THE WORLD

Iran and the US Trade Strikes Across the Gulf; Graham Dead at 71

↩ Developing story — first reported Jul 08 · previously Jul 10, Jul 12

• Iran-US strikes escalate — US hit ~140 Iranian military targets over the weekend; Iran retaliated across Jordan, Kuwait, Qatar, and Oman; Hormuz closure contested but effective. Source
• Sen. Lindsey Graham dead at 71 — South Carolina Republican died Saturday of aortic dissection; the last of the Senate's "three amigos" defense hawks. Source
• World Cup semifinals confirmed — Argentina joins England, Spain, and France; all four bracket spots set with no clear favorite emerging. Source
• Bangkok bar fire kills 27 — Bar fire in Chatuchak district; obstructed emergency exits trapped victims at the rear; 63 injured, 22 critical. Source

The public hearing on Seattle's 2026 Transit Measure is today at City Hall — a 9:30 a.m. remote session and a 5 p.m. in-person session — on Mayor Wilson's proposed 0.3% sales tax renewal. The select committee is expected to vote on amendments Thursday, July 16. Source

Wildfires continue to burn across Eastern Washington. Volunteer firefighter Benjamin Stockdale lost his home in Vantage while he was out fighting the nearby Ryegrass Coulee fire. Source

ON THE ROAD — Aurillac, France

• Stage 10 departs tomorrow (Bastille Day) — Caravan leaves Aurillac at 11:00; riders roll at 13:25. 167 km, 100% Cantal, seven categorized climbs including the Puy Mary/Pas de Peyrol. Arrival at Le Lioran forecast around 17:12.⁷
• Road closures begin TONIGHT at 18:00 — RD67 access to Le Lioran is blocked from this evening. Both the Pas de Peyrol and Col du Pertus approach roads also close at 18:00 tonight (July 13) and stay closed until the peloton has passed tomorrow.
• City center cordoned from 06:00 tomorrow — Rue Pierre Marty (descending), avenue Gambetta, and rue des Carmes close at 06:00 on July 14. Visitors advised to park at Parkings du 8 Mai or the Gare Routière.
• RN122 closed from noon tomorrow — The Saint-Jacques-des-Blats to Murat section closes at noon on July 14; deviations via Montsalvy toward Maurs will be signed.
• High heat tomorrow — Aurillac: 94°F high tomorrow, overcast, 5% precip chance. Carry water if you are watching roadside.
• Village départ animations underway today — Fan activities and barricade installation are ongoing Monday in central Aurillac; barriers on avenue Aristide Briand, boulevard du Pont Rouge, and cours d'Angoulême go up this afternoon.

↑ Back to top

THE PELOTON

The Descent That Opened the Gap; Vingegaard's Year He Almost Quit

↩ Developing story — first reported Jul 09 · previously Jul 10, Jul 11, Jul 12

AURILLAC, France — The Tour village is camped in the city center, and by tomorrow morning the peloton will roll out of here heading north into the Cantal on Bastille Day. Stage 10 to Le Lioran: 166.6 kilometers, 3,800 meters of climbing, seven categorized ascents.¹ The Puy Mary Pas de Peyrol and the Col de Pertus both return — the same two climbs that produced one of the race's more memorable finishes two summers ago, when the yellow jersey contenders arrived at the summit in a different order than they'd left the valley.

The day after the Puy Mary is historically combustible. Every French rider in the peloton treats Bastille Day as a personal mandate, which produces frantic, often badly paced racing on the lower slopes. The accumulated fatigue of seven categorized ascents means the final cat. 3 into Le Lioran will hurt more than the number suggests.

UAE sport manager Joxean Fernández Matxín used Monday's rest day to explain, in unusual tactical detail, exactly how Stage 6 worked. He told Velo that the team had circled the Tourmalet stage months before the Grand Départ as the race's decisive moment — not because of the summit, but because of what comes after it.

