THE WORLD

Hormuz: Trump Drops the Fee, Iran Raises the Stakes

↩ Developing story — first reported Jul 13 · previously Jul 14, Jul 15, Jul 16

• Hormuz — Trump replaced the 20% transit fee with Gulf investment pledges; Iran threatened Bab el-Mandeb closure.¹
• World Cup — Spain and Argentina meet in Sunday's final; Argentina advanced via Messi's two assists over England.²
• Lebanon-Israel — US-brokered talks in Rome concluded Wednesday; Lebanon's president Aoun invited to Washington next week; one million Lebanese remain internally displaced.¹

• Three Queens Fire — A wildfire broke out Thursday northwest of Lake Kachess, east of Snoqualmie Pass — roughly 25 acres between Three Queens and High Box mountains. USFS has air and ground crews on scene; no containment yet.³
• WA income tax on the ballot — The Secretary of State certified Initiative 645 on July 15, sending the state's new 9.9% tax on household income above $1 million to a November vote. Yes repeals it; no puts it into force in 2029.⁴

ON THE ROAD — Adast, Hautes-Pyrénées

• AlTi'Bus — The Communauté de communes Pyrénées Vallées des Gaves runs a summer shuttle from Argelès-Gazost to Col du Soulor, Hautacam, Lac d'Estaing, and Pic du Midi (running since July 1); practical for high-altitude day trips without driving the mountain roads yourself.⁵
• Weather window — Friday 77°F / 25°C, overcast, 20% precip. Saturday clears to 81°F and just 3% — the best day of the window. Sunday risk jumps to 43% precip at 84°F; plan exposed routes for Saturday.
• Campan mounaques — The 33rd Journées des Mariolles returned to Campan as of July 17, filling the village with giant straw figures (mounaques); a short drive up the valley from Adast and worth the detour if the weather holds Saturday.⁶

↑ Back to top

THE PELOTON

Two Broken Collarbones and a Breakaway Battle: Stage 13 Rolls into the Vosges

↩ Developing story — first reported Jul 13 · previously Jul 14, Jul 15, Jul 16

DOLE, France — Fifteen riders had cleared the peloton with 160km still to race when Lidl-Trek came to the front and started winding up the chase. The concern was not Tom Pidcock — best-placed rider in the break, tenth on GC, 11:49 down, comfortable as a passenger. The concern was Jasper Philipsen. Also in the break, trailing Mads Pedersen by only 46 points in the green jersey competition, with the intermediate sprint still more than 100km away. Lidl had every reason to shut it down.¹

Stage 13 is 205.8km from Dole to Belfort, the longest stage of this Tour, crossing only two categorized climbs — the Col des Croix and the category-one Ballon d'Alsace near the finish. The first 150km offers nothing for pure climbers, which made the break the obvious story before the flag dropped. Pogačar was asked before the start whether UAE would chase. "That depends on who is in the break," he said. The definitive move eventually coalesced with Philipsen, Pidcock, Ben Healy, Maxim Van Gils, Matej Mohorič, Tim Wellens, Brandon McNulty, Søren Hagenes, Stefan Schmid, Romain Grégoire, Julian Alaphilippe, and Kevin Jégat among the fifteen, alongside others. Earlier, a five-man move of Kwiatkowski, Asgreen, Vervaeke, Kirsch, and Zimmermann had briefly dangled 20 seconds before being caught.¹

The morning count at sign-on was three short. Fernando Gaviria (Caja Rural-Seguros RGA) did not start after breaking his left collarbone in the Stage 12 sprint crash at Chalon-sur-Saône — Caja Rural described him as "our big hope for the win." Jenno Berckmoes (Lotto-Intermarché) also broke his collarbone and will have surgery in Belgium.² Frits Biesterbos (Picnic PostNL) was the third DNS. Liam Slock (Lotto-Intermarché) is racing on with abrasions. Dorian Godon (Netcompany Ineos), who fell heavily near the front of the peloton, is continuing under twice-daily concussion monitoring.² Søren Wærenskjold, Jonas Abrahamsen, and Anthon Charmig from Uno-X reported cuts and abrasions, no fractures, and all started.

Stage 13 is a prelude. The stages that will decide this Tour — and whether Vingegaard still has a GC race to ride — are Saturday's Stage 14 to Le Markstein and Sunday's Stage 15 at Plateau de Solaison. Vingegaard sits 3:36 behind Pogačar, a gap that did not move on the Tourmalet and did not move in the Massif Central. He said after Stage 10 that the short, punchy climbs there were "not what suits me the best" and that coming away with a smaller loss was "something I can be happy with."³ Visma has repeated some version of this every time the deficit held or widened. The line requires validation now.

