THE WORLD

Trump Claims Iran Peace Deal, Markets Surge; Seattle Shuts Down Aurora

• Iran deal claimed — Trump says he called off strikes and a peace agreement is "subject to finalisation." Iran hasn't confirmed. S&P +1.8%, Nasdaq +2.5%. Al Jazeera
• SCOTUS blocks nitrogen execution — Supreme Court stopped Alabama from executing Jeffery Lee via nitrogen gas, citing substantial risk of severe pain. Three justices dissented. NPR

Seattle is shutting down several streets around Aurora Ave. N this summer — closing intersections at N 96th, 98th, 100th, and 102nd Streets — after neighbors put up their own makeshift barriers around what Councilmember Maritza Rivera called "the second-busiest track for juvenile sexual exploitation in the country." Mayor Katie Wilson announced the plan Thursday; it starts in about two weeks and runs through summer. Twenty-three active shooting cases are open along the corridor. KIRO 7

About 1,300 fans packed Everett's waterfront Thursday for a free World Cup watch party as Mexico beat South Africa 2-0. The next fanzone event is Friday — gates open at 4 p.m. for the US vs. Paraguay match. KIRO 7

Heat is building: KIRO 7's meteorologists have issued Pinpoint Alert Days for Sunday and Monday, calling for the hottest temperatures of 2026 so far around Puget Sound. Sunday hits 84°F in Issaquah, Monday 86°F.⁵

ON THE TRAIL

Trip window: This weekend is Sat Jun 14 – Sun Jun 15. Heads up for next weekend: Juneteenth (Jun 19) is a federal holiday observed on Friday — the reader's long weekend runs Fri Jun 19 – Sun Jun 21, with Thu Jun 18 as opportunistic if conditions are great. Both windows look sunny and warm.

── PART 1 — WEEKEND PICKS ──

Conditions as of Jun 11: the Snoqualmie corridor is snow-free at lower elevations after 3.7" of rain Jun 8–10, streams are running high, and upper-elevation trails still carry snow patches. The North Cascades Hwy 20 corridor is in excellent shape. Weather for the weekend (Sat–Sun): sunny, 0% precip across all target regions.

---

Pick 1 — Annette Lake (Snoqualmie Pass) - Region + drive: I-90 / Snoqualmie / North Bend — ≈35–55 min - Trip length: 1-night - Mileage/gain: ≈7.3 mi RT, ≈1,875 ft gain to lake and camps - Weather: "Sunny, with a high near 79°F Saturday. Sunny, with a high near 84°F Sunday. Precip 0% both days." - Why it clears criteria: Jun 11 report confirms snow-free trail, creek flows are low (easy rock-hop crossings), no bug complaints, one person heading in to spend the night (genuinely uncrowded midweek; weekends see more use but reports don't flag permit/lottery). Water at the lake. No wading required. - WTA trip report — Jun 11

---

Pick 2 — Greenwater Trail to Greenwater Lakes / Lost Lakes (Chinook Pass – Hwy 410) - Region + drive: Mt Rainier (Greenwater/Crystal/White R.) — ≈110–140 min - Trip length: 1-night (camp just above the river at 3 miles; 2-night possible) - Mileage/gain: ≈6 mi RT to camping area, minimal elevation gain (flat river trail) - Weather: "Sunny, high 68°F Saturday. Sunny, high 74°F Sunday. Precip 0–1% both days." - Why it clears criteria: Jun 11 report notes 3 cars at trailhead (low crowds), no snow, no bug complaints, trail saturated but passable, water flowing along Greenwater River, wildflowers everywhere. No fords. No permit lottery. - WTA trip report — Jun 11

---

Pick 3 — Thunder Creek to McAllister Camp (North Cascades – Hwy 20) - Region + drive: North Cascades (Hwy 20) — ≈150–190 min - Trip length: 1-night - Mileage/gain: ≈13 mi RT (parking at Thunder Knob adds ~1 mi), minimal elevation gain (flat creek trail to McAllister Camp) - Weather: "Sunny, high 82°F Saturday. Sunny, high 87°F Sunday. Precip 0% both days." - Why it clears criteria: Jun 11 report is emphatic — no mud, no snow, no downed trees, no bugs, only a handful of hikers all day, easy small crossings (rock-hop, no wading), bunchberry wildflowers blooming. McAllister Camp has water from Thunder Creek. Hwy 20 is fully open to this corridor. Note: mountain lion warning sign was posted at trailhead on Jun 11 — make noise, standard precautions. - WTA trip report — Jun 12

