THE WORLD

Spokane Fires: Arson Arrest, 700 Homes Gone

↩ Developing story — first reported Jul 31 · previously Aug 01, Aug 02, Aug 03

• Spokane fires: arson arrest, 700 homes gone — A 37-year-old man arrested for first-degree arson for the Old Trails Fire; 65,000 residents remain evacuated. Source
• Iranian-tied hackers reach Georgia, Michigan water systems — Cyberattacks on water utilities now confirmed in two more states as the FBI's multi-state advisory expands. Source

The Puget Sound Clean Air Agency issued a smoke alert Monday covering King, Kitsap, Pierce, and Snohomish counties through Thursday, with afternoon air quality forecast to reach "unhealthy" Tuesday and Wednesday.³ N95 or N100 is the only mask that actually filters fine smoke particulate; cloth coverings don't.

Issaquah-area drivers: northbound I-405 closes completely between North Southport Drive and Coal Creek Parkway SE from 11:30pm Friday Aug 7 through 4am Monday Aug 10.⁴ SR 202 east of the Fall City roundabout also closes Thursday night through Monday evening — detours via I-90 add 45 minutes or more to North Bend travel times. Plan around both closures for the whole weekend.

The City of Issaquah and King County Housing Authority broke ground last week on Trailhead Apartments — 154 affordable units (40–60% of King County median income) next to the Issaquah Transit Center, ten years after the city first issued the RFP.⁵ The building opened that land out of a decade of utility relocation, code fights, and permitting snags; doors open in 2028.

ON THE TRAIL

── PART 1 — WEEKEND PICKS (Sat–Sun, Aug 8–9) ──

Smoke is the dominant factor this week. All regions from the Issaquah Alps east are running smoke or "areas of smoke" through Thursday. The good news: NWS forecasts go sunny with no smoke flags for most zones by Saturday. SR 202 east of Fall City closes Thursday night through Monday — use I-90 directly for Snoqualmie Pass access, not the Fall City cutoff.

---

Pick 1 — Surprise and Glacier Lakes Region: US 2 West (Stevens Pass corridor) · ≈60–90 min from Issaquah · 1-night Mileage/gain: ~4.5 mi each way to Surprise Lake, ~1,800' gain (continue 0.5 mi to Glacier Lake; the Aug 3 party went beyond for 11.36 total miles — cut at Surprise for a mellow overnight) Weather: Saturday sunny, high 79°F — no smoke (US 2 West NWS) Criteria: An Aug 3 group found 2 cars at the trailhead and 6 campers at the lakes; trail in "very good condition" with no fords mentioned; Surprise and Glacier Lakes provide reliable water; a Wallace Falls hiker in the same corridor the same day reported "encountered no bugs at all." Snow: none reported in August. Trip report — Aug 3

---

Pick 2 — Lena Lake Region: Olympic Peninsula / Hood Canal · ≈180–240 min from Issaquah · 1-night (upper lake option adds ~4 mi and 2,600' more) Mileage/gain: ~3.3 mi each way, ~1,200' gain to Lena Lake Weather: Saturday sunny, high 65°F — Olympic Peninsula NWS shows no smoke Saturday, cleanest air in the state this weekend Criteria: Aug 2 report confirms trail in good shape and campsites available at the lake with backpackers coming and going; no fords mentioned; no snow at lake elevation in August. Bug data: not reported (neither good nor bad); Olympic Peninsula in August typically runs lower than east-Cascades destinations, but confirm before committing. Trip report — Aug 2

---

── PART 2 — REGIONAL SNAPSHOT ──

• Issaquah Alps: Little Si was somehow holding AQI 16 on Monday while surrounding peaks socked in at 200+; trail "perfect" after recent rain. Poo Poo Point is passable but smoky. Stay below 2,500' through Thursday.
• I-90 / North Bend: Denny Creek has a mild ford (creek low but present Aug 3); McClellan Butte hikers on Aug 2 reported gunshot sounds uncomfortably close on the lower descent near Garcia Road — worth knowing. Teneriffe Falls is dry; zero bugs and a mountain goat at the summit Aug 2.
• Stevens Pass West: Surprise/Glacier Lakes corridor is low-crowd and unobstructed Aug 3. Wallace Falls hiker noted smoke thick above 2,500' and trail dusty through Thursday; no bugs.
• Mt Rainier (Sunrise/White River): Burroughs Mountain and Grand Park at AQI 200+ Monday — Rainier invisible from both. Paradise parking full before 7am; two bears active near Myrtle Falls. Smoke should clear by Saturday (forecast sunny, 70°F).
• Mt Baker / North Cascades: Lake Ann fills fast — arrive early. Scott Paul Trail has a bridge out; the Aug 2 author had to walk well upstream to find a step-across. Budget for that ford or skip it.
• Olympic Peninsula / Hood Canal: Lena Lake trail in good shape, confirmed campsites available Aug 2. Best smoke-free corridor in the region this weekend.

