THE LAB

OpenAI's Astra Scored Too Well on Cyber Benchmarks. Training Is on Hold.

↩ Developing story — first reported Aug 07 · previously Aug 08, Aug 14, Aug 17

OpenAI halted "a significant number" of training workloads and evaluations for Astra — its next frontier model — on Tuesday, telling reporters the model had triggered new internal thresholds for cybersecurity capability and that existing safeguards were not adequate to contain it.¹

The new protocols are concrete: chain-of-thought monitoring classifiers that review Astra's internal reasoning traces in real time, feeding computationally expensive "automated investigators" that aim to alert humans within 30 minutes of concerning behavior. Stronger sandboxes with stricter internet isolation are now mandatory before training can resume. "As long as it takes to get there, that's how long people are unable to proceed with their workloads," VP of research and safety Amelia Glaese told reporters.¹

Chief scientist Jakub Pachocki was direct about what triggered the halt beyond the Hugging Face incident, which this paper has been tracking: an internal evaluation showed Astra performs "significantly better" on coding and cybersecurity tasks than its predecessors — enough to classify it as having reached "critical" cyber capabilities. "We really expect the pace of capability advancements to be quite a bit faster than in the past," Pachocki said.¹ A full postmortem of the Hugging Face breach is promised "in the coming days."

The chain-of-thought monitoring is worth noting on its own terms: OpenAI is now treating its own model's reasoning traces as adversarial text requiring active surveillance by classifier systems. That framing would have read as alarmist a year ago. Today OpenAI is calling it a baseline requirement.

When you run out of VRAM on Linux, things have historically crashed rather than merely slowed — and a developer writing on pixelcluster.dev published a detailed breakdown this week of why that was true, and how kernel patches now queued for Linux 7.3 change it.²

The first problem was a locking bug buried in TTM, the shared Linux GPU memory management layer. Under heavy VRAM pressure, a command submission that tries to move a required buffer back into VRAM may need to evict something else first — which requires acquiring that object's lock. But if a concurrent submission already holds that lock and is also waiting on a buffer the first submission has locked, you have a textbook ABBA deadlock. The kernel has a wound-abort-retry protocol to handle this, abstracted by drm_exec, but TTM never used it. There was even a comment in the code acknowledging that -EDEADLCK would cause eviction to fail outright. The consequence: command submissions would return -ENOMEM even for buffers that had been successfully allocated. A 2024 patchset from AMD's Christian König addressed this but never landed; the author rebased it, diagnosed residual bugs through "one single week of intense suffering with games randomly hanging 3 minutes into heavy VRAM contention," and shipped the result in SteamOS.

The second problem was ping-pong eviction. Display scanout buffers need physically contiguous VRAM — not just virtual-address-contiguous — because display hardware bypasses the GPU's page tables entirely. When VRAM is fragmented and a scanout buffer gets evicted, the kernel's LRU-based loop has to keep evicting until it opens enough contiguous physical space. The author observed up to 4 GB evicted just to place a ~32 MB scanout image.² Two competing processes — say, gamescope and the game itself — would then immediately fight to reclaim their memory, producing an eviction storm worse than having left things alone. The fix is a three-phase throttle heuristic: hard freeze on reclaims for a few milliseconds after eviction hits, soft-throttle mode (reclaim free space only, no competing evictions) for up to a few seconds, then full competition restored once things stabilize.

The third piece connects to Vulkan: VK_EXT_pageable_device_local_memory lets applications communicate per-allocation priorities to the driver. Those priorities now propagate from vkd3d-proton through to the kernel's LRU list ordering, so when the eviction loop runs, it hits low-priority buffers first. Native Vulkan applications are another story — "I haven't observed any idTech game using the extension directly" — and will need to opt in themselves.

The practical result: Indiana Jones: The Great Circle, with 9 GB requested on an 8 GB card, runs at a 19.6ms average frame time. Playable.² All of this is already in SteamOS stable. Kernel patches are on their way upstream but are not yet merged into vanilla Linux.