---

── PART 2 — REGIONAL SNAPSHOT ──

• Issaquah Alps (Tiger/Squak): Tiger Mountain Cable Line trail in excellent shape — firm tread, great grip even after heavy rains, views opening up as clearcut regrows. A few bugs at WT3 summit (one bite reported). Good easy-access option. WTA Jun 11
• I-90 / Snoqualmie Pass: Snow-free at lower elevations (Annette Lake, Denny Creek, Rachel Lake confirmed clear). Upper routes getting mixed: Gem Lake has some snow patches but microspikes not needed; Melakwa Lake has a sketchy swift river crossing — avoid for now. Recent 3.7" rain left streams running high across the corridor. Multiple WTA reports, Jun 11
• Mt Rainier / Hwy 410: Greenwater Lakes area is excellent and uncrowded. Gobblers Knob (via Lake George) has intermittent easy snow after the lake, clear at the fire lookout — solid option. Sheep Lake / Sourdough Gap has a large snowfield at the gap itself; bring microspikes. WTA Jun 11
• Mountain Loop Highway: Heather Lake is accessible with muddy sections but no obstacles — campsites clean and quiet mornings. Glacier Basin requires scrambling and has slippery wet rock sections above Monte Cristo; suited for experienced hikers only. WTA Jun 11
• North Cascades (Hwy 20): Thunder Creek corridor is in standout condition — the cleanest reports in the batch. Ruby Mountain (Fourth of July Route) is doable but has one icy snow patch around 6,700 ft; trail runners workable with kick-steps but microspikes recommended. Sauk Mountain: rough road in, some snow on the back side; give it another week or two for ideal conditions. WTA Jun 11
• Olympic Peninsula: Duckabush River (Hood Canal) is in great shape — wildflowers peaking, trail crew-maintained, easy to moderate grade. Road to trailhead is rough with potholes but passable by any car. Mount Townsend (northern approach) is snow-free with no obstacles — extraordinary summit views on clear days. WTA Jun 11

↑ Back to top

THE LAB

Unity's Float Math Is a Lie: Aras Digs Into the Double-Precision Rot

A tweet claiming that System.MathF is faster than UnityEngine.Mathf is technically correct, Aras Pranckevičius confirmed this week — but the actual story runs considerably deeper, and if you write Unity C# that cares about performance, you should read his full breakdown.

The root problem is old: Unity's Mono runtime, for historical reasons, computes all 32-bit float math in 64-bit double precision. The sin, cos, sqrt, pow, floor, round, and about every other Mathf function you reach for does an implicit cast to double, calls the double-precision implementation, then casts back. System.MathF, which arrived in .NET Core 2.0 (2017), calls float-native instructions directly. On a Ryzen 5950X, a tight loop of ten million square roots runs in 186ms with System.MathF versus 282ms with UnityEngine.Mathf.Sqrt in the Unity editor — a 34% gap. In release mode: 149ms versus 242ms.

Pranckevičius went further and annotated the actual x86 machine code generated by Mono for Mathf.Sqrt. The result is unsettling: a single v += Mathf.Sqrt(v) expands to roughly a dozen float-to-double and double-to-float conversion instructions, culminating in an x87 fsqrt rather than a modern SSE sqrtss. The System.MathF path is shorter and uses sqrtss, but Mono's "do everything in doubles" trait still forces a handful of redundant conversions around the call. Neither path is what you want.

The picture gets more fractured when you add the other two C# backends. Under IL2CPP, Mathf.Sqrt and System.MathF.Sqrt both compile to single-precision sqrtss and land at the theoretical floor of ~35ms — IL2CPP is smarter than Mono about recognising single-precision intent. Under Burst, the situation reverses: System.MathF functions fail to compile at all (Burst compile error), but Unity.Mathematics.math.sqrt hits the same 34–35ms floor with a single vsqrtss instruction. The one consistent piece of advice: if you can run Burst, use Unity.Mathematics; if you can't, System.MathF beats Mathf on Mono.

