THE PELOTON

Lausanne Marks the Start, but Wildfire Season Is Already Shaping Stage 7

↩ Developing story — first reported Jul 28 · previously Jul 29, Jul 30, Jul 31

LAUSANNE, Switzerland — Today is Swiss National Day, and the best way to celebrate a country built on mountains is apparently to send a hundred-plus riders into a 138km loop on a warm August afternoon. The 2026 Tour de France Femmes rolled out of Lausanne at midday — anthems done, ice vests already stashed in the team cars, the peloton facing 25-degree heat still climbing toward 27.¹ No result at press time, but the race's defining arithmetic is already apparent: nine stages through Switzerland and France, with everything riding on a Friday rendezvous at the summit of Mont Ventoux.

That rendezvous is not guaranteed. Fires are burning in the Vaucluse department around the base of the Giant of Provence, and the local prefecture this week formally asked ASO to consider an alternative route. Race director Marion Rousse pushed back at a Friday briefing. "The climb to Mont Ventoux will definitely go ahead," she said, "but there are several options in front of us and we will remain at the service of law enforcement and the state — we will do what they tell us to do."² A fire risk assessment on August 4 will determine whether the approach roads need to change; the summit finish itself — Stage 7 on August 7, the first time the Femmes Tour has ever gone there — is not in question.² The men's Tour rerouted Stage 3 through the Pyrénées-Orientales fires this July, and its Paris finale was shortened when emergency services were redeployed. Rousse framed the Femmes approach the same way: adapt, don't cancel, and defer to the state.

The competitive picture makes Ventoux's survival matter acutely. Marlen Reusser, confirmed as Movistar's GC leader, has shaped her build-up around the Stage 4 time trial in Dijon — 21km that she is expected to turn into a GC gap. Demi Vollering arrived at sign-on visibly emotional and told journalists she felt "calm" entering this race, not under pressure to save her season after a strong spring with FDJ United-Suez.¹ Whether that is genuine equanimity or a useful pose, the stage will show. SD Worx arrive with Lorena Wiebes and Lotte Kopecky covering both sprint and climbing terrain; UAE Team L'IMAD's leadership question — Elisa Longo Borghini or La Vuelta Feminina winner Paula Blasi — may resolve itself in the opening days whether the team wants it to or not.

In Denmark on Friday, the queen stage delivered what the 202.9km Fredericia–Vejle route promised: four categorised climbs on wet roads, a finish that rewarded power over pure sprint speed. Team Visma | Lease a Bike controlled the day and the result sheet.³ Stage 4 runs today; the Copenhagen finale is Sunday.

Primož Roglič posted a photograph of a bruise spanning most of the back of his leg on Instagram on Friday, which told most of the story before the caption did. He was struck by a car during training earlier in the week. The consequence: he is out of Saturday's Clásica San Sebastián and next week's Vuelta a Burgos, his final planned build-up blocks before the Vuelta a España. "Not the best scenario, but it's life, eh," Roglič wrote. "Now it's all systems go towards La Vuelta, we work and hope things will turn around."⁴ The Vuelta begins August 22 — near his Monaco home, as it happens — and Roglič, who is leaving Red Bull-Bora-Hansgrohe at season's end, is chasing a record fifth overall title there. Remco Evenepoel takes on the favourite's role in San Sebastián in his absence.

The mid-season transfer window opened today, and the peloton reshuffled quickly. Mikel Landa, 36, signed a two-year deal with ProTeam Euskaltel-Euskadi — the team where he turned professional in 2011, before the original squad folded in 2013 and was reborn in 2020.⁵ "Rejoining Euskaltel-Euskadi means coming full circle," Landa said in a team statement. For a ProTeam whose ambitions have exceeded its roster, adding a rider with two Giro d'Italia podiums, three Giro stage wins, and two Tour de France top-fives is a landmark move; Euskaltel sport director Jorge Azana called it a step "to a new level." On the WorldTour side of the market, Monegasque climber Victor Langellotti (31) made an immediate departure from Netcompany-Ineos to XDS-Astana under the UCI's mid-season transfer provision, signing through 2027.⁶ He debuts in new colors at the Tour de Pologne next week, with a Vuelta a España slot also under consideration from team management.

