Cold Open — The Executive Who Built Your AI Strategy Has Never Opened ChatGPT

Saturday, July 19, 2026. We scanned 2,207 items off the wire; five made the cut — and the lead is the kind of corporate confession that has been circulating in consultants' group chats for months, finally documented in public.

This post is the print twin of today's Cold Open podcast episode — the day's AI news before your coffee cools.

---

The lead · AI mania is eviscerating global decision-making

Nik Suresh, a consultant who works inside large organizations, published a piece this week that is difficult to summarize without quoting directly. The headline: AI mania is not just overhyped — it is actively breaking how big companies make decisions.

The most pointed anecdote, verbatim:

"In one extreme case, I have seen an executive confess that they had never even used ChatGPT or any AI tool in their life, immediately after producing a technical strategy for an organisation with $2B+ in revenue which was entirely centered around AI."

That is not an outlier. Suresh documents a pattern: executives mandating AI transformation without understanding the tools, and organizations contorting themselves to comply. Simon Willison, who surfaced the piece on his newsletter, describes it as "crammed with spicy anecdotes from anonymous sources" — the kind that circulate privately but rarely make it to print.

A second example from an engineer inside one of these organizations:

"Checking out a parallel copy of our Go repository and telling an AI agent to make changes and commit them to it — just to generate tokens on the leaderboard."

The company had installed a token-consumption leaderboard to measure AI adoption. The result: engineers found the most efficient path to tokens, which turned out to be generating meaningless code commits. The metric ate itself.

Why it matters

This is the accountability gap that builders have been watching widen for the past two years. The pattern is consistent: organizations adopt AI as a signaling behavior before building the internal literacy to use it well. The cost is not just wasted budget — it is bad strategy produced at scale, and engineering cultures warped around vanity metrics instead of actual output.

The Suresh piece is worth a full read not because the anecdotes are shocking, but because they are normal. Anyone embedded in a large organization will recognize the dynamics. What makes this moment notable is that these stories are reaching print.

The fine print

Suresh's anecdotes are anonymous, which makes them impossible to verify individually. The value is the pattern, not any single data point. And the piece is clearly a perspective from a consultant who works with struggling organizations — healthy teams doing good AI work are not his client base, so the sample is skewed toward dysfunction. That said: skewed toward dysfunction or not, the pattern is real.

Sources: simonwillison.net · ludic.mataroa.blog

---

02 · Claude Code is running on Bun written in Rust — and barely anyone noticed

Since June 17th, Claude Code v2.1.181 and later have been shipping on the Rust port of Bun. Jarred Sumner, Bun's creator, noted in his rewrite announcement that startup got 10% faster on Linux — and his next line is the honest one: "barely anyone noticed. Boring is good."

Simon Willison verified it by running strings ~/.local/bin/claude | grep -m1 'Bun v1' — the output returned Bun v1.4.0 (macOS arm64), confirming Bun is bundled inside the Claude Code binary itself.

Why it matters. Infrastructure improvements to the tools you use daily usually ship silently. A 10% faster startup on Linux is not a headline feature, but for developers who open and close Claude Code dozens of times a day it compounds. More interesting is the engineering call Bun made: rewrite from Zig to Rust, migrate all active dependencies into the new codebase quietly, ship. No fanfare, no migration tax for users. That is a well-executed transition.

Sources: simonwillison.net · bun.com/blog/bun-in-rust

---

03 · NYC may require landlords to disclose AI-generated property images

NYC Mayor Zohran Mamdani is pushing rules that would require landlords and realtors to disclose when AI is used to generate images in property listings. The issue: AI-generated listing photos can make a small apartment look spacious, a dark unit look bright, and a dated kitchen look renovated. Buyers and renters are making decisions based on images that do not represent the property.

Why it matters. AI disclosure in commercial contexts is becoming a regulatory category. Real estate is the first sector where AI image generation has a direct, documented financial harm vector — misleading listings drive real decisions. If this passes, it is a template that other cities and other sectors will reference. The HN thread (531 upvotes, 229 comments) shows this is touching a nerve well beyond New York.