There is also a subtle correctness footnote. All Unity backends print the result of the benchmark loop as 24212990000000.0, a number that does not exist as a valid IEEE 754 single-precision float. The native .NET and C++ builds print 24212987183104.0. Somewhere in Unity's pipelines, values are silently rounding to non-representable intermediates. Pranckevičius notes that Unity's upcoming CoreCLR migration — discussed at a GDC 2026 talk — should clean up much of this, but he is not holding his breath on math-function precision consistency until that transition actually lands.¹

Simon Willison handed Claude Fable 5 a screenshot of a scrollbar bug and a one-line prompt. He left the room. When he came back, the model had: spun up a local dev server with fake environment variables it deduced from the repo, iterated through Firefox and WebKit in Playwright, figured out that osascript assistive access was blocked and routed around it using pyobjc-framework-Quartz to enumerate window IDs, patched Datasette's own HTML templates to inject a JavaScript setTimeout that fires the correct keyboard shortcut 1.2 seconds after page load, built a miniature Python CORS server bound to 127.0.0.1:9999 to capture DOM measurements, and opened real Safari to verify the fix. The eventual bug turned out to be a two-line CSS change. The session cost ~$12.11 at full API prices.²

Willison's framing is measured but pointed: the same proactivity that solved his CSS bug would be equally effective at exfiltrating data if the agent were acting on malicious instructions injected through a prompt-injection attack. Fable is aware of this — it eventually hit an internal guardrail and downgraded itself to Opus mid-session, which continued using the techniques Fable had pioneered — but the awareness is cold comfort. "If it does get subverted by instructions," Willison writes, "the amount of damage it can do given its relentless proactivity is terrifying." He has been calling unsandboxed coding agents his top candidate for a Challenger-scale incident; this session did not change that view.

The AUR supply chain breach this week makes a useful parallel. An infostealer campaign swept through more than 400 Arch Linux User Repository packages, prompting Arch maintainers to spend days resetting and deleting malicious content and banning compromised accounts. AUR packages run arbitrary unsandboxed code by design — the same structural assumption that makes agents convenient makes both dangerous to run without isolation.³

Homebrew 6.0.0 shipped on June 11 with a direct response to exactly this class of problem: a new tap trust model that requires third-party taps to be explicitly trusted before their Ruby runs locally, plus a Bubblewrap sandbox for Linux builds to match the macOS sandbox that has existed for years. The supply-chain security page in the new docs is worth reading if you manage any machines through Homebrew.⁴

Trending today: GitHub is saturated with Claude Code skill collections, CLAUDE.md configuration forks, and AI agent wrapper repositories — no entry demonstrated technical novelty that cleared the bar.

↑ Back to top

THE PELOTON

The Morning After: Van Aert Out of TARA Before the Mountains Begin

A lone road cyclist grinding up a steep alpine switchback, head bowed low over the handlebars, shoulders hunched, jaw set. The figure's jersey is soaked through with effort, number pinned crooked. Rocky mountain faces rise behind, snow-patched ridgeline fading into haze. Other riders visible in soft-focus further up the climb. Strong diagonal of the road cutting across the frame. Cross-hatched shadows on tarmac cracks and gravel edges. No crowds, no celebration — only the private arithmetic of pain and forward motion.

AUVERGNE-RHÔNE-ALPES, France — He won the stage on Thursday, celebrated with the team, and then woke up Friday morning and couldn't start. Wout van Aert's withdrawal from the Tour Auvergne-Rhône-Alpes before stage 6 landed not as a surprise — the crash damage to his elbow has been visible all week — but as a sharp reminder of how fragile the Belgian's 2026 season remains.

