THE PELOTON

A Slipped Front Wheel Ends Vingegaard's Tour, and a 2 A.M. Knock Reopens an Old Fight

↩ Developing story — first reported Jul 16 · previously Jul 17, Jul 18, Jul 19

PLATEAU DE SOLAISON, France — Jonas Vingegaard remembers the roundabout, and not much else. Twenty-one kilometres from Sunday's finish, with his Visma-Lease a Bike team massed at the front to chase down the day's break, the road bent right around a traffic island.¹ "We maybe came into this roundabout a little too fast, and then my front wheel just slips. I can't do anything," he said afterward. "That's how cycling races are sometimes. It's just super annoying for us, to be honest."

He was taken by race ambulance to the medical truck in the valley below the summit finish, then sat on the team bus for more than half an hour before he would speak to anyone.¹ When he finally did — to Danish broadcaster TV2, having declined the reporters waiting outside — the collarbone break was confirmed, and so, within hours, was the diagnosis that mattered most: surgery.¹ Visma's statement said only that "due to the severity of the collarbone fracture, surgical intervention has been recommended and will be performed in the coming days," a walk-back from team director Marc Reef's earlier line that the club was still weighing whether an operation was needed at all.²

Vingegaard's own account was strikingly unsentimental for a rider who has spent five Julys locked in what has become the sport's defining rivalry. "It's going surprisingly well. It's of course very disappointing for us to end up like this, and for the Tour to end like this. But that's how cycling is. Luckily, it's only a broken collarbone," he said. "It's no use hanging your head. It just makes it worse."¹ Reef was less measured, calling the crash scene a "strange situation" and pointing at a lead motorbike he said had panicked and braked in front of the bunch. Vingegaard wouldn't go that far. "Maybe the motorcycle makes a mistake, but we can't trust the motorcycles. We have to have our own judgment," he said. "People make mistakes, and we can't change that now."¹

The crash landed on top of a subplot that had already been simmering since Saturday night. Vingegaard was roused for a surprise anti-doping control at roughly 2 a.m., and Pogačar for one of his own around 5 a.m., ahead of Sunday's stage.³ Riders pointed out that the testing landed in the middle of the only recovery window a three-week race allows.⁴ Asked directly whether the disrupted sleep played any part in his crash, Vingegaard wouldn't rule it out but wouldn't claim it either.¹ "It is clear that it has not done anything good for it, you could say. Whether that is the reason, I dare not speculate," he said. "I don't want to blame it on that, because it's a bicycle race, and crashes happen in bicycle races."¹

His rival showed no such restraint. In his own post-stage remarks, Evenepoel called the overnight testing "inhuman" and said the riders' union has to intervene.⁴ "I can't speak for the guys about their focus and how they manage competition, I only think I can say it is very disrespectful to wake up guys at 2 AM and 5 AM," he said. "The CPA has to do something about that. It is inhuman to wake us up, we had a long transfer, and to wake them up like that is pretty disrespectful… Sleep is the only moment we have to recover. This is the hardest race in the world, it makes it even harder. It's something we could and should not accept."⁴

Tom Pidcock's third straight day trying to force his way up the general classification from the break ended in the commissaires' office instead. Twice on the stage 15 climbs, the Pinarello-Q36.5 rider pushed off a car — each offence carrying its own 200 CHF fine, 10-second time penalty and 4-point KOM deduction, for a combined 400 CHF, 20 seconds and eight KOM points gone by the finish.⁵ It capped a stretch in which Pidcock has gone from a virtual fourth place on GC mid-stage to crossing the line well down, worn out by three days of racing on the attack while also defending a top-ten overall position.⁵ "I think I'm getting quite good at getting the breakaways… Today can break loads of records, but so can I," he said afterward, before admitting the final kilometres had been a crawl: "I was completely empty."⁵

Away from the Alps, the Baloise Ladies Tour wrapped its four-stage run on a flat, cobbled circuit through Mechelen, the last of six laps taking in the Keizerstraat five times.⁶ Zoe Bäckstedt carried the leader's jersey into the finale after building her advantage almost entirely against the clock — a prologue win followed by the stage 3a time trial — and the Canyon-SRAM team, alongside the sprinters' squads of Visma-Lease a Bike and Fenix-Premier Tech, spent the final circuits controlling a five-rider move that briefly opened a 2:20 gap before it was hauled back inside the closing kilometres.⁶ "We had the leader's jersey and we had a really strong team. I feel that we had to be in control and we really played it well every day," Bäckstedt said.⁶ Nienke Veenhoven timed her sprint to beat Lotte Kopecky by half a wheel for her third stage win of the race.⁶