The pressure comes from two directions. Behind Vingegaard in the standings, Evenepoel, Juan Ayuso, and Paul Seixas are stacked within a minute of his position. Any further losses in the Vosges would not only extend the chase to yellow but invite a fight for the podium from his rear wheel. Stage 14 compresses four categorized climbs into 155km to Le Markstein. Stage 15 ends at Plateau de Solaison — 11km at nine percent, the same summit where Isaac del Toro sealed his Tour Auvergne-Rhône-Alpes victory last month.³ Pogačar said Thursday that Saturday and Sunday are "the bigger days for the GC" and that Stage 13 is about saving energy for them. UAE has been quiet for 48 hours.

The story of how the gap materialized traces back to the Col du Granon in 2022. UAE sport manager Joxean Fernández Matxín told Velo that the collapse there — when Vingegaard and Jumbo broke Pogačar on the final Alpine day — triggered a wholesale program overhaul. "When you win a Tour with a very young rider, and then you win another one, you think it's perhaps easier than it really is. You get carried away a little by the situation without fully thinking it through." The following winter: overhauled nutrition, expanded cooling protocols, larger and deeper support squad. George Bennett, who was inside UAE during the rebuild, put the change plainly: "The Granon woke up the beast. You could say he won his first Tour while eating pizza, drinking beer and playing PlayStation."⁴ Pogačar's own accounting focused on body temperature as the most measurable gain. "During this Tour de France, my body temperature is simply lower than it was in 2022 or in the Tours before that."⁴ He already has three stage wins and 3:36 on Vingegaard. Matxín's summary: anyone waiting for a Granon repeat is chasing a version of the rider that no longer exists.

As this paper reported Thursday, the co-leadership friction between Evenepoel and Florian Lipowitz at Red Bull-Bora-Hansgrohe flared publicly after Stage 6 at Gavarnie. Sports director Patxi Vila said this week the situation is now "totally fine" — the pair resolved it between themselves. Evenepoel sits third at 4:06, Lipowitz sixth at 38 seconds further back.⁵ Whether the Alps favor one over the other is something Vila says he cannot yet predict. "The road will decide," Jai Hindley said. Vila gave the team's Tour performance eight-and-a-half out of ten and still believes both riders are on track for the podium.⁵

↑ Back to top

THE LAB

Kimi K3: The Largest Model Anyone Has Released, One Reasoning Level, a Hidden System Prompt

Moonshot AI announced Kimi K3 on Wednesday, claiming the title of largest published model at 2.8 trillion parameters — they round it to "3T-class" in their marketing, topping DeepSeek V4 Pro's previous record of 1.6 trillion.¹ Open weights don't land until July 27; the API is live now. Moonshot's self-reported benchmarks put K3 above Opus 4.8 max and GPT-5.5 high on most tasks, behind only Fable 5 and Sol. Arena.ai places it first on their Frontend Code arena, ahead of Fable 5; Artificial Analysis's private long-horizon evaluation gives it an Elo of 1,547, a gain of 732 points over K2.6 and second only to Fable 5. Pricing is $3 per million input tokens and $15 per million output, the most expensive model a Chinese lab has released, on par with Anthropic's Sonnet line and about half the price of Opus 4.8 per task.¹

Simon Willison ran his pelican SVG benchmark against K3 and turned up several things the headline numbers don't surface. The model currently has only one reasoning effort level — maximum — which is expensive: a prompt asking for an SVG of a pelican riding a bicycle consumed 13,241 reasoning tokens to produce 3,417 tokens of actual output, totalling 25 cents for a trivial task.¹ More unusual: that nine-word prompt registered at 95 input tokens. Saying "hi" to the model runs 86 tokens, pointing to an 85-token hidden system prompt the model declines to expose. Vision quality is solid — Willison's alt-text follow-up against the rendered SVG came back accurate and detailed. He also notes that his pelican benchmark has effectively decoupled from frontier model quality at the top end; GLM-5.2 produces better SVG pelicans than Fable 5 or Sol, which doesn't make it a better model. What the test still measures usefully: whether a prompt goes through, and what a trivial task costs.

Puter released a project this week that compiles Firefox to WebAssembly so the whole Gecko browser runs inside another browser. They chose Firefox for its strong single-process support. All network traffic routes through Puter's servers over WebSocket using the Wisp protocol — a requirement because browser-executed WASM can't open arbitrary TCP connections — and end-to-end encryption is preserved for HTTPS traffic (Willison verified this by inspecting WebSocket messages). The team estimates the build work consumed roughly $25,000 worth of Claude tokens, though the actual bill was much smaller because they used a Max subscription plan.² The repo is HeyPuter/firefox-wasm. A parallel effort, theogbob/WebkitWasm, compiles WebKit to WASM but currently has no accessible demo.