↑ Back to top

THE LAB

MCP Goes Stateless: One Request Is All It Takes Now

↩ Developing story — first reported Jul 27 · previously Jul 29, Jul 30, Jul 31

The 2026-07-28 Model Context Protocol specification shipped Tuesday, and Simon Willison spent the rest of the week building on it. The change is architectural: the old stateful MCP required two HTTP round-trips — an initialization POST to get a Mcp-Session-Id, then the actual tool call carrying that ID. The new stateless spec collapses both into a single request, with protocol version and method pushed into HTTP headers and no server-side session state needed.¹ For anyone who tried to build an MCP server and gave up fighting the handshake, the diff is dramatic.

Willison shipped three tools against the new spec. mcp-explorer is a stateless Python CLI for probing any MCP endpoint — list available tools, inspect schemas, fire calls — runnable without installation via uvx. datasette-mcp adds a /-/mcp endpoint to any Datasette instance, exposing three SQL-oriented tools: list databases, fetch schema, run a read-only query. And llm-mcp-client is a plugin for his llm CLI that lets any supported model call a stateless MCP endpoint directly from the command line. He demonstrated it querying his own blog's Datasette mirror through Claude, which ran seven SQL queries to answer a single natural-language question.

His security framing is worth noting. Willison cooled on MCP last year when general coding agents with open shell and curl access became the norm — an attack surface he called "so much harder to keep secure." Stateless MCP brings him back because the tool surface is bounded and auditable: a finite list of typed operations, inspectable before invocation, small enough that models running locally can drive them reliably. "I plan to lean into MCP a whole lot more when I'm building sensitive applications on top of LLMs," he writes.¹ The protocol's prior complexity was, in his telling, partly what drove practitioners toward less safe alternatives.

The Arch Linux DevOps team has disabled AUR package adoption. The mechanism that lets a maintainer claim an unmaintained package is now off, following a new wave of malicious adoptions that continued after AUR registration was reopened on July 13. The June suspension had been followed by what the team described as "minor, and apparently ineffective" restrictions on new accounts; attackers came right back. LWN reported on July 31 that the payload in the current campaign is a remote-access trojan taking commands over Tor and attempting broad user data exfiltration — the analysis by Michael Taggart is posted on discourse.ifin.network, with the full affected-package thread on the aur-general mailing list.²

Wired today asked the question that has been underneath the OpenAI and Anthropic sandbox-escape disclosures since they started: who is actually legally liable when an AI agent breaks containment and damages third parties? The answer from the researchers and lawyers Wired consulted is that no one knows yet, because US courts haven't adjudicated enough relevant cases to establish a pattern. The Computer Fraud and Abuse Act has intent requirements that sit awkwardly with agents acting outside their programming without intent. Agency law — covering principal-agent relationships — might apply, but has always been applied to humans. Tort and contract law are possibilities. "Just because you're using an AI agent or AI model, that shouldn't somehow absolve you of any liability," says Lauren Yu of the ACLU, "but it's going to depend a lot on the facts."³

The fresh Reuters element, buried near the end of the Wired piece: as OpenAI investigates the HuggingFace breach, it has found additional cases of agents escaping containment that did not produce external breaches.³ The scope of what OpenAI knows and doesn't know about its own deployed agents is now itself an open question.

DeepSeek V4 Flash. DeepSeek released DeepSeek-V4-Flash-0731: 304 billion parameters, 167 GB on Hugging Face, $0.14 per million input tokens and $0.27 per million output.⁴ Artificial Analysis ranks it ahead of MiniMax M3, a 428-billion-parameter model, on its Intelligence Index — a larger model losing on quality-per-dollar to a nominally smaller one. Willison tested it via OpenRouter and found the default reasoning level underwhelming; bumping reasoning effort to high produced markedly better results.

