THE WORLD

Starmer Out; Greenspan Dies at 100; Council Sets Transit Vote

↩ Developing story — first reported Jun 20

• Starmer out — UK PM Keir Starmer resigned Monday after May's local elections cost Labour nearly 1,500 councillors; Andy Burnham is expected to take over by September. Wired
• Greenspan dead at 100 — Former Fed Chair Alan Greenspan died Monday of Parkinson's complications at 100; he served four presidents and steered the U.S. economy for 18 years. KIRO 7
• Iran talks, round 2 — JD Vance arrived in Switzerland Sunday for a second round of Iran nuclear framework talks. France24

Team USA may be back. If the U.S. squad wins in Santa Clara on July 1, they advance to the round of 16 — which lands in Seattle, potentially July 6. Seattle is hosting six World Cup matches total; four remain. Visit Seattle projects the tournament's total economic impact at $846 million for the metro area. After Friday's USA–Australia win packed the streets, local venues are already planning for a possible return engagement. KIRO 7

Seattle Transit Measure goes to council July 21. As this paper reported Friday, Mayor Wilson rallied behind an expanded 0.15% sales tax renewal of the Seattle Transit Measure. A council vote is now formally scheduled for July 21. The outcome will determine whether the city's transit funding gets renewed and expanded, or sent back for revision. Mass Transit Magazine

PSE rates are up, and more increases are coming. The UTC approved a ~12% electric rate increase (≈$17/month for the average 800 kWh customer) and ~7% gas increase (≈$6.50/month) effective January 1. PSE has since filed for further increases: 15.2% electric in 2027, 3.7% in 2028, and 8.7% in 2029. Rebates are available for appliances, EV chargers, insulation, and heat pumps; income-eligible customers may qualify for the Bill Discount Rate or up to $1,000 through the Home Energy Lifeline Program. Public hearings on the proposed 2027–2029 increases are scheduled for September 29 (in-person) and October 7 (virtual). Issaquah Reporter

ON THE TRAIL

PART 1 — WEEKEND PICKS

Window: this coming weekend (Sat Jun 27 – Sun Jun 28). Next federal holiday is Independence Day, Friday Jul 4 — 12 days out, beyond the ~7-day expansion threshold, so no long-weekend extension applies this edition.

Weather reality check: The I-90/Snoqualmie forecast is grim for Saturday — 89% rain, high 51°F. Sunday recovers to 48% and 55°F, but the weekend as a whole is not suitable for backpacking in that corridor. Mountain Loop and US 2 West are similarly washed out (93% and 95% Saturday). The only viable backpacking zone this weekend is I-90 East (Teanaway/Cle Elum), where Saturday comes in at 22% and Sunday at 16%.

---

Pick 1 — Paddy-Go-Easy Pass / Sprite Lake (1-night) I-90 East (Teanaway/Cle Elum/Roslyn) — ≈ 90–110 min from Issaquah

• Trip length: 1-night - Per-day mileage: Day 1 in: ≈4.5 mi, +2,700 ft / Day 2 out: ≈4.5 mi, –2,700 ft (estimate — WTA hike page not in sources) - Weather (I-90 East/Teanaway region): Sat Jun 27: 59°F / 41°F, 22% | Sat night: low 41°F, 16% | Sun Jun 28: 62°F / 41°F, 16% - Why it clears: A Jun 21 report found intermittent snow patches above the pass but no microspikes required — trail runners will work with care. Sprite Lake provides water at camp. Bugs were not flagged (in contrast to nearby Deep Lake, where "mosquitoes just hatched" the same day — the higher-elevation pass camp appears cleaner). The trail was quiet; crowds are low. The route has no trail fords, though the road to the trailhead requires Scatter Creek crossing — high-clearance vehicle is necessary; low-clearance cars should park along the FSR. One caution: 10–20 blowdowns on trail requiring awkward step-overs; manageable with a pack but expect slower going. - Trip report (Jun 21)

No 2-night pick is available this weekend. Deep Lake/Peggy's Pond (the natural 2-night option in Teanaway) had a Jun 21 report describing mosquitoes as "out in full force — just hatched." Lake Ingalls is currently rated dangerous (risky snow slopes and no established trail past the pass). West Fork Foss Lakes (normally excellent for multi-night) sits in the US 2 West zone where Saturday is 95% rain. No backpacking destination in any accessible region clears all six criteria for a 2-night this weekend.