Separately, Thinking Machines Lab — Mira Murati's company — released open weights for Inkling, a 975B-total / 41B-active MoE with a 1M context window and controllable reasoning effort; ALSO NOTED has the full picture.

Trending today: GitHub is dominated by AI skills files, agent-wrapper repositories, and curated collections — the one retrocomputing outlier worth a look is microsoft/comic-chat, the source code for the Microsoft Comic Chat IRC client, surfacing this week as a historical curiosity.

↑ Back to top

THE LONG READ

The Pledge Nobody Kept: An AI Researcher's Account of How Google Signed the Deal

At 11:45 p.m. in late April, Alexander Turner learned via a Signal group chat that Google had signed. The classified deal allowed the Pentagon to use Google's AI for "any lawful government purpose." The ethical guardrails, such as they were, said Google's systems "should not" be used for autonomous weapons or mass surveillance — language Turner recognized immediately as non-binding.¹ Google never made an internal announcement. The building, when he next went in, felt like a memory. He left.

What Turner published this week is not a typical resignation letter. It is a date-stamped account of a five-month campaign by a single research scientist to stop one of the world's most powerful technology companies from signing away its AI ethics commitments. The essay names names — Jeff Dean, Demis Hassabis, Sundar Pichai, Stuart Russell, Yoshua Bengio, Geoffrey Hinton — and documents, with careful sourcing and explicit acknowledgment of what cannot be verified, how nearly every senior person with both the stated convictions and the institutional leverage to act chose not to.

The story begins in January 2026, when DHS officers shot and killed at least two people during immigration enforcement operations. Turner, a research scientist at Google DeepMind, learned that Google Cloud was part of the DHS supply chain. He decided to find the most effective lever he could pull.

His analysis of leverage was precise. He dismissed mass petitions — Google had already ignored one signed by nearly a thousand employees. He ruled out strikes and sit-ins on the grounds that Google had likely hardened itself against those tactics. He concluded that AI talent is top-heavy: you don't need a hundred engineers, you might need one. That one was Jeff Dean.

Dean is Google's 30th employee, Chief Scientist, co-lead of the Gemini project, and, in Turner's description, "considered a saint at Google."¹ More to the point, Dean had signed the Future of Life Institute's lethal autonomous weapons pledge in 2018, along with Demis Hassabis, Shane Legg, and Google DeepMind as an organization. The pledge was explicit: "we will neither participate in nor support the development, manufacture, trade, or use of lethal autonomous weapons." Turner's argument was that if Google signed an "all lawful use" deal with the Pentagon, Dean would logically be bound by his own pledge to quit. Turner wanted Dean to know he had cover.

He organized a petition, collected roughly 250 signatures from GDM and Research staff, and asked Dean to meet for lunch.¹ Dean agreed.

In late February, the Pentagon issued an ultimatum to Anthropic: remove the red lines from Claude's existing contract — the clauses prohibiting lethal autonomous weapons and AI-enabled mass surveillance — or face designation as a supply chain risk, which would force all military contractors to stop using Claude. Turner was in Paris at an IASEAI conference, the International Association for Safe and Ethical AI, chaired by Stuart Russell.

He expected the room to be discussing the news. Nobody was. "People busied themselves with the usual abstractions," he writes: public choice theory, coordination problems.

Turner interrupted Russell's lunch conversation. Russell was unambiguous: he would convene an IASEAI vote, announce it at closing, get Yoshua Bengio and Geoffrey Hinton on board. At the closing session, Russell called the Pentagon's ultimatum "an extortion racket."¹ He promised a member poll that, if it cleared two-thirds, would produce a public statement.

Turner paid the $75 membership fee that evening.

By Thursday morning, the poll had not materialized. Mark Nitzberg, IASEAI's interim executive director, explained that the organization would need to act after Anthropic's Friday deadline. Then Anthropic made its own statement, holding the red lines. IASEAI concluded it no longer needed to act. The poll was never held, no statement was ever issued, and Nitzberg eventually stopped replying to Turner's messages.

Yoshua Bengio's office declined to make a statement without explanation. Hinton was attending remotely; Turner never reached him directly.

The pattern Turner documents throughout is less dramatic than outright betrayal and more dispiriting for it: commitments made in public contexts evaporating under the friction of actual decision-making. Stuart Russell, who presented the "Slaughterbots" videos to the United Nations and delivered over two hundred talks against autonomous weapons, fell silent at what Turner calls "the first real collision between modern AI and military use." Turner distinguishes between the public frankness in the conference hall — "extortion racket" — and the absence of any statement the world would see.

