PI at Ghent University and VIB.AI. Manifold learning, contrastive learning, self-supervised learning, genomics, transcriptomics, interpretability. Excess mortality, statistical forensics. Born but to die and reas'ning but to err.
https://dkobak.github.io
Dmitry Kobak
"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."
We 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...
Last 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
OpenAI'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-...
Meet 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
one 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.
Top-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.
This 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.
I 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!
(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...
We 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
Video
An 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.
We extracted (parts of) 12 books in experiments with 4 frontier-lab, production LLMs.
We prompted the LLMs with a short prefix of a book and asked them to complete the rest. For Harry Potter and the Sorcerer’s Stone, we extracted 95.8% of the book from jailbroken Claude 3.7 Sonnet.
