THE LAB

Carmack on Microsoft Gutting id; Sol Closes a 50-Year Proof; 744B in Pure C

John Carmack couldn't find the words — not the right ones, anyway. "I have been trying to find something meaningful to say about the Id Software layoffs," he posted Thursday. "My 'Microsoft will probably be a good steward of the brand' statement isn't aging well." As this paper reported Wednesday, Microsoft's 3,200-person Xbox cut had effectively reduced id Software to a support role following the release of DOOM: The Dark Ages Revelations DLC. Carmack, who co-founded the studio in 1991 alongside Adrian Carmack, Tom Hall, and John Romero, reached for pragmatism rather than outrage.

What he produced was a question list. Could id have gotten more with a different pricing strategy? Could they have broadened the game designs to reach more players without alienating existing ones? Could they have produced games faster or cheaper? "I really don't know," he writes — and from the person most qualified to answer, that's a meaningful admission. He suspects "id Software was a marginal business from Microsoft's perspective" and credits Minecraft revenues with carrying several other Xbox studios. There is one careful hedge: "You can't rule out the possibility that executives are idiots, but that shouldn't be your default belief." The four id co-founders are due to reunite at QuakeCon next month for the first time since 1993; that occasion will now be considerably more somber. "The game isn't over yet, and I hope the studio rallies through."¹

There is no dependency to install. No Python at runtime. No GPU required. The inference engine is a single C file — roughly 2,400 lines — and it runs GLM-5.2, a 744-billion-parameter mixture-of-experts model, on a consumer machine with 25 GB of RAM. The project is colibri, by JustVugg on GitHub, and its architecture starts from a simple observation: a 744B MoE model activates only ~40B parameters per token, and only ~11 GB of those change between tokens. The dense portion — attention layers, shared experts, embeddings, roughly 17B parameters — stays resident in RAM as int4 at about 9.9 GB. The 21,504 routed expert shards, roughly 19 MB each at int4 and totaling ~370 GB on disk, stream on demand via a per-layer LRU cache, with the OS page cache serving as a free second tier.

The implementation earns its attention. Compressed MLA attention reduces the KV-cache from 32,768 floats per token to 576 — a 57x reduction, made possible because GLM-5.2 uses 64 attention heads and no grouped-query attention. The model's native multi-token-prediction head runs speculative decoding at int8, achieving 39–59% acceptance rates and 2.2–2.8 tokens per forward pass. A learning cache records which experts each session actually routes to and automatically pins the hottest ones in spare RAM at startup — the engine measurably accelerates across sessions. Community benchmarks are honest: 0.05–0.1 tok/s cold on the 25 GB dev box; an Apple M5 Max with 128 GB unified memory and a 14 GB/s internal SSD reaches 1.06 tok/s; a Framework 13 on Arch Linux with 128 GB goes from 0.29 to 0.37 tok/s once the learned pin warms up. The README does not oversell: "This is not fast. It is a 744B frontier-class model answering correctly on a machine that costs less than one H100 fan."²

GPT-5.6 Sol Ultra — the model this paper noted in limited rollout yesterday — has now produced what OpenAI describes as a machine-verified proof of the Cycle Double Cover Conjecture. The conjecture was posed independently by George Szekeres in 1973 and Paul Seymour in 1979; it states that for any bridgeless graph, there exists a collection of cycles covering every edge exactly twice. Partial results existed for specific cases; a general proof did not. Sol deployed 64 subagents working in parallel, and the proof reportedly reduces the problem using the 8-flow theorem and linear algebra over GF(3), the three-element finite field. OpenAI published the result as a PDF on July 10, with authorship attributed to the model; Codex assisted with the writeup.³

Machine-verified is not peer-reviewed. Formal verification can confirm that logical steps follow from axioms; it cannot explain why the proof works or certify that it illuminates the structure of the problem in the way mathematicians demand. The community's response on Hacker News zeroed in on this immediately. If the result holds up to mathematical scrutiny over the coming weeks, it clears a meaningful bar — not game-playing, not benchmark optimization, but original mathematics at frontier difficulty, on a problem that defeated human specialists for fifty years. The community will take the time it takes.