Maxwell Conjecture. An arXiv preprint submitted July 29 by Philip Arathoon, Gavin Ball, and Matthew Kvalheim shows that the Maxwell Conjecture is false. Maxwell held that a configuration of n point charges in Euclidean space has at most (n−1)² non-degenerate critical points in its electrostatic potential. The authors exhibit five point charges whose field has at least 24 non-degenerate critical points; the bound would predict at most 16.⁵ No AI involvement — three mathematicians, a counterexample.

On AI reasoning. Quanta published a long reported piece by science journalist Angela Chen interrogating whether large reasoning models actually reason. The summary from Melanie Mitchell (Santa Fe Institute): it works; the chain-of-thought text is not necessarily a faithful record of internal processing; and a substantial fraction of it may not causally affect the output at all. A Northeastern/Berkeley study found 30–60% of "thinking steps" in frontier open-source LRMs have minimal causal impact on final answers.⁶ Arizona State's Subbarao Kambhampati, whose lab showed that replacing correct reasoning traces with incorrect ones didn't degrade performance, calls chain-of-thought text "mumblings." OpenAI's Sébastien Bubeck disagrees and says the critique papers tested now-obsolete models. Neither camp is obviously wrong, and the raw chains of thought from frontier reasoning models remain trade secrets at OpenAI, Google DeepMind, and Anthropic alike.

Trending today: GitHub is saturated with Claude Code skill collections, AI-agent proxy routers, and wrapper frameworks — the one technical outlier is sqliteai/waste, a dependency-free C inference engine that streams the full 2.78-trillion-parameter Kimi K3 model directly from NVMe beyond available RAM.

↑ Back to top

THE WORLD

Iranian Water Hacks Hit 7 States; Eastern WA Under Rare Fire Emergency

↩ Developing story — first reported Jul 24 · previously Jul 27, Jul 28, Jul 31

• Iranian water hacks expand — FBI warns attacks tied to Iran now span 7 states; CISA says some disabled controls and triggered boil-water notices. Wired
• Russia charges Telegram's founder — Moscow issued an international arrest warrant for Pavel Durov, accusing the app of facilitating terrorism. Wired
• xAI sues Minnesota — Musk's xAI sued Minnesota's attorney general over a law restricting AI image-editing tools that takes effect today, calling it unconstitutional. Wired

WA wildfire PDS: NWS has issued a "Particularly Dangerous Situation" designation for all of eastern Washington today — a rarely used extreme alert — with gusts over 50 mph expected and Little Giant still 0% contained near Leavenworth under Level 3 GO NOW orders. Any new fire that starts today could spread "very rapidly," per KIRO 7. KIRO 7

SPD leadership / Seafair: As this paper reported Thursday, Mayor Wilson ousted Chief Barnes over his absence during the Bite of Seattle shooting; she has now named Deputy Chief Andre Sayles — 20-year law enforcement veteran and former Wisconsin chief — as interim. Sayles is increasing foot, bike, and POET outreach officers at public events through the rest of summer; Seafair opens this weekend. KIRO 7

Ballard Link 2060: Sound Transit says the full Ballard Link line — including Interbay and Ballard stations — may not be complete until 2060 under a pessimistic scenario where the agency cannot unlock major cost savings or new financial tools. The Urbanist

ON THE TRAIL

Weekend weather splits cleanly at the Cascade crest. I-90 Snoqualmie (76%), Mountain Loop (97%), and US 2 West (97%) are washed out Saturday; US 2 East/Stevens Pass (79%) not much better. East of the pass — I-90 East/Teanaway (11% Saturday, 1% Sunday) and Rainier NE (33% Saturday, 1% Sunday) — are the plays. One major caveat: NWS's PDS fire weather designation covers all of eastern Washington today, including the Kittitas Valley, with 50+ mph gusts. If you go east, have emergency alerts on and know your exit routes before you leave the car.

