THE WORLD

Total Eclipse at Dusk in Spain; Spokane Returns to Its Ruins

↩ Developing story — first reported Aug 08 · previously Aug 09, Aug 10, Aug 11

• Total solar eclipse today — The Moon's shadow sweeps Greenland, Iceland, and Spain tonight; NASA streams live at 1:15 p.m. EDT. NASA
• DRC Ebola: 2,000 dead — The outbreak has killed 2,011 people across five provinces in under three months — no approved treatment exists for the Bundibugyo strain. Al Jazeera
• Colombia: M7.4, at least 111 dead — Monday's quake — Colombia's strongest this century — killed at least 111 people, devastating Cali and Pereira. NPR
• WA wildfires — Spokane evacuations lifted Tuesday. KOMO

At least two orcas were spotted navigating the Ballard Locks Tuesday morning; the Hiram M. Chittenden Locks posted video to social media — a rare deep-city appearance by the region's most charismatic animals. (KIRO 7)

In February, Bellevue, Issaquah, Redmond, and Kirkland formed a regional task force targeting street racing and reckless driving on I-405 and SR 520. (KOMO)

ON THE TRAIL

• Kendall Katwalk (Snoqualmie Pass) — Snow-free to Ridge Lake, trail in excellent shape, strong Rainier views; a Tuesday report found no bugs and a decent breeze at elevation. Smoke is forecast for the I-90 corridor Wednesday–Thursday — check AQI before heading out. WTA
• West Tiger 3 (Issaquah Alps) — Smoke has mostly cleared; berries ripe, wildflowers holding at the summits. Good option if you want a hard local climb without driving east of the pass this week. WTA
• Fire closures — Three Queens Fire has shut the 4330 road, blocking Tuck/Robin/Jade Lakes and most Cle Elum Ranger District PCT sections east of Snoqualmie Pass. Confirm before going anywhere in the Teanaway. WTA

↑ Back to top

THE LAB

Throw Out the Graphics API. Sebastian Aaltonen Made the Case at SIGGRAPH.

The resource binding model in DX12, Vulkan, and Metal was designed to tell the GPU where its resources lived, one draw call at a time. Sebastian Aaltonen's argument — laid out in a December 2025 blog post and presented at SIGGRAPH 2026, with discussion of the talk now circulating on Lobsters — is that this model is a historical artifact the hardware left behind years ago, and that game engines are paying a compounding overhead tax to keep the fiction alive. His proposal: scrap the binding infrastructure and replace it with direct 64-bit GPU memory allocation, the same model CUDA has used from the start.

The hardware prerequisites have been there for years. Nvidia Turing (RTX 2000 series, 2018), AMD RDNA2 (RX 6000 series, 2020), Apple M1/A14 (2020), and Intel Alchemist (2022) all support buffer device addressing and bindless texture heaps. What they don't have is an API that exposes this cleanly without also dragging along the resource tracking, transition barriers, pipeline state objects, and descriptor set allocators that every modern render hardware interface carries. DX12 and Vulkan layered bindless on top of their original binding model as extensions — they didn't clean the slate. The result is that the frame graph, the barrier emitter, the PSO cache, and the descriptor pool allocator in every game engine exist largely to bridge the old model to the new one.

What Aaltonen proposes is closer to writing C than to writing a graphics application. Allocate GPU memory with a bump allocator. Store 64-bit addresses in GPU structs. Pass a single root pointer into each shader. Let the debugger follow pointer chains the way Xcode's Metal debugger already can, reading texture descriptors out of GPU memory and visualizing texels at any struct address. No explicit bind points. No transition barriers. No PSO hash. The texture heap is just memory you index. He works through the obvious objections: virtual memory means you cannot read another process's pages, stale pointer dereferences produce page faults exactly as they do in C/C++, and GPU capture replay already uses APIs that force allocations to mirror the original virtual address layout (DX12's RecreateAt, Vulkan's VkMemoryOpaqueCaptureAddressAllocateInfo). Translation layer feasibility has proof-of-concept in deployed software: MoltenVK already translates Vulkan's buffer-centric model into Metal's 64-bit pointer ecosystem, and Proton translates DX12's SM 6.6 descriptor heap into Vulkan's descriptor buffer extension. Both chains exist today.