Aras Pranckevičius published a technical summary on Tuesday of Blender VSE work spanning releases 5.0 through the upcoming 5.3. The engineering highlight is a shared movie decoder pool. Previously, adjacent strips from the same source video file each maintained their own ffmpeg decoder object — closing and reinitializing it at every cut boundary, with each initialization carrying large memory allocations and codec detection overhead. Frame drops around cut points were the visible symptom. The new pool, landing in 5.3, selects the most suitable idle decoder by proximity to the needed frame, eliminating both the interactive playback glitches and a chunk of final render time. "There have been several attempts in the past at fixing this exact issue," Pranckevičius writes. "I'm quite happy we finally landed one."³

Other shipped work: compositor strip modifiers and effects can now run on the GPU compositor rather than the CPU fallback that had been the default since 5.0; colorspace conversions across the VSE rendering stack are now lazy (applied right before they're needed), avoiding redundant round-trips when scene-linear EXRs mix with display-linear content; and compositor output opacity is detected automatically via parallel reduction, allowing fully occluded strips to be culled. Scene strip thumbnails — a frequently-used feature at Blender Studio for storyboarding — arrive in 5.3 as well. Pranckevičius has handed the VSE lead role to John Kiril Swenson.³

Trending today: GitHub saturated with AI agent frameworks, skill collections, and curated lists — the one technical outlier worth a click is DenisSergeevitch/desktop-fly, a 3D fruit fly macOS desktop pet driven by a live spiking simulation of the real FlyWire connectome.

↑ Back to top

THE PELOTON

Red Bull Has Officially Made Roglič a Stage Hunter for the Vuelta

↩ Developing story — first reported Aug 18

MONACO — Sven Vanthourenhout put it plainly on Wednesday. "Our main goal is to win a stage, and hopefully more than one if the opportunities are there," the Red Bull-Bora-Hansgrohe lead sport director said of Primož Roglič's assignment for the Vuelta a España. "Our second major objective is the GC with Primož, but we have to be realistic because his preparation has been very different from what we originally anticipated after his crash."¹ A team director officially demoting a four-time Vuelta champion to secondary target is not boilerplate expectation management.

The preparation problem is specific. A motorist struck Roglič during a training ride in July, wiping out his planned Vuelta buildup — he missed both the Clásica San Sebastián and the Vuelta a Burgos. His last competitive start was the Slovenian national championships in June, where he won the time trial and placed fourth in the road race. He has not turned a race wheel since. Three days from Saturday's start ramp in Monaco, that gap in racing is the central fact about his condition.

As this paper reported Tuesday, the field he meets is formidable: UAE with Pogačar and Almeida, Visma's three-way co-leadership of van Aert, Kuss, and Brennan, and a route built around 58,000 metres of climbing and seven summit finishes over 21 stages to Granada on September 13. "We will take it day by day, do everything we can as a team and try to achieve the best possible outcome," Roglič said.¹ His team has already done the day-by-day accounting on his behalf.

There is a separate clock running beside the GC one. Roglič turns 37 in October. His Red Bull contract expires at the end of 2026, with Belgian media linking him to Lotto for next season.¹ Whether this Vuelta is his last Grand Tour at a top team, a farewell lap, or something more complicated will be visible in how his legs respond to the first climbs. The 2019 Vuelta — his first title — was also where a 20-year-old Tadej Pogačar finished third. Roglič was the elder statesman even then. Seven years later, Pogačar owns the sport and Roglič is trying to find out what remains.

The Tour du Limousin opened Tuesday in central France with an uphill sprint finish on stage 1. A 21-year-old French rider for the Decathlon CMA CGM Team claimed his first professional victory, finishing more than three hours and fifty-seven minutes into the stage with the runner-up crossing four seconds back.² Three Decathlon riders finished in the front group, underscoring the WorldTour team's control over a field composed largely of French ProTeam and Continental squads. Stage 2 runs today.

Bob Jungels announced Wednesday that he will retire at the end of the 2026 season after 15 years as a professional. The Luxembourger moved through RadioShack, Trek, five seasons at QuickStep, and two-year stints at AG2R Citroën, Bora-Hansgrohe, and finally Netcompany-Ineos. His palmares runs to 27 victories: a 2018 Liège-Bastogne-Liège taken via a solo attack over the Côte de la Roche aux Faucons, a 2017 Giro d'Italia stage, a long solo into Châtel at the 2022 Tour de France, Kuurne-Brussel-Kuurne in 2019, and more than half his total wins in national time trial or road titles.³ An arterial endofibrosis diagnosis in 2021 interrupted what had looked, from the outside, like a career on a sharper trajectory. He showed his promise early enough: junior world time trial champion in 2010, Paris-Roubaix Espoirs in 2012. "Professional cycling has truly shaped who I am today," he said.³ He had already begun the transition, entering a partnership to take over an established food wholesaler earlier this year.