Turner's lunch with Dean in Mountain View, where he arrived in a dress shirt and slacks, is described with characteristic restraint. He proposed that Dean head a Defense AI Review Body — a 25-page governance framework Turner had drafted on vacation days, vetted by military and surveillance law experts, designed to create binding oversight of which government use cases Google would and would not serve. Dean did not take up the idea. Turner declines to characterize the conversation further, pointing instead to Dean's public conduct: "He tweeted and signed an amicus brief in support of Anthropic. Google later signed the deal. Jeff is still at Google, despite his pledge."

Dean had, notably, publicly signed the amicus brief filed by Protect Democracy, breaking from Google's institutional silence in a way that attracted attention and apparently added to Pentagon officials' concerns about the company. Turner credits this as a genuine act. He simply concludes that it wasn't enough, and that if Dean had threatened to walk, the classified deal would look different: "I would expect the classified deal to contain at least some binding provisions."

Turner sent the framework to Demis Hassabis directly. Hassabis routed it to senior policy staff, who left the message on read. Turner followed up; he was told to circle back in a few months. He offered to fly from San Francisco to London to answer questions. The message was left on read again.

The deal was reported as signed on April 27th.¹ Google's contract terms were weaker than OpenAI's. The "should not" language on autonomous weapons carries no enforcement mechanism.

The essay's sharpest passage concerns Hassabis's public rationalization. When Google acquired DeepMind in 2014, the company agreed never to use its AI for military or weapons purposes. In February 2025, Hassabis co-authored a post announcing updates to Google's AI principles; the update removed the explicit prohibitions on weapons and surveillance. In subsequent interviews, he said "nothing's changed about our principles" and described the guiding standard as: "we've got to thoughtfully weigh up the benefits, and they've got to substantially outweigh the risk of harm."

Turner's response is precise: "That's not a principle. A principle is something you commit to in advance so that you can't talk yourself out of it later, even when the benefits seem to outweigh the harms. One cannot violate a 'principle' of 'I'll decide when I see it.'"

He also notes Hassabis's philosophy of governance: after failing to negotiate a semi-independent legal structure for DeepMind, Hassabis concluded that "trustless" governance was impractical and that the answer was mutual trust and a seat at the table. Turner's counterargument is structural. The framework he proposed required trusting only a single person (the Chief Scientist) long enough to seat a review body; after that, transparency and contract substituted for trust. Hassabis's concern was that a governance board "probably wouldn't do the right thing when it came to the crunch." Turner's reply: a person at the table faces exactly the same crunch, but with worse incentives — equity, social bonds, self-image tied to the company — and no transparency whatsoever.

The result of the experiment, Turner writes, is visible: "When profit and pressure met ethical commitment at Google DeepMind, pressure won and pledges lost. When profit and pressure met ethical commitment at Anthropic, ethics won."

He is not credulous about this distinction. He notes that Anthropic's CEO still does not know what role, if any, Claude played in the bombing of an Iranian girls' school — an "all lawful use" deal eliminates the visibility that would even make the question answerable. But the institutional distinction holds: one company structured its commitments as rules; the other structured them as judgment calls. Under pressure, judgment called in favor of signing.

Turner closes with a direct accounting of who still holds the pledge and what they are doing while holding it.¹ He sees three honest options for a pledge-signer who stays: explain publicly how staying is consistent with the pledge, say plainly that you no longer hold it and why, or quit. "Wearing the pledge while saying nothing isn't one of them."

He declined outreach from the OpenAI safety team. He is currently unemployed.

↑ Back to top

FROM THE ARCHIVE

The Tiger on Main Street: July 17, 1955

A tiger and a panther broke loose from a circus parade and staged what reporters called a "furious death struggle" on Main Street, USA.¹ That was July 17, 1955 — Disneyland's opening day.

Disneyland workers would come to call it Black Sunday. The Associated Press filed that Walt Disney had, "probably for the first time in his career," disappointed thousands of youngsters.¹ Disney told the press: "We'll settle down and get this place operating. It may take a month before everything's going smoothly."

It took about that. The stagecoach ride in Frontierland was shut down in the first weeks after the coaches proved too top-heavy and prone to flipping.¹ Nearly all 36 cars on the Autopia — Disney's utopian miniature freeway, where children were supposed to learn "respectful rules of the road" — were wrecked by aggressive drivers who simply crashed into each other.¹

None of it mattered. Seven weeks after Black Sunday, a million people had paid to walk through Disneyland's gates.¹ Within a few years, the park had surpassed the Grand Canyon and Yellowstone in popularity. By 2015, sixty years in, more than 750 million people had passed through its turnstiles.¹

Disney had built something so right in concept that it survived everything wrong with the execution. The tiger helped.