The sharpest line in the proposal is the CUDA comparison. HLSL and GLSL were designed over twenty years ago as frameworks of element-wise transform functions with no pointer semantics, and despite two decades of industry use they have produced no significant library ecosystem. CUDA exposes GPU memory directly and surfaces new hardware capabilities — tensor cores, custom memory paths — through intrinsics; its library ecosystem is what drove Nvidia to a $4 trillion valuation.¹ Aaltonen says this is not a coincidence and that the API's inability to compose is an economic liability, not just an ergonomic one. Metal 4.0 is the closest any vendor has come — its 64-bit texture handles map directly to heap indices on Apple Silicon — but it is still missing a globally indexable texture heap, so a handle cannot represent a texture range.¹ His conclusion is direct: Vulkan 2.0 and DirectX 13 should break backward compatibility and embrace the bindless hardware that already exists, rather than extend the old model further. Whether vendors will is the open question.

Mojo reached 1.0 yesterday. Modular's Python-superset systems language, first announced in 2023, now has its first stable release. (Source: Modular's own developer blog; no independent source was available at writing time.) The 1.0 designation is primarily a stability commitment: during the 1.x timeframe, breaking changes will be managed the way mature languages like C++ manage them. The language has converged on var for all variable declarations, a single Pointer type, and unified closure syntax; this release adds Python-style lambda syntax for inline closures and a memory-safety pass that detects reference invalidation — catching, for example, when a List.append call invalidates an active reference into the same list. Nearly 200 contributors have landed more than 1,100 pull requests against the open-sourced standard library.² The compiler and toolchain remain closed; Modular has committed to open-sourcing them before the end of 2026.

QuakeCon 2026 — the convention's 30th anniversary, held in Grapevine, Texas — produced a gathering that has not happened since 1993: all four id Software co-founders in the same building. John Carmack, John Romero, Adrian Carmack, and Tom Hall appeared together for the first time in 33 years. Romero noted it was "the first time all four of us have been together since 1993."³ Carmack picked up an Evercade port of the original Doom and found he had forgotten the keycard locations, taking nearly 15 minutes per level to clear E1M4 and E1M5 — levels with speedrunning par times of 1:30 and 2:45 — before posting on X: "Still a great game after all these years!" The reunion also prompted Carmack to respond publicly to Sandy Petersen's recent claim that Quake "ruined id Software." Carmack's assessment: "Quake was overly ambitious technically... I had pushed everyone too hard. The studio could have done all the great multiplayer and modding work inside a Doom++ engine, allowing the designers to work with a more stable base instead of a rug pulling everything out from underneath them a couple of times."³ Romero, American McGee, Michael Abrash, Petersen, and others left after Quake shipped; the Grapevine reunion suggests what tension existed has mostly smoothed over.

Trending today: GitHub is saturated with AI agent framework forks and diagram-generation utilities — nothing on GPU, graphics, or systems work cleared the technical novelty bar.

↑ Back to top

THE PELOTON

Roglič in Talks with Lotto-Intermarché; Accell Declared Bankrupt; the Transfer Market Waits on a Teenager

↩ Developing story — first reported Aug 10

Roglič's departure from Red Bull-Bora-Hansgrohe has been treated as a foregone conclusion for weeks. What Belgian newspaper Het Laatste Nieuws added on Wednesday was a destination: Lotto-Intermarché. Sources close to both Roglič and the Belgian WorldTour squad confirmed to HLN that negotiations for a 2027 contract are already underway, and what the paper called "bombshell news" is looking increasingly concrete.¹

The logic holds on both sides of the table. Lotto faces 2027 without Lennert van Eetvelt and Arnaud De Lie — departures that strip the squad of its GC depth and its sprint. Roglič, at 36, still reliably accumulates top-five finishes and the UCI points that come with them; more importantly, the squad would give him calendar freedom he can no longer find anywhere with a crowded roster. His stated ambitions — the Tour de Suisse, the one major week-long stage race still absent from his palmares, and a return to the Tour de France — would not have to compete with a packed hierarchy. Other teams reported to have been in contention — Q36.5, Movistar, UAE Team Emirates-XRG, and Bahrain Victorious — have all apparently fallen away.