"We know that Tadej on that terrain is much stronger," Matxín said. "That's why we expected that once he reached the bottom of the Tourmalet, the gap would continue to grow, which is exactly what happened."²

The numbers bear out the logic. The race leader gained more time on the descent and the 17-kilometer shallow drag to Gavarnie than he did on the Tourmalet itself. Matxín pointed to physics: "On those shallower gradients, absolute power becomes much more important, and Tadej weighs about six kilos more than Jonas. That's where the gap really opened up."²

UAE's approach was two-pronged. On the way up, the team shut down every breakaway attempt and denied Visma the chance to position a domestique up the road. Once the attacks began on the Tourmalet, Isaac Del Toro dropped back into the chase group with a specific task: stay glued to the chasers and prevent any organized pursuit from forming. "It's also true that the other leaders behind in that chase group were never going to work together to close that gap in an efficient manner," Matxín said. "Isaac was an important weapon in that situation."²

Writing for CyclingNews on Monday, analyst Philippa York noted the historical pattern that made the stage predictable in retrospect: every multi-winner in Tour history — Anquetil and Indurain in time trials, Merckx and Hinault on the first big mountain test — landed the decisive psychological blow on the first proper GC day and controlled the race thereafter. The surprise, York observed, is that anyone was surprised.³

Vingegaard himself provided the rest day's most striking disclosure, telling Danish outlet TV2 that at the end of 2025, he came close to quitting cycling entirely.

"I said last year that if this was how it was going to be, I couldn't be in it any more," he told TV2. "That's why we changed some things."⁴

The 29-year-old described reaching a mini-burnout with the relentless demands of WorldTour racing — the constant weight management, the training load, the months away from home. Unlike most modern pros, Vingegaard has never relocated to Nice or Andorra. He forced a compromise from Visma-Lease a Bike for 2026 without specifying terms, though the outline is visible: he won the Giro d'Italia on debut this spring and has substantially cut altitude camp time.

"I'm much more motivated this year, and I'm much happier being a cyclist," he said. But he added a caveat: "We have taken a step in the right direction, but it is clear that it is only a step."

His comments carry context beyond personal biography. Visma lost director Grischa Niermann and Vingegaard's coach Tim Heemskerk this spring — Heemskerk said he was struggling to apply his creativity at the team and now works with Remco Evenepoel.⁴ The roster erosion that has quietly hollowed out Visma's depth this Tour is partly traceable to a squad managing the fallout from its most important rider's discontent. Vingegaard is under contract through 2028.

Two pieces of team and incident news from the first rest day:

Soudal-QuickStep announced at a press conference Monday that Safety Jogger will become co-title sponsor from 2027, ending QuickStep's 24-year run as title sponsor. The Belgian shoe and PPE company has been a minor backer since 2022; it stepped up to bib shorts branding at the start of this season, and last year's special yellow-and-black rest day kit suggested the direction. QuickStep will retain a presence on the racing kit through 2030. The team has won 1,000 victories under the QuickStep title, including eight Tours of Flanders, six Paris-Roubaixs, and the 2022 Vuelta a España.⁵

UAE heads into the rest day having claimed €69,980 in prize money through nine stages — more than the combined total of the 12 lowest-earning teams, and more as an individual-rider tally for the race leader than nine full teams have earned collectively.⁶

Separately, eight fans were injured — two taken to hospital — when a car carrying L'Équipe journalists crashed into the barriers protecting the final kilometer of Stage 9's finish straight in Ussel. French authorities have opened an investigation. Barriers were re-erected and the stage concluded normally, but the incident renewed scrutiny of vehicle management in the race convoy. It is not the first time: in 2013, an Orica GreenEdge team bus became wedged under the finish gantry at that year's opening stage.⁷

↑ Back to top

THE LAB

Fable Punts Again; A Four-Line Change That Buys 4× Loop Throughput

↩ Developing story — first reported Jul 02 · previously Jul 07, Jul 10

The extension is now three-for-three. Anthropic pushed Fable 5's deadline again yesterday — now through July 19 — and this time the competitive rationale was made explicit. Simon Willison notes that GPT-5.6 Sol is "clearly a Fable/Mythos class model," which makes Anthropic's original story about needing time to gauge compute demand harder to sustain. Terms are unchanged: Fable 5 consumes up to half a subscriber's weekly usage allowance, with usage credits covering whatever remains.¹

OpenAI moved the opposite direction the same day. Thibault Sottiaux announced removal of the 5-hour cap for Plus, Business, and Pro plans; efficiency improvements to GPT-5.6 Sol that will lower apparent usage consumption (exact numbers pending); and a milestone: 6M active users, with a usage reset rolling within the hour. Willison's verdict is direct: "Anthropic should change track and keep Fable permanently available on those plans. OpenAI are winning users simply due to the uncertainty that surrounds Fable access."¹

A post on purplesyringa.moe demonstrates a case where a semantically useless if-statement — one that the compiler sees as dead code — quadruples loop throughput.