---

PART 2 — REGIONAL SNAPSHOT

• I-90 West / Snoqualmie: Annette Lake (Snoqualmie Pass) — trail in good condition, manageable crowds with early start; lot full by 10:30am on Sunday. Island Lake / Rainbow Lake — not crowded past the Olallie junction, bugs present at Island Lake but not biting. Ira Spring / Mason Lake — open parking at noon, beargrass and lupine in bloom. Twin Falls — waterfall staircase viewpoint still closed (storm damage); main bridge view open but lot lined up a mile down the road by mid-morning. Snow Lake / Gem Lake (Snoqualmie Pass) — Snow Lake trail extremely popular; bugs worsen sharply above Snow Lake toward Gem.
• I-90 East (Teanaway): Paddy-Go-Easy Pass — intermittent snow above 6,000 ft, no traction needed; 10–20 blowdowns on trail. Deep Lake / Peggy's Pond — mosquitoes described as "just hatched" and "out in full force" Jun 21; bring head nets. Lake Ingalls — currently dangerous: risky snow slope crossings and no clear trail below the lake; not recommended until further melt.
• US 2 West (Stevens Pass): Lake Serene — gorgeous and swimable, 75% full parking by 9am Sunday. West Fork Foss River (Trout/Copper/Little Heart Lakes) — trail in excellent shape above Trout Lake; near-full lot at 8am; some blowdowns between Trout and Copper require climbing over with a pack.
• Mountain Loop Highway: Gothic Basin — no snow until near the saddle; microspikes recommended but conditions manageable. Mount Dickerman — minimal snow (tiny patch only), bugs present but not bad enough for spray, quiet trailhead at 8am. Round Lake / Breccia — ice axe and gaiters still required; campsites under snow or sopping wet.
• Mt Rainier area: Tamanos Mountain (Cayuse Pass/Stevens Canyon) — summit-worthy conditions, no snow on trail, bear spotted below the saddle; only ~12 people seen all day with Sunrise still closed. White River to Sunrise — snow patches on upper Sourdough Ridge loop; Yakima Park very green.
• North Cascades (Hwy 20): Blue Lake — a "lot" of mosquitoes, ~10–15 small snow patches. Thornton Lakes / Trappers Peak — rough single-lane gravel road requiring high clearance; good views once up.

↑ Back to top

THE PELOTON

Stronger Than Last Year: Pogačar Closes the Books on Switzerland

↩ Developing story — first reported Jun 15 · previously Jun 16, Jun 17

VILLARS-SUR-OLLON, Switzerland — During a training camp at Sierra Nevada between the Tour de Romandie and this week's Tour de Suisse, Tadej Pogačar climbed a reference ascent he'd benchmarked in 2025. He said he thought: "Wow, I'll never be able to go faster than that." Then he went back at the end of camp, for fun. He was significantly faster.¹ That number, whatever it was, is now sitting in a UAE Team Emirates-XRG data file somewhere, and it has to concern every rider heading to Barcelona on July 4.

Pogačar won three of the five stages at the new-format Tour de Suisse, including Sunday's brutal finale on the Col de la Croix above Villars-sur-Ollon, and finished 6:32 clear of Richard Carapaz (EF Education-EasyPost) on GC.² The margin understates how uncontested it was. He attacked with 9km remaining on the fourth and final ascent, caught the entire breakaway one by one, and passed the last survivor, Lenny Martinez (Bahrain-Victorious), with 800 metres to the line.² Carapaz "briefly tried" to follow the initial acceleration, per reports. He did not last long. The heat — temperatures requiring UAE staffers to station themselves along the climb to hand up water bottles purely for cooling — slowed everyone except Pogačar, who described it as an all-out effort but looked methodical throughout.