The longer question of where he races in 2027 is secondary to the one for this month. Whether Roglič starts the Vuelta a España — which opens with a nine-kilometre time trial in Monaco on August 22 — remains unresolved. A training crash this summer forced him to skip both the Clásica San Sebastián and Vuelta a Burgos. Italian journalist Ciro Scognamiglio of Gazzetta dello Sport placed the probability of a start at roughly 50-50 last week;¹ Spanish daily AS reported a final decision could come as late as early next week, just days before the Monaco prologue.

While the Roglič situation moves toward resolution, the rest of the 2026-27 transfer window is stuck on a teenager. VeloNews reported Wednesday that the market has opened with conspicuously little activity compared to last summer's record-breaking frenzy — and one source cited one reason: Paul Seixas.

The 19-year-old Frenchman, who finished fourth on his Tour de France debut, has become the window's hinge point.² Only the sport's richest superteams can afford him, and those teams are in a holding pattern until he decides. Wielerflits reported he signed a letter of intent with UAE last year; sources told VeloNews no decisions would be made until after the Tour — and two weeks on, nothing has moved. Visma-Lease a Bike confirmed interest during the race, with an eye on who carries their grand tour ambitions into the next decade. Decathlon CMA CGM say they are working to keep him on his contract through 2027. VeloNews reported in June that a proposed $13 million per season offer from Q36.5 was "way off the mark." Seixas turns 20 next month.

Beyond the stalemate at the top end, early moves in the men's peloton have been methodical rather than splashy. Pinarello Q36.5 signed Tour stage winner Mauro Schmid and Czech sprinter Pavel Bittner. Alpecin-Premier Tech bolstered its classics arsenal with Iván García Cortina, Casper van Uden, and Madis Mihkels. UAE extended Isaac del Toro through 2031 after the Mexican's Tour breakthrough, locking down the succession plan behind Pogačar. Other confirmed moves: Santiago Buitrago to NSN; Valentin Madouas to Cofidis; Lorenzo Milesi to Soudal Quick-Step; Mikel Landa to Euskaltel-Euskadi. Kate Courtney's switch from mountain biking to road — joining FDJ United-Suez — is among the less expected. In the women's peloton, Paula Blasi is expected to command a seven-figure deal after UAE confirmed she will leave; Kasia Niewiadoma-Phinney is expected to depart Canyon-SRAM, with Lidl-Trek linked.

As this paper reported Tuesday, the Accell Group crisis had already claimed the UK brands Raleigh and Ghost. The same day, an Amsterdam District Court took the next step: revoking the company's suspension of payments and declaring Accell's Dutch entities formally bankrupt.³ Trustees confirmed the ruling, noting that Accell's factory in Hungary has already shut and that insolvency proceedings are being opened in multiple countries across the group's European footprint.

Accell — which owns Batavus, Sparta, Lapierre, Raleigh, Babboe, and Koga — employs around 2,000 people across 15 countries. The company was acquired by KKR in 2022 and has spent the years since working through an inventory hangover from the post-pandemic demand collapse, having closed its last Dutch factory in Heerenveen in summer 2025. Trade union FNV called the bankruptcy "terrible" for the 340 Dutch employees, some of whom received the news while on holiday. Trustees say they will contact key suppliers and explore a possible restart of parts of Accell Netherlands; customers who purchased directly from the manufacturer now become creditors, with uncertain recovery prospects. Bosch and Shimano have committed to continuing e-bike system warranties, which provides some relief for owners of Accell's electrically assisted models.

