Supercomputing for AI — Foundations, Architectures, and Scaling Deep Learning. [804-page masterpiece] Read it online: https://
jorditorresbcn.github.io/supercomputing
-for-ai-book/
… Buy it: https://
amzn.to/4qS4pFz GitHub repo: https://
github.com/jorditorresBCN
/supercomputing-for-ai
…
SYSTEMS
-
Supercomputing for AI — Foundations, Architectures, and Scaling
By
–
-
Full motion transformer trained in 3 days on 128GPUs at 10000x speed
By
–
This full motion transformer was trained in 3 days on 128GPU at 10.000x faster than wall clock speed.
— Linus ✦ Ekenstam (@LinusEkenstam) 25 février 2026
in english: this AI model controls the motion of the robot.
It supports, text to command, remote teleportation and much more.
Amazing work by Jim Fan et al. 🫰 https://t.co/ynXhrNwu5iThis full motion transformer was trained in 3 days on 128GPU at 10.000x faster than wall clock speed. in english: this AI model controls the motion of the robot. It supports, text to command, remote teleportation and much more. Amazing work by Jim Fan et al.
-
LLM Performance Degradation Over Extended Context Windows
By
–
Performance degrades over the course of 100k tokens even, let alone the whole currently supported window… After a few turns of coding Python, it just can't reliably use its tools anymore. Requires constant jumping back and/or offloading.
-
Workshop: Building MCP Servers for AI Agents in Production
By
–
Look! >> @PacktPublishing is hosting a workshop on 29th March on “Building MCP Servers in Production”, led by Peder Holdgaard Pedersen, for engineers designing the secure, scalable systems that support AI Agents. Find more details and register here: https://
eventbrite.com/e/building-mcp
-servers-in-production-tickets-1982519419953?aff=kirk
… -
Contextual UI vision from 2014 now enabled by on-device LLMs
By
–
Contextual UI is a game changer…. The catch? I wrote this in 2014 But the promise of a fully contextual OS and machine is more promising today than ever before. On device LLM's running in secure and private enclaves, orchestrating behind the scenes. Helping the OS.
-
Fast KV Compaction via Attention Matching for Long-Context LLMs
By
–
"Fast KV Compaction via Attention Matching" The current problem is that long-context LLMs would choke on KV cache memory, so systems either drop tokens or write lossy summaries. As a solution, the paper shows you can shrink an entire prefix to ~50x fewer KV slots by solving
-
Agentic AI Fundamentals: Planning, Memory, Tools, and Multi-Agent Systems
By
–
Agentic AI, simplified: • Agents are systems
• Plan → Act → Reflect loops
• Memory = compounding intelligence
• Tools = real-world impact
• Multi-agent > single-agent If it only chats, it’s not agentic. Via @ingliguori -
New AI System Design Book: 10 Domain Driven Blueprints Released
By
–
New release! "Mastering AI System Design: Architect, Build and Deploy AI Systems Using 10 Domain Driven Blueprints and Interview Strategies" available at http://
amzn.to/3NFNOWF 𝙏𝙖𝙗𝙡𝙚 𝙤𝙛 𝘾𝙤𝙣𝙩𝙚𝙣𝙩𝙨:
1. Introduction to AI System Design
2. Crafting Intelligent Systems -
Encoding Context at Data Source for AI Effectiveness
By
–
For AI to be effective, context must be encoded as far "left", or as close to the source, as possible. When you define data models at the edge using tools like MQTT and Sparkplug, you ensure that every downstream system, including AI agents, understands exactly what that data… pic.twitter.com/irUbXI2DKk
— Lucian Fogoros (@fogoros) 21 février 2026For AI to be effective, context must be encoded as far "left", or as close to the source, as possible. When you define data models at the edge using tools like MQTT and Sparkplug, you ensure that every downstream system, including AI agents, understands exactly what that data
-
Operating Systems Will Be Rebuilt Using Differentiable Programming
By
–
Prediction: we will end up rebuilding Operating Systems using automatically differentiable programming primitives. It won’t be a monumental task with LLMs, and it will lead to gradients back-propagating into the OS kernel. < 7 years.
→ View original post on X — @reza_zadeh, 2026-02-21 03:39 UTC