THE PELOTON

Switzerland Signs Off: Reusser Delivers a Masterclass at Villars-sur-Ollon

↩ Developing story — first reported Jun 19 · previously Jun 20

VILLARS-SUR-OLLON, Switzerland — The same mountain, the same finishing climb, the same day: Sunday on the Col de la Croix became a clinic in how to ride a stage race when you already hold the lead. In the women's finale, the GC leader came into stage 5 having won the time trial the day before and left the decisive climb with the race settled — a third career Tour de Suisse Women title, second consecutive, decided with an acceleration in the final kilometre that neither of the two riders still in contention could answer.¹

The stage had opened up at the midpoint when Femke de Vries (Visma-Lease a Bike) attacked and drew out a front group of five: de Vries, Megan Arens (Picnic PostNL), the two GC contenders — Kasia Niewiadoma-Phinney (Canyon-SRAM) and the leader — plus Cédrine Kerbaol (EF Education-Oatly), who bridged after a short chase. Arens was shed on the descent of the Col de la Croix, leaving four going into Villars-sur-Ollon. Niewiadoma-Phinney attacked with 8.8km remaining; only one rider could hold her wheel.¹ Kerbaol was over 30 seconds down at one point, then clawed her way back in the final kilometre and immediately attacked for the stage — only to be outjumped from behind. "I could profit from the situation we had in the GC," the winner said at the finish. "Kasia did a lot of work on this climb, and that helped me a lot."¹

"I'm super happy with how everything went," she added. "It's not that I came into this race with the biggest confidence, so I'm super happy that it worked out."

The win carries an edge of timing beyond the sporting. The victorious rider competes for Movistar, whose title sponsor Telefónica is reported by Bloomberg to be seeking an exit from its €25 million-per-year contract — a story that has hung over the team since it broke earlier this month.³ The women's squad has been the team's most visible success story in recent years, and one of its standard-bearers just won a race for the third time. Whether that matters to a telecoms company reorganising around its Spanish operations is a separate question.

On the men's side, the final stage was rolling at press time with Tadej Pogačar (UAE Team Emirates-XRG) holding a 4:22 lead over Richard Carapaz (EF Education-EasyPost) at the gun.² UAE Team Emirates-XRG controlled pace on the first ascent of the Col de la Croix. The gap barely merits discussion — a four-and-a-half minute deficit with a single mountain stage to go is, in practical terms, already a final result. Pogačar won Saturday's time trial by less than a second from Mathieu van der Poel at 53.4 kph.² Stage 5 results are in the results field above.

Elsewhere in the sport, Netcompany-INEOS is now past the Tour de Suisse — the race whose conclusion the team had identified as the trigger point for its decisions on Josh Tarling (broken collarbone) and Oscar Onley (dislocated shoulder, ravine crash). Both riders have been at the pre-Tour camp; team doctors were expected to rule on Tour starts after Sunday. Thirteen days remain before the July 4 Grand Départ.

↑ Back to top

THE LAB

Anthropic's Own RCT Shows AI Makes Engineers Worse at Engineering

The study researchers at Anthropic ran on recruited software engineers is the one that should make people uncomfortable. Fifty-two software engineers were given a basic coding task and divided into two groups: one used an AI assistant, one did not. The result, reported today in Nature, is that the AI-assisted group scored significantly lower on comprehension afterward — the skills the task was supposed to build did not form.¹ "No lesson learnt" is the subheading the Nature editors chose.

The colonoscopy data is equally blunt. A study of Polish physicians — each with at least 2,000 colonoscopies behind them — gave them access to an AI system that flags adenomas in real time, but only on some days. Before AI: adenoma detection at 28.4%. After several months with the tool, detection on AI-off days dropped to 22.4%.¹ In three months, seasoned endoscopists lost six percentage points of diagnostic accuracy on the days the crutch was gone. Yuichi Mori, a physician-researcher at the University of Oslo who co-authored the study, told Nature: "There is no established solution against deskilling right now."¹

What makes this a different kind of story is the source. This isn't an outside researcher poking at OpenAI's user data. Researchers at Anthropic designed and ran the coding RCT on recruited engineers, and published the result. The finding is that engineers who used an AI assistant during the task showed measurably lower comprehension of the material they had just worked through — at a level that was detectable in a study of 52 people. The sample is small enough that the authors will hedge in fine print, but the direction is clear.