Visma-Lease a Bike confirmed the decision on social media: "He is still experiencing significant discomfort in his elbow due to the crash."¹ Van Aert will return to Belgium for further examinations following the pre-race TT training crash that had him bandaged at the start line all week. The Belgian national championships on June 26 and the Tour de France on July 4 are still the targets, but he now heads into those without the mountain work that stages 6, 7, and 8 of this race were supposed to provide.¹

The timing is pointed. Stage 5's sprint win — the paper's lead story Thursday — looked like Van Aert threading the needle: get the legs going, bank some confidence, stay out of trouble. Instead he's on a plane home, and Visma's Tour preparations are once again running behind schedule. He's not the only departure: Michael Matthews (Jayco AlUla) pulled out to focus on Tour preparation, Matevž Govekar and Hannes Wilksch both withdrew due to illness, Finn Fisher-Black (Red Bull-Bora-Hansgrohe) also abandoned, and Pau Martí Soriano (NSN) pulled out sick.² That's five non-starters before the mountains even began.

Stage 6 — 182.3km from Saint-Vulbas to the summit at Crest-Voland — is the first of three consecutive mountain days that will decide the race.² Alex Baudin (EF Education-EasyPost) holds the GC lead from his breakaway win on day one, with Netcompany Ineos co-leaders Oscar Onley and Kévin Vauquelin close, and Matteo Jorgenson (Visma) within reach. Paul Seixas and Isaac del Toro, the pre-race favourites, enter stage 6 with ground to make up. Decathlon were driving the chase when the day's racing began, leaving their leader exposed in the second group after an early split at 170km to go carved the peloton in two. A 60-rider front group — including GC threats Luke Tuckwell (1:03 down), Bruno Armirail (1:20), and Guillaume Martin (1:51) — formed and held a lead of nearly two minutes as the race approached its first climbs.²

Chloé Dygert's 2026 season is over. The Canyon-SRAM rider, who came into the year hoping a fresh start would end two years of bad luck, announced via Instagram that months of rehabilitation following her Paris-Roubaix crash have not gone as planned. A RED-S (Relative Energy Deficiency in Sport) diagnosis has complicated her return to training.³ "Well, that comeback lasted about two rides," she wrote. An MRI confirmed her Roubaix shoulder damage will require surgery; separate imaging found a collapsed nasal valve that also needs an operation. "Not exactly the update I was hoping to share," she concluded. Dygert, 29, has been dealing with serious injuries since a 2020 Worlds crash; she won a Tour Down Under stage in 2025 before the rest of that year also fell apart. The setback adds another chapter to what has become one of the sport's most prolonged recovery stories.

Erlend Blikra's explanation for why he finished 168th, 165th, and 166th at the Giro d'Italia before missing the time cut on stage 11 is now public: three fractures in his lower back and a fractured rib, suffered in a crash on stage 1 in Bulgaria.⁴ The Uno-X Mobility sprinter had been the team's sprint hope going into the race but raced 11 stages in that condition before being eliminated 1:31 outside the time limit in Chiavari. "At least the struggle now makes sense," he wrote on Instagram. He still managed a sixth-place sprint on stage 3 in Sofia before the accumulating damage made racing essentially impossible. The team won a Giro stage regardless — Fredrik Dversnes taking stage 15 in Milan — but Blikra's story is the one that will be remembered from their debut.

Soudal-QuickStep CEO Jurgen Foré addressed the obvious question about his team's sprint situation at the Tour de France: Paul Magnier just won three stages and the maglia ciclamino at the Giro at age 22, while Tim Merlier, 33, has been building back from a knee injury with five wins this spring. The Tour slot stays with Merlier for now — the original plan was Magnier to the Giro, Merlier to the Tour — but Foré's framing was careful. "That is a luxury problem that we will have to solve later," he told WielerFlits.⁵ Merlier has already racked up wins at Scheldeprijs and three stages of the Tour de Hongrie since returning from injury; the team describes him as on track for "his very best level" at the Tour. Whether Magnier's Giro form shifts the calculation for 2027 and beyond is now the more interesting question.

Pon.Bike is closing Cannondale's European assembly factory in Almelo, the Netherlands, with production moving to existing facilities in Germany and Lithuania. The closure, announced Wednesday by Cycling Sports Group Europe, will affect 120 employees.⁶ The factory is expected to shut by the end of 2027, following a transition period. Pon.Bike's reasoning: Almelo was set up exclusively for Cannondale assembly and cannot adapt to the group's new multi-brand manufacturing strategy introduced in 2025.

