THE PELOTON

Forty Riders Went Up the Road in Belfort, and the GC Picture Came Back Different

↩ Developing story — first reported Jul 14 · previously Jul 15, Jul 16, Jul 17

BELFORT, FRANCE — By the time the race hit the lower slopes of the Ballon d'Alsace on Friday, the day's breakaway numbered forty riders — so large and so stacked with GC-team lieutenants that Decathlon CMA CGM and Soudal-QuickStep were the only two outfits left without a man up the road.¹ Tom Pidcock had talked his way into it too, with three Pinarello-Q36.5 teammates around him,² and by the summit it was obvious the longest stage of this Tour had stopped being a formality for the breakaway and started being a genuine GC problem for everyone who wasn't in it.

Pidcock started the day 10th overall, a rider nobody expected to be allowed off the front on a 205.8km stage built for sprinters' teams to control. Instead he found himself in a 40-strong split that built a lead approaching eight minutes before Bahrain-Victorious even worked out they needed to defend a top-ten placing.² "We thought always from the start that we'd try to get in the break if we can, stay near the front, do the minimum," Pidcock said afterward. "From doing that, there was a split, 40 of us were in the front, 40 guys with four of us, it was the perfect situation and from there we just rode."²

The group thinned hard on the climb itself — Maxim Van Gils went with 34km to go, Luke Plapp countered, and Pidcock had to close both moves himself before cresting the summit with a select group that included Mauro Schmid.¹ "Tom is such a fighter and he's really turned this race upside down and made it really exciting again," Pinarello-Q36.5 manager Doug Ryder said at the finish. "He's such an impact guy and an impact rider."² Behind, the maillot jaune group was content to let the whole scene play out; Tadej Pogačar had no interest in spending matches on a stage he'd said he wanted to take easy before the Alps, and the moves that mattered wouldn't be for the stage win.¹

What it means for the race is the story now. Pogačar's cushion over Jonas Vingegaard held through the stage — there were no time bonuses up for grabs among the top three on the day¹ — but Pidcock's ride reshuffled everything below it, pushing him from mid-pack into serious GC contention with a hard week still ahead. He wasn't pretending it changes his ceiling. "I'm going to lose a lot of time in the time trial, that's for sure. I'll keep doing what I'm doing," he said. "The next few days are going to be GC days, so we'll race them as hard as we can."¹ Ryder was similarly careful not to overpromise: "We've got a hard two days left, this is a tough three-day block before the rest day and an incredibly hard week. You can lose 10 minutes on the last stage in Paris. So there's lots of opportunities for him to still do something significant but I think we'll just take it day by day."²

Today's stage is where that caution gets tested. Mulhouse to Le Markstein packs just over 3,800 vertical meters into 155.3km, with three category 1 climbs and a cat. 2 ascent before a finish on the Col du Haag that race organizers have flagged as one of the explosive points of the route — the summit sits just six kilometers from Le Markstein, leaving any GC rider dropped there nowhere to hide.³ The final six kilometers average over ten percent.⁴ Sunday's stage to the Plateau de Solaison is being talked about in the same terms.⁴ Between the two, this weekend is shaping up as the real test of whether Vingegaard or Remco Evenepoel can put a dent in Pogačar's advantage before the Stage 16 time trial.²

Pogačar, for his part, let something slip after Friday's finish that had nothing to do with the next two days. Asked about a comment from Prince Albert of Monaco — who told L'Équipe that Pogačar "had said he would be at the start" of the Vuelta a España, which opens in Monaco next month — the Slovenian smiled and didn't deny it. "If the prince said, then there is a high chance," he said.⁴ Only three riders have ever won the Tour and the Vuelta in the same season: Jacques Anquetil in 1963, Bernard Hinault in 1978, and Chris Froome in 2017.⁴ Pogačar last rode the Vuelta in 2019, his first Grand Tour, taking two stages and finishing third overall before winning his first Tour the following year; he's stayed away from Spain every August since, favoring an autumn program that includes the Worlds.⁴ A Tour-Vuelta-Worlds treble in the same year has never been done by anyone.⁴