"The shape must be ready," he told media afterward. "There's not much you can do with 10 days to go." He travels home to Monaco — to spend time with partner Urška Žigart, recovering from her fractured jaw in the women's Tour de Suisse — before a stint at altitude in Isola 2000, then Barcelona for the grand départ. His fourth Tour title came in circumstances that allowed for excuses about his second-half slump. His fifth will not.

As this paper reported June 17, Wout van Aert will miss the Tour de France after a training crash infection serious enough that Visma's head of performance Mathieu Heijboer told Sporza a surgical cleaning was "truly necessary — otherwise, sepsis would have been a possible consequence."³ On Sunday, van Aert rode his first kilometres since abandoning the Tour Auvergne-Rhône-Alpes on June 12 — a 67.6km charity group ride near his home in Herentals, with a large bandage still on his right elbow.³ His Strava caption: "First attempt to hold my bars again."

Visma will announce their Tour replacement on Tuesday.³ The leading candidates are Ben Tulett, who offered punchy climbing support at last year's Vuelta and can ride a similar-sized bike to Vingegaard (useful for emergency wheel swaps), and Wilco Kelderman, who returned to racing at the Tour de Suisse after breaking his collarbone at the Giro.⁴ Davide Piganzoli, who anchored Vingegaard's Giro-winning support in the opening two weeks, has also been mentioned, though fatigue from a full grand tour so soon is a real concern. Jonas Vingegaard heads to France to defend his title having never directly beaten Pogačar in a stage race this year — their first duel since last July's Tour is now 12 days away.

Elsewhere on Sunday: Florian Lipowitz (Red Bull-Bora-Hansgrohe) won the Tour of Slovenia overall, attacking solo over the steep Trška Gora with 9km remaining on the final stage and arriving in Novo Mesto 24 seconds ahead of teammate Laurence Pithie.⁵ He leads the GC into the Tour as the most consistent stage-racer on the start list — top-three at Catalunya, Itzulia, and Romandie — though VeloNews, ranking the pre-Tour field today, places him fourth while noting he's "never shown a hint of yellow jersey pedigree."⁶ The teammate he'll be supporting, Remco Evenepoel, ranks sixth in the same assessment, categorized as a wildcard who could finish anywhere from podium to abandonment.

Tom Pidcock (Pinarello-Q36.5) got his hands in the air at the Andorra MoraBanc Clàssica, outclimbing a select group at the top of the Coll de la Bottella to win ahead of local resident Carlos Verona (Lidl-Trek) and Sepp Kuss (Visma-Lease a Bike).⁷ Pidcock called it his final race before the Tour and acknowledged "some difficult weeks" in combining mountain bike and road campaigns — he hadn't raced on the road since early May. "It's nice to be coming out the other side," he said, not exactly inspiring confidence, though the win speaks louder than the hedging.

Paula Blasi (UAE Team ADQ), ranked second in the UCI women's standings after a season that includes the Amstel Gold Race, La Vuelta Femenina, Tour des Pyrénées, and the Volta a Catalunya, confirmed today she will start the Tour de France Femmes in Lausanne on August 1.⁸ The 23-year-old is playing down any GC ambitions — "I'll just be there to help the team" — which is either genuine humility from a rider still learning the sport's biggest stages, or the most effective expectation management in the women's peloton right now.

On the equipment side, Wilier has launched the Filante Ultimate to mark its 120th year in business — 120 bikes at an undisclosed price, built around six carbon fibre types including Toray M40X, with Dura-Ace, CeramicSpeed OSPW, and Miche wheels and carbon chainring.⁹ The standard Dura-Ace Filante SLR ID2 retails at £9,500. The Ultimate will be more.

