THE WORLD

Third Suspect at Large in Seattle Shooting; Iran Says No Talks

↩ Developing story — first reported Jul 19 · previously Jul 21, Jul 24, Jul 27

• Iran Holds the Line — Tehran says "no negotiations at present" as Oman convenes regional diplomacy; Saudi Arabia intercepted Iraq-launched drones; Netanyahu flies to meet Trump. CNN
• Berlin Pride Attacker Shot Dead — Van-and-machete assailant Abdul Ballout, one dead and 29 hurt, was tracked to a Berlin suburb and shot by police Sunday. PBS NewsHour

A 15-year-old is in custody in connection with the Seattle Center mass shooting that killed three; a 19-year-old believed to be a second shooter was found dead at the scene, and police are still searching for a third suspect. Mayor Wilson and Chief Barnes defended a response that included an erroneous initial statement about two arrests when only one had occurred, and a press conference that didn't start until 11 p.m. — nearly five hours after the first shots — with Barnes noting he had been attending a law enforcement conference in Dallas at the time. KIRO 7

The Little Giant Fire near Leavenworth nearly doubled overnight to 22,296 acres with 0% containment, as three new fires — Bradeen Hill in Stevens County (3,548 acres, 0% contained), Signal Peak in Yakima County (3,000 acres, 0% contained), and the Ptarmigan in the Pasayten Wilderness (10,842 acres, 0% contained, trail closures expanding) — joined the Kaiser Canyon and Modrite fires burning across eastern Washington under a continuing Red Flag Warning. Fox 13 Seattle

Mercer Island's City Council voted unanimously last week to adopt zoning changes — density increases near the Sound Transit light rail station, plus new permissions for emergency shelter and supportive housing — to satisfy a 2025 state housing ruling just days before a July 31 state deadline. A Growth Management Hearings Board compliance hearing is set for September 15, despite more than 1,000 residents signing a petition opposing the supportive-housing provisions. The Urbanist

ON THE TRAIL

WEEKEND PICKS — Sat–Sun, July 30–31

Teanaway / Lake Ingalls area (I-90 East, 90–110 min): A July 26 report found the route "wonderfully uncrowded," with clear views of Adams and Rainier.⁷ Weekend forecast: sunny with 1% precip Saturday and Sunday — the most reliable weather window in the region this weekend.

Kendall Katwalk + PCT corridor (I-90, 35–55 min): July 26–27 reports show the trail in good shape — one manageable blowdown easily bypassed, wildflowers and early berries coming in, creek crossings easy with poles. Overnight camping is available along the PCT past the Katwalk at Gravel and Ridge Lakes. Weekend forecast: 27% precip both days; bring a rain layer.

Foss Lakes / West Fork Foss River (US 2 West, 60–90 min): A July 27 report found bugs "finally mellowing out," water at every lake and falls, no snow beyond Copper Lake, and only about ten people all day. Weekend forecast: 47% Saturday, 54% Sunday — Saturday looks solid; have a rain plan for Sunday.

REGIONAL SNAPSHOT

• Issaquah Alps: No bugs on Cougar Mountain (Jul 27); storm-damaged trails in the NE corner repaired with new bridges. Good local option for a quick shakeout. - I-90 / Snoqualmie Pass: Kendall Katwalk solid; Denny Creek trail zero blowdowns with crossable streamflow to upper Melakwa Lake. Rattlesnake Ledge draws heavy weekend crowds — arrive before 8 a.m. or skip it. - Mountain Loop Highway: Goat Lake reported bug-free at the lake (Jul 26); Mount Pilchuck had morning fog but views clearing by early afternoon. Weekend rain chance 49%/50% — technically under the disqualification threshold but bring full rain gear. - Teanaway / I-90 East: Lake Ingalls uncrowded with open alpine views; Johnson Mountain has head-high brush through burned sections, tough in heat. - US 2 West (Stevens Pass): West Fork Foss bugs mellowing and water available; Frog Mountain light crowds and mostly bug-free (Jul 26), consistent grade with high wow-factor views. - Mt Rainier area: Burroughs Mountain in excellent condition with peak wildflowers (Jul 27); Naches Peak Loop gorgeous with minimal bugs and only a slight smoke smell from the Yakima Valley, not intrusive. Mystic Lake backcountry — no snow, no bugs, wildflowers at peak (Jul 26). - North Cascades / Hwy 20: Hidden Lake Lookout road closed 2.5 miles from the trailhead — adds 5+ miles round trip, budget the extra time. The Ptarmigan Fire (10,842 acres, 0% contained) is expanding in the Pasayten Wilderness with active trail closures; check the Okanogan-Wenatchee NF website before heading into the area.