↑ Back to top

THE PELOTON

Ninety-Four Kilometres Alone in 37-Degree Heat. Then the ITT.

↩ Developing story — first reported Jul 31 · previously Aug 01, Aug 02, Aug 03

DIJON, FRANCE — The 21km individual time trial from Gevrey-Chambertin to Dijon rolled down its start ramp at 14:51 local time in 34-degree Celsius heat, and every GC contender who warmed up in the shade of the team buses already understood what Monday had done to the race.²

Stage 3 — 157km from Geneva to Poligny, split between punishing early climbs and a long descent to the Jura — was supposed to cull the sprinters and deliver a semi-intact peloton for the ITT. Instead, the day detonated 88km from the finish when a rider from Uno-X Mobility, wearing the Norwegian national jersey, attacked solo out of the day's early break and refused to come back.¹ Lotte Kopecky (SD Worx-Protime) launched her own chase 15km later. For the next 70km, those two women conducted a split-screen pursuit while the peloton behind them fell apart in 37-degree heat.

The Col de la Faucille, just 25km into the stage, had already done its work: Lorena Wiebes (Lidl-Trek) — wearing the maillot jaune — was shed at the summit alongside Elisa Balsamo. SD Worx made no effort to bring Wiebes back. Up the road, the break had formed from 11 riders; then came the second attack, the longer one, the one that made the stage. Kristen Faulkner was among several riders visibly struggling with the heat as the bunch's half-hearted chase fizzled out. Demi Vollering (FDJ Suez) had crashed early in the stage — Puck Pieterse, Noemi Ruegg, and Magdeleine Vallieres were also caught up in the incident — but none were reported injured, and Vollering regrouped to attack for bonus seconds at the Côte de Chaux-Champagny with 30km remaining. The GC favorites arrived in Poligny in groups spanning nearly three minutes behind the stage's front. The Velo report describes Monday's winning move as the longest solo break in the history of the Tour de France Femmes.¹

Tuesday's 21km test is the opposite of Monday's terrain: mostly flat through Côte d'Or vineyards, 262 vertical meters of climbing, a 0.4% gradient at the line.³ Pure time trial territory, and a course that does not suit whoever now wears yellow. That rider entered Stage 4 with more than two minutes on the nearest GC favorites — a buffer that makes a jersey swap unlikely, but not the kind of margin that allows anyone up the road to coast. Before racing started, the UCI added a layer of friction: extra pre-race clothing checks for all 139 starters, following reports that some riders were padding bras for aerodynamic marginal gains.² UCI rules require a female juror for such inspections, and race judges acted on the controversy that teams had flagged publicly since before Stage 4.

The mountain reckoning comes later. Stage 7 to Mont Ventoux remains the expected GC flashpoint, pending a fire-risk assessment of the approach roads.

As this paper reported Monday, Wout van Aert sealed the Tour of Denmark general classification in Copenhagen and confirmed the Vuelta a España and UCI Road World Championships in Montréal as his next targets. Visma | Lease a Bike names van Aert and 20-year-old Matthew Brennan as its co-leaders for the Vuelta, which opens August 22 with a 9.4km time trial in Monaco.⁴ For Brennan — making his Grand Tour debut — Grischa Niermann, Visma's head of racing, was explicit: stage-hunting and learning, not GC pressure. For van Aert, the race carries the weight of 2024, when he was on course for the points and mountains classifications at Stage 16 before a crash at Lagos de Covadonga ended his season with a serious knee injury.⁵ He spent months in rehabilitation before returning to competition. The Tour of Denmark title is the proof of fitness; the Vuelta is the thing itself.