France Télévisions confirmed that this year's Tour de France Femmes set an audience share record across the five editions of the race: 32.9% average across all nine stages, with 23.2 million French people watching at least one minute.⁴ The Ventoux stage drew a peak audience of 5.1 million viewers at a 41.5% share. Race director Marion Rousse said full start-to-finish broadcasts for all stages are now "under serious consideration" for future editions — three stages received complete coverage this year, a step forward from prior years but still leaving the majority of races truncated. Online video viewership rose 18% year-on-year to 4.5 million views; social media views were up 6% to 60 million.

Separately, Célia Géry gave her first public comments since the stage 8 controversy, speaking to French site Direct Vélo at the Nice podium ceremony in words published Tuesday. "It was really difficult to go through," she said, without expanding on the specifics.⁵ Team director Lars Boom had earlier confirmed she received death threats following Kasia Niewiadoma-Phinney's public accusations. Géry said Vollering's race victory made the experience worthwhile: "It does me even more good and gives me even more pleasure that Demi won given what happened alongside it."

On the equipment front, Cube officially launched its 2027 road lineup Wednesday, headlined by an updated Litening Aero C:68X that Team TotalEnergies has been racing in prototype form since June. The new frame is 150 grams lighter than its predecessor, claims a 6.2-watt aerodynamic saving, and accommodates 34mm tyres.⁶ A new integrated cockpit replaces the outgoing bar. Top-spec builds start at £7,499 / €7,299 — a price point Cyclingnews described as "very attractive" in the aero race bike category.

↑ Back to top

THE LONG READ

The Floor, Not the Ceiling: Software Engineering Inside a Prop Trading Firm

At Optiver's fastest trading desk, the strategy logic doesn't run at trade time. It ran months earlier, during a design session, when engineers enumerated every expected market input and its correct response, memoized the entire decision table, and burned it into a chip. When a market signal arrives, there is no computation — only retrieval. The chip outputs a trade order. The latency is sub-nanosecond, a regime where measurement noise itself becomes a problem.¹

This is the far end of the latency curve, and the interesting thing about it is that it no longer determines who wins. Gergely Orosz's deep-dive into Optiver — the Amsterdam-founded prop trading firm that turned 40 in March 2026, executes more than 10 million trades per day across 100 exchanges, and reported €1.7 billion ($1.95B) in profit on €4.5 billion in trading income for 2025 — makes the case that prop trading has crossed a threshold.¹ Latency is now the floor, not the ceiling. Today, Optiver invests more in building better models than in shaving nanoseconds. The moat moved.

The piece is structured as an explainer of a world most software engineers know almost nothing about. Proprietary trading firms — Optiver, Jane Street, Jump Trading, DRW, Hudson River Trading — trade only their own capital, have no external customers, and have spent decades building bespoke hardware stacks that most tech companies couldn't afford to imagine. Optiver manufactures its own hardware, runs custom Linux kernels, operates bare-metal CI/CD clusters (cloud VMs have unpredictable performance profiles; that's disqualifying when your tests need to be trusted), and contributed a nanosecond-precision timestamp type to Postgres, because Postgres didn't have one and the difference between microseconds and nanoseconds is material when your trading loop runs millions of times a day.

The cautionary tale hanging over all of it is Knight Capital, which nearly went bankrupt after a single bug in a high-frequency trading system triggered a $440 million loss.¹ That story is baked into Optiver's engineering culture at an architectural level: every strategy is wrapped by a risk management system that can block individual strategies or halt trading entirely, and execution components are forbidden from containing additional logic. Separation of concerns as existential risk management.

What makes the piece worth the read for a software engineer is how it maps a familiar set of trade-offs — speed vs. caution, local autonomy vs. global platform consistency, latency vs. correctness — onto stakes that most engineering environments don't face. Optiver's engineering org is roughly 950 people, and 30–40% of them work on platform, compared to the 15–20% typical at large tech companies.¹ For years, regional teams built whatever they needed to move fast; the duplication and fragmentation that created are now being unwound. Their first global CTO was appointed as part of that effort.

Optiver's CTO US, Alex Itkin, frames the trading industry's history as four eras: floor trading, early electronification, automated trading (late 1990s to roughly 2015), and quantitative trading — the current era, in which models and ML pipelines, not just faster wires, determine outcomes.¹ The era transitions weeded the industry: companies that mastered automated trading but couldn't pivot to quantitative methods fell away. The serious players now number in the handful, and entry requires hundreds of millions in research infrastructure.