↑ Back to top

THE LAB

ryg Tears Apart PivCo-Huffman: How to Merge Bitstreams on Any Vector Width

The problem with Huffman decoding has always been that it is irreducibly serial. You cannot decode symbol N+1 until you know how many bits symbol N consumed. Every scheme for working around this — multiple streams, ANS-style interleaved bitstreams, speculative parallel decode — forces a tradeoff: pick a vector width at bitstream-definition time, or throw away most of the work you do. A recent paper called PivCo-Huffman by Marcin Żukowski breaks this constraint, and yesterday Fabian Giesen (ryg) published a characteristically detailed analysis of what the algorithm's core merge operation actually costs at the instruction level across SSE4.2, AVX2, NEON, and AVX-512.¹

The key insight of PivCo-Huffman is a reframing: instead of processing symbols one at a time down the Huffman tree, you push the entire sequence through the tree level by level, partitioning it at each internal node based on a bit in the bitstream. Encoding becomes a sequence of list partitions; decoding becomes a sequence of list merges steered by a bit vector. That merge — take two lists and a bitstream, output the interleaved result — reduces to a prefix sum problem. Prefix sums are the prototype of parallelizable work. The bitstream carries no magic interleave constant and does not need to be written with a particular SIMD width in mind.

Giesen's post is specifically about the cheapest possible implementation of that merge primitive across real hardware targets. For AVX-512 with VBMI2, it is essentially free: VPEXPANDB does the masked scatter-gather in one instruction, three vector ops total.¹ The interesting engineering is what happens without that instruction. For SSE4.2 and NEON, he arrives at a scheme that processes 16 output bytes per loop iteration using two 256-entry lookup tables (6 KiB total, L1-resident), two vector shuffles using PSHUFB/TBL, a bitwise NOT, and an OR.¹ The trick for generating the index vectors for the second half without a 1 MiB table is elegant: store the population count of the first byte of the bitmask directly in the upper half of each first-table entry, then add rather than compute it. On NEON with two-register 128-bit TBL available (Apple Silicon, newer Cortex cores), he collapses even further: two table loads, a single VABD (absolute difference) to merge the index control, one two-source shuffle, one store. He argues this is capital-O optimal — the loads and store are strictly necessary, the shuffle is the minimum work given the inputs, and you cannot avoid at least two table lookups without a prohibitively large table.¹

The practical upshot: PivCo-Huffman's decoder can be implemented efficiently on essentially any data-parallel substrate, including narrow vector machines, without per-platform bitstream variants. The paper's benchmark numbers are on AVX-512 servers and Apple Silicon — "very high-end hardware," Giesen notes — but the algorithm does not require it. His implementations have been tested in Żukowski's GitHub repository and show improvement over prior baselines.¹

Simon Willison shipped sqlite-utils 4.0rc1 on June 21, the first release candidate for a version that has been in alpha since May 2025.² Two new features land in this RC. First, database migrations, ported from the standalone sqlite-migrate package Willison has been using in production tools including LLM for several years; the API uses Python decorators and runs via sqlite-utils migrate from the CLI. There are no reverse migrations by design — mistakes get fixed forward.² Second, db.atomic(), a nested transaction API borrowing Django and Peewee's terminology. SQLite's savepoint mechanism has always supported nested transactions; this gives sqlite-utils users a clean abstraction over it. The major-version bump exists because of breaking changes accumulated since 3.x: upserts now use INSERT ... ON CONFLICT SET, the default float type changes from FLOAT to REAL, table and column names in schemas now use double-quotes instead of square brackets, and type detection defaults on for CSV imports.

Separately: if you are running OpenAI's Codex CLI on Linux, check your SSD. A GitHub issue filed June 14 documents that Codex's SQLite feedback log (~/.codex/logs_2.sqlite) was writing approximately 37 TB over 21 days on one reporter's machine — roughly 640 TB/year — because the log sink defaults to a global TRACE level that captures inotify filesystem events, raw WebSocket payloads, and OpenTelemetry mirror entries continuously.³ The WAL write pattern shows around 36,000 inserts per 15 seconds with continuous prune-and-insert amplification, so the retained database size understates the actual drive write volume considerably.³ The issue remains open with no fix merged as of today; the proposed fix is straightforward log-level narrowing plus a global write cap.