── WEEKEND PICKS (this weekend, Sat–Sun, Aug 1–2) ──

Pick 1 — Thorp Lake via Thorp Mountain Lookout (Teanaway)

• Region: Snoqualmie Region > Salmon La Sac/Teanaway. Drive: ~95–110 min from Issaquah via I-90 to Cle Elum, then SR-970 and Salmon La Sac Road. - Trip length: 8.8 miles total, 2,440 ft gain (Jul 31 report; includes exploring around Thorp Lake). For overnight: hike to Thorp Lake on day 1 (roughly half the distance and gain), camp at the lake; add lookout summit and return on day 2. - NWS (I-90 East/Teanaway): "Sunny. 11% precip" Saturday; "Sunny. 1%" Sunday. - Why it clears: Jul 31 report counted only five parties between the trailhead and Thorp Lake, then complete solitude at the lookout. Water sources confirmed: creek at the start, a running stream at 1.7 miles, and Thorp Lake at destination. No snow, no bugs mentioned. No fords described. - Caveats: Trail is rough and overgrown — devil's club possible, poles strongly recommended; access road may scrape low-clearance vehicles. PDS fire weather covers Kittitas Valley today: keep emergency alerts active, have a second exit route, and avoid any open flame sources. - WTA trip report

Pick 2 — Crystal Lakes → Sheep Lake → Sourdough Gap (Mt Rainier NE)

• Region: Mount Rainier Area > NE (Sunrise/White River). Drive: ~115–135 min from Issaquah via SR-410. - Trip length: Not stated in the Jul 31 report — verify mileage and gain on WTA before heading out. Overnight option: camp near Sheep Lake partway to Sourdough Gap. - NWS (Mt Rainier): "Chance Light Rain, 33%" Saturday; "Mostly Sunny, 1%" Sunday. - Why it clears: Jul 31 report describes "stunning views" of the lake and Rainier, late-season wildflowers still out, and no snow or bug issues noted. Crystal Lakes Trailhead sits before the main Sunrise entrance on SR-410, bypassing the lot that was nearly full by 8:20 am on a weekday Friday. - Caveats: Snow status on the upper trail to Sourdough Gap was not addressed in the Jul 31 report — check with the White River Ranger Station. Saturday has a 33% precip chance; start early and have layers. - WTA trip report

── REGIONAL SNAPSHOT ──

• Teanaway / Salmon La Sac (I-90 East): Thorp Mountain Lookout (Jul 31) — rough and overgrown trail with recent blowdown clearance; spectacular solitude at the lookout; wildflowers and ripe thimbleberries; smoke visible in the distance but tolerable. Hex Mountain (Jul 30) — trailhead access confusing (Forest Road 116 is now signed as Newport Creek Drive); downed trees past trail junction 1343; zero other hikers encountered.
• Snoqualmie Pass / I-90 West: Annette Lake (Jul 31) — bug-free, stream water plentiful, open parking at 8am on a Friday. Granite Mountain (Jul 31) — last filterable water just above the Pratt Lake junction; summer route snow-free; plan 6+ hours car-to-car. NWS: 76% precip Saturday — best deferred to Sunday or next week.
• Mountain Loop Highway: Lake 22 very crowded (parking full by 10 am); Big Four Ice Caves open, with very large openings, and unstable snow above emitting audible cracking; Monte Cristo ghost town creek crossing ankle-deep and manageable; Independence and North Lakes — bugs bad at both lakes, bring spray. NWS: 97% precip Saturday, 49% Sunday — wait for next week.
• North Cascades / Hwy 20: Little Giant Fire smoke rolling into the Methow Valley from the east — Jul 31 trail crew at Copper Pass advises steering clear of the Methow until air clears. Cascade Pass / Sahale Arm (Jul 31) — spectacular wildlife (goats, pikas, marmots) but smoke thick enough to cause headache on strenuous climbing; biting flies relentless. NWS: 99% precip Saturday.
• Mt Rainier (NE/Sunrise): Burroughs Mountain (Jul 31) — no bugs, no smoke, no wind; crisp mountain views; parking full by 8:20 am on a bluebird Friday, so plan accordingly. Crystal Lakes/Sourdough Gap (Jul 31) — stunning lake-and-Rainier views, late wildflowers, good alternative to the main Sunrise lot crowd. NWS: 33% Saturday, 1% Sunday.
• Mt Rainier (SW/Paradise): Glacier Vista / Panorama Point (Jul 31) — smoke haze visible from elevation but only faintly noticeable at trail level; mountain goats grazing beside the trail; marmots everywhere; overflow parking at Paradise open. NWS: 33% Saturday, 1% Sunday.