↑ Back to top

THE LAB

A 2022 Email Surfaces: Altman Wanted an Open Model to Starve Rivals of Funding

Before OpenAI ever set the terms of the modern AI race, it was worried about being outflanked — and it had a plan to make sure that didn't happen. An October 2022 email Sam Altman sent to OpenAI's board, surfaced this week in the Musk v. Altman litigation and republished by Simon Willison, lays it out without much subtlety: "We have been having extensive discussions around open source strategy. We will discuss it more at our next board meeting, but one thing we'd like to do soon is to create a language model with the approximate capability of GPT-3 that can run locally on consumer hardware and release that. We'd like to do it soon, before Stability or someone else does. In general, we think this helps discourage others from releasing similarly-powerful models, and makes it harder for new efforts to get funded."¹

Willison's post is short — a single quoted paragraph — but it's a rare unfiltered look at the calculus behind OpenAI's occasional open-weight releases. The stated goal wasn't research openness or community goodwill; it was denying oxygen to Stability AI and anyone else who might attract funding by shipping a capable open model first.¹ That framing is worth keeping in mind the next time a lab frames an open release as a gift to the ecosystem rather than a competitive maneuver.

Simon Willison also spent part of the weekend reverse-engineering his own Claude Code install, chasing down a claim from Bun creator Jarred Sumner's post on rewriting Bun in Rust: that Claude Code v2.1.181 (released June 17th) and later ship on the Rust port of Bun. Willison didn't take the vendor's word for it — he ran strings ~/.local/bin/claude | grep -m1 'Bun v1' against his own binary and got back Bun v1.4.0 (macOS arm64), a version number that doesn't exist in any tagged Bun release (the latest is v1.3.14, from May 12th) but does match a commit that bumped package.json to 1.4.0 back in May and has sat there, unreleased outside of canary builds, ever since.² A second check — preloading a script that prints Bun.version before invoking claude --version — confirmed the same 1.4.0 string. Anthropic, in other words, is running an unreleased canary build of Bun's Rust rewrite in production across, per Willison's own estimate, millions of installs, and nobody noticed until someone went looking.² Sumner's own account is the vendor's: 10% faster startup on Linux, otherwise invisible. Willison's independent poke at the binary is the part worth trusting.

Mojang's own snapshot notes are the only source on this one, so take it as vendor-reported: in a snapshot posted Thursday — a day late, because "peak vacation season" hit the Swedish team's usual Tuesday cadence — Mojang confirmed it has switched Minecraft: Java Edition's window management, input handling, and platform integration from GLFW to SDL3.³ For a game with Minecraft's install base, swapping out the windowing layer is not a small change: keybindings now map to physical key positions via SDL scancodes instead of layout-dependent codes, Wayland becomes the preferred backend on Linux, exclusive fullscreen is gone on macOS, and the same snapshot slips in new core shaders for order-independent transparency.³ It's the kind of unglamorous infrastructure migration that a game this old and this widely deployed rarely gets to do cleanly.

Smaller, but worth a look if you write embedded firmware: Wren6991 — the same developer behind the RISCBoy RISC-V handheld — shipped CodeSizer, a static code-size profiler that answers the perennial "why is this binary so big" question properly. It runs objdump and addr2line against an ELF file to unwind the inlined call stack at every instruction address, then attributes code size to the correct node in an interactive HTML call tree — solving the real problem with LTO'd, heavily-inlined embedded builds, where a single function symbol can hide dozens of inlined callers and flat symbol-size tools tell you nothing useful.⁴ The sample report, run against the Raspberry Pi RP2350's bootrom ELF, is on GitHub if you want to see what the output actually looks like before pointing it at your own build.⁴

Trending today: the list is saturated with AI-agent wrappers, coding-assistant CLIs, and "awesome-llm-apps"-style curated lists — nothing there cleared the novelty bar; today's repo pick came from Lobsters instead.