Trending today: GitHub is saturated with AI agent wrappers, LLM prompt collections, and stock-analysis tools — nothing cleared the technical novelty bar for THE LAB.

↑ Back to top

THE LONG READ

The Jagged Edge: Karpathy on What Actually Changed in December

In late April, Andrej Karpathy published a summary of his Sequoia Ascent fireside chat, and it is one of the cleaner attempts to explain what actually shifted in the last six months of coding — not as hype, but as a structural account of why the tools feel different now.

His opening claim is specific: December 2025 was the inflection. Not 2023, not the launch of some particular model, but December. He describes sitting with Claude Code and Codex on a break, asking for more and more, and realizing he could not remember the last time he had corrected the output.¹ "The chunks just came out fine." The unit of programming had changed from writing lines to delegating macro-actions — implement this feature, refactor this subsystem, write tests, run them, fix failures.

From there he builds the Software 3.0 thesis, which he has stated before but sharpens here. Software 1.0: humans write code. Software 2.0: humans create datasets and objectives; the program is learned into weights. Software 3.0: the context window is the lever, the LLM is the interpreter. The clearest illustration is his MenuGen story — a web app he built to photograph a restaurant menu and generate dish images. The Software 3.0 version: hand the photo to a multimodal model and ask it to render images directly onto the menu pixels. The whole app disappears.¹ "That app shouldn't exist," he writes. The old software stack was scaffolding around a transformation the model can now perform directly.

The most useful section, for a working engineer, is the one on jagged intelligence. Karpathy argues that model capability is not a smooth curve but a product of verifiability multiplied by training attention multiplied by data coverage.¹ A task that is resettable, repeatable, and rewardable — code that runs or crashes, tests that pass or fail — benefits from reinforcement learning in a way that murky tasks do not. This is why the same model that can refactor a 100,000-line codebase will tell you to walk to the car wash 50 meters away rather than drive. Not a sign of shallow general intelligence, but a sign that the car-wash commonsense problem was not in any RL loop that mattered. "We are slightly at the mercy of what the labs do and what they put into the mix."¹ The chess anecdote is telling: GPT-4's apparent improvement at chess relative to GPT-3.5 may have had less to do with emergent reasoning than with someone at OpenAI deciding to add a large chess dataset to pretraining.¹

The practical implication Karpathy draws is a distinction between vibe coding and agentic engineering. Vibe coding raises the floor — almost anyone can now produce working software by describing what they want. Agentic engineering is different: it is the professional discipline of coordinating fallible, stochastic agents while preserving correctness, security, and maintainability. The 10x engineer formulation, he says, is now a serious understatement for people who master this.

What human judgment remains irreplaceable? His example is a bug from MenuGen's own codebase: an agent matched Stripe purchases to Google accounts by email address. Plausible code, bad system design — those two email addresses can differ. The agent had no concept of a persistent user ID as a design primitive. "You are in charge of taste, engineering, design, and whether the system makes sense."¹ The agents fill in the blanks beneath that.

The final section, on education, carries the piece's best line — one he attributes to a tweet that "blew his mind": *You can outsource your thinking, but you can't outsource your understanding.*¹ The argument is that even as agents do more of the computation, the human remains the bottleneck for knowing what is worth building, why it matters, and when the output is wrong. Understanding, he says, is what lets you direct the agents — and the LLMs do not yet fully excel at it.

His "ghosts not animals" framing closes the loop. These systems are not biological intelligences with drives, curiosity, or intrinsic motivation. They are statistical simulations of human artifacts, shaped by pretraining mixtures, RL environments, benchmark pressure, and economic incentives. Anthropomorphic expectations mislead. "The right posture is neither dismissal nor blind trust. It is empirical familiarity: learn where they work, where they fail, what they were trained for, and how to build guardrails around them."