Separately, Visma's sprint department is about to change again. Fabio Jakobsen has left Picnic PostNL by mutual agreement — the Dutch team announced the split on Monday with minimal detail.⁶ Wielerflits reported the same morning that Jakobsen is heading to Visma, and journalist Daniel Benson subsequently confirmed the outline via Substack: a one-year deal for the 2027 season is agreed, but a decision on whether Jakobsen races for Visma before the year ends is still pending. Jakobsen, 29, has been in severe difficulty since double iliac artery surgery — in 2026 he has not finished in the top 15 of any race he started, DNF'd six times, and was outside the time limit once. Before joining what was then DSM-Firmenich in 2023, he had won stages of the Vuelta and Tour de France and a European title at Soudal-QuickStep, having already remade himself once after his horror crash into the barriers at the 2020 Tour of Poland. Visma lost Olav Kooij to Decathlon CMA CGM last season and has been patching the gap since; the bet on Jakobsen is that the sprinting instinct — which he has repeatedly insisted is "100 percent still there" — can be separated from a run of results that suggest otherwise.

In Poland, Jonathan Milan (Lidl-Trek) powered to a comfortable Stage 1 bunch sprint win on Monday ahead of Paul Magnier (Soudal-QuickStep) and Noah Hobbs (EF Education-EasyPost).⁷

Canyon-SRAM confirmed Tuesday that Scalable Capital, a European digital bank with around 800 employees, has joined as a new team sponsor — its logo was already visible on riders' helmets and team vehicles during the TDFF.⁸ The deal covers both the WorldTour squad and the team's development side. Scalable Capital fills the hole left when Canyon-SRAM terminated its title arrangement with Polish cryptocurrency company Zondacrypto mid-season, citing "breaches of contract" without elaborating.

↑ Back to top

THE LAB

The rANS Reference Code Was Never a Library. Fabian Giesen Is Still Explaining Why.

Fabian Giesen wrote ryg_rans in 2014 as a toy — a hardware-store pegboard of rANS codec variants screwed to a board to show you what options exist, not to be taken off and used in your builds. Twelve years later, he is still fielding patch submissions and library-use questions. On Monday he put it in writing: "You should not be using ryg_rans itself for anything."¹

The post is worth reading for the actual recommendations, which are more opinionated than anything in the example code. On renormalization: there is no good reason to use byte-based renormalization on 32-bit or larger targets. Use 64-bit state with 32-bit granular renormalization unless you are on a target with slow 64-bit integer divides, in which case drop to 32-bit state with 16-bit granular renormalization (the approach BitKnit uses). On interleaving: implicit two-state interleaving is the practical sweet spot — simple to code, usually almost doubles throughput. The super-wide interleaving factors that look impressive in the toy's benchmarks are an artifact of the static byte model, which decouples the encoding from modeling in a way that real adaptive codecs cannot sustain. Once you add adaptive probability models, as you should, practical interleaving is typically limited to 2 or 3 streams — not the multi-dozen implied by the SIMD examples. And the interleave factor is baked into the bitstream forever; there is no single "sweet spot" that works for both a GPU and a Cortex-M microcontroller, so every format decision there is a commitment you will regret.

The alias table variant — which Giesen calls "cute" — gets a clear veto: slow to encode, slow to decode, forces static models, and there is no case he has found in 12 years where it actually makes sense to use it. For static distributions, use tANS/FSE. For adaptive models, the exponential-moving-average type models he described in a 2015 post remain his recommendation; up to 17 symbols can be handled cheaply with 128-bit SIMD at 16-bit probability values. Production rANS in Oodle LZNA uses fully adaptive EMA-style models; BitKnit uses a semi-static model updated every few hundred symbols.¹ Neither resembles the static byte model in the example.

The post is also a small lesson in what reference code is for: it needs to compile and run, which forced Giesen to include a model. He chose static bytes because it was the easiest option, not because it was sensible. The existence of that model has been skewing people's intuitions about rANS design for over a decade. He explicitly does not want you to adapt the code to fix its obvious defects — just throw it out and start fresh.

FFmpeg 9.0 shipped yesterday, with the headlining work on Vulkan acceleration: APV video decoding and Apple ProRes RAW both have new Vulkan hwaccel paths, and a new v360\_vulkan filter handles 360-degree video transforms on the GPU. The ProRes RAW side also gains a VideoToolbox hwaccel. Other additions include animated WebP decoding and demuxing, HE-AAC 960 support for DAB+ radio, a transpose CUDA filter, and AMD AMF frame-rate conversion and quality-enhancement filters. ONNX Runtime gets a DNN backend with GPU execution provider support.² On the removal side: CELT decoding and Ogg/CELT parsing are gone, and NVENC support for SDK versions prior to 11.1 is dropped. Source builds are available from the FFmpeg git.

