After 8 years at Meta, I'm starting a new chapter at CEU.

TL;DR: After 8 years at Meta, I'm starting a new chapter as Professor of Practice at Central European University (CEU) in Vienna — teaching, building, and staying hands-on with AI. CEU is a rare institution: global in outlook, committed to academic freedom and critical thinking. This is the move.

linkedin 553 reactions · 119 comments Read on LinkedIn →

On Anthropic's "Does AI help or hurt learning?" paper.

This paper from Anthropic is so valuable in my role as Professor of Practice at CEU. My first reading: using AI while learning a new technical skill can measurably reduce actual understanding, even when it doesn't reliably reduce performance. That gap — between what you can do with the tool and what you actually know — is the whole pedagogical problem of this moment.

linkedin 266 reactions · 18 comments Read on LinkedIn →

When I got my first management role, I didn't know what the job actually was.

Nobody told me the job was running 1:1s, aligning roadmaps, writing growth plans, sitting in calibration meetings arguing over whether someone's work was 'senior-level' or just very strong mid-level. I figured it out eventually — by being a mediocre manager for longer than I'd like to admit, while good people depended on me to be better. So I taught a two-day intensive at CEU compressing that learning, and open-sourced the whole course.

linkedin 94 reactions · 13 comments Read on LinkedIn →

My last day teaching Data Science 4 at CEU.

Yesterday was the last day of my Data Science 4: Deep Learning and Topics in Applied AI course at Central European University. Teaching deeply technical, cutting-edge topics was an incredible experience — made possible by a group of students who were just genuinely game.

linkedin 73 reactions · 2 comments Read on LinkedIn →

Using AI agents to iteratively improve XGBoost models.

Szilard Pafka and I just published a piece on using AI agents to iteratively improve XGBoost models. Everyone knows Claude can write code — but it turns out it can research, try, evaluate, keep what works, discard what doesn't, and push model quality upward across repeated experiments.

linkedin 51 reactions · 9 comments Read on LinkedIn →

No magic — but first, a transformers PR.

I'm teaching an applied deep learning course at CEU this semester: six weeks, free Kaggle GPUs, and one cumulative problem — a long-tail 113-class classifier with a 2,666:1 class imbalance. Week 3 needed OLMo 2 sequence classification in HuggingFace transformers; the implementation didn't exist. So I wrote it. PR merged in 24h.

linkedin 36 reactions Read on LinkedIn →

Memory masquerades as reasoning.

I keep encountering a failure mode in LLMs that is subtler than hallucination. It's that memory masquerades as reasoning. I asked Claude to figure out if the sunset would be visible over the ocean from a specific beach in Thailand. Within minutes, it built a working tool with NOAA solar calculations — and confidently produced an answer that turned out to be wrong in an instructive way.

linkedin 32 reactions · 4 comments Read on LinkedIn →

There should be more medleys.

Last week I posted 'There should be more medleys' on BlueSky. Then I spent today actually doing something about it. I have 350+ guitar tabs backed up from Ultimate Guitar. The problem: how do you build a medley that actually flows? Not just 'these songs are in the same key' but actually makes musical sense?

linkedin 27 reactions · 7 comments Read on LinkedIn →