bodt. arxiv.org/abs/2508.15929 1279
Reposted by Dmitry KobakAlex Diaz-Papkovich @alexandr.bsky.social · 27/08/2025one of the more fun topics stemmed from a discussion with @hippopedoid.bsky.social over what the oldest 2D PCA visualization we could find was. After some scouring, we settled on a 1960 paper from researchers at the Université de Montréal about turtle carapaces, which we recreated. 172
Reposted by Dmitry KobakAlex Diaz-Papkovich @alexandr.bsky.social · 27/08/2025Last year I met a bunch of great researchers who work with high-dimensional data at a Dagstuhl seminar. This week we put out a preprint about the history and philosophy of low-dimensional embedding methods, their applications, their challenges, and their possible future arxiv.org/abs/2508.15929 1157
Dmitry Kobak @hippopedoid.bsky.social · 24/04/2025Hi Ben, I've been trying to contact you by email but keep failing. Wanted to DM you here but your DMs are closed. If you either follow me or open your DMs, I'll message you! Cheers. 000