FP4 Training: DeepSeek V4, NVIDIA Blackwell, and the End of FP16
FP4 training was a research curiosity in 2024. By 2026 it ships in production frontier models. What changed and what tradeoffs remain.
9 min read
Agentic AI & LLMs
Agentic AI & LLMs archive page 256 of 631
Agentic AI & LLMs
Agentic AI, LLM engineering, and the models behind modern automation — multi-agent systems, LLM evaluation and comparisons, RAG, fine-tuning, AI infrastructure, security, and production AI engineering.
From the blog
Latest articles
KV-Cache Offloading Strategies: CPU, GPU, and NVMe Tradeoffs in 2026
KV-cache is the dominant memory cost in long-context inference. The 2026 offloading strategies that make 1M-token serving practical.
8 min read
Long-Horizon Agent Tasks: Why 90% Fail Past Hour Three (and How to Fix It) — Long horizon tasks
Long horizon tasks: long-horizon agent runs collapse for predictable reasons. A 2026 teardown of failure modes and the architectural patterns that actually keep agents on track.
9 min read
MCP Server Ecosystem 2026: The 50 Most-Used Model Context Protocol Servers
An inside look at the Model Context Protocol server ecosystem in 2026 — the official, community, and enterprise servers driving real production agent workloads.
9 min read
Mixture of Experts Beyond Sparse: Granite, DeepSeek-MoE, and Mixtral Patterns
MoE evolved beyond simple top-k routing. The 2026 patterns from Granite, DeepSeek-MoE, and Mixtral that make MoE practical at scale.
8 min read
Multi-Agent Debugging: Finding the Bug Across 12 Concurrent LLM Calls
Multi-agent systems break in ways single-agent systems never do. The 2026 debugging stack and the patterns that turn opaque failures into reproducible bugs.
8 min read
MXFP4 Quantization Explained: The Microscaling Format Behind 2026 Inference
MXFP4 is the quantization format powering 2026 inference on NVIDIA Blackwell, AMD MI355X, and Intel Gaudi 3. What it does, why it works, and what it costs.
9 min read
The Orchestrator-Worker Pattern: Anthropic's Research Architecture Explained
Anthropic's published multi-agent research architecture is a clean orchestrator-worker design. What it does, why it works, and how to adapt it.
9 min read
Paged Attention and Its Descendants: Memory-Efficient LLM Serving in 2026
PagedAttention launched a family of memory-management techniques that make modern LLM serving possible. The 2026 descendants and what they fix.
8 min read