Sources: petapixel.com · HN discussion

---

Also on the radar

• Operational excellence — Cortex released DRIVE, a framework for measuring and improving engineering performance in AI-accelerated teams. Worth a look if your org is trying to separate signal from mania.
• Dev tooling — Simon Willison built a SQLite Query Explainer (with help from Claude Fable) that runs SQLite in Pyodide in the browser and adds a human-readable explanation layer to EXPLAIN QUERY PLAN. Inspired by Julia Evans asking "maybe one day I'll learn to read a query plan." Now you can.

---

Trends in dev tools

What moved this week in the tools engineers actually ship with.

• Silent infrastructure upgrades are how serious tools ship. Bun's Rust rewrite reached Claude Code v2.1.181 in June without anyone noticing. That is the gold standard: a foundational rewrite that users experience as "it feels a bit snappier" rather than a migration event. More Rust-in-the-toolchain moments are coming. (bun.com)
• AI-assisted debugging tools are proliferating at the dev layer. The SQLite Query Explainer is one instance of a broader pattern: LLMs being used to translate opaque technical output (query plans, compiler errors, profiler traces) into plain language explanations. The pattern generalizes cleanly to any tool with terse, expert-only output. (simonwillison.net)
• Operational excellence frameworks are being rebuilt for AI-paced teams. DRIVE by Cortex is not the first, but its positioning around AI-accelerated engineering specifically signals that the old DORA metrics are being revisited. Token leaderboards are what you get when you do not have a real measurement framework — DRIVE is an attempt at the real thing. (cortex.io/drive)
• AI disclosure is moving from ethics to compliance. NYC's proposed rule is developer-adjacent in an important way: if disclosure becomes legally required in one industry, the demand for reliable AI-watermarking and provenance tooling goes from a nice-to-have to a requirement. Watch this space. (petapixel.com)

---

Popular skills

The agent-skills wave — portable instruction sets that let a coding agent pick up domain expertise on demand — keeps spreading across engineering contexts.

• The "AI strategy" gap Suresh documents is, at root, a skills problem. The organizations failing at AI transformation are not failing on tooling — they are failing on the human capability to evaluate, prompt, and steer the tools. Agent skills files are one answer at the individual level: encoding what "good use of AI looks like in this domain" as a reusable artifact rather than tribal knowledge. The corporate version of this problem is unsolved.
• SQLite explainers built by AI agents point toward a new dev-skill category. When an AI agent can build a domain-specific explanation layer on top of an existing tool in a session, the ceiling of what a solo builder can prototype in a weekend moves again. The skill is not knowing SQLite query plans — it is knowing how to ask an agent to build you a tool that explains them.
• Operational measurement frameworks need a skill-layer translation. DRIVE's framework exists as a document. The version that actually helps a team is one where an agent can ingest the framework and surface the right metric for the current context on demand. Nobody has built that yet. Opportunity.

---

AI fun fact

The first industrial AI system to deliver documented, large-scale ROI was XCON (eXpert CONfigurer), built by Digital Equipment Corporation and Carnegie Mellon University in 1980. XCON automatically configured customer computer orders — a task that previously required an expert who knew 420 components and 17,500 rules. By 1986 it was saving DEC an estimated $40 million per year and had processed over 80,000 orders. It ran on VAX hardware, was written in Lisp, and worked reliably for over a decade. The executives who approved it used it every day. (CMU research)

---

Sources

• AI Mania Is Eviscerating Global Decision-Making — Nik Suresh, ludic.mataroa.blog
• Simon Willison on AI mania — simonwillison.net
• Claude Code uses Bun written in Rust — simonwillison.net
• Rewriting Bun in Rust — bun.com
• NYC landlords and AI image disclosure — petapixel.com
• SQLite Query Explainer — simonwillison.net
• DRIVE framework — cortex.io
• XCON expert system — AAAI AI Magazine