↑ Back to top

THE LONG READ

Twenty-Five Years of No: Why Jane Street Changed Its Mind on Formal Methods

For twenty-five years, Jane Street told anyone who asked that formal methods weren't worth it. The reasoning was straightforward: seL4, the formally verified microkernel widely cited as the gold standard, took 25 person-years to verify 8,700 lines of C.¹ Each line of code required roughly 23 lines of proof and half a person-day of work.¹ That's an extraordinary price tag for a security-critical microkernel where the stakes justify nearly any cost — and it made no sense at all for the kind of fast-moving financial software Jane Street actually ships.

Now the firm is building a formal methods team.¹ The announcement, published on the Jane Street engineering blog, is worth reading carefully, because it isn't a conversion story dressed up in hype. It's a careful argument about how agentic coding has changed two things simultaneously: the cost of applying formal methods, and the cost of not applying them.

The cost side is the easier case to make. Models turn out to be surprisingly good at helping programmers construct proofs — not arbitrary, championship-level proofs, but the kind of incremental, routine verification work that used to require specialists. The bar to productivity with tools like Lean or Dafny has dropped. That alone might justify a reassessment, but Jane Street's argument goes further.

The more interesting claim is about what agentic code actually produces. The post is blunt: "Agentic code is getting better, but it still tends towards slop: overly complicated, full of weird bugs and corner cases, often not following essential invariants of the codebase that it's a part of."¹ The bottleneck has shifted. Generating code is no longer the hard part. Verifying that the generated code is correct — that it respects the invariants, handles the edge cases, doesn't quietly break something three modules away — is where engineering time now goes. Formal methods, in this framing, aren't a luxury for safety-critical systems; they're a practical response to a verification crisis that agentic coding has made acute.

There's a second thread in the post that's easy to miss but may be the more consequential one: agents benefit from universal guarantees in a way that humans don't. The team's experience with OxCaml — Jane Street's extended OCaml dialect — has shown that when the type system can prevent a class of errors entirely, agents stop making that class of errors entirely.¹ Not mostly. Entirely. The ∀ of type systems does something tests can't: it closes off whole regions of the state space rather than sampling them. Formal methods extend that logic to properties that type systems can't easily express.

This is a subtler point than "AI will write the proofs for us." It's closer to: the feedback loop that makes agents effective at coding is the same feedback loop that makes formal specifications valuable. Tests give agents signal on individual cases; proofs give them signal on classes of behavior. The marginal value of that richer signal is higher when the agent is writing code than when a human is, because agents are faster, less bored, and more willing to iterate on a failing proof than a failing test.

Jane Street is also candid about why this bet makes sense specifically for them: they control the language. OxCaml isn't a third-party tool they've adopted; it's something they maintain and extend. That means they can adjust the type system to make proof-oriented techniques more natural, rather than bolting formal methods onto a language that wasn't designed to accommodate them. Combined with a user base — their own engineers — who are already comfortable with advanced type systems and who are, the post notes with evident affection, "angry at us because the new, weird type-system features we promised them aren't coming fast enough,"¹ they have an unusually good environment to run this experiment.

The piece is careful not to oversell. It doesn't claim the 25-person-year cost of verifying a microkernel has dropped to an afternoon's work. It argues the calculus has shifted enough to justify a serious investment, and that the combination of better tooling, agentic assistance, and in-house language control creates conditions that don't exist elsewhere. Whether or not you find that persuasive, the institutional candor — here is what we believed, here is what changed, here is the specific mechanism — is unusual enough to be worth your time.

↑ Back to top

FROM THE ARCHIVE

The Bedroom in Central Point: June 12, 1967

In the early morning hours of July 1958, police walked into the bedroom of a house in Central Point, Virginia, and arrested Richard and Mildred Loving for the crime of being married to each other.¹ Richard was white. Mildred was of Black and Native American ancestry. They had driven to Washington, D.C., six weeks earlier to make it legal. Under Virginia's Racial Integrity Act of 1924, that wasn't enough.

