Latest · Development Jul 7. 2026

Frontier Frugality

Frontier-tier models like Claude Fable 5 cost $10/$50 per million tokens — 10x the price of Haiku 4.5 — and an agentic coding session re-sends its whole history on every turn, so the bill grows roughly quadratically with session length. This post works through the levers that tame that bill: prompt caching, delegating bulk reading to cheap subagents, difficulty-based routing, verification asymmetry, and — the lever with the longest half-life — distilling the expensive model's judgment into rules files that cheap models can execute after your frontier access ends.

Michael Wan

Machine Learning, AI & Agents

Paper explainers, model architectures, agent systems, and the mechanics behind modern AI

Jul 4. 2026 explainer

How Reasoning Models Learn to Think

Pretraining teaches a model what to say; it does not teach it how to think through a hard problem. Reasoning models like OpenAI o1 and DeepSeek-R1 add a second training phase: reinforcement learning against rewards that can be checked mechanically, such as a math answer or a passing test suite. This post explains that recipe — the verifiable reward signal, the GRPO algorithm that makes it cheap, the behaviors that emerge, and the new scaling axis it unlocked at test time.

Michael Wan

Jun 25. 2026 explainer

GLM-5.2: The Open-Weight Model That Beats GPT-5.5

GLM-5.2 is an MIT-licensed open-weight model that extends its context window to 1M tokens and ranks as the highest open-source model on Z.ai's evaluations, beating GPT-5.5 on long-horizon coding. Its headline trick, IndexShare, reuses sparse-attention indices across groups of layers to cut per-token FLOPs by 2.9× at 1M context. This article explains the architecture, the core mechanic, the real benchmarks, and the community's reaction.

Michael Wan

Development

Tools, workflows, and practical engineering insights

Business & Industry

Compensation, industry trends, and career strategy through data

Jul 1. 2026 explainer

The AI Investment War

By mid-2026, the AI industry has organized itself into four layers — applications, models, compute, and silicon — stitched together by a dense web of equity stakes and hundred-billion-dollar supply contracts. The striking feature is not the size of any one deal but the circularity: a chipmaker invests in a lab, the lab pays a cloud, the cloud buys the chips. This interactive map lets you trace every major deal — click a company to see its full position in the war, or click an arrow for the terms — and the essay decodes the deal structures and asks what would tell the bulls from the bears.

Michael Wan