Away from the road, Rohan Dennis is facing the possibility of prison again. The former world time trial champion pled guilty this week to driving while disqualified, after police pulled him over last month with two children in the vehicle despite a five-year driving ban imposed following the death of his wife, Melissa Hoskins, in December 2023.⁵ Dennis had already avoided a custodial sentence for that case — a suspended 17-month term and a two-year good behavior bond — after pleading guilty to a lesser charge over her death.⁵ Prosecutors are now asking for both the new driving offense and the breach of his good behavior bond to be referred to the District Court.⁵ He's due back in court in September.⁵

Further north, the Baloise Ladies Tour delivered its own GC swap. Stage 2's 130.4km run through Zulte came down to a bunch sprint after a late four-rider breakaway — Anne Knijnenburg, Noémie Thomson, Aileen Schweikart and Eline Van Rooijen — was reeled in inside the final five kilometers, with SD Worx-Protime driving the chase.⁶ The overall lead changed hands, with Charlotte Kool taking over from overnight leader Zoe Bäckstedt.⁶

↑ Back to top

THE LAB

Thinking Machines Opens Up Inkling; Turso Runs Doom Inside SQLite's Bytecode VM

Mira Murati's Thinking Machines Lab put its first open-weights model on the internet on July 16: Inkling, a mixture-of-experts transformer with 975 billion total parameters and 41 billion active, trained on 45 trillion tokens of text, images, audio, and video, released under Apache 2.0.¹ Simon Willison, who tests every notable model release with the same absurd benchmark, ran Inkling through it: generate an SVG of a pelican riding a bicycle, then feed the rendered image back to the model and ask it to describe what it drew. Inkling produced a serviceable cartoon bird on a bike — and then, looking at its own output, described it as "a stork or seagull."¹

Willison's read on the release is generous but clear-eyed. Thinking Machines is upfront that Inkling isn't a frontier model — the company's own model card calls it "not the strongest overall model available today, open or closed," pitching it instead as a base for fine-tuning on the company's Tinker platform.¹ The model card itself is thin, and the linked training-data documentation says almost nothing beyond boilerplate about scraping "publicly available content." What's notable is the release itself: Apache 2.0, multimodal, and — per Willison — competitive with the open-weight models coming out of China, giving the US open-weights field a third real contender alongside NVIDIA's Nemotron and Google's Gemma 4. A smaller Inkling-Small (276B total, 12B active) is promised once testing finishes.¹

Turso, the startup that spent years rewriting SQLite in Rust, says it's now going after Postgres — and, according to the company's own blog post (there's no independent verification of this yet, so take the framing as Turso's own), it's already running Doom to prove the architecture works. SQLite, the post explains, isn't really a database so much as a virtual machine: it compiles SQL into its own bytecode language, VDBE, and executes that instead of interpreting SQL directly. Turso wrote a small C-to-VDBE compiler and got real Doom — compiled down to that bytecode — running frame-by-frame as VDBE opcodes, playable in a browser tab via WASM, each rendered frame literally a query result row.²

The bigger claim built on top of that stunt is architectural: Turso wants to become "the LLVM of databases" — one bytecode core with pluggable SQL frontends. Postgres is the second frontend, following an internal proof-of-concept called pgmicro that's now merged into the main tree.² Turso's pitch is that despite Postgres and SQLite differing wildly on disk layout — heap files, TIDs, and vacuum versus SQLite's clustered B-trees — the actual operations both engines perform are things VDBE can already do, or can be extended to do. The company is candid about how far this is from finished: no published packages yet, no PL/pgSQL support, extensions loaded through a WASM sandbox that isn't guaranteed to stick around, and full Postgres wire-protocol compatibility still to come. MIT-licensed, in-tree from day one, per the post's FAQ.²

Gwern's predictions have a good track record — his 2020 essay on scaling anticipated the GPT-3-to-ChatGPT arc two years before it happened³ — so his newest post, arguing frontier labs are training models the wrong way, is worth taking seriously even though it's mostly gone unnoticed outside a twelve-comment Hacker News thread. Sean Goedecke lays out the argument in a piece published today: grokking — the phenomenon where a model trained far past apparent convergence suddenly jumps in capability, first documented by OpenAI in 2022 on simple math tasks — happens when a model is forced to compress a small dataset into fewer parameters than it needs for rote memorization, at which point it discovers the underlying rule instead.³ Frontier labs currently do the opposite: relatively modest models (Goedecke pegs frontier open-weight MoEs like Kimi-K3 at roughly 3 trillion total, 50 billion active parameters — trainable in a couple of days on a big cluster) gorged on oceans of data. Gwern's proposal is to invert that: a hundred-trillion-parameter model trained on a deliberately small dataset, overtrained until it groks language the way it groks arithmetic.³ Goedecke's honest caveat is the same one that will keep any lab from trying it — the run looks like a failure for weeks or months before (if it works at all) it doesn't.