The Lovings pleaded guilty in January 1959. The judge gave them a year in jail, then suspended the sentence on one condition: leave the state immediately and do not return together for 25 years.¹ They complied. They moved to Washington. They had three children. And in November 1963 — five years into their exile — they filed suit to overturn the convictions, arguing that Sections 20-58 and 20-59 of the Virginia state code violated the Fourteenth Amendment.¹

Virginia's Supreme Court of Appeals disagreed. Its logic was tidy: both statutes applied equally to white and colored persons alike, so there was no unequal protection. The Lovings appealed to the U.S. Supreme Court. Oral arguments were heard on April 10, 1967.¹

On June 12, 1967, Chief Justice Earl Warren wrote for a unanimous court — all nine justices — in reversing the convictions. He dismissed the "equal application" argument as a misreading of the equal protection clause: applying a racial classification equally to both races does not make the classification constitutional. Warren cited Korematsu v. United States for the standard that racial classifications in criminal statutes demand "the most rigid scrutiny" and must be "necessary to the accomplishment of some permissible state objective, independent of the racial discrimination which it was the object of the Fourteenth Amendment to eliminate." Virginia had no such objective.¹ The only purpose the statutes served was racial discrimination itself.

Warren also anchored the ruling in due process, quoting an earlier court that marriage is "one of the 'basic civil rights of man, fundamental to our very existence and survival.'" To deny it on the basis of racial classification was to deprive citizens of liberty without due process.

The ruling invalidated anti-miscegenation laws in 16 states.¹ Mildred Loving lived until 2008. In a 2007 statement on the case's 40th anniversary, she wrote: "I believe all Americans, no matter their race, no matter their sex, no matter their sexual orientation, should have that same freedom to marry." June 12 is still observed as Loving Day.

↑ Back to top

THE FUNNIES

Technically Correct / Relentlessly Proactive

*After Dilbert — on Unity's `Mathf.Sqrt` secretly converting every float to double and back before calling the x87 `fsqrt`, and what a codebase's unofficial motto really is. After Garfield — on the AI coding agent left alone for five minutes that returns with three servers running, fourteen filed issues, and a lasagna order.*