↑ Back to top

THE LAB

The Bun Rewrite Was 'Done' Eleven Weeks Ago. The Machine Is Still Running.

On July 27, Thomas Lockwood cloned the Bun repository and started counting. The last release tag — bun-v1.3.14 — was stamped May 12, 2026. Eleven weeks ago. The only comparable gap in Bun's history was a six-week skip in late 2022.

What Lockwood found in the interim: 2,475 open pull requests from robobun, the account used by Claude Code. On July 9, when the breathless takes about Bun's AI rewrite peaked, that count was 1,277. Eighteen days later it had nearly doubled. The CI pipeline — Buildkite, running continuously since the rewrite merged to main — takes roughly 40 minutes per PR. At current queue depth, clearing it would take 86 continuous days of pipeline time. Separately, Lockwood's commit analysis shows Anthropic employees appearing directly in the Rust work, and usage data points to Claude activity ramping up again rather than winding down.

The original claim, from Bun lead Jarred Summner, was that the rewrite happened over 11 days in May at a cost of $165,000 in API calls — about $15,000 a day in tokens.¹ Lockwood isn't disputing that window. His point is what it excluded: CI infrastructure costs, Anthropic employee labor, and the ongoing Claude work still queued in nearly 2,500 unmerged PRs. If the machine is still burning $10,000 a day, the total is approaching $800,000 — and the project has yet to cut a release. Lockwood frames his conclusion in the quietest possible register: "In my tiniest mouse voice ever I want to ask probably the most relevant question one can ask about a Business Thing: Was it worth the money and are the companies worth the valuation?" Given that Anthropic owns Bun and has an obvious interest in demonstrating Claude Code's industrial-scale viability, the question is legitimate regardless of the answer.

A FermiSense writeup describes the same playbook applied across eleven production deployments: take an open-source base model, apply proprietary domain labels, run a reinforcement-learning stage against a scored version of the actual workflow. The numbers from three named examples are blunt.

Bridgewater Associates trained a model on labels from its expert investors to judge document relevance — "relevant" by Bridgewater's internal standard, which no amount of prompting got frontier models to absorb reliably. The fine-tuned model makes roughly 30% fewer mistakes than the best frontier model, at a fraction of the inference cost.² Harvey built a legal agent for transaction due diligence and memo drafting using RL on an open-weight base; it outperforms both GPT-5.5 and Claude Opus 4.8 on Harvey's rubrics. Intercom post-trained a vertical support model, Fin Apex, on billions of customer-service interactions; it resolves more issues than frontier alternatives while costing less per call. The piece's appendix collects eight more deployments following the same template. The underlying logic is simple: frontier models are general, and for narrow workloads, "knowing what Bridgewater means by relevant" outweighs raw parameter count.

On the open-weights side, Moonshot AI shipped the Kimi K3 weights to Hugging Face on July 27 — 2.8 trillion parameters, 1.56TB.³ Simon Willison flagged the license evolution. K2 used a self-described "modified MIT" that required attribution from companies above 100 million monthly active users or $20M in monthly revenue. K3 drops the "open source" framing entirely: any entity running a Model-as-a-Service business whose aggregate revenue exceeds $20M over any 12-month period must negotiate a separate agreement with Moonshot before commercial use. Willison notes, without irony, that Moonshot consistently uses "open weight" rather than "open source" in its own materials — a distinction the rest of the industry has been slow to enforce on itself. OpenRouter is already serving K3 from seven providers, with most at $3 per million input tokens and $15 per million output, matching Moonshot's own API pricing.