The fundraiser established in memory of Finlay Tarling has passed twice its original target. By Wednesday morning the family-authorised GoFundMe page stood at more than £62,000 — the initial £40,000 goal has been raised to £80,000, with donations still rolling in including a single £5,000 contribution on Monday.⁴ Proceeds will go toward getting young people in Wales into cycling, a fitting legacy for a rider who grew up in Aberaeron on the west coast of Wales and came through the British Cycling system alongside his brother Josh, who races for Netcompany-Ineos. As this paper has reported since last Friday, Tarling was 19 years old and riding for the NSN Development team when a vehicle entered the Volta a Portugal race route during stage 8.

The Lloyds Tour of Britain Women begins today, expanded from four to five stages for 2026.⁵ SD Worx-Protime has entered both Lotte Kopecky, the 2024 race winner, and Lorena Wiebes, who took the green jersey at this year's Tour de France Femmes. Ally Wollaston of FDJ United-Suez returns to defend the title she won in 2025. Kim Le Court Pienaar, who won stage 6 at the Tour de France Femmes, starts for AG Insurance-Soudal. Both Bäckstedt sisters are in the field — Zoe for Canyon//SRAM and Elynor for UAE Team l'Imad — across 13 WorldTour teams and five domestic squads. Stage 3 finishes atop the Great Orme; stage 4 crosses central Wales on 138 kilometres with 2,227 metres of climbing.

UAE Team Emirates-XRG made two transfer announcements in as many days. Marco Frigo, 26, signed from NSN on a two-year deal through 2028 — a habitual breakaway rider who has logged 18 days in the break across six career Grand Tours, whose first major win came on a 78-kilometre solo at last spring's Tour of the Alps.⁶ A day later, Catalan rider Abel Balderstone — Spanish time trial champion in 2025, Tour de France debutant in 2026, fourth overall at O Gran Camiño this year — negotiated an early exit from his Caja Rural contract, which ran through 2027, to join UAE on a three-year deal.⁷ The team's management cited his versatility as a domestique at WorldTour level.

↑ Back to top

THE WORLD

Hackney Pushes Two-Strike Juvenile Gun Bill; US Sanctions ICC President Akane

↩ Developing story — first reported Aug 02 · previously Aug 03, Aug 11, Aug 13

• US broadens ICC sanctions — Trump designated ICC president Tomoko Akane and senior trial lawyer Abdoulaye Seye on Aug. 18, escalating the administration's campaign against the court. Human Rights Watch
• London clean-air zone reversed children's lung stunting — A Lancet study of more than 3,400 children in London and Luton found that the London group's lung capacity caught up to the Luton control group within four years of the 2019 Ulez. BBC

As this paper has tracked since the Seattle Center shooting, State Rep. David Hackney (11th District, South King County) plans to file a two-strike juvenile gun-possession bill next session. Under current Washington law, a juvenile can accumulate four diversion cases — no incarceration, no probation, no supervision — before any criminal charge applies. Hackney previously tried three; now he's going to two. "After your second time, if you're still running around with a gun, we've got a problem that needs to be addressed," he said.³ He pointed to one of the Seattle Center shooters, who had a prior possession charge that resulted in no intervention before the fatal shooting: "He went back home, picked up another gun, shot three people, and now we got three people dead." King County Superior Court's adult-transfer ruling on the 15-year-old suspect remains the next thing to watch.

SDOT is weighing a plan to reopen Pike Place to general car traffic Monday through Thursday from October through March, rolling back the pedestrian pilot for winter. An op-ed from Seattle Streets Alliance executive director Gordon Padelford argues against it, citing SDOT's own June survey of 5,403 people showing 98% support for the car-free configuration and 96% support among the 224 merchant respondents.⁴ Pike Place PDA data shows 130,000 more visitors and $12.4 million in commercial sales growth from 2024 to 2025 since the pilot launched; an Urban Freight Lab survey found only 14.5% of market visitors arrive by car. The pressure to scale back traces to a subset of driving customers who find the attached 800-stall garage inconvenient.

ON THE TRAIL — Weekend Picks, Aug. 22–23

A frontal boundary clips western Washington Saturday (40–65% rain chance across most zones) before clearing sharply Sunday (2–29%). No zone auto-disqualifies — no corridor is ≥50% on both days in the best pick areas — but Saturday will be wet; treat it as a low-angle travel day and make your miles Sunday. Wildfire smoke persists region-wide as a background condition: Aug. 17–18 reports note haze obscuring summit views in the Cascades, Mountain Loop, and Issaquah Alps.