The problem is a sequential lookup loop: j = next_j[i][j]. Each iteration depends on the previous one's memory read, so the CPU cannot pipeline across iterations. The loop is latency-bound. The fix is to convert it into a branch-predicted form: if (j != next_j[i][j]) { j = next_j[i][j]; }. When the branch is biased toward "no change" — true in the common case, where most chunks keep the same encoding — the processor speculatively skips the assignment and sees no dependency between iterations. Mispredictions flush and restart from the correct state. The result: 320μs down to 80μs on the synthetic benchmark.² A 2× gain in more realistic workloads, which the author calls "still worthwhile" despite LLVM leaving some performance on the table.

The catch is that any optimizer recognizes the guarded assignment as equivalent to the unconditional one and eliminates it via CSE. Keeping the branch requires a volatile cast on the conditional read: j = *(uint8_t volatile *)&next_j[i][j], which forces the compiler to treat the two reads as independent. A Lobsters comment after publication surfaced an LLVM-specific option: [[unlikely]] also prevents the branch from being removed by the optimizer. The volatile version remains the more portable choice — it generates better code and works on GCC as well.

George Hotz published a short essay yesterday separating the technology from the narrative around it. He is, by his own account, genuinely enthusiastic: last week he set up a Linux box running opencode against a local GLM-5.2, called it "the Year of the Linux Desktop,"³ and walks back some of the skepticism from an earlier post about models and programming. The hedges are careful — "you have to be really careful, they can increase cognitive fatigue" — but the overall read is that models are useful in the way compilers and Stack Overflow are useful. What he objects to is the "window closing" framing and the jump from "fancy autocomplete" to "it's gonna own the whole light cone." His summary: "AI is the continuation of the computer revolution. I love computers so much."

Ray Myers at The Coding Agency has written a longer piece using Andrew Kelley's response to Bun's Zig-to-Rust migration as its spine. Myers' thesis is that the Bun write-up reads as marketing for Anthropic's Fable model rather than a real technical decision document — motivation is present, options considered are thin, and pros/cons are absent entirely. The most pointed observation is about style guides: Bun's post dismisses them as hard to enforce, then in the same breath claims agentic review successfully audited the full rewritten codebase against stringent guidelines. Myers calls this self-contradicting. The broader argument: "Anthropic needs people to believe that all problems can be solved by adding more agents, even the problems that agents introduce. Instead of better tools we buy tokens. Instead of learning skills we paste a SKILL.md."⁴ Myers draws his own line: "You can adopt coding agents without making the Dark Software Factory your platonic ideal."

On the AI×game-dev front: a project page for InfiniteDiffusion describes breaking what the authors call the content-generation trilemma — where you can have at most two of: infinite extent, stateless generation, and learned realism. Diffusion models achieve realism but are bounded; procedural noise is infinite and stateless but cannot learn; autoregressive outpainting requires shared global state. InfiniteDiffusion reformulates any diffusion model as a lazy computation indexed by seed and coordinates, with an LRU cache as a performance hint only. The claimed result: O(1) random access, full determinism, and embarrassing parallelism across unbounded terrain, with little quality degradation against MultiDiffusion.⁵

Trending today: GitHub saturated with AI agent wrappers, CLAUDE.md skill collections, and proxy tools repackaging subscription models as APIs — nothing on the list passes the technical novelty filter.

↑ Back to top

THE LONG READ

Twenty-Seven Years and Two Hours: Terry Tao Finally Builds the Thing He Abandoned in 1999

In 1999, Terry Tao started writing Java code for a spacetime diagram tool — "Inkscape, but in Minkowski space," as he describes it now. The concept was a visualization environment for special relativity: a full interactive canvas, but in Minkowski geometry rather than Euclidean. The code complexity overwhelmed him. He shelved it. He was, at the time, already among the most gifted mathematicians alive. He would go on to win the Fields Medal, solve problems across a half-dozen fields, and become widely regarded as the greatest living mathematician. The spacetime diagram sat unfinished.