The practical question for any engineer reading this is the one Kevin Crowston, an information scientist at Syracuse University, puts bluntly in the piece: "Just being aware that this phenomenon exists hopefully provokes some self-reflection about which skills people want to maintain and which they're willing to outsource."¹ That framing — outsourcing versus losing — is doing a lot of work. Skills you deliberately choose not to exercise are different from skills you didn't notice eroding. The problem is that AI tools make the erosion invisible until the day the tool is unavailable.

A different kind of careful work: a reverse-engineering post published June 18 by the blog calif.io dissects what Apple actually means when it says it has "started writing parts of the core operating system kernel in Swift." The headline at WWDC implied a memory-safe kernel rewrite. The reality, uncovered by grepping through the iOS and macOS 27 kernelcache, is narrower and more interesting than either the hype or the dismissal.

The author found Apple's Embedded Swift runtime inside com.apple.kec.pthread on macOS 27 — 2.4 KB of __TEXT_EXEC.__text, statically linked.² The effort has a name internally: KernelKit. It gets its own Mach-O platform constant (25 on macOS, 26 on iOS — previously undocumented; the author extracted the table by calibrating against known entries in the Xcode 27 beta linker). The linker hard-codes /System/KernelKit/usr/lib/swift as a search path. There is a TARGET_OS_KERNELKIT preprocessor conditional. The toolchain can already target it. Apple just hasn't published the SDK.

What the Swift runtime in the kernel actually does right now: essentially nothing. The author scanned all 370 macOS kernelcache fileset entries for swift_* references and found that 23 of the 37 Swift symbols are global exports — and not a single other component in the kernelcache imports them.² Only swift_dynamicCast calls swift_release to clean up after itself; zero other kexts link against the runtime. XNU itself remains entirely C/C++: the KDK's DWARF compile units for the M5 Pro SoC break down as 1,855 C11 files, 249 C++14 files, 2 assembler files, zero Swift.² The author verified this on t6041 (M4 Max) and the KASAN build as well — it can't be hiding behind stripped symbols.

The read-out: Apple shipped the SDK plumbing and the runtime blob first so they can watch it not break anything for a beta cycle or two before landing actual Swift kernel components. iOS 27 gets the platform ID (26) but not yet the linked runtime. macOS comes first, then iOS later. This is the DriverKit playbook — its own directory under /System, its own SDK, its own platform constant — repeated seven years on. The call to action in the post's closing line is characteristically understated: "Also if anyone at Apple wants to leak KernelKit.macosx.sdk I will treat it with the respect it deserves."

A deadline worth noting before Tuesday: three Microsoft-signed certificates that anchor the Secure Boot chain of trust expire June 24.³ These are the cryptographic linchpins of UEFI Secure Boot on both Windows and Linux — the chain that verifies firmware signatures at boot time before the OS loads. Systems that haven't received the relevant updates will lose Secure Boot protection. Wired's coverage today explains the full threat model (UEFI bootkits, LoJax, MosaicRegressor), but the operative question for most people is whether their systems have received the Microsoft update pushing new certificates. Server hardware and custom Linux setups with manual Secure Boot key management are the highest-risk category. Check your UEFI settings and confirm your certificate store is current before the 24th.

OCaml 5.5.0 shipped June 19 with a set of language additions worth noting: module-dependent functions (lightweight functors passable as arguments with dependent types), inline higher-rank polymorphism via explicit type annotations, external types for FFI (type int_gmp = external "mpz_t" — provably distinct from other types even outside the current module), and generalized local definitions. The GC gets a new idle phase and a sweep-only startup phase, both aimed at reducing latency on small heaps. There are also two CVEs patched — CVE-2026-28364 in intern.c and CVE-2026-41083 in Filename.quote_command on Windows.⁴ The relocatable compiler is the quality-of-life win most people will notice first: you can now copy or move an OCaml installation without bytecode interpreter mismatches.

Trending today: GitHub is saturated with AI image/video wrappers and Claude skill fork repositories — nothing on the trending list this weekend cleared the technical novelty bar.

