Newsletter Archive

Dell, Nokia, and Lenovo—dot-com casualties—are rallying on a combined $1.7 trillion market value surge as AI infrastructure spending transforms legacy hardware makers into AI-era winners.

Bloomberg Tech · 1 min read Market

Vector embeddings excel at semantic similarity but silently fail on negation, exact identifiers, and domain-specific acronyms in RAG systems—the post maps these predictable failure modes and explains what retrieval techniques to use instead.

Towards Data Science · 26 min read Tools

Microsoft and Nvidia are partnering to build AI PCs running autonomous agents rather than traditional Copilot assistants—marking a pivot toward systems that act independently on user tasks instead of just responding to prompts.

The-decoder · 2 min read Industry

SoftBank plans to invest up to €75 billion to build AI data centers in France, marking one of the largest foreign investments in European AI infrastructure and a significant shift in global AI compute distribution.

Bloomberg Tech · 1 min read Industry

Washington lawmakers are intensifying efforts to regulate AI, but the reporting lacks specifics on which bills, agencies, or enforcement mechanisms are driving the debate.

Kcentv · 1 min read Policy

Microsoft's SkillOpt optimizes natural-language prompts for frozen LLM agents using trajectory-driven edits and validation gates, treating skill refinement like neural network training (epochs, batch size, learning rates) without touching model weights. Ships as deployable best_skill.md artifacts.

GitHub Trending · GitHub repo Repos

A 7,500+ line interactive textbook with 22 explainers that walks you through building a GPT-class model from tokenizer to inference, every line commented and explained without jargon. Two narrative walkthroughs trace a sentence through the entire architecture step-by-step.

GitHub Trending · GitHub repo Repos

TokenSpeed, a new LLM inference engine optimized for agentic workloads, achieves 580 tokens/sec on Qwen3.5-397B using Blackwell GPUs. Built with static compilation for multi-GPU parallelism, type-safe KV cache management, and one of the fastest MLA implementations, it matches TensorRT-LLM performance with vLLM's ease of use.

GitHub Trending · GitHub repo Repos