Three shorter items worth a coffee-length read. Julia Evans spent part of yesterday debugging why an FTS5 query against a 4,000-row SQLite table was taking five seconds — the fix was running ANALYZE, which dropped it to 0.05 seconds, a reminder that SQLite's query planner needs the same statistics tending a "real" database does.⁴ Wired's Steam Machine review found a real gap between Valve's marketing and its own Steam Machine Verified benchmark: the "4K at 120Hz" pitch relies on FSR upscaling, while the actual Verified label is benchmarked at 1080p — and unsupported titles like Crimson Desert run acceptably at 1080p but drop to 20–25fps at native 4K even on Ultra with FSR engaged.⁵ And on GitHub's trending list, Robbyant's lingbot-map is a feed-forward 3D reconstruction model that turns streaming video into a live point cloud at roughly 20fps using a "Geometric Context Transformer" with paged KV-cache attention — a legitimately different approach to streaming SLAM-style reconstruction than the iterative optimization methods it benchmarks against.⁶

↑ Back to top

THE WORLD

Transit Measure Clears Committee Intact; Kaiser Canyon Fire Forces Evacuations

↩ Developing story — first reported Jul 12 · previously Jul 13, Jul 15, Jul 16

• Iran Rejects Oman's Hormuz Corridor — IRGC says only Iran-designated shipping routes are authorized, rejecting Oman's new safe-passage corridor near Larak Island. Middle East Monitor
• World Cup Final: Spain vs. Argentina — Sunday's MetLife Stadium final pits two unbeaten teams; a win would be Messi's last World Cup match. ESPN

Seattle / Eastside

• Level 3 Evacuations as Kaiser Canyon Fire Tops 2,000 Acres — The Okanogan County fire, possibly lightning-caused, has burned over 2,000 acres and sits at 0% containment; SR 155 is closed near Nespelem and Red Cross shelters are open in Omak and Coulee Dam. KIRO 7
• Denny Blaine Stays Open, Judge Orders City to Crack Down — A King County judge found public nudity and sex acts at the clothing-optional beach a public nuisance but rejected neighbors' bid to close the park, ordering Seattle to add staffing, signage, and sightline barriers instead. KIRO 7
• Transit Measure Clears Committee Intact — Seattle's Select Committee kept Mayor Wilson's 0.3% sales-tax transit levy largely as proposed Thursday, voting down Councilmember Kettle's scaled-back 0.225% amendment 2-7.⁵ The full council votes Tuesday, July 21, ahead of the August 4 King County ballot filing deadline. The Urbanist

ON THE ROAD — Adast, Hautes-Pyrénées, France

• 15th "Fiesta mômes" Draws Crowds Despite the Heatwave — Argelès-Gazost's Parc thermal hosted its annual kids' festival July 5 — inflatable slides, circus workshops, face painting — with turnout holding up even in the heat. lourdesactu.fr
• No other dated news from the Adast–Argelès-Gazost–Cauterets–Lourdes valley cleared today's research. Watch for Lourdes' 200-panel signage overhaul along the Robert Hossein Space–Pic du Jer route, which starts July 20.

↑ Back to top

THE LONG READ

The Chip Faggin Built Out of Spite Just Turned Fifty

Federico Faggin did not leave Intel because he had a better idea. He left because he was tired of fighting for the ideas he already had. He'd shipped the 4004, then the 8008, then fought a bloody internal battle just to get the 8080 approved — delays that cost Intel nine months of lead time while competitors caught up.¹ So in 1974 he quit, took Ralph Ungermann — then the head of Intel's own microprocessor division — with him, and started a company with no product and, for a while, no plan.¹ He briefly considered building a microcontroller before deciding the margins were too thin for a fabless startup.¹ What he settled on instead was petty and brilliant: build a better 8080. He called it, informally, "Super 80."¹ Zilog called it the Z80, and this month it turns fifty.