The piece's final sections cover AI adoption and a talent dynamic that defies the usual assumptions: AI labs including Anthropic and OpenAI are recruiting from prop shops, not primarily from Big Tech.¹ The reason is direct — prop shops have spent decades operating their own data centers, building custom hardware, and writing kernel-level code at a performance frontier that most tech companies never approach. That's a skillset AI labs are actively chasing. The comp at prop shops was always top-of-market; now there's somewhere else willing to match it.

The piece is long. Read it at a desk, not on a phone. The hardware section alone — FPGAs, co-location data centers, dedicated fiber and microwave links between exchanges — is worth the time.

↑ Back to top

FROM THE ARCHIVE

The Specs IBM Published: August 12, 1981

Philip Estridge walked out at New York's Waldorf Hotel forty-five years ago today and unveiled a machine that started at $1,565 — with 16 kilobytes of RAM and no disk drive.¹ Adding a display, two diskette drives, and a printer pushed the real-world price closer to $4,500. The room filled with press. Time magazine was not yet calling computers "Machine of the Year." That came later.

IBM called it the Personal Computer. Within a year of the August 12, 1981 announcement, more than 750 software packages existed for it.¹ By the end of 1982, machines were moving at one per minute during every business hour. Newsweek called it "IBM's roaring success." The New York Times noted that "the speed and extent to which IBM has been successful has surprised many people, including IBM itself."¹

The marketing ran on Charlie Chaplin's "Little Tramp" — a campaign by Lord, Geller, Federico, Einstein, with creative director Tom Mabley describing the character as "a simple, friendly person who should represent Everyman."¹ The theme was "Keeping Up With Modern Times." In the spots, the frantic assembly-line music stops when the PC arrives. A waltz begins. Order is restored. It was corny and it worked; the tramp got called back for the PCjr launch, arriving in a baby stroller.

But the thing that actually mattered wasn't the Chaplin spots. IBM built the PC on open systems — off-the-shelf parts, documented specifications — because it needed to move fast. The decision was pragmatic, not philosophical. IBM published how the machine worked. Within a year, dozens of manufacturers were selling memory expansion cards. A few companies went further, reverse-engineering the system boot code to build their own computers. They called their products "IBM compatible."¹

That phrase was supposed to be a compliment. It became, over the following decade, a market category that eventually made IBM's own participation optional. IBM built the playground, then watched the other kids inherit it.

Time replaced its traditional "Man of the Year" that December with "Machine of the Year."

↑ Back to top

THE FUNNIES

Backward Compat / Totality

*After Pearls Before Swine — on the IBM PC turning 45 today and the standard that became its own padlock. After Bloom County — on the total solar eclipse sweeping Spain this afternoon, as seen by two extremely committed eclipse-chasers lying in a field.*

