🌟 Meet #DeepSeekMoE: The Next Gen of Large Language Models! Performance Highlights: 📈 DeepSeekMoE 2B matches its 2B dense counterpart with 17.5% computation. 🚀 DeepSeekMoE 16B rivals LLaMA2 7B with 40% computation. 🛠 DeepSeekMoE 145B significantly outperforms Gshard,

My Third Post Implications of the H20 It will have profound change on the dynamics of US-China AI race especially in the RL/inference age

Let's be honest, the raw capabilities of the models are still quite far from the greatest human geniuses, today we are more around 115/120 IQ (at best) with still a lot of optimization to be done on agency, and avenues of research such as lifelong learning and more.

We've entered a new phase where progress in chatbots is starting to top out but progress in automating AI research is steadily improving. It's a mistake the confuse the two.

Introducing GLM-4.5V: a breakthrough in open-source visual reasoning GLM-4.5V delivers state-of-the-art performance among open-source models in its size class, dominating across 41 benchmarks. Built on the GLM-4.5-Air base model, GLM-4.5V inherits proven techniques from

À voir.

1/n I’m thrilled to share that our OpenAI reasoning system scored high enough to achieve gold 🥇🥇 in one of the world’s top programming competitions - the 2025 International Olympiad in Informatics (IOI) - placing first among AI participants! 👨‍💻👨‍💻

In my opinion, the most important takeaway from this result is that our OpenAI International Math Olympiad (IMO) gold model is also our best competitive coding model. 🧵

you always need money. you need money for compute. you need money for hard-to-get data. you need money for researchers today and brand emissaries tomorrow. you need money for when the algorithmic advances tap breaks its laminar flow

This story is so insane, dream narrative for burgers and their pure-hearted vassals, that I might well cook up my own, also based on half-baked rumors, experience in an authoritarian society and just a little bit of sleuthing. DeepSeek was a breakout hit, but patronage networks

Are frontier AI models really capable of “PhD-level” reasoning? To answer this question, we introduce FormulaOne, a new reasoning benchmark of expert-level Dynamic Programming problems. We have curated a benchmark consisting of three tiers, in increasing complexity, which we call

Rohan Paul False, intelligence still scales logarithmically with compute. And it doesn’t make sense to call them LLMs when they’re natively multimodal. Just models – it’s cleaner.

Debunking yet another study of many that claim benefit of "brain-inspired" mechanisms without doing proper controls - comparing to reference transformer of same size. Doing reference comparisons is the way to avoid "brain-inspired" further sliding towards red flag for scams.

China's Top 19 Open Model Labs We ranked all the organizations in China releasing open models, from the top of DeepSeek to small, newer academic labs making waves with tech reports and niche models. interconnects.ai/p/chinas-top-1…

every day that nvidia stock is down I glance at the kaplan et al scaling law chart and laugh confidently into the future

Paper TLDR: - Reasoning LLMs fail as task complexity rises (e.g. more Hanoi’s tower rings). - The reasoning chain starts fine then diverge and falls apart. My Guess: - It’s not the task complexity, it’s the reasoning chain length. - Train to make longer chains: ceiling goes up.

Introducing ComputerRL, a framework for autonomous desktop intelligence that enables agents to operate complex digital workspaces skillfully. arxiv.org/abs/2508.14040 ComputerRL features the API-GUI paradigm, which unifies programmatic API calls and direct GUI interaction to

«L’histoire ne nous jugera pas avec clémence.» Une conversation avec Gabrielius Landsbergis🇱🇹, ancien ministre lituanien des Affaires étrangères, signée Maria Tadeo. A lire absolument—à discuter. legrandcontinent.eu/fr/2025/08/20/…

Introducing DeepSeek-V3.1: our first step toward the agent era! 🚀 🧠 Hybrid inference: Think & Non-Think — one model, two modes ⚡️ Faster thinking: DeepSeek-V3.1-Think reaches answers in less time vs. DeepSeek-R1-0528 🛠️ Stronger agent skills: Post-training boosts tool use and