↑ Back to top

THE WORLD

Spain Wins the World Cup; a King County Drug Bust Nets 38 Pounds of Meth

• Spain won its second World Cup title, beating Argentina 1-0 in extra time on Ferran Torres's 106th-minute goal in New Jersey.¹ - Argentina's Enzo Fernández was sent off for a second yellow card late in regulation, forcing his team through extra time a man down.¹ - It may have been Lionel Messi's final World Cup; Spain shut him down all night, and he left MetLife Stadium in tears.¹

Inflation dominated conversation at Sunday's "ReUnion on Union" block party in the Central District, where the Africatown Community Land Trust handed out housing and small-business resources alongside the music and food. Seattle-area inflation is up 4.5% over the past year, with produce prices up nearly 12%, clothing up nearly 15%, and fuel up 25.4%; residents told KIRO 7 they want elected officials to feel it too.²

King County Sheriff's deputies broke up a drug trafficking operation spanning King, Snohomish, and Pierce counties, serving warrants at two Lynnwood distribution hubs Saturday. The bust yielded 38 pounds of methamphetamine, 4.4 pounds of fentanyl powder, 1.7 pounds of heroin, more than $95,000 in cash, and three arrests.³

High bacteria counts have closed swimming at Gene Coulon Memorial Beach, Houghton Beach in Kirkland, Mount Baker in Seattle, and Salt Creek near Port Angeles; toxic algae has separately shut Lake Marcel near Carnation and Palmer Lake in Pierce County. County health departments say the beaches reopen once testing clears.⁴

KIRO 7 meteorologists have called Pinpoint Alert Days for Tuesday and Wednesday, with highs touching 90°F in Seattle and the mid-90s farther south as onshore flow shuts down and wildfire smoke drifts in from British Columbia. The heat breaks Thursday as marine air returns.⁵

Pierce County's council narrowly approved a Unified Regional Approach to homelessness in June, aiming to avoid the governance mess that has hollowed out the King County Regional Homelessness Authority.⁶ Advocates call it a genuine step forward but warn the county-controlled structure lacks independent accountability and rests on notoriously incomplete data — last year's official point-in-time count found 2,955 homeless residents against outreach estimates of 10,000 or more.⁶

ON THE ROAD — Adast, Hautes-Pyrénées, France

A lightning-sparked wildfire above Barèges — in the same stretch of the Pyrenees as Adast — reignited this past week after wind gusts fanned embers that had smoldered since the fire first broke out July 9. Firefighters redeployed a water-bombing helicopter to hit flare-ups in the steep, rocky terrain, and hikers are being told to stay off the massif because of rockfall risk; officials are also asking residents not to jam emergency lines reporting smoke that's already been spotted and logged.⁷

Separately, the Vallées des Gaves announced their first chamber-music festival for the 2026 season, a new addition to the valley's cultural calendar.⁸

Beyond those two items, local coverage for the reader's last day in the valley was thin — no other dated news turned up from the outlets serving Adast and the surrounding Gaves valleys.

↑ Back to top

THE LONG READ

Zero Successes in a Year and a Half, and Nobody Will Say So Out Loud

A customer called Mitsubishi about a car problem and got a very good robot. The voice was natural, the response was fast, the promise was clear: someone would call back. Nobody did. Six months later, that customer — an AI consultant who has spent the past year and a half watching corporate AI projects fail — was still waiting, and had quietly decided not to buy another Mitsubishi.¹ Somewhere in a dashboard, that call almost certainly reads as a resolved ticket, because a request that vanishes doesn't register as a failure. It just isn't there.

That anecdote is the small version of the argument in "AI Mania Is Eviscerating Global Decision-Making," an essay from the consultant behind ludic.mataroa.blog. The big version is blunter: in a year and a half of engagements — some run directly, some observed in passing while doing unrelated work — the success rate on corporate AI projects has been 0%. Not mostly succeeding. All of them failed.¹ The author is careful to note this isn't really an indictment of what LLMs can do; it's an indictment of what large organizations do to any sufficiently ambitious software project, with AI's novelty as an extra unpriced risk stacked on top of the usual ones.