↑ Back to top

THE WORLD

Cormorant Cove Closed; Heat Advisory Monday

• Puget Sound advisory — King County's West Seattle pump station leaked 13,000 gallons of untreated wastewater June 17; Cormorant Cove closed through June 24.
• Heat advisory Monday–Tuesday — NWS issued the advisory for the Seattle area; 90-degree weather is likely Monday and Tuesday, then temperatures back down toward the end of the week.
• Lake Washington drowning — A man swimming near Holmes Point off O.O. Denny Park went under around 1:30 a.m. Saturday; search has pivoted to recovery.

ON THE TRAIL

• West Tiger 3 / Seattle View Trail (Issaquah Alps) — Foxgloves are at peak on the Seattle View Trail right now — "the entire hill is foxgloves," with penstemon, columbines, and tiger lilies on the W Tiger 3–2 saddle. Afternoon parking was easy Saturday. 7 miles. (WTA, Jun 20)
• Kendall Katwalk via Gravel Lake (I-90 / Snoqualmie Pass) — Trail to the Katwalk is snow-free; slushy snowfields begin about a half-mile past it. Bring poles and plan your water early — virtually nothing after mile 3. Arrive by 7 a.m. for the main lot. (WTA, Jun 20)
• Taylor Mountain (Issaquah Alps) — An 8-mile loop with a bear sighting Saturday — mama and cub spotted from about 300 yards. Well-maintained, forested throughout, ~1,000 ft gain. (WTA, Jun 20)
• Weather outlook — Get out today (77°F, sunny in Issaquah) or hold off until the heat breaks. Monday and Tuesday bring a Heat Advisory with highs near 83–87°F; next weekend (Jun 28–29) looks wet, with rain likely Saturday (66°F in the Issaquah Alps).

↑ Back to top

THE LONG READ

The Tree That Held Him: Oscar Onley, a Ravine, and the Tour de France

↩ Developing story — first reported Jun 17 · previously Jun 18

The television cameras missed it entirely.¹ Team staffers riding in the car behind spotted Oscar Onley's bike lying in the road and stopped — but Onley wasn't there. He had gone over the barrier and simply disappeared.

What they found, eventually, was a 23-year-old Scottish climber suspended in the crowns of a tree above a ravine dozens of meters deep. Staff from the team car and race organization had to hoist him back up by hand. Onley crossed the finish line 29 minutes and 17 seconds behind stage winner Maxim van Gils, then didn't start the next day.¹

The crash happened on the descent of the Côte d'Héry-sur-Ugine, with less than 10km to go in stage 6 of the Tour Auvergne-Rhône-Alpes on June 13. Dave Brailsford described what happened to Dutch outlet Wielerflits the following morning. "Fortunately, Oscar landed in the crowns of a tree during his fall," he said. "As a result, he didn't end up in a ravine dozens of meters deep. Our men from the team car, together with people from the organization, were able to very carefully hoist him back up from that tree. He was incredibly lucky. It could have ended much worse."¹

Geraint Thomas, speaking to Cyclingnews, kept it brief: "I think it was a left-hander, he just overshot slightly. We're staying positive. It could have been much worse." Matteo Jorgenson, who was descending behind him, told Eurosport he saw Onley go over the edge: "That kind of spooked me a little bit." Isaac del Toro — the eventual race winner — was overheard at the finish asking other riders if there was any news.

That Onley dislocated a shoulder, suffered multiple wounds, and is already back at Netcompany-Ineos's pre-Tour training camp tells you something about the alternative outcome everyone around that descent was contemplating.¹

The timing could not be worse, or more consequential. The 2026 Tour de France starts in Barcelona on July 4. As of this paper's last report, the Netcompany-Ineos roster for the Grand Départ will not be finalized until after the Tour de Suisse concludes today — doctors need to assess both Onley and Josh Tarling, who underwent surgery Saturday for a fractured left collarbone suffered in a separate crash on the same day.

A team preparing for the Tour with its two biggest question marks sitting out the final tune-up race is not ideal. Filippo Ganna, Carlos Rodríguez, Michał Kwiatkowski, Kévin Vauquelin, Dorian Godon and Thymen Arensman are all considered likely starters.¹ That's a credible eight if you add two from the reserve pool. But Onley is different. He finished fourth at the 2025 Tour — a result that, for a 22-year-old, generated the kind of attention that reshapes a team's dynamics heading into the next year.¹ None of the likely starters carry that particular weight of expectation into July.

