Essays and field notes on AI, software engineering, design, and the craft of building product teams that ship. Written by the engineers doing the work.
DeepSeek V4 lands open-weights under MIT at $1.74/$3.48 per million — 1/20 the cost of Opus 4.7 — and reframes the FY27 cost-tier routing slot entirely.
Cursor orchestrates, Claude Code and Codex execute, cross-provider review verifies — the AI coding stack composes into three layers instead of consolidating.
Gemini 3.5 Flash shipped at I/O 2026: 76.2% Terminal-Bench, 289 tok/s, $1.50/$9 per million — and beats last cycle's Gemini 3.1 Pro on coding.
Microsoft's June 8 MAI launch shipped seven in-house models — Thinking-1, Code-1-Flash, Image, Voice, Transcribe — the OpenAI dependency thins.
Reinforcement learning with verifiable rewards is the dominant post-training paradigm — GRPO replaces PPO and rewards become audit-ready.
Arcade closed a $60M Series A on June 15 led by SYN Ventures, with Morgan Stanley and Wipro, lifting total funding to $72M for its secure agent-action layer.