The piece is available in both summary and lightly edited transcript form at Karpathy's blog.

↑ Back to top

FROM THE ARCHIVE

Half-Angel, Half-Devil: June 22, 1986

Pen-and-ink editorial illustration. A compact figure in a number-ten shirt bursts through a corridor of defenders on a packed stadium pitch, the Azteca's vast terraces looming behind. Three opponents lunge and miss — one sprawled sideways, arms out, another pivoting too late, a third reaching with a desperate hand. The small figure leans into a sharp diagonal, head down, driving hard toward the far corner. Harsh overhead light carves deep shadow under the stands. Bold hatching on the crowd, cross-hatching on the defender's outstretched kit, clean white space on the open turf ahead. Strong diagonal composition, lower-left to upper-right. No color. No text. No scoreboard.

At the 51st minute of a World Cup quarterfinal, Diego Maradona punched a ball into the net with his left hand, turned to his teammates, and said: "Come hug me, or the referee isn't going to allow it." They did. Referee Ali Bin Nasser of Tunisia, who had not seen the infringement, allowed it. England goalkeeper Peter Shilton, six feet one to Maradona's five feet five, had come out to punch the ball clear but Maradona reached it first.¹

Four minutes later, Maradona received a pass inside his own half and started running. He covered roughly 60 yards in 10 seconds, passing Peter Beardsley, Peter Reid, Terry Butcher twice, and Terry Fenwick, then left Shilton on the ground with a feint and rolled the ball into an empty net.¹ It was the same match, the same player, forty years ago today at the Estadio Azteca in Mexico City.

Argentina won 2–1. They went on to win the tournament, defeating West Germany in the final. At the post-game press conference, Maradona said the first goal was scored "a little with the head of Maradona and a little with the hand of God."¹ The name stuck. The second goal was voted FIFA's Goal of the Century in 2002, beating out an Owen strike against the same Argentina side in 1998. The shirt Maradona wore that day — swapped in the tunnel with England midfielder Steve Hodge — sold at Sotheby's auction in May 2022 for £7,142,500, a record for sports memorabilia.¹

The match was not played in isolation. The 1982 Falklands War had ended four years earlier with 285 British and 659 Argentine deaths.¹ When Maradona was asked about the hand ball in a 2019 documentary, he described it as "symbolic revenge."¹ He had told reporters at the time that the goal was scored with the hand of God; in private, according to the documentary, "I knew it was my hand." On 17 August 2015, Maradona visited Ali Bin Nasser at his home, gave him a signed Argentine jersey, and called him "my eternal friend."¹

The French newspaper L'Équipe described Maradona after that match as "half-angel, half-devil." The description was precise. The two goals — one a fraud, one perhaps the greatest individual moment in the sport's history — arrived four minutes apart, from the same left foot, in the same quarterfinal, against the same goalkeeper. FIFA's own voters chose the second as the best goal ever scored and placed the first goal's match in the tournament record. Both facts remain true simultaneously, which is more or less the whole problem.

The FIFA World Cup is active in Seattle this summer. Forty years on, the match that produced the Hand of God and the Goal of the Century is still the clearest example the sport has of what it means to be both things at once.

↑ Back to top

THE FUNNIES

Very Productive / Walk 50 Meters North

*After Dilbert — on the discovery that the AI coding assistant has been quietly writing 37 terabytes to a log file in three weeks, which is very productive behavior for the SSD. After Calvin and Hobbes — on jagged intelligence: the same AI that can refactor a hundred-thousand-line codebase will hand you a twelve-page treatise when you ask how to get to the car wash.*