American Andrew August, who is racing the Tour de Suisse and has posted his first two professional victories this season, is among the reserve riders who could be called up if Onley or Tarling miss the start.¹

Onley has not spoken publicly about the crash. His presence at camp is the signal the team is willing to share. The decision comes today.

↑ Back to top

FROM THE ARCHIVE

The First Program: June 21, 1948

Interior of a 1948 university computing laboratory, pen-and-ink. A massive machine stretches across the room — seventeen feet of steel racks, cathode ray tubes, patch cables, and hand-soldered components salvaged from wartime surplus. A lone engineer in a rumpled jacket leans toward a glowing Williams tube, one hand resting on a panel of toggle switches, pencil tucked behind his ear. Cross-hatched shadows pool beneath the racks; the tube's phosphor screen radiates a small halo of light in the dark room. Overhead industrial lamps illuminate coiled cables snaking across the floor. Through a tall window behind, the gothic stonework of a Victorian university building is visible in pale winter light. No text, no color — only confident black linework, dense hatching on the machine's metal surfaces, and stark white space around the engineer's intent, solitary figure.

The program was 17 instructions long. It had one job: find the highest factor of a given number. And on June 21, 1948, at the Victoria University of Manchester, it ran — for the first time in history — on a machine that actually remembered it.¹

The machine was the Manchester Small-Scale Experimental Machine, which its builders called Baby. The name was a joke. At 17 feet long and nearly a ton, it was only "small" compared to the ENIAC, which stretched 95 feet and weighed 27 tons and, notably, could not store a program at all.¹ What Baby proved was that a cathode ray tube — the Williams-Kilburn tube, developed by Freddie Williams and Tom Kilburn — could serve as random-access memory. Every computer program you have ever written depends on that proof.

Kilburn wrote the program. It found the highest factor of 2^18. The answer is 131,072, and Baby arrived at it by running 3.5 million calculations in 53 minutes.¹ The hardware was war surplus, some of it salvaged from Bletchley Park — the same codebreaking infrastructure that had spent the war cracking German ciphers was now being repurposed into the seed of the stored-program computer.

The demonstration wasn't a product launch. It was a concept proof. But the concept, once proved, moved fast. By August 1948, construction of the Manchester Mark 1 had begun.¹ By 1949 it was operational. By 1951, the Ferranti Mark 1 — its commercial descendant — became the world's first commercially available computer.

Seventy-eight years later, the question Baby answered in 53 minutes takes a modern chip a few nanoseconds. The interesting thing is that the architecture underneath — a machine that fetches instructions from memory and executes them in sequence — is the same architecture. Kilburn's 17-instruction program is ancestral to every binary you've ever shipped.

↑ Back to top

THE FUNNIES

The Funnies: Deskilled / First Run

*After Pearls Before Swine — on the Anthropic RCT finding that engineers who delegate their code to an AI assistant leave the session understanding it less than when they arrived. After Bloom County — on the Manchester Baby's seventeen-instruction program, which ran on June 21, 1948, and spent 53 minutes doing what a skeptic reckoned he could have done in his head.*

