Fetching from the wire…
01
02
03
04
05
06
07
08
Top 5 · 2026-02-21 · source-backed
Anthropic engineer Thariq Shihipar disclosed that Claude Code treats cache hit rate drops as production incidents — the entire product architecture was designed around prompt caching constraints. Plan Mode exists as callable functions (not tool permission toggles) specifically to avoid breaking the cached prefix. Tool Search uses lightweight stubs with defer_loading flags. Cached reads cost $0.30/M tokens (10% of standard $3.00/M). This is the most significant "how the product is actually built" disclosure from Anthropic, revealing that cost optimization drives Claude Code's design as much as capability does. Implicator.ai | Simon Willison
Each link below shares sources, entities, or timing with this story.
Anthropic: Prompt Caching as Production Architecture — The most significant engineering disclosure from Anthropic this year. Claude Code's static-first prefix structure, Plan Mode as callable functions, Tool Search with defer_loading, and compaction maintaining identical syste...
Willison's prolific output continues: he covered ggml.ai joining HuggingFace, Taalas 17K tok/s, SWE-bench Feb 2026 leaderboard (Opus 4.5 leads at 80.9%, 4 Chinese models in top 10), Paul Ford's NYT essay, Martin Fowler's "LLMs eating specialty skills" observation, and Thariq S...
Willison published his AI Engineer World's Fair conversation with Anthropic's Cat Wu and Thariq Shihipar, covering Claude Code, Claude Tag, and Fable. Primary-source practitioner conversation with the people who actually build the thing, rather than secondary coverage of a pre...
The IDE market is fragmenting, and this week drew the sharpest lines yet. Cursor 3 launched as a rebuilt agent-orchestration platform in Rust and TypeScript, replacing the VS Code fork with an Agents Window for dispatching and monitoring multiple AI coding agents. Anysphere hi...
Simon Willison's Blog — 11 posts across Feb 18-20 covering the full AI landscape: ggml.ai/HF acquisition, Taalas hardware, SWE-bench February leaderboard (Opus 4.5 leads at 80.9%), Paul Ford's NYT disruption essay, Fowler's "LLMs eating specialty skills," and Thariq Shihipar's...
This one hit my inbox and I had to read it twice. Anthropic announced a partnership with SpaceXAI for the entire Colossus 1 data center in Memphis. 220,000 NVIDIA GPUs. 300+ megawatts. That's the largest single compute acquisition by any AI lab. Full stop. But the part that ma...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.