Trending today: GitHub is saturated with Claude Code skill forks, AI-agent wrappers, and CLAUDE.md collections — no meaningful technical novelty cleared the bar.

↑ Back to top

THE PELOTON

The 2026 Tour Was the Fastest Ever. Don't Expect That to Last.

↩ Developing story — first reported Jul 16 · previously Jul 25

The finishing average was 43.328 kilometers per hour — the fastest Tour de France ever recorded, beating last year's 42.849 kph by more than half a kilometer per hour. Eight road stages exceeded 45 kph.¹ This was not an anomaly. It was the outcome of several compounding forces that have been building since COVID and show no sign of reversing.

The fueling numbers are the clearest explanation. Sports science long held that 120 grams of carbohydrate per hour was roughly the ceiling: beyond that, performance gain flattens and gut-rot risk rises. Riders are now ignoring this. Social media posts during the race showed Jonas Abrahamsen targeting 160 grams per hour — 640 calories, the equivalent of a share-size bag of gummy bears, every hour, for up to five hours of racing.¹ Jonas Vingegaard and Lennert van Eetvelt were working at 140 grams (560 calories) on the hardest days. Multiple WorldTour nutritionists told Velo that eating below 120 grams per hour has become uncommon at Grand Tours. The mechanism is not about immediate wattage — excess carbohydrate beyond the threshold does not produce more power in the moment. What it does is accelerate overnight glycogen recovery in a race that runs for three weeks. The riders who can hit Stage 19 as hard as Stage 4 are the ones who never let their debt compound.

The other accelerant was structural. UAE Team Emirates-XRG arrives at every Tour as the sport's largest budget, around $60 million a year, and for the first time it was not alone at the top. Lidl-Trek, Decathlon CMA CGM, and Red Bull-Bora-Hansgrohe all came to the 2026 race with resources and depth built to compete at the same level. That compressed the ambition across the peloton — every team arrived believing a stage win was achievable, which made every stage fast from the gun. Sepp Kuss, who has ridden 17 Grand Tours and six Tours de France, was direct about it: "This year was one of the toughest. The course was maybe not the hardest, but the way we raced every stage made it really tough. The level is so high in the peloton now that everyone wants to win a stage and every stage then becomes crazy fast."¹

The trend line on average Tour speed since COVID reads: 39.9, 41.2, 42.1, 41.1, 41.8, 42.8, 43.3 kph. Nothing in that sequence suggests the 43.3 record holds for long. Budgets will rise. Fueling targets will push north again. Heat training is now a non-negotiable preparation block. The same forces that made this the fastest Tour on record are still accelerating in the same direction.

Wout van Aert will start the Tour of Denmark tomorrow in Aalborg — his first race since elbow surgery ended his spring. The five-stage race through August 2 is his calibration block before the Vuelta a España, which opens August 22, and the Road World Championships in Montreal on September 27.⁶

UAE Team Emirates-XRG has confirmed its Vuelta squad without Tadej Pogačar or Juan Ayuso. Pogačar will target the Worlds and Il Lombardia rather than the Vuelta. Adam Yates and João Almeida will co-lead the UAE effort at the Vuelta.⁵

Paul Seixas, fresh off his Tour de France debut at 19, will take two weeks of rest before building toward an autumn block: GP Quebec and GP Montreal in September, the Worlds road race on September 27, and Il Lombardia in October. Seixas holds a contract with Decathlon CMA CGM through the end of 2027. Decathlon has not waited on the squad-building side: Pavel Sivakov, Ben O'Connor, and Laurens De Plus were signed for 2027 to build around Seixas,³ and Romain Bardet was brought in as sporting manager. The team is also reportedly exploring a Swiss registration to raise its salary ceiling. Seixas, asked in L'Équipe about his future, was measured: "I don't have to act quickly. If I stay, I stay, if I leave, I leave. I don't know how far it's going to go but I have faith in the team and the team has faith in me."³

Former Giro d'Italia stage winner Danilo Napolitano has been handed a 20-year ban by Italy's anti-doping authority, NADO Italia, in a case tied to the Salus Ciclistica Seregno junior development team, where he was working as a coach. The suspension runs to April 7, 2046.² Former team president Marco Moretto received eight years. NADO Italia cited violations of Articles 2.6 and 2.8 of the WADA Code — covering possession and administration of prohibited substances — but released no further details. Napolitano raced professionally from 2004 to 2017.