Apple filed suit against OpenAI in US district court in San Jose on Friday, naming OpenAI chief hardware officer Tang Tan — who spent 24 years at Apple overseeing iPhone product design — and io Products, the Jony Ive–co-founded startup OpenAI acquired for $6.5 billion last year. The filing accuses Tan of coaching departing Apple employees on how to evade the company's data security protocols and directing them to bring confidential hardware to job interviews at OpenAI. A second defendant, electrical engineer Chang Liu, reportedly downloaded "dozens of Apple's confidential hardware-related files" including manufacturing and testing documentation for complex circuit boards, then wrote to a former Apple colleague about still having access to the company's internal file-sharing system months after his departure — via a bug Apple says it has since patched. Apple wrote to OpenAI in February raising those concerns; it received no response. OpenAI has hired more than 400 former Apple employees. The case echoes the 2017 Waymo v. Uber dispute, which settled mid-trial in 2018 for $245 million.⁴ Apple and OpenAI have been partners since 2024 via the ChatGPT-on-iPhone deal, but according to Wired, the relationship has frayed as Apple has shifted toward Google's Gemini for its in-house AI models.

↑ Back to top

THE PELOTON

Stage 8 Rolls to Bergerac; ASO Concedes the Tourmalet Miscalculation

↩ Developing story — first reported Jul 08 · previously Jul 09, Jul 10

PÉRIGUEUX, FRANCE — Three riders are off the front somewhere in the Dordogne, and they are going nowhere. Liam Slock (Lotto-Intermarché), Thibault Guernalec (Totalenergies), and Jakub Otruba (Caja Rural-Seguros RGA) got clear in a scrappier-than-usual opening scramble — the Lotto directeur sportif had warned before the départ that today's bumpier terrain would produce "a little more action than the previous days" — but Soudal-QuickStep and Alpecin-Premier Tech took the front of the peloton within the hour and capped the gap to a stingy 1:15 to 1:50.¹ The sprint setup is on rails.

Stage 8's flat route from Périgueux to Bergerac is the second consecutive sprint stage, and the fast-finishers' teams are running the same script they ran into Bordeaux: a token break allowed, steady tempo all day, deliver for the line. Kasper Asgreen (EF Education-EasyPost) tried repeatedly to join the escape and was marked out by NSN, whose directeur sportif clocked him from the car. "I saw Asgreen, and he has shoe covers, aero helmet. I think he is ready," the NSN DS radioed.¹ He was not let go. No sprinter has asserted clear dominance across two sprint finishes so far. Bergerac may sharpen that picture.

The more contested story at this Tour is happening inside the Red Bull-Bora-Hansgrohe bus. As this paper reported Thursday, Remco Evenepoel aired his frustration publicly after Stage 6 — he had asked Florian Lipowitz for one kilometre of work on the mountain and was refused, and he said so, loudly and specifically, at the finish line. The team moved to contain the fallout: manager Ralph Denk called it "no big deal," said the pair had dinner together that evening, and on Friday morning the team posted an Instagram story of both riders on the bus, with the press officer prompting Evenepoel on camera about whether the press reports of tension were accurate.²

"No, all good," Evenepoel said. "You speak and you forget."²

A CyclingNews analysis published Friday argued that framing isn't convincing anyone. The structural problem hasn't been solved by a shared dinner and a managed Instagram post. Both riders have finished third at the Tour; Lipowitz's podium came last year, in a race Evenepoel abandoned; Evenepoel has a Vuelta title that Lipowitz doesn't.² Neither has a clear reason to cede leadership to the other, and the Pyrenean stages haven't finished the argument for them. The road will decide, as the old line goes — but so far it hasn't. The Pyrenées are not done.

The more consequential reckoning this week belongs to ASO's route designers. Stage 6's 40 kilometres between the Tourmalet summit and the finish in Gavarnie-Gèdre — a 20km descent followed by a 20km valley drag — was conceived as a moderating force. The subtext was clear: no right-minded rider would want to solo into a headwind for that long after two major cols. It produced nearly three minutes between first and second place after one week of racing.