Last week, Tao finished it. It took a couple of hours.¹

He was already doing something more modest: porting two dozen old Java 1.0 applets to JavaScript. The applets dated from 1999, built for complex analysis and linear algebra courses — the honeycomb visualizer he wrote with Allen Knutson, the Besicovitch set tool, a handful of others. Web standards had dropped Java 1.0 support years ago, and the applets had been dead since. He asked an AI coding agent to port them. The agent did it in hours, with graphical upgrades: the Besicovitch set applet gained color, replacing his original monochrome version.¹

The bug count is the detail worth dwelling on. LLM coding agents are known for subtle and blatant errors alike. Across two dozen ported applets, Tao found one minor bug — a drag event in a complex analysis applet misbehaved when dragging outside the main box. The agent, for its part, identified two bugs in Tao's original 1999 code that he hadn't known about. Net wash.¹

Having finished the ports, he turned to the abandoned spacetime project. A few hours of what he calls "vibe coding" — conversational back-and-forth with the agent — and the 1999 vision was finally running. He published a transcript of the process, edited down to remove "a large number of tedious technical implementation reports."¹ He is treating the result as an alpha, asking for feedback, and acknowledging that LLM-generated code almost certainly contains bugs he hasn't found yet. He is also unbothered by this, because the visualizer is a secondary aid, not a proof. The asymmetry matters to him: the downside risk of a bug in an interactive applet is a misdrawn diagram; the downside risk of a bug in a mathematical argument is different in kind.

Before the post was finished, he thought of another use: a visualization tool for the Gilbreath conjecture, to accompany a paper he'd just written about it. That one also took a few hours.¹ He wrote that he may add interactive visualizations as supplements to future papers.

The piece is short. Tao is not writing a manifesto. He is reporting, with characteristic precision, what happened when he tried a thing. The evidence is specific — applet counts, bug counts, hours logged, a 27-year gap closed. He is not claiming AI has changed mathematics; he is noting that it has changed what he can afford to build as a side project. The distinction is important, and he makes it without laboring it.

For anyone who works in software and has heard the AI productivity arguments recycled endlessly through hype cycles, Tao's account is a different kind of testimony. Not a vendor claim, not a benchmark, not an accelerationist forecast — just the world's greatest living mathematician describing what he actually shipped last week, and what he had been meaning to build since 1999.

Read the full post at terrytao.wordpress.com

↑ Back to top

FROM THE ARCHIVE

First Goal, Wrong Whistle: July 13, 1930

Pen-and-ink editorial illustration. A French footballer in 1930s-era knickerbockers and laced leather boots pivots through a contested penalty area, boot making contact with the ball — the moment the first goal in World Cup history is struck. Two Mexican defenders lunge in behind him, arms outstretched. The Estadio Pocitos in Montevideo looms in the background: wooden terracing half-filled with figures in fedoras and double-breasted jackets, a low corrugated roof against flat overcast sky. Strong confident linework, dense cross-hatching on the shadows beneath boots and the underside of the stand roof, bold diagonal composition driving left to right. No color, no gradients. Generous white space in the upper sky.

Jules Rimet had to improvise. The first FIFA World Cup was to be staged in Uruguay in 1930, but when the Depression made transatlantic travel expensive — and when amateur players worried their jobs wouldn't survive a months-long absence — most European associations declined the crossing. Italy, Spain, Germany, Holland, and England, which had won three consecutive Olympic gold medals and had never lost a competitive international on home soil, all stayed home.

Rimet managed to convince four European nations to attend only after Uruguay agreed to cover the traveling costs. Romania required a different kind of persuasion entirely. King Carol selected the squad himself, granted every player a three-month leave from their employers, and personally guaranteed their jobs would be waiting when they returned. That is how Romania — and France, and Belgium, and Yugoslavia — got to Montevideo.

