Albert Catalán Tatjer (@actatjer) header
Albert Catalán Tatjer (@actatjer) profile picture

ML efficiency PhD @ELLISInst_Tue & @MPI_IS before @uni_tue & @unibarcelona

Joined November 2024
Photo by Albert Catalán Tatjer (@actatjer)Photo by Albert Catalán Tatjer (@actatjer)Photo by Albert Catalán Tatjer (@actatjer)Photo by Albert Catalán Tatjer (@actatjer)Photo by Albert Catalán Tatjer (@actatjer)Photo by Albert Catalán Tatjer (@actatjer)Photo by Albert Catalán Tatjer (@actatjer)
He did it again
We can finally talk about it: We found a way to extract hidden reasoning of frontier models using a vulnerability in the APIs of every frontier AI company. We verified that our reasoning token count matches billed API thinking tokens 1:1 for most of the prompts we queried.
9
232
🚨 Quantization robustness isn’t just post-hoc engineering; it’s a training-time property. Our new paper studies the role of training dynamics in quantization robustness. More in 🧵👇
3
7
3
35
21,483
7️⃣ 🧰 Practical takeaways for model builders: Maintain higher learning rates longer when possible Prefer higher weight decay (λ) if performance is equal Use weight averaging to counteract brittleness Study quantization during pretraining, not only after
1
6
580