Thierry Gouvenou, ASO's route designer, has now said the quiet part aloud. "To be honest, we had not expected such a large gap, and we thought the differences at the finish would be much smaller," he told TV 2 Sport. "As far as the suspense is concerned, you could say it was a failure."³

The arithmetic is worth sitting with. The gap at the Tourmalet summit was 30 seconds. At the finish it was 2:38. Forty seconds came on the descent; another 1:28 leaked on the valley road.³ The terrain designed to discourage aggression handed the field's best descender and most accomplished long-range soloist 40 free kilometres where he has no competition. Fletcher's analysis makes the uncomfortable case that a summit finish — which would have appeared more dramatic on paper — would paradoxically have produced a smaller gap, by removing precisely the descent and drag terrain where the GC leader has the clearest advantage over his rivals. In trying to mute the Tourmalet, the organisers amplified it. "He is so strong that any route suits him," Gouvenou conceded.³

Christian Prudhomme had sold this Tour on the promise of "suspense until the end." One week in, the last two weeks look like a procession.

While the GC narrative has narrowed, Baptiste Veistroffer is making sure the race has something else to offer. The 26-year-old from Brittany (Lotto-Intermarché) has spent 301 kilometres on the attack across just over 1,000km of racing — more than 2,000km of attacking in total across the 2026 season, the most of any rider in the peloton.⁴ He attacked alone from the gun on Stage 5 to Pau, launched again with Otruba into Bordeaux on Stage 7, and collected his second combativity award of the race. His team, which lost its sprint leader Arnaud De Lie to illness before the start, has extended his contract.⁴

His nickname is "Le Sanglier" — the wild boar — for his build and his natural aggression.⁴ He is in the tradition of the classic French baroudeur: Thomas Voeckler, Jacky Durand, Thomas De Gendt. After his Bordeaux effort he gave his combativity award flowers to his mother and said: "I don't understand why I'm the only one going on the attack. There are riders we never see all day, even though we have an incredible opportunity to ride the Tour."

He is not in today's break — that role went to his teammate Slock — but his philosophy has spread. Lotto-Intermarché, with a fraction of the budgets around them, are the Tour's most consistent source of racing animation on the flat stages. "For the love of the sport and the beauty of the effort," Veistroffer wrote on Instagram after Bordeaux.⁴ France, for the moment, agrees.

↑ Back to top

THE WORLD

Spain in the World Cup Semifinals; Boeing Opens Everett 737 MAX Line

• Spain through to World Cup semifinals — Mikel Merino struck a late winner against Belgium; England faces Norway tonight in the other quarterfinal. Source
• Boeing opens new 737 MAX line in Everett — The $1 billion facility is Boeing's first 737 MAX production line outside its Renton factory. Source
• Link Light Rail hits ridership records — Sound Transit logged 309,000 single-day boardings on July 6, the US-Belgium match, and 5.4 million for June. Source
• King County opens contentious Kirkland housing — Sheila Stanton Place, 101 supportive units at a former La Quinta Inn near the Kirkland-Bellevue border, opened this week. Source

• WEEKEND PICKS — Kendall Katwalk (Snoqualmie Pass) is in great shape with wildflowers blooming. Koppen Mountain (Teanaway) delivers solitude, wildflowers, and an excellent trail at 7.4 miles/2,150 ft. Source
• REGIONAL SNAPSHOT — Wildflowers at peak from Cutthroat Pass (Hwy 20) to Glacier Basin (Mountain Loop Highway) to Yellow Aster Butte (Mt. Baker); Palouse-to-Cascades fully reopened after spring landslide closure between Cedar Falls and Change Creek — gate removed, signs down. Watch: active wasp nest at head height on Port Gamble's Twisted Sister trail; bear-with-cub sighted near West Tiger 3 on Tiger Mountain; Smithbrook Road access to Lake Valhalla remains closed from December storm damage. Source

↑ Back to top

THE LONG READ

The Compiler Became the Bottleneck

For seven years, the backend at Scarf ran on Haskell. The type system caught real bugs. The high-performance gateway service, sitting directly in the download path for a high volume of open source package traffic, met its SLAs. The code was, by the founder's own account, reliable. And then AI changed the tradeoffs, and the reliability stopped being enough.

