DeepSeek just dropped the single best end-to-end paper on large model training. It covers — Software (MLA, training in FP8, DeepEP, LogFMT) — Hardware (Multi-Rail Fat Tree, Ethernet RoCE switches) — Mix (IBGDA, 3FS filesystem) DeepSeek's engineering depth is insane. Must read.

I am pleased to announce a new version of my RL tutorial. Major update to the LLM chapter (eg DPO, GRPO, thinking), minor updates to the MARL and MBRL chapters and various sections (eg offline RL, DPG, etc). Enjoy! arxiv.org/abs/2412.05265

Huh. Looks like Plato was right. A new paper shows all language models converge on the same "universal geometry" of meaning. Researchers can translate between ANY model's embeddings without seeing the original text. Implications for philosophy and vector databases alike.

This AI Agent can think, code, reason and browse in a single loop just like humans. Outperforms other AI Agent frameworks like Manus AI, Genspark AI, and OpenAI Deep Research. And it's 100% Opensource.

New Lens on RAG Systems RAG systems are more brittle than you think, even when provided sufficient context. Great work from Google and collaborators. Good tips for devs included. Here are my notes:

🚨This week's top AI/ML research papers: - Spurious Rewards - FLUX.1 Kontext - Learning to Reason without External Rewards - Reasoning LLMs are Wandering Solution Explorers - VLM-3R - Silence is Not Consensus - Beyond Markovian - The Entropy Mechanism of RL for Reasoning LMs -

Releasing the Energy-Book 🔋 from its first appendix's chapter, where I explain how I create my figures. 🎨 Feel free to report errors via the issues' tracker, contribute to the exercises, and show me what you can draw, via the discussion section. 🥳 github.com/Atcold/Energy-…

🔥 Midjourney video is almost here... And it has, somehow, that incredible artistic aesthetic that MJ is famous for. Yes, these are all MJ video generations!!! 🧵👇

🚨Should We Still Pretrain Encoders with Masked Language Modeling? We have recently seen massively trained causal decoders take the lead in embedding benchmarks, surpassing encoders w/ bidirectional attention. We revisit whether Bert-style encoders are a thing of the past? (1/N)

How do you build an LLM-evaluator / LLM-as-Judge? The book for "AI Evals for PMs and Engineers" has a chapter devoted to it (35% discount: maven.com/parlance-labs/…) First, we need to define the right metrics. For example, we can start by listing the failure modes from our error

🗣️ Manuel Vallecillo Galeano: un agente de la autoridad que trafica con odio ✍🏼Romàn🔻 diario-red.com/articulo/espan…

My NYU Center for Data Science colleague, Carlos Fernandez-Granda, released the 700-page textbook «Probability and Statistics for Data Science», where he condenses 10 years of teaching experience at @NYUniversity. 200 exercises, 102 notebooks, 115 videos! 🥳🥳🥳 ps4ds.net

My work here is truly done. Nobody with intellectual integrity can still believe that pure scaling will get us to AGI. GPT-5 may be a moderate quantitative improvement (and it may be cheaper) but it still fails in all the same qualitative ways as its predecessors, on chess, on

Stanford Deep Learning for Computer Vision taught by Professor Fei-Fei Li (Fei-Fei Li) and Assistant Professor Ehsan Adeli. Such an enjoyable YT series. (link in comment)

The paper turns LLM role play into a retrieval task so the agent stays in character. When the researchers tested their system against "jailbreak" attempts (where someone tries to push the model out of its assigned role), the model relied more heavily on its reference examples

I want to frame this whole article.

Lamentem comunicar-vos el traspàs del professor José Bernabeu Alberola. En nom de tota la comunitat de la Facultat de Física volem transmetre el nostre més sentit condol als seus familiars, amics i col·laboradors, i acompanyar-los en tan difícils moments. Descanse en pau.

Microsoft just killed the GPU mafia! 🤯 They've open-sourced bitnet.cpp, a blazing-fast 1-bit LLM inference framework optimized for CPUs. This is a major step forward for running large models locally, without expensive GPUs or cloud costs. Demo app + repo + paper in 🧵 ↓

Reading a math book without doing the problems is like reading a workout plan without doing the exercises.