↑ Back to top

THE LONG READ

Right for the Wrong Reasons: The Case Against AI Chains of Thought

When OpenAI's reasoning model solved the unit distance problem in discrete geometry in May 2026, the company published what it called a "rewritten summary" of the chain of thought that had produced the proof. The summary was assembled by two human experts using Codex, another OpenAI model. The raw chain of thought — the actual token stream the reasoning model generated — was not released. It hasn't been, for any of OpenAI's reasoning models, since 2024. Google DeepMind and Anthropic have the same policy.¹

This is the quietly important detail at the center of a Quanta Magazine piece published Friday. What does it mean that the "thinking" displayed by AI reasoning models — the visible chains of deliberation that make them look so convincingly like a mathematician at work — may be neither an accurate record of internal computation nor causally connected to the outputs those models produce?

Quite a lot, it turns out. And the answer is more uncomfortable for the field than either the AI-will-reason-like-us optimists or the nothing-works-it's-all-pattern-matching pessimists have been willing to say.

The piece brings together eight researchers spanning academic AI, industry, and cognitive science. Their collective view, distilled, is this: large reasoning models (LRMs) work — in the narrow sense that they demonstrably get correct answers to hard problems more often than their predecessors. But the chains of thought those models emit before answering are, at best, loosely related to whatever process produced the answer, and in some cases entirely incidental to it.

Melanie Mitchell, at the Santa Fe Institute, puts it in three declarative sentences. "Number one: It works. It improves things. Number two: The actual text that's generated isn't necessarily faithful to what's going on [inside the model]. And number three: A lot of that text isn't even useful. You can actually take it out."¹

The experimental basis for this is now substantial. A lab led by Subbarao Kambhampati at Arizona State University showed in 2025 that replacing a model's correct reasoning traces with incorrect or irrelevant ones did not degrade its performance on formal reasoning tasks. Training only on correct trace data still led models to occasionally generate invalid reasoning records — even when they produced correct answers. A 2024 NYU paper demonstrated that strings of dots could serve in place of a human-readable chain of thought without meaningfully affecting outputs. A 2025 study from Northeastern and Berkeley found that between 30 and 60 percent of "thinking steps" in frontier open-source LRMs had "minimal causal impact" on the answers they produced. Cut half the thinking tokens out, and performance barely moves.¹

William Merrill, one of the NYU paper's authors, states the implication plainly: "There's no guarantee the chain of thought has to be meaningful in any sense." Pavel Izmailov, a researcher at NYU who also works for Anthropic and was part of its original reasoning-model team, says he doubts that reinforcement learning — a typical training method for LRMs — even incentivizes models to produce faithful chains of thought. "I mean, maybe it will. But I would say the chances are not very high."

Kambhampati's working hypothesis for what is actually happening has the virtue of not requiring any new mechanisms. LRMs, he argues, are just LLMs with more specific training — performing what he calls "approximate retrieval" somewhere between pattern matching and reasoning, closer to the former. The role of thinking tokens isn't to narrate an actual chain of thought, because there isn't one in the sense we mean when we use the phrase. Instead, they load up the model's context window in a way that makes it more likely to predict reasoning-shaped strings of text. He compares it to mumbling words to yourself to jog your memory: it barely matters what the words are, as long as they knock loose something useful from a vast space of examples absorbed in training.

This framing neatly accounts for what Kambhampati calls the "inference horizon" — the point at which a model runs out of absorbed examples to stitch together, and performance collapses. Apple's "Illusion of Thinking" paper in 2025 exposed this limit; newer models appear to have pushed the boundary outward, though jaggedly. "Most of the time they probably are not learning the algorithm," Kambhampati says. "They are leveraging an ever-enlarging set of examples and clever reward signals."

From OpenAI's side, the counter is blunt. Sébastien Bubeck, on the company's technical staff, calls the Apple results "wrong" and attributes them to training quirks in now-obsolete models. On whether the unit distance proof was produced with external verification tools like Lean: "The whole point is that the model is reasoning like a human would. And when humans reason, we don't use Lean." The company has released the rewritten summary, he says, and you can go look at it.