What makes the essay worth forty minutes rather than a skim is that it doesn't stop at the failure-rate claim — it explains why almost nobody inside these organizations is able to say it out loud. Executives who tell the truth get removed. Employees who are honest get "randomly" selected for layoffs.¹ So the org chart routes around candor. Engineers who do their jobs competently without a chatbot now lie and say Claude did it — "AI-washing," in the author's phrase — because their managers are unhappy if the work wasn't visibly AI-touched. Elsewhere, staff are graded on "token leaderboards" where more spend reads as better performance, so the people with the skills to actually optimize a system instead set two LLMs looping against each other, walk away, and watch Netflix. Nobody has been caught, even when they privately think the output isn't fit to ship.¹

The essay's sharpest scene is a live demo of Snowflake's Cortex — an AI layer that, per the author's account of a briefing from Snowflake's own staff, runs at something like 92% accuracy on production data, which sounds impressive until you translate it into one wrong number in ten on a CFO's dashboard.¹ The consultant's team stopped pitching it. But on the occasions they showed it anyway, at a client's insistence, the room's mood would flip in real time — skeptical, budget-conscious buyers turning, in the author's words, into people plunging their hands into their own chests to present their "still-beating credit cards."¹ The team ended up refusing sales to any lead who showed more than passing interest in AI, because the pattern that interest predicted — magical thinking about vendor claims, willingness to buy anything with the right label — created legal and reputational exposure the consultancy didn't want.

The most useful section, though, is the one that explains why executives keep repeating numbers they privately doubt. It isn't simple dishonesty, or not only that. A Fortune 500 executive, speaking without microphones in the room, told the author that if a vendor's executive publicly contradicted a customer's inflated productivity claims, it would read as an attack on that customer's credibility and risk the contract. So the vendor stays quiet. But that vendor is also, in other deals, the customer — and faces the same pressure from someone below it in the chain. The result, as the author frames it, is a prisoner's dilemma played out across entire industries: cooperate (repeat the nonsense) and keep your job; defect (tell the truth) and get replaced by someone who won't. An anonymous CISO quoted in the piece put the mood among his peers plainly — "quietly skeptical but afraid to speak up" — and compared the atmosphere to cloud adoption, except with "a cult-like atmosphere to it that you didn't see with the cloud."¹

The essay's back half turns practical: how to raise concerns about a specific project without getting shot as a heretic (one-on-one conversations, anonymous polls on success likelihood, talking to the people who actually use the tooling day to day), and, for readers who've concluded the fight isn't winnable at their company, how to survive — contract instead of taking a salary, stop reading AI news, start a job search quietly the moment you're handed 2,000-line AI-generated pull requests to review.¹ None of it is triumphant. The essay's closing note is that this particular mania will pass, the way crypto's did, but that the underlying organizational credulity won't — it'll just wait for the next trigger.

The reader of this paper has almost certainly sat in a version of the rooms this essay describes. What the piece offers isn't a novel observation — plenty of people have noted that AI mandates produce theater — but a rare, granular, on-the-record account of the mechanism: who lies to whom, why the lie is individually rational, and why the aggregate effect is that large institutions currently cannot make a sound decision about the technology they've bet their strategy on.

↑ Back to top

FROM THE ARCHIVE

The Alarm That Did Not Force an Abort: July 20, 1969

Five minutes into the descent burn, 6,000 feet above the Sea of Tranquility, the guidance computer in the lunar module Eagle threw a 1202 program alarm.¹ Neil Armstrong and Buzz Aldrin had no idea what it meant. Neither, for a few seconds, did most of Mission Control. In Houston, a 24-year-old computer engineer named Jack Garman told guidance officer Steve Bales it was safe to keep going, and Bales relayed the call up the chain in time for Armstrong to fly through it.¹

The alarm wasn't a bug in the arithmetic. It was an executive overflow — the Apollo Guidance Computer's real-time scheduler running out of cycles because a rendezvous radar switch had been left in the wrong position, feeding it a second stream of position data it had no use for during a landing.¹ Software engineer Don Eyles later traced the root cause to an electrical phasing mismatch that made the stationary radar antenna look, to the computer, like it was oscillating — spurious interrupts stealing cycles the landing software needed.¹ Margaret Hamilton, who ran the Apollo flight software effort at MIT's Draper Lab, had built the AGC to shed low-priority tasks and preserve the ones that mattered rather than crash or force an abort. It did exactly that, several times, during the descent, and kept flying.¹

It nearly wasn't enough anyway. As Armstrong took semi-automatic control to steer clear of a boulder field near what would later be named West crater, Eagle burned through propellant faster than the flight plan allowed for.¹ At 100 feet, with only about 90 seconds of propellant remaining, the margin for an abort was closing fast.¹ He landed with about 25 seconds of that margin left, according to the numbers Mission Control was watching in real time — though post-mission analysis put the true figure closer to 50 seconds, the shortfall an artifact of propellant sloshing in the tanks and uncovering a sensor early.¹ Subsequent Apollo lunar modules carried anti-slosh baffles because of it.¹