The lineage matters more than the anniversary date. A blog post making the rounds this week — written by a hobbyist who built his own Z80 machine as a teenager and then fell down a Computer History Museum oral-history rabbit hole researching it — traces the chip back further than most retrospectives bother to go: to the Datapoint 2200, a programmable terminal built by the Computer Terminal Corporation out of a "cemetery" of individual TTL chips. CTC contracted both Intel and Texas Instruments to shrink that TTL logic onto a single die. Neither finished on time.¹ By the time Intel's chip was ready — originally just "1201" under the company's numbering scheme, later renamed by marketing to the 8008 — CTC had already shipped its terminal on the old TTL design and moved on.¹ Intel sold the orphaned chip anyway.

The 8008 was a strange, constrained thing: 3,500 transistors, a DIP18 package so cramped that address and data lines had to share pins, a -9V supply alongside +5V, and — because the Datapoint design assumed serial memory too slow to host a stack — no memory-based call stack at all, just an internal 8-deep return address register.¹ Faggin, brought over from the 4004 project, wanted to fix all of it immediately. Management wanted to see how the market reacted first.¹ The result was the 8080: a real stack pointer, a 16-bit address space, software-maskable interrupts, and — since Intel had decided along the way to break binary compatibility with the 8008 — a cleaner instruction set.¹ It also needed three separate supply voltages and two phase-shifted clocks, which is the kind of thing that looks fine in a datasheet and miserable on a breadboard.

The Z80, when Faggin and Masatoshi Shima (also poached from Intel, who'd poached him from Busicom) finished it, fixed the breadboard problem and then some. Single 5V supply, single clock, and — because Intel had copyrighted its assembly mnemonics, forcing Zilog to write its own syntax from scratch — an instruction language that's arguably still more readable today than Intel's.¹ The chip added two index registers borrowed in spirit from Motorola's 6800, a bank-switched shadow register set for fast interrupt handling, self-repeating block-transfer instructions, and enough exposed control signals that a hobbyist could build a working computer with little more than a 74-series logic chip and an EPROM.¹ The whole project cost about $400,000 against a $500,000 budget from Exxon, and it came in on time — a sentence you don't get to write about many chip programs, then or now.¹

Two facts in the piece are worth sitting with. First: the Sharp LR35902 inside the original GameBoy is a Z80 derivative, meaning a chip lineage that started as a fix for a Texas terminal company's TTL mess spent the 1990s running Tetris in millions of kids' hands.¹ Second, and more startling given the piece is being published to mark a fiftieth anniversary: Zilog didn't discontinue the Z80 until June 2024. Two years ago.¹ Not a museum piece — a part you could still order for an embedded design until very recently, outliving the 8080, the 8085, and most of the machines it was originally built to power.

The chip's afterlife is its own small tragedy of corporate politics. Exxon's investment — meant to build a computing empire to rival IBM, with stakes in typewriter and word-processor makers built around Zilog silicon — is, per the piece, one of the reasons IBM passed on Zilog for the original PC and went with Intel's 8088 instead.¹ Faggin's own partner Ungermann left before Zilog became a full Exxon subsidiary in 1980; the company split off again in 1989, went public in 1991, and has since bounced through private equity into its current home under Littelfuse.¹ The processor born out of one man's frustration with a large company ended up owned by a long chain of other large companies anyway. It just took the silicon another fifty years to notice.

↑ Back to top

FROM THE ARCHIVE

The Scoreboard Wasn't Built for a Perfect Score

The board flashed 1.00. Not 10.00 — the Montreal Forum's scoreboard had never been built to display a fourth digit, because no one who designed it imagined a gymnast would ever need one. On July 18, 1976, 14-year-old Nadia Comăneci finished her uneven bars routine in the team competition, and the judges gave her the score the machine couldn't render.¹

She hadn't gone in trying to be first at anything historic — every gymnast in the event was required to perform the same compulsory routine, which meant the only room to separate herself was in execution. "I knew I was good at what I was doing, but I didn't know how good I was compared to others," she said afterward. "I added amplitude to every skill."¹ It worked: a perfect 10, the first ever awarded in Olympic gymnastics. Comăneci went on to collect six more that Games, on her way to three golds and the title of youngest all-around champion in Olympic history. Bela Karolyi, the Romanian coach who had scouted her at age six and would later defect to the U.S. and build the American program that produced the "Magnificent Seven," said simply: "Nadia has courage." Within a week she was on the covers of Time, Newsweek, and Sports Illustrated simultaneously.¹