The Tour de France Femmes avec Zwift begins Saturday with a start in Lausanne. Canyon-SRAM, one of the only women's WorldTour teams without a corresponding men's squad, will debut a new kit tied to the 'We Femmes' campaign — a cycling-outreach initiative backed by Canyon, SRAM, and Zwift, with a stated aim of bringing 1,000 new women per year to organized rides over the next decade.⁴ The team won the 2024 Tour de France Femmes with Kasia Niewiadoma-Phinney.

↑ Back to top

THE LONG READ

One Character. Eighteen Months.

The Nova Scotia Court of Appeal put it plainly: "Mr. Klayme is factually innocent of the offences. He should never have been charged, let alone convicted." By the time those words were written, Brandon Klayme had already served eighteen months in prison for crimes he did not commit.¹

The reason he was there at all: a single underscore.

In 2018, a 12-year-old girl in Madison, Wisconsin, was trading messages with an adult male on the Kik messaging service. Her mother found an "inappropriate" photo on her phone and called the Dane County Sheriff's Department. A forensic search turned up 125 Kik messages between the girl and a user with the handle "fus__ro_dah" — a reference to the Unrelenting Force dragon shout from The Elder Scrolls V: Skyrim, spelled with two underscores after "fus."¹

To identify this person, investigators served a subpoena on Kik. Somewhere in the paperwork, a character was dropped. The subpoena requested records for "fus_ro_dah" — one underscore. Kik complied, returning an email address associated with that account. That email address belonged to Brandon Klayme, in Halifax, Nova Scotia. Google records placed the email at a Canadian IP address. Halifax Regional Police took the IP to Bell Aliant, which traced it to Klayme's physical address. He was arrested in 2020.¹

Investigators found no evidence of the crime on Klayme's digital devices. He was convicted at trial anyway, in 2023. Sentenced in 2024. Served the full eighteen months.¹

It was only while preparing his appeal that the single-character discrepancy surfaced. "Fus__ro_dah" was not "fus_ro_dah." The perpetrator the Wisconsin investigation was hunting had two underscores. The man Canadian police arrested had one.

The question the ruling raises is not whether the courts ultimately did right — they did, and their language is unambiguous. The question is how a chain of institutions with the original forensic report, the subpoena, and Kik's response in hand failed to notice, across six years of legal proceedings including a full criminal trial. Nobody compared the usernames character by character. Not the Wisconsin investigators who wrote the subpoena, not the Halifax officers who received the handoff, not the prosecutors who built a case on the IP match, not the defense at trial.

Software engineers have a name for this category of mistake: off by one. The canonical version involves loop indices and array bounds. This version involved a string that was never verified against its source. The right username was sitting in the forensic report the entire time — "fus__ro_dah," two underscores, in writing — while every subsequent step in the chain ran on the wrong one.

That is the structural problem the case exposes. Physical evidence has an implicit verification loop built in: a photograph of a suspect can be held next to the suspect. Digital evidence chains do not. When police hand you a username, Kik hands you an email, Google hands you an IP, and Bell Aliant hands you a subscriber address, each link looks solid on its own. The record matches the email. The email matches the IP. The IP matches the address. The address belongs to Klayme. What no one checked — apparently for the entire life of the case — was whether the opening query matched the forensic report that motivated it.

The discrepancy emerged during appeal preparation. That is how the system found it: not through prosecutorial disclosure, not through an investigative audit, not through the trial itself, but through an appeal the defendant filed himself — while investigators and courts had six years to look and did not.