Eagle touched down at 20:17:40 UTC on July 20, 1969.¹ Armstrong's full transmission to Houston — "Houston, Tranquility Base here. The Eagle has landed." — opened with an unrehearsed swap: Tranquility Base, the landing site's new name, in place of Eagle, the ship's call sign, a tell to Mission Control that the landing was real.¹ Six and a half hours later, at 02:56 UTC on July 21 by the clock but still "July 20" in the mission's frame and the world's memory, he stepped off the footpad and onto the regolith.¹

The interesting part of that story was never really the flag or the plaque. It was that the software failed safely under load it hadn't been designed to expect, degraded gracefully instead of crashing, and a 24-year-old caught the failure mode fast enough to say "go" instead of "abort." That's a systems-engineering result, not just a historical one.

↑ Back to top

THE FUNNIES

Zero Out of Three Hundred, and a 2 A.M. Spit Cup

*After Pearls Before Swine — on the long read's zero-for-three-hundred corporate AI success rate, and the quiet art of not saying so. After Calvin and Hobbes — on the Tour's nighttime anti-doping wake-up calls, imagined as a kid's bedtime routine nobody consented to.*

↑ Back to top

ALSO NOTED

Also Noted

• Merlier's Tour Ends Short of Paris — Soudal-QuickStep's three-time stage winner climbed off during Stage 15's second consecutive brutal mountain day, after his own director had flagged the time-cut risk before the start. cyclingnews.com Jul 19, 2026
• Del Toro Chose to Crash Rather Than Collide With Pogačar — When the peloton overshot a corner on Stage 15, Isaac del Toro deliberately hit the deck instead of taking down his UAE Team Emirates-XRG leader, then got paced back up by a grateful Pogačar. cyclingnews.com Jul 19, 2026
• Why Palantir Assigns a 1979 Improv Textbook to New Hires — Software engineer Sean Goedecke argues that Keith Johnstone's *Impro* — canonical reading across Silicon Valley founders and onboarding decks — reads less like an acting manual and more like a literal playbook for running a cult, mask-induced trance states included. seangoedecke.com

↑ Back to top

THE QUESTION

The Only Reliable Audits Are the Ones Nobody Asked For

Two disclosures landed in today's paper that only look unrelated. Both are about the same failure mode: an institution's official account of what it is doing can diverge from the truth of what it is doing for years, and the gap closes only when someone with no stake in the official story bothers to go looking.

The first came out of a courtroom, not a shareholder letter. As THE LAB reports today, an October 2022 email from OpenAI's CEO to its board — surfaced only because litigation forced it into evidence — lays out the real reasoning behind the company's occasional open-weight releases: ship something capable enough to "discourage others from releasing similarly-powerful models" and "makes it harder for new efforts to get funded."¹ That is a materially different story than the one told publicly at the time, about openness and ecosystem contribution. Nobody volunteered the real version. It took a subpoena.

The second came from an engineer checking his own machine instead of trusting a blog post. A developer's claim that Anthropic's coding tool now ships on an unreleased canary build of a competitor's runtime should have been unverifiable without inside access — except that Simon Willison ran strings against his own installed binary, found a version number that matches no tagged release, and confirmed it with a second, independent test.² Anthropic never announced the swap. It took a stranger's curiosity and a command-line utility to find it, running — by Willison's own estimate — across millions of production installs.

Put next to each other, the two stories describe the same structural fact from opposite ends: the account an organization gives of its own system and the actual behavior of that system are not the same document, and nothing inside the organization is designed to reconcile them. This paper's long read today makes the mechanism explicit from the other side of the table — a consultant's account of what happens inside large organizations trying to adopt AI, where the people closest to a system's real performance have the strongest incentive not to say what they've seen. Candor gets you removed. Silence gets you promoted. The report writes itself in whatever direction keeps the org chart intact.

Which leaves an uncomfortable asymmetry. Verification happened today in both directions — but only because a plaintiff's lawyer wanted an email and an engineer wanted to know why his terminal felt a little faster. Neither is a system. Neither scales. For every claim that gets checked because litigation forces it into discovery or because someone happened to run the right diagnostic, there's an unknown number that don't — no lawsuit pending, no engineer curious enough to open the binary. The honest answer to "how do you know what this system is actually doing" is, right now, "you mostly don't, unless someone with nothing to lose decides to check." That's not a comfortable place to build a year's worth of decisions on. It's the place most organizations are building them anyway.

↑ Back to top