↑ Back to top

THE FUNNIES

Obviously. / Opening Day Had Not Gone Entirely as Planned

After Peanuts — on the pelican benchmark that stopped measuring anything useful once the dog got too good at it. After The Far Side — on what Disneyland's opening day, July 17, 1955, looked like from the perspective of the review committee.

↑ Back to top

ALSO NOTED

Also Noted

• Fred Wright's Long Wait — The two-time British champion recounts how a thorn lodged in his knee at a Chile altitude camp derailed his first Classics campaign with Pinarello-Q36.5, and why he believes the elusive first international win is close. cyclingnews.com Jul 17, 2026
• Inkling: Thinking Machines Lab's First Open Weights — Mira Murati's company released Inkling, a 975B-total / 41B-active MoE with a 1M context window, controllable thinking effort, multimodal audio and vision, and a SWEBench Verified score of 77.6%, available for fine-tuning on their Tinker platform. thinkingmachines.ai Jul 15, 2026
• Microsoft Comic Chat Goes Open Source — Microsoft released the source for the 1996 IRC client that rendered conversations as comic panels with AI-selected poses and expressions, presented at SIGGRAPH '96, shipped with Windows 98 in 24 languages, and introduced the world to Comic Sans. opensource.microsoft.com Jul 16, 2026
• How a 64k Intro Is Made — A detailed post-mortem of Guberniya — a 61.8kB Windows demo released at Revision 2017 — covering raymarched signed distance fields with hg_sdf, GPU-simulated wire physics via Verlet integration, a Paulstretch ambient synthesizer built without external libraries, and GNU Rocket animation, all in 8,300 lines of C++. lofibucket.com
• Seattle Transit Measure: What's in the Saka Package — Chair Rob Saka's amendment package for the 2026 Seattle Transit Measure — voted on by the Select Committee Thursday ahead of a full council vote July 21 — adds safety and security, affordability, and accessibility provisions to Mayor Wilson's 0.3% sales-tax proposal. council.seattle.gov Jul 15, 2026

↑ Back to top

THE QUESTION

When the Benchmark Outlives Its Calibration

A benchmark doesn't signal the moment it stops measuring the right thing — it just keeps running while the frontier it was calibrated against has moved somewhere else. Two stories in today's edition arrive at this problem from different domains, and neither has an obvious remedy.

Simon Willison's pelican SVG test — a prompt asking a model to draw a bird riding a bicycle, then scoring the result — has, by his own accounting, mostly severed its connection to frontier model quality at the top end. GLM-5.2 produces better SVG pelicans than Fable 5.¹ The benchmark still measures two things: whether a prompt gets through, and what a trivial task costs (Kimi K3 charges $0.25 for nine words).¹ But the thing it originally indexed — relative capability across frontier models — has moved past it. The subject improved; the test didn't retire.

THE PELOTON carries a parallel case in different vocabulary. Joxean Fernández Matxín, UAE's sport manager, described what the 2022 Col du Granon collapse set in motion: overhauled nutrition protocols, expanded cooling capacity, a deeper support squad. The result is a rider whose body temperature "is simply lower than it was in 2022."² His explicit conclusion: any tactical framework built on the Granon collapse is benchmarking an athlete who no longer exists. The 3:36 gap is a real number.² The inference you draw from it — about what it would take to close it, about which climbs are opportunities — depends on a model of Pogačar that was calibrated on 2022 data.

The structural problem is identical in both cases: measurement frameworks get built on snapshots, and snapshots date faster than frameworks do. Software test suites fail this way — a green suite written against the old implementation keeps passing through a refactor because behavior was preserved, but the assertion no longer tracks what it claims to track. Pelican benchmarks fail this way. GC gap analysis calibrated on a crack that happened under a different rider's physiology fails this way. The test doesn't announce when it crosses from signal to noise.

The question worth carrying today is the one both stories circle without resolving: how do you recognize when your measurement is still running but no longer tracking the thing you built it to track? Willison's honest accounting is that the pelican benchmark retains two narrower, valid uses — prompt throughput and cost-per-trivial-task — which is scope revision under observation. That's not a failure; that's what it looks like when someone notices in time. The harder case is when the framework keeps producing confident-looking numbers and nobody has checked whether the calibration still holds.

↑ Back to top