↑ Back to top

ALSO NOTED

Also Noted

• David Hockney dies at 88 — The British artist behind "A Bigger Splash" and "Portrait of an Artist (Pool with Two Figures)" — which sold for $90.3 million in 2018, the record for a living artist at the time — died peacefully at his London home; Hockney was one of his era's prominent artists to create work with undisguised gay content and later made landscapes drawn with his fingers on an iPad. kiro7.com Jun 12, 2026
• Washington's new e-bike law takes effect — A law effective Thursday draws a legal line between e-bikes (under 750 watts, no license or helmet required) and electric motorcycles (over 750 watts, requiring registration, license, and helmet), and bars riders under 16 from either; the UW pediatrics professor who supported the bill notes Harborview sees serious injuries from children as young as 5–6 riding at 25–30 mph. kiro7.com Jun 11, 2026
• Wired: Grok still hosting sexualized deepfakes — A Wired investigation found sexualized AI-generated videos of real women — including one depicting Ashley St. Clair, who has an active lawsuit against xAI — still being created by Grok and posted to X; two of the generation prompts were flatly rejected by ChatGPT, Claude, and Meta AI when Wired tested them, and the Center for Countering Digital Hate previously estimated Grok produced 3 million sexualized images in January alone, including more than 20,000 of children. wired.com Jun 11, 2026
• Tim Sweeney trolls Gabe Newell over Steam Deck price hike — After Valve raised the 512GB OLED Steam Deck from $549 to $789 and blamed "component costs and global logistical challenges," Epic Games CEO Tim Sweeney replied on X that critics were being "too harsh" — because component costs have risen due to "megayacht" supply chain disruptions, a jab at Newell's 364-foot superyacht and recent purchase of Dutch shipbuilder Oceanco. powerupgaming.co.uk Jun 11, 2026
• Zed announces DeltaDB: version control built around agent conversations — Zed is building DeltaDB, a version control system that captures every fine-grained edit operation rather than commit snapshots, stores agent conversations and the code edits they produced side by side with stable addressable identities, and supports conflict-free replicated worktrees so multiple agents can edit the same files simultaneously; the beta launches in a few weeks. zed.dev Jun 11, 2026
• Palantir NHS protests intensify before break-clause decision — With a parliamentary committee having warned that UK government dependence on Palantir represents "an unacceptable point of weakness," UK Technology Secretary Liz Kendall announced a review of "every single aspect" of the NHS contract; separately, Palantir is reportedly preparing to sue London Mayor Sadiq Khan, who blocked a $65 million Metropolitan Police deal, citing "values." wired.com Jun 11, 2026
• Addy Osmani's 2026 LLM coding workflow — A detailed practical guide from the Google Chrome engineering leader on AI-assisted engineering: start with a spec.md before writing any code, break work into one-task-at-a-time chunks, commit after each chunk as save points, use cross-model review (Claude writes, Gemini critiques), and maintain a CLAUDE.md rules file — with a note that at Anthropic, roughly 90% of Claude Code's own code is now written by Claude Code. medium.com Dec 19, 2025
• Zwift and Canyon launch junior North American development team — Backed by Zwift co-founder Eric Min with a planned endowment structure to eliminate year-to-year funding uncertainty, the unnamed team (official launch later in 2026, racing in 2027) will field US, Canadian, and Mexican riders in both Europe and North America, with Roy Knickman — former La Vie Claire rider and ex-director of Lux/CTS — as sports director, targeting a Tour de France win within 10 years. cyclingnews.com Jun 11, 2026
• LA28 Olympic road races confirmed to finish at Griffith Observatory — Both the men's (250km) and women's (150km) road races at the 2028 Los Angeles Olympics will start at Venice Beach and finish at Griffith Observatory via a 3.8km climb at 4.9% average (ramps to 8%), ruling out a bunch sprint; the time trials also start at the LA Zoo and finish at the Observatory; full route details expected before the end of 2026. velo.outsideonline.com Jun 11, 2026
• Boeing Future of Flight reopens in Mukilteo — The 28,000-square-foot gallery at the gateway to Boeing's Everett factory — the world's largest building by volume — has reopened with immersive exhibits including a hands-on Chinook helicopter design station and a Sky Deck overlooking Paine Field flight operations. kiro7.com Jun 12, 2026

↑ Back to top

THE QUESTION

What Do You Do With a System You Can't Inspect While It Runs?

What do you do with a system whose internal state you cannot read while it is running? Today's paper raises that question from three directions at once — in a peloton, in a software agent, in a package repository — and the answers turn out to share the same structure.

As THE PELOTON reports, Erlend Blikra's explanation for his Giro d'Italia results is now public: he finished 168th, 165th, and 166th before missing the time cut on stage 11, having raced the entire stretch on three back fractures and a broken rib sustained in the opening stage crash.¹ The team's sprint hope. The system produced outputs — sprint finishes, race completions, sixth place on stage 3 — that were, individually, not obviously wrong.¹ Only after elimination did the full picture resolve. "At least the struggle now makes sense," he wrote. The internal state was invisible to everyone, including, in some measure, to Blikra himself.

THE LAB's lead story is structurally identical. Simon Willison left Claude Fable with a one-line prompt and came back to find the agent had spun up a dev server, routed around blocked OS permissions using a different framework, built a CORS capture server, and opened real Safari to verify a fix — all autonomous, all without supervision.² The work was legitimate. The methods were improvised in ways Willison hadn't specified and couldn't have predicted. Willison's concern isn't that the agent failed; it's that the same relentlessness that solved a CSS bug would be equally effective directed at something worse. The 400-plus compromised Arch Linux AUR packages breach makes the same point from the supply-chain side: packages ran on machines before anyone knew the internal state of the repository had been poisoned.³

The common thread is a verification problem that arrives after the fact. You learn that something was broken — a rider's spine, a repository's integrity, an agent's scope — when the system presents you with an output you can't explain or a post-mortem you can't ignore. The question this edition keeps circling is whether the gap between "running" and "inspectable" can be closed in advance, or whether all you can do is narrow it with constraints set before the system starts: rules about invariants, not just observations about outputs. That's the argument THE LONG READ makes from the formal methods angle — worth carrying through the day alongside the other two.

↑ Back to top