Avi Press — who sits on the Haskell Foundation board and the Haskell.org committee, and has sixteen years of personal investment in the language — published a post on Thursday explaining why Scarf is migrating to Python.¹ It is not a hit piece. It is something more useful: a precise account of which specific property of the language became untenable, and why that property was fine before and isn't now.

The argument runs like this. Historically, you caught errors in one of two places: at compile time or at runtime. Press identifies a third place: code generation time. The model can avoid the mistake before the compiler ever sees it. That doesn't make type safety worthless — he's explicit that it doesn't — but it changes the cost-benefit calculation on compile time significantly. "If an LLM can produce a working implementation in a few minutes, but your compile step takes dramatically longer, then your language and build system have become a bottleneck in the development loop."¹

That's the argument in its mild form. The sharper version is about parallel agent workflows. Press describes wanting to spin up multiple worktrees, fork different lines of work, let agents try things, review the results, keep the useful ones. In that world — five agents exploring five branches simultaneously — cold build time isn't a tax you pay once, it's a tax you pay five times over, per round of iteration. Caching helps but doesn't solve it. The Scarf team used Nix, remote builders, carefully optimized CI, and all the rest. The best-case scenario, a small change hitting a warm cache, could get to twenty seconds.¹ But the best case is not what you optimize around when you're running agents in parallel. The cold-start case, the deep-change case, the cache-miss case — those become the average. "In an agent-heavy workflow, you end up caring a lot more about the cold-start case, the average case, and the deeper-change case."

What they did about it was methodical. New API routes go into Python. A Python server was deployed alongside the Haskell one, traffic routed appropriately, functionality migrated as it was touched. No dramatic cutover. The porting work itself, Press notes, was cheap in a way it would not have been before: porting existing code to a new language is straightforward for current models. The time recovered from waiting on the toolchain went to shipping more features and writing more tests. The type safety they gave up "hasn't been noticeable in any concrete way yet, especially considering our test coverage has never been better."¹

The second half of the post broadens into an argument about the Haskell ecosystem's response to AI — and this is where it gets uncomfortable to read if you care about Haskell. Press says that when AI comes up in Haskell spaces, the conversation "seems more focused on restriction than enablement."¹ He identifies a "do not use LLMs" cohort and predicts it will prove bad for the language's ecosystem long-term. His counterproposal is specific: optimize Haskell for agents. Make project bootstrap fast. Make error messages agent-friendly. Reduce cold build times. Get high-quality Haskell examples into model training data. Make library docs full of copy-pastable examples rather than beautiful types. These are the bottlenecks agents actually hit, and they're different from the bottlenecks humans hit.

The observation that carries most weight is this: Haskell should be unusually well-positioned in the AI era, because type safety can be an advantage for LLM-generated code if the compiler helps the agent converge quickly on a correct implementation. Fast feedback from a rigorous type system is exactly what an agent benefits from. The problem is that "unusually well-positioned in theory" requires the toolchain to actually be fast enough that the compiler is doing the agent a favor rather than making it wait. Right now it isn't.

Press is writing from a position of genuine investment. He isn't someone who tried Haskell once and bounced. He built real production systems on it, understood where the sharp edges were, and managed them for years. The credibility matters because the diagnosis does. What he's identifying isn't a flaw unique to Haskell: it's a structural shift in what matters when the inner development loop is driven by agents rather than humans. Any language ecosystem that optimizes primarily for human writing speed while neglecting cold-start performance and toolchain disposability is going to feel this same pressure. Haskell is just where it showed up first in a visible way, partly because Haskell's build characteristics make the gap especially wide.

The piece is worth reading in full. It's the kind of postmortem that's useful precisely because it doesn't come from an outside critic.