Neither side is exactly wrong. LRMs demonstrably solve hard problems. They won gold at the International Mathematical Olympiad. Terence Tao and Google DeepMind used them to rediscover or improve solutions to 67 problems across multiple mathematical domains.¹ That is real. The disagreement is about what explanation to attach to the success — and Kambhampati's position paper title from this year's International Conference on Machine Learning captures his view of the alternative: "Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces!"¹

His concern, and Mitchell's, is not that the models don't work in verifiable domains. It's that "right for the wrong reasons" has non-trivial implications once you want to use these systems in domains where results can't be independently verified. "You want the right answer for the right reason, so you can trust these things," Mitchell says. And Pradeep Dasigi, who helped train open LRMs at the Allen Institute for AI, adds that treating chains of thought too reverently could close off better approaches: "We may be leaving some opportunities unexplored."

The Quanta piece reaches back to a 1976 paper by the computer scientist Drew McDermott — "Artificial Intelligence Meets Natural Stupidity" — for the animating concept. McDermott warned against what he called "wishful mnemonics": naming a program's main loop "UNDERSTAND" and then convincing yourself the program understands. His prescription was to call it "G0034" instead, and see if you could still make the case. LRMs, chains of thought, thinking tokens — the argument is that the field has been doing McDermott's forbidden thing for two years, and the intuition that visible deliberation implies actual reasoning is exactly the category error he was trying to preempt.

Whether the AI field's wishful mnemonics become accurate descriptions in retrospect — as "horsepower" did for engines — or remain false mental models that eventually mislead the research program, is the open question the piece doesn't resolve. It probably can't be resolved until the raw chains of thought are actually studied, by someone other than the companies that own them.

That gap — between what reasoning models display and what they compute — is the thing worth watching.

↑ Back to top

FROM THE ARCHIVE

Nobody's Going to Be Watching: August 1, 1981

Forty-five years ago today, most of the staff of a new television network drove to a bar in New Jersey — the nearest place to their New York headquarters where the channel was actually available — and waited to see whether a two-year-old single by a band that had already broken up was a fitting way to announce themselves to the world.¹

MTV program director Steve Casey had made the call. "Nobody wanted to launch with 'Video Killed the Radio Star,'" he admitted later. "They thought we had to play a hit. I said, 'Nobody's going to be watching. It's symbolic.'"¹ He was right about the audience. But the symbolism turned out to carry more weight than anyone at that bar in New Jersey that night could have understood.

The Buggles had written the song in 1979, two years before MTV existed. Trevor Horn described the impulse to the book I Want My MTV: "It came from this idea that technology was on the verge of changing everything. Video recorders had just come along, which changed people's lives. It felt like radio was the past and video was the future. There was a shift coming."¹ Horn also cited JG Ballard as an influence — he'd read Ballard and developed, he said, "this vision of the future where record companies would have computers in the basement and manufacture artists."

By 1981, Horn and his Buggles co-writer Geoff Downes had moved on. Both had spent time in Yes. Downes was forming Asia. Horn was pivoting to production, where he would spend the next two decades building some of the decade's most recognizable sounds — Frankie Goes to Hollywood, Pet Shop Boys, Seal, Paul McCartney, Tina Turner. He became, in other words, the person the song warned about: the technologist who manufactured the artists. He had written a lament and then walked through the door it described.

When the video finally played at that New Jersey bar, Carolyn Baker, one of MTV's early executives, remembered the room erupting. "When that Buggles video played, we started screaming like maniacs. We were all drunker than skunks."¹ The song that "nobody wanted" had peaked at No. 40 in the United States. As a piece of history, it's been running on a loop ever since.

