The Honest Benchmark Is the One You Can't Inspect
Specific Labs took ten real engineering tasks out of private production codebases, licensed from the companies that own them, and ran eight frontier model-and-harness combinations against all of them: 640 graded rollouts, a best score of 38.8%, and an outright zero on one task no configuration ever solved. The code is invisible on purpose. Hacker News spent the day asking who gets to check the scorer.
Agents Made Writing the App Twice the Cheap Option
Shopify went all-in on React Native in 2020 so it would never build the same feature twice. This week its engineering blog said the opposite: with coding agents translating between Swift and Kotlin, the Shop app was rebuilt as a fully native app in 12 weeks, its 300-screen flagship is mid-migration, and three of its React Native libraries are being handed off or archived. The Hacker News thread asked the question the post never answers: what did the tokens cost?
Everyone Is Working Around the Same Bugs
Dan Luu makes the case for bug blindness: most people hit the same broken software every day and stop noticing, and much of what we call computer literacy is really a library of invisible workarounds. The Hacker News thread proved his point by finding a bug in his own blog.
Virtual iPhones Are Here, Built on Apple's Own Framework
A small Hacker News post this morning showed a full iPhone operating system - kernel, userspace, and all - booting inside a virtual machine on a Mac. The tool is built on Apple's own Virtualization.framework and the research VM infrastructure behind Private Cloud Compute, and it does what Apple never shipped: a real iPhone OS image running as a VM.
Fast Software Is Now a Few Sentences Away
Dan Luu argues the cost of performance optimization just collapsed - and HN is split on whether anyone will actually use it.
A Write That Vanished Into Thin Air
Tailscale's months of outages traced back to a data race in SQLite that had lurked for 16 years - a committed write, gone without a trace, no error raised.
The Code Map That Cuts Token Waste by 38x
A Python tool called code-review-graph uses Tree-sitter to build a persistent code map for AI tools — and the claimed 38x to 528x token reduction is backed by real benchmarks.
Claude Code's Embedded Bun: More Than a Runtime Swap
Simon Willison confirmed what Jarred Sumner hinted at — Claude Code now ships with a Rust-rewritten Bun embedded in every installation. The implications go beyond a faster JavaScript runtime.