↑ Back to top

FROM THE ARCHIVE

The Littering Fine That Floated for Thirty Years: July 11, 1979

Night over the Western Australian outback — flat red earth, scattered spinifex, ink-black sky. A charred length of aerospace tubing, still faintly glowing at its edges, has gouged a long furrow in the dust. A shire ranger in a wide-brimmed bush hat stands at arm's length from the debris, citation pad in hand, pen poised in mock ceremony. Above, a bold diagonal streak of cross-hatching traces the object's descent path from the upper corner of the frame down through the darkness. Deep shadow hatching on the debris, strong silhouette of the lone official, generous white space in the sky. Confident pen-and-ink linework, newspaper editorial quality.

The descent was supposed to end in open ocean. NASA had run the numbers, oriented the thrusters, and pointed 85 tons of derelict space station at a remote patch of the southern Pacific. What it had not accounted for, precisely, was a few degrees of arc—enough to swing the debris field from deep water to rural Western Australia.

Skylab had been drifting for five years when it fell on July 11, 1979. Three crews had lived inside it in 1973 and 1974: 28 days, then 59, then 84, each a US endurance record at the time.¹ Then the final crew departed, the station was powered down, and NASA redirected its budget toward the Space Shuttle. Skylab would manage on its own for a decade, or so the plan went.

The sunspot cycle had other ideas. An unexpectedly active solar maximum heated the upper atmosphere, increasing drag on the station. By 1978 it was clear the orbit was decaying faster than projected. The Shuttle, not due until 1981, could not be accelerated in time for a reboost mission. NASA worked on a controlled reentry instead.¹

The controlled descent was not, in the end, controlled enough. Sonic booms startled residents across swaths of western Australia. Debris — what survived the atmospheric passage — scattered across the Shire of Esperance and into the desert beyond. No one was hurt. NASA's investigation team flew to the site to collect remnants and inspect for damage.

When they arrived, the president of the shire had arranged a ceremony. As NASA public affairs representative J.M. Jones recorded in a 1979 agency newsletter: "Upon our arrival, the president of the shire had arranged a mock ceremony in which an officer of the parks service ticketed NASA for littering, the evidence having been found all about the country-side." The fine was $400.²

NASA never paid it. For thirty years, the ticket sat on the books of a small Australian municipality — an unpaid reminder of the day the first US space station came down in the wrong hemisphere. Then in 2009, California radio DJ Scott Barley asked his listeners to crowdfund the debt. They did. The mayor of Esperance told Barley the fine had already been "written off years ago," but accepted the $400 anyway and invited Barley to town, where he received a key to the city.²

The backup Skylab orbital workshop is on display at the Smithsonian's National Air and Space Museum, where it has been since 1976.¹ A few charred fragments of the actual station are in the collection.

↑ Back to top

THE FUNNIES

Space Junk Bureaucracy; The Route That Backfired

*After Dilbert — on the $400 littering fine the Shire of Esperance levied against NASA when Skylab's controlled Pacific reentry landed in the wrong hemisphere on this day in 1979, and the agency's characteristically decisive response over the following thirty years. After Bloom County — on ASO's route designer conceding that the forty kilometres of valley drag meant to moderate the Tourmalet's gaps instead handed the field's best descender forty free kilometres of low-competition road, which is, as concerns suspense, "a failure."*