↑ Back to top

FROM THE ARCHIVE

The Three Generals: July 28, 1932

At 4:30 in the afternoon on July 28, 1932, Major George Patton kicked his heels against his mount and the horse lurched forward, signaling the cavalry line. The riders drew their sabers. Behind them, five Renault tanks — seven-ton relics from the same war the men in the encampment ahead had survived — clanked into position on Pennsylvania Avenue. Infantry fixed bayonets. A machine-gun crew set up on the street.¹

Up ahead, in a shantytown of tarpaper and scrap tin on the Anacostia Flats, roughly 20,000 WWI veterans watched the Army come for them.

They had arrived in Washington that spring from all over the country, riding freight trains and hitchhiking, to lobby Congress to pay a soldiers' bonus that had been authorized in 1924 but deferred until 1945. The Great Depression had shredded that timetable. They called themselves the Bonus Expeditionary Force. They were broke, hungry, and they had nowhere else to go. The House passed their bill in June. The Senate killed it 62 to 18, and the veterans who had crowded the Capitol galleries walked outside and sang "America" and didn't leave.²

They were still there six weeks later when Army Chief of Staff Douglas MacArthur, his chest bright with medals, stepped out of a staff car on Pennsylvania Avenue and gave the order to clear them.

MacArthur was not supposed to be there. Military protocol called for a commanding general to remain at headquarters, especially one whose role was administrative. His aide — 42-year-old Major Dwight Eisenhower — reportedly advised against it.¹ MacArthur didn't listen. He took personal command of the eviction, deploying gas grenades to disperse the veterans while cavalry rode down those who held their ground with bricks and iron bars.

One gassed bystander ran up to MacArthur's staff car, face streaming, and yelled that the American flag meant nothing to him after what he'd just witnessed. MacArthur told the nearest officer to arrest him if he opened his mouth again.

By evening, the Army had cleared the downtown area without firing a shot. Hoover, who had watched the red glow of the fires from the White House, twice sent word that troops were not to cross the bridge to the Anacostia encampment. MacArthur waved off the presidential orders — saying he hadn't time to be bothered by people pretending to bring instructions — and sent troops across anyway. They drove the veterans out with gas and fire, burning the camp to the ground. One baby, left behind in the confusion, died from gas inhalation. Two veterans had already been shot dead by police during the earlier confrontation. The day's final count: three dead, 54 injured, 135 arrests.¹

A postwar Veterans Administration study found that 94 percent of the marchers had actual service records. MacArthur had estimated the figure at 10 percent.¹

Among the marchers Patton's cavalry routed was Joseph Angelo, a former doughboy from New Jersey. Fourteen years earlier, in the Meuse-Argonne, Angelo had pulled a wounded Patton out of a foxhole to safety. He wrote that it was "a most distasteful form of service" and that most of the Bonus Marchers were "poor, ignorant men, without hope, and without really evil intent."¹

Newsreel cameras caught the eviction. Theater audiences across the country booed the footage.

Franklin D. Roosevelt, then governor of New York and already confident he'd beat Hoover in November, read the morning papers and told an aide: "He should have invited a delegation into the White House for coffee and sandwiches." Already confident of success, Roosevelt now felt victory was certain — this was a black eye no one could overcome. That November, Roosevelt carried 42 states and 472 electoral votes. Hoover won 59.¹

MacArthur, Patton, and Eisenhower — the three men on Pennsylvania Avenue that afternoon — would each play defining roles in the war that followed twelve years later. History rendered its verdict on the Bonus March mostly through what those men said about it afterward. MacArthur eventually recast the marchers as a "vanguard of a starved band" and claimed credit for supplying them with tents.¹ Eisenhower waited until after MacArthur's death in 1964 to reveal that he'd counseled against the whole thing. Patton said what he saw and left it at that.

Three years after the eviction, Congress passed the soldiers' bonus over FDR's own veto. The veterans got their money. The two shot dead by police were buried at Arlington with full military honors.