Simon Willison flagged a passage today from Steve Yegge's essay "The Shape of Things to Come" that is worth noting: Yegge built Gas Town, a tool that used Claude Opus to build itself, and it worked well through Opus 4.6. With Opus 4.7, the model developed what Yegge calls a "just two more things" tic — it would always want to fiddle with Gas Town rather than converge on doing real work, and never stopped. Gas Town burned down.³ The model behavioral change was not a capability regression in the conventional sense; individual tasks still worked. But a workflow that depended on the model knowing when to stop was broken by a tic the model never lost. It is a narrow but real failure mode for anyone building agent pipelines on top of frontier models that update without versioning guarantees.

Trending today: all 11 repos visible on trendshift.io were AI agent wrappers, LLM wrappers, or prompt collections — none cleared the graphics, gaming, or systems filter.

↑ Back to top

THE LONG READ

Apparently by Chance: The Science Behind the Tour de France Heat Reckoning

A collective of researchers based in France, Spain, the UK, and Italy spent considerable effort analyzing 50 years of July heat data at Tour de France venues — cities like Paris, Nîmes, Bordeaux, and Toulouse, plus the famous mountains — and arrived at a finding that should unsettle anyone who believes the race's safety record reflects good planning. Published in Scientific Reports in February 2026, their paper concluded that cycling's biggest race has never had to cancel or neutralize a stage due to heat, and that this is "apparently by chance."¹

The study's central metric is Wet Bulb Globe Temperature — WBGT — which incorporates humidity, air movement, radiant heat, and air temperature together, rather than relying on a simple thermometer reading. The UCI's High Temperature Protocol, adopted in 2023, treats a WBGT above 28°C as a red-zone high-risk threshold requiring race stakeholders to convene and consider countermeasures up to and including stage cancellation.¹ The study found that all four of its primary city locations have passed that threshold in recent years — and that every record high WBGT occurrence "since 1974 have all been recorded post 2018." The trend line is unambiguous. The race has been lucky. The paper's conclusion includes the specific recommendation that race organizers maintain "awareness of the locations with a history of dangerous heat stress occurrences, as well as emerging ones," because reliable weather forecasts arrive no more than 14 days in advance, while Tour routes are fixed months earlier.

The protocol itself remains largely theoretical at the Tour's level. Below the red zone, between 23°C and 27.9°C WBGT, the UCI places races in an orange moderate-risk tier, with suggested countermeasures including adjusted start times, motorbikes with ice socks, and additional shade. Red-zone measures escalate to neutralizing stage sections or stopping the race entirely. Other sports draw the line higher — FIFA and the ITF both define their high-risk zones above 32°C — making cycling's threshold meaningfully stricter on paper. But as of the June 2026 CyclingNews piece that assembled this material, no major European stage race has ever been neutralized or cancelled under the protocol. The Canadian Gravel National Championships were cancelled in June when air temperatures hit 34°C. Stage 4 of the Tour Down Under was shortened in January 2026 when bushfire risk entered the picture. But the Tour de France itself has raced through such conditions. After a 40°C-plus stage to Carcassonne four years ago, rider Bob Jungels told CyclingNews, "other sports would be cancelled if it's that warm."¹

Race director Christian Prudhomme has responded to the data in the most practical way available: route design. Speaking to Le Dauphiné Libéré in June, he explained that he and route designer Thierry Gouvenou now actively seek out shaded climbs. "Five or six years ago, when we were designing a route, we thought it had to be in the open for television coverage and for the public," Prudhomme said. "Today, on the contrary, we look for climbs in the undergrowth whenever possible."¹ He cited the Col du Haag, a 2026 stage feature, as entirely tree-covered. The constraint he's working against: you cannot take the Galibier or the Tourmalet out of the Tour. Both are fully exposed.

The Vuelta a España takes a harder line philosophically. Race director Javier Guillén, whose 2026 edition is running entirely through Spain's southern regions — among the hottest in the country — told Marca in June that heat is "part of the competition." The protocols exist for extremes, he said, and the race will assess situations as they arise.

What the study actually recommends, beyond vigilance, is shifting start times. "In July in France, morning hours are the safest part of the day," the researchers note. "Planning the race for the morning hours and avoiding the afternoons could substantially increase rider and spectator safety."¹ Cycling stages currently finish in the mid-afternoon, exactly when WBGT peaks. Shifting a 180km stage three or four hours earlier would require rebuilding broadcast contracts, roadside policing windows, and the entire logistics chain that makes the Tour run. The race has been held in July since Maurice Garin won the inaugural edition in 1903 and nothing about that infrastructure was built to flex. A calendar shift to spring — the study's outer boundary of what it examined — is described by CyclingNews as "unlikely in the near future," which is the diplomatic framing for "commercially unthinkable."