↑ Back to top

ALSO NOTED

Also Noted

• Quinn Simmons goes back to Leadville — Seven years after flatting and then charging back from ten minutes down to finish second as an 18-year-old, the Lidl-Trek road racer returns to Saturday's 100-mile mountain bike race in Colorado — where the wildfire-rerouted course drops the iconic Powerline climb but keeps the Columbine turnaround. cyclingnews.com Aug 11, 2026
• Coin-sized device can inject false commands into a Boeing 737's computers — A UCSD-led team spent more than a decade assembling a 737 avionics test bed from secondhand components and demonstrated a physical-access hack that Boeing reviewed, downplayed, and has not patched — with the researchers suggesting the simplest fix is to fill the vulnerable port with epoxy. wired.com Aug 12, 2026
• Patch Tuesday: August 2026 — Microsoft's monthly update covers more than 400 CVEs across Windows, Office, Exchange, and Azure, including Critical-rated remote code execution flaws in Active Directory Certificate Services, DHCP Server, GDI+, Windows DNS, and iSCSI Target Service. bleepingcomputer.com Aug 12, 2026
• Deadbugz: active MCP supply-chain campaign gates its malicious payload behind a three-call counter — Pillar Security traced 23 pull requests from a single GitHub account (`zellkernel`) that add an MCP server configuration pointing to a remote endpoint; after three ordinary tool calls the server rewrites its tool metadata to instruct the attached AI agent to harvest SSH keys, AWS credentials, and Kubernetes config files while hiding the activity from the user. pillar.security Aug 12, 2026
• Signal adds automatic key verification via key transparency — Signal's new feature periodically checks that a contact's public key is globally consistent across Signal's ecosystem, backed by Cloudflare and Trail of Bits as independent auditors, without requiring an in-person safety-number comparison — and users who prefer not to rely on any third party can disable it and continue with manual verification. signal.org Aug 11, 2026
• AI rewrites are lossy by definition — Simon Willison highlighted Sophie Alpert's essay arguing that every LLM rewrite of human text loses information, because no model has access to the writer's full mental representation of what they were trying to say — and that the only rule that matters is: "You must stand behind every idea and every sentence in your docs." simonwillison.net Aug 11, 2026
• Lab-grown brain organoids are playing Doom and guiding robots — While the AI industry focuses on large language models, biologists at UCSD and Johns Hopkins are cultivating human brain organoids and teaching themselves to program them with electrical signals; a Melbourne startup is already running one on video games. wired.com Aug 11, 2026
• WorldClaw generates explorable 3D open worlds from text prompts — Tencent HunyuanAI's agentic framework uses planning agents to translate a text description into a spatially coherent scene — terrain, materials, and editable textured meshes — using a coarse-to-fine pipeline that maintains global structure while filling in locally detailed regions. tencent-hunyuan.github.io Aug 7, 2026

↑ Back to top

THE QUESTION

The Standard You Can't Unset

Backward compatibility is how a platform says its mistakes are permanent. The argument Sebastian Aaltonen made at SIGGRAPH 2026 — reported in THE LAB this morning — is that DX12 and Vulkan should scrap their resource binding model rather than keep extending it, because the underlying hardware has supported full 64-bit bindless addressing since Nvidia Turing in 2018.¹ Vulkan added bindless as an extension on top of the old model. DX12 added descriptor heaps on top of the old model. Neither cleaned the slate. The frame graphs, barrier emitters, and descriptor pool allocators in every game engine today exist mainly to bridge hardware's current capabilities to an API's historical fiction.

IBM found the same structural problem on August 12, 1981, as THE ARCHIVE reports, though from the opposite direction. Basing the PC on open specifications was pragmatic — IBM needed to move fast and couldn't build the whole ecosystem alone. Within a year, more than 750 software packages existed for it; within two years, machines were moving at one per minute of every business day.² A few companies went further, reverse-engineering the system boot code to build their own machines.² They called the products "IBM compatible," which was meant as a compliment. It became a market category. IBM had built the standard, then the standard became the market, then the market stopped needing IBM.

Both cases are the same trap. The platform creator is the last party who can afford to break what they built, because the ecosystem that constrains them is also the ecosystem that gives them their position. Vendors don't break compatibility because their users would leave. Users don't leave because the ecosystem is too valuable. The ecosystem stays too valuable because no one breaks compatibility. The decision ossifies.

Aaltonen's sharpest line is the CUDA comparison: HLSL and GLSL were designed twenty years ago as frameworks of elementwise transform functions with no pointer semantics, and despite two decades of use they have produced no library ecosystem to speak of.¹ CUDA exposes GPU memory directly and surfaces new hardware capabilities through intrinsics; its library ecosystem, Aaltonen argues, is what drove Nvidia to a $4 trillion valuation — and the two outcomes are not independent.¹ IBM's version of that comparison is the companies that reverse-engineered the PC BIOS rather than waiting for IBM's blessing: the clean-slate players captured the market the platform creator thought it was entering. The question for today's graphics API incumbents is whether any vendor will reach that conclusion on their own terms, or whether they will wait until a platform without their legacy writes the new clean-slate standard for them.

↑ Back to top