↑ Back to top

THE FUNNIES

One Underscore Short / The Candy Warehouse Stage

After Peanuts — on the single missing underscore that sent Brandon Klayme to prison for eighteen months while everyone in the chain looked right past the typo. After Calvin and Hobbes — on the 600-calorie-per-hour fueling arms race that made the 2026 Tour the fastest in history, and what the human body has to say about it around kilometer 180.

↑ Back to top

ALSO NOTED

Also Noted

• FDA Clears First OTC Acetaminophen-Naproxen Combination — The agency approved a fixed-dose tablet of 650mg Tylenol and 220mg naproxen sodium — evaluated across eight clinical studies — that Kenvue says begins working in under 30 minutes and lasts 12 hours, with three-year OTC exclusivity. kiro7.com Jul 28, 2026
• The Astronaut Who Watches Themselves Come Home — Flight surgeons who debrief returning ISS crews describe a weeks-long perceptual state of observing their own lives from outside the frame often enough that it has become a recognizable readjustment pattern, driven by vestibular recalibration, the orbital hypervigilance that kept them safe in space, and a residue of the overview effect that doesn't switch off at splashdown. spacedaily.com Jun 24, 2026
• Shared Claude Chats Are Being Indexed by Google and Bing — Wired found that Anthropic's robots.txt tells crawlers to stay out, but individual shared-chat pages don't include the "noindex" HTML tag that both Google and Bing require to actually prevent indexing — the same gap that drew coverage in September 2025 and apparently still isn't fixed. wired.com Jul 27, 2026
• Dario Amodei: Anthropic Has Never Advocated for Banning Open-Weights Models — In a post responding to the open-weights debate sparked by US officials considering restrictions on Chinese models, Amodei lays out three actual Anthropic policy positions: keep advanced chips out of authoritarian hands, stop industrial-scale distillation attacks, and require mandatory pre-release safety testing for all sufficiently capable models whether open or closed. anthropic.com Jul 27, 2026
• Ethan Mollick's AI Tool Guide Has Shifted to Agentic Systems — Simon Willison flags that Mollick's updated guide has moved from chat-model comparisons to tools that "do the equivalent of many hours of real human work in one go," with Gemini now off the list since Google lacks an established agentic product, and a useful observation that ChatGPT Work on mobile and desktop are effectively different tools. simonwillison.net Jul 27, 2026

↑ Back to top

THE QUESTION

When General Excellence Becomes Table Stakes

An arms race doesn't stop when both sides reach the same ceiling — it reconfigures around whatever that ceiling didn't cover.

Today's paper reports two accounts of that reconfiguration happening simultaneously in different industries. In the peloton, the Tour just finished at a record 43.3 kph, driven in part by budget parity at the top of the sport and by fueling science that has pushed carbohydrate intake past a threshold that sports science long treated as roughly fixed.¹ When multiple teams operate at the new frontier and multiple riders target the same fueling targets, the individual edge migrates somewhere else: depth of mountain support, heat preparation, the ability to field a second GC contender. In enterprise AI, the same shift is visible from the other direction: domain-specific models trained on proprietary labels are outperforming frontier LLMs on narrow workloads by margins that raw scale can't close, at a fraction of the inference cost.² The advantage isn't in the general capability — it's in whoever can precisely define what "correct" means for their specific workflow and train against that definition.

The structural question isn't which domain is further along. It's whether specialization is a durable edge or a temporary arbitrage. In cycling, the fueling regime that once looked like a ceiling is now the floor — the teams that built an early advantage have watched the peloton absorb it. In AI, the logic is explicit: Bridgewater's edge is its labels, not its model architecture, and any competitor with comparable domain data can run the same playbook.² The specialist beats the generalist until the generalist learns the specialization, at which point the game resets at a higher baseline and the question becomes: what's the next thing the generalist doesn't yet know?

The question worth carrying: when a capability becomes table stakes — universally available, universally adopted — does the cycle of commoditization and re-specialization accelerate, or does the cost of the next edge eventually outrun the return on it?

↑ Back to top