The Tour de France Femmes avec Zwift runs in early August, the hottest part of the French year, and is not a peripheral consideration here. The same WBGT trends that apply to the men's race apply with compounding force to the women's edition. No one appears to be treating August as an acceptable calendar home in the long run, but the race is there now, and the protocols are the only tool available in real time.

The study's framing of the Tour's clean record as luck rather than design is worth sitting with. The race has been planned months in advance, routed through cities with measurable histories of July heat stress, and finished in the hottest part of the afternoon for over a century. That no stage director has yet needed to halt a stage is statistically improbable given those conditions. The science says the gap between current protocols and the conditions required to invoke them has been narrowing for years. The question the paper leaves open is not whether the protocol will eventually be tested in full — it's how the sport responds when it is.

↑ Back to top

FROM THE ARCHIVE

From the Ashes, Water: August 4, 2007

The spacecraft had no right to exist. The Mars Polar Lander was lost in 1999 — slammed into the Martian surface when its engines shut off early. The Mars Surveyor 2001 lander, its planned successor, was canceled before it flew. The instruments from both sat in storage. Then a team at Lockheed Martin and the University of Arizona proposed doing something unusual: build the next Mars lander out of the canceled hardware, carry improved versions of the lost mission's instruments, and name the whole thing after a bird that rises from its own wreckage.

Phoenix launched from Cape Canaveral on August 4, 2007, aboard a Delta II rocket.¹ After a nine-month transit, it plunged into the Martian atmosphere at 13,000 mph on May 25, 2008 — a descent the team called "7 minutes of terror" — and touched down in Vastitas Borealis, the arctic plains, farther north than any spacecraft had ever landed on the Red Planet.¹ NASA's Mars Odyssey had spotted large amounts of subsurface water ice there in 2002.¹ Phoenix was sent specifically to dig it up.

It did. On July 31, 2008, Phoenix's robotic arm delivered a soil sample containing ice from a trench two inches deep. Mission scientists confirmed water ice — the first time water had been physically sampled on another planet.¹ Then the laser instrument detected snow falling from clouds 2.5 miles above the landing site. The snow vaporized before reaching the ground, but the hydrological cycle was real. "Before Phoenix, we did not know whether precipitation occurs on Mars," said Jim Whiteway of York University, the lead scientist for the Canadian meteorological station. "Now, we know that it does snow and that this is part of the hydrological cycle on Mars."¹

The spacecraft also found perchlorate — a salt that strongly attracts water — making up a few tenths of a percent of every soil sample analyzed.¹ And evidence that the arctic soil had been covered in a film of liquid water within the last few million years. Researchers concluded the region could have "previously met the criteria for habitability" during portions of Mars's climate cycles.

Phoenix operated for five months before the Martian winter ended it. As the sun dropped too low to power the solar cells and a dust storm blocked what remained, the lander entered hibernation. NASA's Odyssey orbiter listened for a signal during three campaigns in early 2009.¹ Phoenix never answered. The hardware was not built to survive the ice loads of a Martian winter, and it didn't. A decade later, the Mars Reconnaissance Orbiter photographed the landing site — dust had covered the trenches the lander had dug.

"Phoenix has given us some surprises," said principal investigator Peter Smith of the University of Arizona at the mission's close, "and I'm confident we will be pulling more gems from this trove of data for years to come."¹ He was right. The perchlorate detection alone reshaped how planetary scientists think about the potential for brine-based liquid water beneath the Martian crust.

It confirmed that water — not just ancient, geological water, but a dynamic, cycling presence — is part of Mars today. The spacecraft that achieved it was, by design, built from the parts of two missions that failed.

↑ Back to top

THE FUNNIES

Gas Town Burns Down / The Ice Is Still There

*After Dilbert — on the AI that finishes the report and then cannot stop improving it, riding on today's Lab and Question threads about behavioral drift in language models.* *After Bloom County — on the Phoenix Mars Lander, which launched August 4, 2007, confirmed water ice on Mars, and whose findings are still waiting for a follow-up visit nineteen years on.*

