🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

At Sesterce, we built automatic failover for massive GPU clusters. Only one other global company has mastered this. While Europe debates AI ethics, we're solving the engineering problems that actually matter. Some build the future, others write reports about it.

The real energy crisis isn't about heating homes. It's about powering AI infrastructure. Europe dismantles nuclear while trying to compete in the most energy-intensive technology race in history.

The secret weapon of Chinese tech giants like Tencent and Xiaomi? They can fund experimental tech through profitable divisions. European companies can't form conglomerates due to our own regulations. We're fighting a war with self-imposed restrictions our competitors don't

Trillion-dollar tech companies aren't born from perfect regulation. They're born from freedom to experiment, fail, and iterate rapidly. Europe has designed systems that make this nearly impossible.

If you don't know what this is you probably suck at coding

We discovered faster way to compute product of matrix by its transpose! This has profound implications for data analysis, chip design, wireless communication, and LLM training! paper: arxiv.org/abs/2505.09814 The algorithm is based on the following discovery: we can compute

It only takes 2.5 kWh of human brain compute to go from birth to a PhD, and a constant 20W for on edge inference.

Re-entering the discussion with an informative panel discussion on 'Infrastructure Gold Rush: Where to Invest Before Agentic AI Goes Mainstream'. #aimsummitlondon2025 Youssef El Manssouri Sesterce Bilal Hafeez Dion Wiggins Al Fardan Ventures Gayatri Sarkar✊🏽🌏 🚀 Advaita Capital

Most decentralized training today follows DDP-style approaches requiring full model replication on each node. While practical for those with H100 clusters at their disposal, this remains out of reach for the vast majority of potential contributors. Delving back into the

We've reached a major milestone in fully decentralized training: for the first time, we've demonstrated that a large language model can be split and trained across consumer devices connected over the internet - with no loss in speed or performance.

NCCL sending the loss value from the last pipeline parallel stage back to rank 0 so the user can print it

RL is very inference heavy and shifts infrastructure build outs heavily Scaling well engineered environments is difficult Reward hacking and non verifiable rewards are key areas of research Recursive self improvement already playing out Major shift in o4 and o5 RL training

Multimodal Large Language Models: A Survey Nice graph depicting the evolution from foundational transformer and diffusion structures. All references in the paper:

Exciting Line-Up coming at the Raise Summit! We're thrilled to share the agenda for the upcoming RAISE Summit, where innovation meets opportunity! Join us to explore groundbreaking ideas and network with industry leaders pushing the frontiers of technology. Must-Attend

torchft + TorchTitan: 1200+ failures, no checkpoints, model convergence. A Llama 3 model was trained across 300 L40S GPUs with synthetic failures every 15s. No restarts. No rollbacks. Just asynchronous recovery and continued progress. 📘 hubs.la/Q03t1Z0b0 #PyTorch

my last 10-15 flights time

I wrote a new blog on TPUs -- it's been fun seeing how different they are from GPUs and also drawing things on excalidraw again✏️ henryhmko.github.io/posts/tpu/tpu.…

“France has 5 GW of unused energy — let’s build sovereign AI datacenters!” At Sesterce, we believe France can be Europe’s AI powerhouse. By 2030, AI needs 100 GW — France must act now. 🎥 Watch Youssef El Manssouri on why France could lead Europe’s AI future: bit.ly/46bmqr1 #AI

Equations that changed the world.

How to sync GPUs in multi-GPU training, clearly explained (with visuals):