Montreal is marking the 50th anniversary this year. Comăneci is due back outside Olympic Stadium on August 1 for a commemorative event, and organizers are reportedly recruiting other women named Nadia — a nod to how many parents named daughters after her in the years that followed — to join her on stage.²

↑ Back to top

THE FUNNIES

A Bird That Doesn't Know Itself, and a Number Too Big for the Board

*After XKCD — on Inkling, Thinking Machines' new open-weights model, drawing a pelican on a bicycle and then, shown its own drawing, calling it a stork.*

*After Krazy Kat — on a Montreal scoreboard built to show three digits meeting a fourteen-year-old's perfect ten.*

↑ Back to top

ALSO NOTED

Also Noted

• Follow-up: Harper's Thumb Injury Was Worse Than First Reported — The crash that ended Chris Harper's Tour on the Puy Mary descent cost him part of his left thumb; he's now headed to Zurich for further specialist assessment. velo.outsideonline.com Jul 16, 2026
• A Bounty to Turn the PS5 Back Into a Computer — Right-to-repair group Fulu, fronted by Louis Rossmann, is offering $10,000 (matched up to $20K total) to whoever can strip Sony's software locks off a PS5 and get Linux running on it, following Sony's early-July move to end physical-disc production for new PS5 games. wired.com Jul 17, 2026

↑ Back to top

THE QUESTION

Every Unprecedented Bet Looks Like a Mistake First

Whether Tadej Pogačar actually starts the Vuelta a España next month, or whether any AI lab ever spends the money to overtrain a model past the point where its own metrics call the run a failure, the two bets share the same shape: for months, doing the unprecedented thing looks exactly like doing the wrong thing.

As this paper's Peloton section reports today, Pogačar didn't deny it when a reporter raised Prince Albert of Monaco's comment that he'd start the Vuelta — "if the prince said, then there is a high chance," he said, smiling.¹ The interesting part isn't whether he does it. It's why almost nobody else does. The physical case against a Tour-Vuelta double isn't new — three weeks of recovery, a second three-week race in the worst heat of the Spanish summer — but plenty of strong riders could survive it. What actually stops them is that trying and coming up short in September reads, to sponsors and to the sport's memory, as diminishment: the story becomes "faded after the Tour" rather than "tried something nobody's tried." The safer path is to protect the one result you already have. A grand tour season is built, contractually and psychologically, around a single peak — everything else is described as "conditioning" for it, because admitting otherwise means admitting you're gambling with the thing that pays the bills.

The Lab section covers a piece published today by Sean Goedecke, laying out an argument from the pseudonymous researcher Gwern that frontier AI labs are training models the cautious way — modest-sized models fed enormous amounts of data — when the more interesting result might come from the opposite bet: a far larger model, trained on deliberately less data, forced to compress what it's given until it stops memorizing and starts generalizing.² Goedecke's own caveat is the one that matters here: a run built on that premise looks like a failure for weeks or months before it either works or doesn't.² No lab metric distinguishes "about to grok" from "wasted the cluster." That's not a technical problem. It's an incentive problem — quarterly investor updates and press cycles don't have a slot for "still looks broken, trust the process."

Put the two next to each other and the pattern is the same one you'd recognize from any organization that rewards visible, near-term wins: the bet worth making is often indistinguishable, for a long stretch, from the bet already known to be a mistake. Cycling has an answer to this problem that AI labs don't — a rider can quietly test the limits in a race that isn't the one people are watching, the way Pogačar tested himself at the 2019 Vuelta before he'd won anything.¹ Frontier labs don't have an equivalent low-stakes proving ground; a hundred-trillion-parameter training run either happens on the flagship cluster or it doesn't happen.²

So the thing worth tracking this month isn't whether Pogačar wins in Spain or whether some lab's overtrained model eventually groks language the way models already grok arithmetic. It's whether either bet gets placed at all — whether a team or a lab is willing to sit inside a result that looks like failure long enough to find out if it isn't.

↑ Back to top