NOTES ON ARTIFICIAL INTELLIGENCE
I’m Flavian, Chief AI Architect at Criteo AI Lab, where my research focuses on performance recommendation and recommendation agents. This is my notebook on how models learn, reason and recommend, written by me and animated with the help of AI tools. It covers foundational topics, new research from the field, and my own published work.
WORDS + MOVING PICTURES10 entries · Latest first
/ 8 VIDEOS · 69 MIN
Why training cheats, and what bounds it. Sound checkers, criteria set before the answer, hidden challenges, uncertainty discounts, audits of attempted cheats and KL budgets, each derived step by step with its limits, then a security view of building RL environments where cheating does not pay.
Read article/ TWO VIDEOS · 19 + 20 MIN
Models write programs well. Can they execute one? A small RL environment that checks every move: Qwen3-14B decides 204 of 240 grammar puzzles (a coin gets 120) but executes the procedure through only 30 times, and when it errs it skips ahead. Six questions on memory, the step, chain of thought and training.
Read article/ 5 VIDEOS · 55 MIN
An advanced five-part series measured on tiny models: circuits, what one forward pass can compute, what training learns, computation as a system, and universality and scaling.
Read article
/ 12:45 VIDEO
Part 2 of Don’t Press the Button. A separation of powers for AI: summable risk, the forbidden corner, four classes of models plus humans with the keys, and the complications: swarms, self-improvement and a shared backbone.
Read article
/ 13:46 VIDEO
Can training make an AI agent reliably refuse? The evidence for and against, what kind of test each 2026 incident actually was, alignment vs control, defence in depth, and four readings of one summer.
Read article
/ 8:14 VIDEO
A review of our work at Criteo AI Lab: why learning from logged or stale data should be pessimistic, from log-IPS and logarithmic smoothing to RoiRL and Lambert-tempered targets, compared with REINFORCE, PPO and GRPO.
Read article
/ 8:32 VIDEO
From APE and OPRO to textual gradients, DSPy and MIPROv2, and GEPA’s reflective, Pareto-guided prompt evolution. Plus the 2026 follow-ups, and how to choose.
Read article
/ 6:40 VIDEO
Self-distillation turns error messages and successful rollouts into token-level credit. Then the 2026 variants that trust the model’s own teacher only where it helps: per question, per difficulty, per token.
Read article
/ 8:20 VIDEO
From demonstrations to reward models trained on human comparisons, the KL-regularized objective and PPO. Plus reward hacking, where LLM-as-a-judge fits, and DPO.
Read article
/ 6:53 VIDEO
From verifiable rewards to group-relative advantages. A visual walk through the learning objective, tool-use masking, and the two importance ratios behind asynchronous training.
Read article
Try another topic or a different search.