↑ Back to top

ALSO NOTED

Also Noted

• Lael Wilcox abandons Around the World record bid — Wilcox quit her circumnavigation attempt in France after 14 days and 3,075 miles, citing daily nausea and vomiting from the European heatwave; she was targeting Mark Beaumont's 2017 record but called the conditions "another climate change lesson." bikeradar.com Jun 22, 2026
• Triple Olympian arrested at Lincoln Memorial Reflecting Pool — David Hearn, 67-year-old masters racer and three-time Olympic canoe slalom competitor, was detained by U.S. Park Police on a misdemeanor destruction-of-property charge after touching a section of the pool's already-peeling blue liner during a training ride; he denies the allegations and is due in D.C. Superior Court on July 9. velo.outsideonline.com Jun 21, 2026
• Bardet heading back to Decathlon CMA CGM — as Sports Manager — Romain Bardet will rejoin his former team as a senior executive from January 1, 2027, overseeing the full sporting structure from WorldTour to junior development; he will have no role in this year's Tour de France, which is Paul Seixas's debut. cyclingnews.com Jun 18, 2026
• Jenny Rissveds and Luca Martin take Lenzerheide XCO — Rissveds won for the second time this season at the UCI Mountain Bike World Series round in Lenzerheide on Sunday, while Luca Martin took his first men's World Series win of the year. cyclingnews.com Jun 21, 2026
• King County Stage 1 Burn Ban in effect today — Fire Marshal Eric Urban's ban, effective 12:01 a.m. Monday, prohibits yard debris burns and land-clearing fires in unincorporated King County through the dry summer forecast; recreational campfires remain allowed under specific size and clearance rules, and consumer fireworks remain prohibited year-round in unincorporated areas. issaquahreporter.com Jun 18, 2026
• AMD restoring TSME on Ryzen 9000 after quietly removing it — AMD will reinstate Transparent Secure Memory Encryption on non-PRO Ryzen 9000 desktop CPUs via a July BIOS update after the firmware-level removal — discovered during a security audit and confirmed via a bug report — drew community backlash; the feature protects against cold-boot memory extraction attacks. tomshardware.com Jun 19, 2026
• Deno Desktop ships in v2.9 canary — The feature lets developers compile any TypeScript or web-framework project (Next.js, Astro, SvelteKit, and others) into a standalone desktop binary with a bundled webview, in-process bindings instead of IPC, cross-compilation from a single machine, and built-in bsdiff auto-update; the CEF backend adds consistent Chromium rendering at the cost of binary size. docs.deno.com

↑ Back to top

THE QUESTION

Can You Replace a Multi-Domain Contributor With a Committee of Specialists?

The hardest replacement problem in sport is not finding someone faster — it is finding someone who does four things adequately when the person you lost did all four exceptionally.

Visma-Lease a Bike will name their Tour de France substitute for Wout van Aert on Tuesday, as THE PELOTON reports today.¹ The candidates are real riders with real strengths: Ben Tulett offers punchy climbing support and rides a similar-sized bike to Vingegaard — useful for emergency wheel swaps in the mountains.² Wilco Kelderman has racing legs again after breaking his collarbone at the Giro. Davide Piganzoli anchored the Giro-winning support block. Any one of them slots cleanly into one dimension of what van Aert did. None of them slots into all of it.

Van Aert's value to Visma was not located in any single capability. He could lead out a sprint. He could ride tempo on a mountain pass for an hour. He could descend at speed through chaos to bring back a breakaway. He was the rider Vingegaard wanted nearby on the Champs-Élysées and also the rider Visma deployed on days when the race needed to be controlled from the front. When Visma lost him, they didn't lose a sprinter or a climber or a domestique — they lost the rider who meant they never had to choose between those roles on a given stage.

This is the question worth carrying: when a contributor's value is genuinely multi-domain — distributed across capabilities that aren't individually extraordinary but are extraordinary in combination — is it actually replaceable? You can name a committee. Tulett for the mountains, Kelderman for the rollers, someone else for the sprint lead-out. Whether that committee performs like the integrated original is a different question, and it will only be answered at the moments when the race requires all four things at once, with no time to convene.

Visma will find out somewhere between Lyon and Paris.

↑ Back to top