On July 13, 1930, the first matches of the first World Cup were played. A French forward named Lucien Laurent scored, and the goal stood as the first in the history of the competition.¹

France did not last long in the tournament. In their second match, they lost to Argentina 1-0, under circumstances the record does not flatter. The referee blew the final whistle six minutes before time.¹ Both sides walked off the pitch. When the error was discovered, Argentine players had to be summoned back onto the field to play out the remainder. France went home in the group stage regardless.

The United States did considerably better. They beat Belgium and Paraguay in the preliminary round, then lost to Argentina 6-1 in the semifinals — at the time the best result the U.S. had achieved in a World Cup.¹

Uruguay, the hosts and preemptive favorites, defeated Argentina 4-2 in the final before 93,000 people on July 30.¹ The Estadio Centenario, built specifically for the tournament, was finished in time for the decisive match. The final was a rematch of the 1928 Olympic gold medal game, and Uruguay won it the same way.

Ninety-six years later, France is in the 2026 World Cup semifinals. Lucien Laurent's name is in the record books. The referee from that second group match is not.

↑ Back to top

THE FUNNIES

Code That Does Nothing; The Whistle That Spoke Too Soon

*After Peanuts — on the four-line fix from THE LAB that made a loop run four times faster by doing, essentially, nothing at all. After The Far Side — on the 1930 World Cup referee who, per THE ARCHIVE, blew the final whistle six minutes before time and watched both teams walk off the pitch.*

↑ Back to top

ALSO NOTED

Also Noted

• Sam Neill dead at 78 — The New Zealand actor, who had declared himself cancer-free in late April after a four-year lymphoma treatment, died suddenly Monday in Sydney; his family described the loss as "sudden and unexpected." theguardian.com Jul 13, 2026
• Albertsons opioid trial opens in King County — Washington AG Nick Brown opened trial Monday against Albertsons, alleging the chain dispensed more than 641 million opioid pills in the state between 2006 and 2022 with documented red flags consistently overridden; the case runs through September before Judge Janet Helson. kiro7.com Jul 13, 2026
• The Pacific temperature chart nobody is covering — A climate essay at lyrebirddreaming.com describes how sea-surface temperatures in the Niño 3.4 region have not just set a new record but departed entirely from the range of every prior year since 1982 — direct satellite and buoy observations, not a model. lyrebirddreaming.com Jul 10, 2026
• Seixas's spring résumé: two WorldTour wins before the Tour — The 2026 UCI calendar confirms Paul Seixas won both the Itzulia Basque Country and La Flèche Wallonne this spring, context for why mid-August transfer negotiations with Decathlon are expected to start from a position of significant leverage for the rider. procyclingstats.com

↑ Back to top

THE QUESTION

The Lag Between the Break and the Signal

At what point does the pressure required to produce elite-level output become the thing that breaks it? The question turns out to be useful in more than one domain today.

As THE PELOTON reports, Jonas Vingegaard nearly walked away from professional cycling at the end of last year — not because of injury or poor form, but because the system had been optimizing him past a threshold he could sustain. Weight management, training loads, months from home, the full administrative cost of being the sport's second-most-important rider. He was podium-level by every external measure. He renegotiated. He's still racing, and says he's much happier. He also adds: "It is only a step."¹ The structural pressure hasn't gone away; he's bought time.

THE LAB carries a structurally identical argument from a different domain entirely. Writing about a major TypeScript runtime's decision to use AI coding agents to rewrite its codebase in a new language, Ray Myers names the failure mode: the "Dark Software Factory."² Agent-generated throughput looks clean in the metrics. The engineering competence required to evaluate, debug, and extend that output may be quietly draining away. The worry isn't a production incident today — it's a capability gap that only becomes legible when it's too late to close. His summary: "Instead of better tools we buy tokens. Instead of learning skills we paste a SKILL.md."²

The lag is the problem both stories are pointing at. Competitive systems — athletic, technological, institutional — are designed to respond to what their external metrics can register. Internal depletion is slow, diffuse, and personal; it rarely triggers any alarm before it surfaces as something dramatic. By then the best case is a renegotiation. The worst case is the person, or the skill, is already gone.

The question worth carrying is not whether these systems are breaking. It's whether the lag between the fracture and the signal is, by design, too long to catch.

↑ Back to top