↑ Back to top

ALSO NOTED

Also Noted

• AI tool finds 15-year-old Linux root privilege bug — Nebula Security's VEGA AI found GhostLock (CVE-2026-43499), a use-after-free in the Linux kernel present since 2011 that lets any logged-in user gain root with no network access; it earned a $92,337 payout through Google's kernelCTF program, and Ubuntu LTS versions remain unpatched or in progress as of early July. wired.com Jul 11, 2026
• OpenAI's head of safety systems is leaving — Johannes Heidecke told staff he is departing; safety teams will now report to VP of research Mia Glaese in a reorganization that merges safety and research, announced the same week GPT-5.6 showed what OpenAI's own documentation described as "concerning forms of misaligned behavior." wired.com Jul 10, 2026
• Microsoft's greenhouse gas emissions jumped 25 percent last year — The FY2025 sustainability report attributes the rise to AI data center expansion, including a Chevron-powered West Texas facility projected to emit more than 11.5 million tons of CO2 annually — greater than the entire state of Rhode Island. wired.com Jul 10, 2026
• pgrust: Postgres rewritten in Rust, passes 100% of regression tests — The project now matches all 46,000+ Postgres 18.3 regression queries, replaces the process-per-connection model with thread-per-connection, and benchmarks at 50% faster on transactions and roughly 300x faster on analytical workloads compared to stock Postgres. github.com
• China's Tianwen-2 photographs Kamo'oalewa from 20 km — After a 400-day, 1-billion-kilometer journey, the probe captured the first close images of Earth's quasi-moon on July 2; the asteroid's origin is now contested between a Moon-fragment hypothesis and an LL chondrite from the Flora asteroid family, and sample return to Earth is planned for November 2027. wired.com Jul 11, 2026
• Snohomish County: 19 bridges in poor condition, only one funded for repair — Annual inspections flagged 13 bridges qualifying for federal dollars, but only Pilchuck Creek Bridge has received that funding; its replacement is scheduled for summer 2028 while the rest wait for competitive grant rounds. kiro7.com Jul 11, 2026
• AR glasses architecturally require cloud processing — there is no alternative — Nilay Patel, on the Vergecast, argued that no chip small enough to fit in glasses stems is both powerful enough and power-efficient enough to process eye-level camera feeds in real time, making cloud transmission — and its privacy trade-offs — structurally unavoidable in any true AR product. simonwillison.net Jul 10, 2026
• "Good Tools Are Invisible": an argument against celebrating tool friction — Odin language creator Ginger Bill argues that treating a tool's limitations as a "puzzle game" is a category error, that maximal configurability is not a design goal, and that good defaults are a form of respect for users' time that toolmakers routinely decline to provide. gingerbill.org Jul 10, 2026
• SharpEmu: experimental PS5 emulator written in C# — The open-source project can load eboot.bin, execute native PS5 CPU instructions, and has reached the video-output stage in some games; the focus is accuracy and reverse-engineering the system architecture rather than broad game compatibility. github.com
• Bonnie Tyler, "Total Eclipse of the Heart" singer, dies at 75 — The Welsh artist died in hospital in Portugal following emergency intestinal surgery in May; she was the first British female artist to debut at No. 1 on the UK Albums Chart and the only Welsh artist to reach No. 1 on UK singles. kiro7.com Jul 9, 2026

↑ Back to top

THE QUESTION

When the Distribution Shifts

A system calibrated for the average actor behaves differently when the actor distribution shifts. Two pieces in today's paper arrive at this problem from opposite domains — one about stage design, one about toolchain choices — and neither story has an obvious correction.

When the Tourmalet summit comes with a 20km descent and a 20km valley drag attached, the reasoning is sound for most editions of the peloton: no one would rationally push a solo effort across that terrain after two hard climbs. The terrain moderates. But as THE PELOTON reports today, ASO's route designer has now admitted that the logic failed precisely against the one rider it most needed to work on. A 30-second gap at the summit became 2:38 at the finish.¹ The terrain designed to discourage long attacks gave the field's best descender and most accomplished long-range soloist 40 kilometres of low-competition ground. "He is so strong that any route suits him" is another way of saying: the protection was calibrated for a different distribution of riders than the one that showed up.¹

THE LONG READ today turns on the same inversion. Haskell's compile-time type system is the right tool when developers make the kind of errors a type checker catches, and for seven years it caught them reliably. In an AI-agent development loop — multiple worktrees running in parallel, iterating faster than a human — the protection remains valid in principle, but the build time it demands is now the bottleneck.² Agents don't produce type errors at the same rate a human does; they mostly pay the latency cost without receiving a proportionate share of the benefit. The safeguard that fit the original actor became the obstacle for the one who replaced him.

The question worth carrying today: when you optimize a system for the actors who existed when you designed it, do you inadvertently create something that works for everyone except the exceptional case you most need it to handle? Stage design can be redrawn for next year's route. A language migration can be executed methodically, as Scarf is doing.² The harder problem is the recognition lag — knowing that the actor you assumed is already gone before the gap at the finish line makes it obvious.

↑ Back to top