Research library
Software engineering
Practical architecture and development decisions for software that people can use, maintain, and adapt.
89 articles · Page 6 of 8
TPU v8 vs Blackwell: How AI Silicon Is Splitting Into Training and Inference Chips
Training and inference need different silicon. TPU v8t/v8i architecture, comparison to Blackwell's unified design, and per-token cost implications.
OpenAI Ships Codex in the ChatGPT Mobile App — The Phone Becomes an Agent Remote Control
Codex now runs in the ChatGPT mobile app as a remote control for macOS agents — the phone supervises while code and credentials stay on desktop.
Compressed Sparse Attention: How DeepSeek V4 Reached 1M Context at 27% of the FLOPs
DeepSeek V4 hits 1M context at 27% of V3.2's per-token compute. How Compressed Sparse Attention and Heavily Compressed Attention combine to do it.
KV Cache: The Hidden Memory Wall in LLM Inference
The KV cache memory wall in LLM inference: the math behind long context costs and architectural solutions (GQA, MQA, MLA, paged attention).
Anthropic Launches Claude Design — Figma Drops 7% the Same Day
Claude Design generates prototypes, slide decks, and mockups from plain prose, imports design systems, and triggered a 7% Figma stock drop.
Anthropic Unveils Claude Mythos Preview and Project Glasswing — A Cybersecurity Alliance Built Around Its Most Capable Model
Claude Mythos Preview found thousands of zero-days in major OSes and browsers. Anthropic gates it via Project Glasswing for vetted security teams.
OpenAI Launches ChatGPT 5.5 and Merges Everything Into a Single Desktop Super App
ChatGPT 5.5 unifies Codex, Atlas, and chat in one desktop app. Better memory and task continuity enable integrated AI workspace for coding and web work.
Anthropic Accidentally Leaks Claude Code's Entire Source Code via npm
A misconfigured npm package exposed 512,000 lines of Claude Code's source—revealing system prompts, KAIROS daemon, and unreleased features.
Gemini 3 Flash: Google's Best Coding Model Isn't Its Most Powerful One
Gemini 3 Flash: 78% SWE-bench outperforms Gemini 3 Pro on coding. Fastest, cheapest model optimized for agentic coding and multimodal reasoning.
NVIDIA Nemotron 3 Super: The Open Model Built for Multi-Agent Systems
Nemotron 3 Super: 120B open model blending Mamba and Transformer, with 4x memory efficiency, 1M token context, and 7.5x throughput advantage.
Anthropic's Claude Code Review: AI-Generated Code Reviewed by AI
Multi-agent code review for Claude Code: filters false positives, ranks bugs by severity, automates GitHub PR feedback with structured inline comments.
Small MoE Models: How Sparse Routing Makes Efficient AI Possible
Small-scale Mixture of Experts: sparse routing lets 47B models match 70B dense equivalents. Mixtral, DeepSeek-MoE, Phi-MoE, efficiency math.