AI Digest — July 19, 2026

Quick Notes

• Consultant Nik Suresh offers a caustic, anecdote-packed take on how AI hype is corroding corporate decision-making — including an executive who authored an AI-centric strategy for a $2B+ company without ever having used ChatGPT, and vendors afraid to debunk “100x productivity” claims lest they lose enterprise contracts. https://simonwillison.net/2026/Jul/19/ai-mania/#atom-everything
• Simon Willison verifies Jarred Sumner’s claim that Claude Code v2.1.181+ (June 17) ships the Rust port of Bun, finding his own install reports Bun v1.4.0 — an unreleased preview — with the Rust build now running in production across millions of devices; startup got ~10% faster on Linux and “barely anyone noticed.” https://simonwillison.net/2026/Jul/19/claude-code-in-bun-in-rust/#atom-everything
• Nate B Jones demos an air-gapped laptop running a downloaded local model that reads a contract, finds and masks private information that shouldn’t touch cloud AI, and refuses to certify unreadable content as safe — framing on-prem/local small models as a large emerging business (citing Discovery Bank and Bayer fine-tuning small Microsoft models for confidential work). https://www.youtube.com/watch?v=5slsNizN6MQ

Structured Summaries

Research insights

Nik Suresh, via Simon Willison, argues that “AI mania” is degrading decision-making inside large enterprises. The piece is full of spicy anonymous anecdotes: an executive who produced an AI-centered technical strategy for an organization with $2B+ in revenue despite never having used any AI tool; an engineer rewriting a Go repository in Zig purely to look busy and keep their job; and, most tellingly, a vendor’s explanation of why nobody pushes back on absurd “100x productivity” claims — contradicting a customer executive’s inflated expectations reads as heresy and can get an enterprise contract cancelled, so honesty gets suppressed. The throughline is that social and commercial incentives, not just sales fluff, are what keep implausible AI narratives unchallenged. Sources: https://simonwillison.net/2026/Jul/19/ai-mania/#atom-everything

Tooling

Simon Willison investigates Jarred Sumner’s claim that Claude Code v2.1.181 (released June 17) and later use the Rust port of Bun. Poking at his own install, he finds it reports “Bun v1.4.0 (macOS arm64)” — a version number ahead of the latest public GitHub release (v1.3.14, May 12), evidence Anthropic is shipping a preview of an unreleased Bun. He confirms the Rust build (later released as Bun canary via bun upgrade --canary) is running in production across millions of devices, delivering roughly a 10% Linux startup improvement that “barely anyone noticed” — echoing Sumner’s “Boring is good.” Sources: https://simonwillison.net/2026/Jul/19/claude-code-in-bun-in-rust/#atom-everything

Product / local-model use cases

Nate B Jones showcases a fully offline laptop that still runs AI: a downloaded local model reads a contract, detects and masks private material that should never go to cloud AI, and declines to mark unreadable content as safe. He frames private, on-premise small models as a major business opportunity that Microsoft is betting on — companies want AI help with contracts, Word docs, and customer files but balk at uploading them. Examples include Discovery Bank, which fine-tuned five variant models on two smaller Microsoft base models for confidential financial functions (cutting response times from ~5–6s to ~1.5–2s), and Bayer, which taught a small model its proprietary crop-label and regulatory data so advisors process 100+ page labels in under 30 seconds instead of hours or days. (Both examples run in Azure rather than strictly on customer hardware.) Sources: https://www.youtube.com/watch?v=5slsNizN6MQ

🔗 View this digest on the web: https://ainews.rusig.com/digests/2026-07-19-morning/