↑ Back to top

ALSO NOTED

Also Noted

• Kirkland e-bike collision — A child was taken to Harborview Medical Center in serious condition after colliding with a car while riding an electric bike at NE 124th St. and 109th Ct. NE in the Totem Lake neighborhood Saturday. kiro7.com Jun 20, 2026
• Seattle takes the dim sum title — On June 18, the Chinatown-International District set a Guinness World Record for the largest dim sum party: 830 verified participants on S. King Street, beating Sydney's 2019 record of 764. kiro7.com Jun 21, 2026
• Where LLMs actually earn their keep — A software engineer's post enumerates use cases where LLMs reduce genuine drudgery: triaging API endpoint alerts from alert to root-cause log analysis (15+ minutes cut to 1–2), embedding-indexed customer call transcripts for product discovery, and stripping long technical videos down to whether they're worth watching. aggressivelyparaphrasing.me Jun 21, 2026
• C++26 cleans up std::format — The C++26 formatting library adds native pointer formatting (no more `reinterpret_cast` to `uintptr_t`), a `std::formatter` for `std::filesystem::path` that correctly handles Windows UTF-16-to-UTF-8 transcoding, `constexpr std::format` for integers and strings, and renames `std::runtime_format` to `std::dynamic_format`. mariusbancila.ro Jun 19, 2026
• Distrobox rewrites in Go — The container tool Distrobox has released v2.0.0-rc.1, a complete rewrite from shell to Go: single binary (dropping separate `distrobox-enter`, `distrobox-create` executables), same CLI interface and config-file compatibility, existing v1 containers work except for exported apps which must be recreated. distrobox.it Jun 20, 2026
• Fourier epicycles on a cube — A visual/mathematical piece constructs a 3D human-face from Fourier epicycles by projecting the face surface to three 2D planes (xy, yz, zx), computing coefficients separately for each, then recombining endpoints — the three sets of rotating circles are displayed on three faces of the cube simultaneously. andreinc.net Jun 15, 2026
• SafeR by the numbers — The UCI's mid-season SafeR report (pre-Tour de France) counts 297 race incidents analyzed in the first half of the year; the most common cause is rider error (29%), with positioning pressure at key points (cobbles, sprints, climbs) at 12%; riders account for 42% of yellow cards issued and team staff for 37%; a gear-ratio cap test is scheduled for Tour of Guangxi in October. uci.org Jun 2025
• Twenty-two years ago today — On June 21, 2004, SpaceShipOne pilot Mike Melvill crossed the Kármán line at 100.124 km above Mojave, becoming the first commercial astronaut in history and the first person to reach space in a privately funded, non-government vehicle; the 24-minute flight led directly to the Ansari X Prize win that October. airandspace.si.edu Jun 21, 2026
• OpenCode archived; project lives on as Crush — The `opencode-ai/opencode` terminal AI coding assistant repository has been archived and the project has moved to `charmbracelet/crush`, developed by the original author in collaboration with the Charm team (makers of Bubble Tea, Lip Gloss, and other Go TUI tooling). github.com Jun 21, 2026

↑ Back to top

THE QUESTION

When You Build a Machine to Hold the Knowledge, Can You Still Hold It Yourself?

The whole point of Baby — the Manchester machine that ran its first program seventy-eight years ago today — was that a computer could hold knowledge in memory so a person didn't have to. That was the proof: store the program, retrieve it, execute it, and the machine does the work. Every stored-program computer since is a direct descendant of that idea. The question this edition raises is what happens when the knowledge being stored is not a factored integer but the skill of the person sitting at the keyboard.

As THE LAB reports today, Anthropic's own researchers ran a randomized controlled trial on 52 software engineers — one group used an AI assistant on a coding task, one did not — and the AI-assisted group showed measurably lower comprehension of the material afterward.¹ Not lower performance during the task. Lower understanding after it. The skill the task was supposed to build did not form. A parallel finding from a study of Polish colonoscopists, published in the same Nature piece, showed that experienced physicians lost six percentage points of adenoma detection accuracy on the days the AI diagnostic tool was unavailable — measurable degradation in expert judgment, in physicians with at least 2,000 procedures behind them, within months of using the tool.¹

This is the version of the problem that the stored-program computer framing makes legible. Baby in 1948 was designed to augment human computation — to hold what humans found expensive to hold, so humans could do more. The original bargain was accumulation: the machine adds capacity, the human retains their own. What the colonoscopy data suggests is that the bargain has a hidden term. When the tool is present long enough, the skill migrates. It doesn't stay in both places. The person offloads it to the machine, and the machine holds it exclusively, and on the days the machine is gone, so is the skill.

Kevin Crowston, an information scientist at Syracuse University, put the operative question to Nature this way: which skills do you want to maintain, and which are you willing to outsource?¹ The distinction between "outsourcing" and "losing" is doing real work in that sentence. Outsourcing implies a decision, a conscious choice to delegate a task while retaining the underlying competency. What the RCT suggests is something different — not delegation, but erosion. The engineers didn't decide not to learn. They used the tool, and the learning didn't happen.

Baby proved that machines can remember. The open question, seventy-eight years on, is whether the thing we built to remember for us has started making us forget.

↑ Back to top