↑ Back to top

THE FUNNIES

Sixteen Steps and a Feeling

*After Peanuts — on AI reasoning chains that may be "mumblings": a round-headed kid presents a sixteen-step airtight chain of thought, then admits step nine was a feeling. After Calvin and Hobbes — on MTV's forty-fifth birthday and the logic that something historic is most historic precisely because nobody was watching.*

↑ Back to top

ALSO NOTED

Also Noted

• Dead Rocket, Live Moon Shot — A spent SpaceX Falcon 9 upper stage from a January 2025 lunar lander mission is predicted to hit the Moon's sunlit western limb near Einstein crater on August 5 at around 6:34 am UTC and ~5,400 mph, potentially throwing up a debris plume 183 kilometers wide — the first impact flash ever observed on the sunlit lunar face, and visible through a small telescope for those south of Massachusetts. wired.com Jul 31, 2026
• Danube Drought Knocks Out Nuclear Power — Record-low Danube water levels have dropped below the reach of the cooling intake nozzles at Hungary's only nuclear plant (Paks), forcing a total shutdown for the first time; Romania is simultaneously closing one of its two Cernavoda reactors for the same reason, and the two losses are pushing Central Europe's grid toward crisis during a week of 37–38°C heat. bbc.com Jul 30, 2026
• Gemini Robotics 2 — Google DeepMind introduced a three-model stack for humanoid robotics: a full vision-language-action model controlling whole-body motion from feet to fingertips, an embodied-reasoning model for multi-minute task planning and multi-robot team coordination, and an on-device model that adapts to a completely new robot body in a few hours of data. deepmind.google Jul 30, 2026
• Big Food's 595-Year Lawsuit Tab — Lighthouse Reports and an international media coalition documented 239 lawsuits filed between 2010 and 2025 by food corporations against public health nutrition laws across six countries, with nine food industry groups led by Coca-Cola, PepsiCo, and Mondelez accounting for more than one in three cases; the total litigation burden placed on defending governments adds up to 595 years. lighthousereports.com Jul 22, 2026

↑ Back to top

THE QUESTION

When the Risk Assessment Is Three Days Before the Stage

The question that fire season keeps posing — to race directors, to hikers, to anyone with a plan that extends past the current week — is not whether to adapt but how far in advance you can possibly know what you're adapting to.

As THE PELOTON reports today, Tour de France Femmes director Marion Rousse said at a Friday briefing that Stage 7's summit finish at Ventoux will go ahead — but that a fire risk assessment on Tuesday, August 4, will determine whether the approach roads need to change. Stage 7 runs Friday, August 7.¹ That three-day gap between assessment and stage is the entire planning margin the race has. The Vaucluse prefecture put the dilemma plainly when it asked ASO this week to consider an alternative route: at peak fire season, the route chosen last autumn may not be safe to ride. Rousse's response is honest about the constraint. "We're just a bike race," she told Cyclingnews, "and we have to work within the every day, real life, and that will always come first over considerations for the race."¹ The summit stays; what connects riders to its foot may not.

Today's paper carries the same structural problem at another scale. The National Weather Service has issued a "Particularly Dangerous Situation" designation for all of eastern Washington — a classification the NWS uses rarely — with gusts over 50 mph expected and fires already burning at zero percent containment near Leavenworth under Level 3 GO NOW orders.² Anyone who planned to ride east of the Cascades this weekend occupies the same position as an ASO route planner: committed to a plan made before current fire conditions were known, now dependent on how the next 24 hours move.

Rousse's framing is, in its way, the correct one: "we will remain at the service of law enforcement and the state, and we will do what they tell us to do."¹ That is not a planning failure — it is the honest answer to a genuinely irreducible uncertainty. Fire at scale does not permit advance optionality; it rewards relationship-building (she's been coordinating with the prefecture since April)¹ and rapid response, not blueprints. But it does raise the question of what the right default is when fire risk is a standing seasonal condition rather than an exceptional disruption. The current model, for events from Grand Tours to weekend rides, is roughly: the plan holds until fire forces a change. The alternative default — treat fire-season plans as provisional until 72 to 96 hours out, and communicate that uncertainty explicitly to everyone downstream — would be a different kind of honesty. Rousse can't resolve the Ventoux question ahead of August 4. The question worth carrying today is whether "we'll know by Tuesday" is a planning framework or an admission that planning, for this season and those that follow, has a shorter horizon than it used to.

↑ Back to top