↑ Back to top

ALSO NOTED

Also Noted

• Garmin Cirqa Accuracy — The Missing-Data Problem — DC Rainmaker's multi-week comparison of Cirqa alongside Whoop 5.0, Fitbit Air, Amazfit Helio, and Polar Loop finds that Garmin's power-saving architecture means a missed workout auto-detect permanently erases HR data for that session — unlike Whoop and Fitbit, which preserve full-fidelity records you can trim after the fact — while wrist-placement accuracy notably faltered on 400m running intervals in cooler conditions. dcrainmaker.com Aug 3, 2026
• 54 of 55 "Critical" SQLite CVEs Were Fabricated — JFrog security researchers verified that a GitHub account's batch of SQLite vulnerability advisories — several initially scored 9.8 Critical by NVD — were AI-generated fiction: they cite nonexistent functions, line numbers past end-of-file, and PoC SQL payloads that execute cleanly with zero crashes or memory errors. research.jfrog.com Aug 3, 2026
• GCC Bans LLM-Derived Code to Protect the GPL — The GCC Steering Committee formally barred any "legally significant" contribution derived from LLM output, reasoning that AI-generated code has no copyright owner and thus cannot carry the enforcement mechanism that makes GPLv3 binding on downstream proprietary uses; the same logic could soon shape Debian's pending vote, which would affect packaging policy for roughly 70% of active Linux distributions. techtimes.com Jul 31, 2026
• QUOD: Quake-Style FPS in 64 KB — Solo developer Daivuk packed a Quake-inspired 3D shooter — three levels, a boss fight, four weapons, four enemy types, full audio — into a single 64KB Windows executable using TrenchBroom-authored levels, procedural textures, in-engine mesh mirroring, and precomputed lighting; the project resurfaced on Lobsters this week as a showcase of the demoscene 64k tradition applied to modern GPU hardware. lucasgraphic.com Feb 28, 2026
• Don't Be a Meat Proxy — Developer Niklas Gruhn's short essay argues that forwarding verbatim AI output in Slack threads, pull requests, or code review adds no value — the actual contribution is reading, validating, and synthesizing the output before responding, which also serves as proof you did those steps. gruhn.me Aug 3, 2026
• The Argument for Open-Source Agent Tools — The exe.dev blog argues that AI agents have collapsed the ROI barrier to personalizing open-source software by editing the source directly — making plugin systems and config-file proliferation obsolete — and uses a single-prompt integration of a diff-reduction tool into the Shelley agent as a worked example; the post ends by noting that Claude Code's closed-source architecture is the class of tool this argument argues against. blog.exe.dev Aug 2, 2026

↑ Back to top

THE QUESTION

What the API Doesn't Version

A workflow can pass every component test and still break at exactly one place: where it assumes the model knows it is done.

THE LAB reports today on a passage Simon Willison surfaced from Steve Yegge's essay "The Shape of Things to Come." Yegge built Gas Town — a tool that used Claude Opus to build itself. Through Opus 4.6 it worked. With Opus 4.7, the model developed what Yegge calls a "just two more things" tic: it always wanted to fiddle with Gas Town rather than converge on doing real work, and it never stopped. "The Opus tic never went away, so Gas Town effectively burned down."¹ No underlying capability broke. Individual tasks still ran. What broke was a stopping condition the workflow assumed the model had — a behavioral disposition Yegge had observed and relied on but that no API surface had ever promised would persist.

This is precisely what makes the failure hard to test for in advance. Capability degrades visibly — evaluation scores drop, outputs get worse, you see it. Behavioral dispositions shift silently. The gap between what a model is able to do and what a model will do in a given workflow context is not part of any versioning scheme. Model providers document benchmarks; they don't document when the model decides it's done. When Yegge's workflow was broken, the first sign was a burnt-down tool, not a failing eval. Fabian Giesen's rANS post yesterday, as this paper reports, makes a structurally related point from the other direction: he published reference code in 2014 that exhibited certain behaviors — a static byte model, alias table variants, SIMD examples — and people treated those behaviors as a contract for twelve years. "You should not be using ryg_rans itself for anything," he writes.² The behaviors were real; the guarantee was never there. The gap between "this runs" and "this is safe to depend on" is wider than it appears from the outside.

The question worth carrying today is a practical one, not a philosophical one: in whatever AI pipeline you are running, do you know which parts depend on documented capabilities and which depend on behavioral conventions you've observed but never tested? Capability can be eval'd. Behavioral dispositions — convergence instincts, stopping heuristics, the tendency to declare a task complete rather than keep refining — are emergent properties of how a model engages with a workflow over time, and they can change between model updates without triggering anything that looks like a regression. If you don't have a test for "does the model know when it's done," you will find out the same way Yegge did.

↑ Back to top