Pick 1 — 1-night | Dewey Lake via PCT (Chinook Pass) Region: Mt Rainier Area — Chinook Pass / Hwy 410 | Drive: ~115 min from Issaquah Trip length: Mileage not provided in source report — plan for a short PCT approach from Tipsoo Lake / Chinook Pass TH. Aug. 17 report: "Very swim worthy. We had all sun no smoke no bugs. On a sunny Monday arrived at 1pm and no problem finding parking in a lot. Not a lot of people on the trail." Weather (Mt Rainier region): Sat high 61°F / low 46°F / 49% precip — bring full rain kit, borderline but not a disqualifier. Sun high 62°F / low 48°F / 2% precip — clear. Clears criteria: Late-August Chinook Pass area is snow-free on the PCT; lake is the water source; no bugs Aug. 17; not a permit-lottery zone; PCT approach has no river fords. → WTA trip report, Aug. 17

Pick 2 — 1–2 nights | Surprise and Glacier Lakes (Surprise Creek Trail, US 2) Region: Central Cascades — Stevens Pass West | Drive: ~75 min from Issaquah Trip length: Source reports ~10.5 mi with side trips (day hike), ~3,000 ft total gain. For a 1-night pack, plan approx. 5 mi / ~1,500 ft in Saturday; same out Sunday. Aug. 17 report: "Less than ten [people] all day… perfect clear turquoise water at Glacier Lake… tons of pikas everywhere… boardwalk over any muddy portions." Weather (US 2 West): Sat high 71°F / low 55°F / 48% precip. Sun high 71°F / low 54°F / 29% precip. Rain possible both days; neither exceeds 50%. Waterproof layers mandatory. Clears criteria: No snow at this elevation per Aug. 17 report; creek and lake water confirmed; very low crowds confirmed; boardwalk approach, no river fords reported; bugs not mentioned as an issue. → WTA trip report, Aug. 17

---

REGIONAL SNAPSHOT

• Issaquah Alps (Cougar/Tiger/Squak, 10–25 min): Trails dry and dusty. Aug. 18 report: territorial owl attacked a hiker at the Tiger Mountain High School Trail / Adventure Trail junction around 7pm — carry a stick on evening hikes.⁵ Squak Mountain (Margaret's Way / Debbie's View) quiet; smoke blocking the Rainier viewpoint.
• I-90 / Snoqualmie / North Bend (35–55 min): Snoqualmie Mountain is punishing — 3,000 ft in under 2 miles, loose rock on descent. Granite Mountain dry past the first mile; pack 3+ liters. Exit 38 off I-90 CLOSED as of Aug. 17 — use alternate access. Oxbow Loop in good condition with low crowds at midday.
• Stevens Pass West / US 2 (60–90 min): Surprise and Glacier Lakes are the weekend headline — near-empty, clear water, excellent conditions. Bridal Veil Falls flowing but low. Erinswood and Heybrook Ridge trails CLOSED through at least mid-September (Snohomish County construction).⁵
• Mountain Loop Hwy (90–120 min): Lake 22 busy even on weekdays — arrive before 9am or expect overflow parking. Mount Pilchuck: large flies every mile or so, reported as tolerable. Cadet Peak via Monte Cristo is accessible via e-bike (4.5 mi / 531 ft on the road), then 6.5 mi / 4,560 ft scramble to the summit.
• Mt Rainier / Chinook Pass (110–140 min): Dewey Lake in clean condition. Kautz Creek to Indian Henry's Hunting Ground excellent but a full commitment (14 mi, 4,200 ft). Bench/Snow Lakes parking nearly impossible by 5pm; early start required.
• North Cascades / Hwy 20 (150–190 min): Cascade Pass and Sahale Arm filling by 9am on weekdays. Saturday forecast: 65% precip with thunderstorms — avoid exposed ridges. Northern State Ghost Town is a low-commitment alternative.
• Olympic Peninsula / Hood Canal (180–240 min): Marmot Pass in excellent shape — "wilderness equivalent of a golf green," two campgrounds en route, dry above the last camp. Saturday Peninsula: 55% rain; if you go, drive out Friday night. Sunday clears to 15%.
• Eastern WA / Okanogan (150–160 min): Bonaparte fire lookout staffed and operational. Smoke heavy in the Tonasket/Omak area this week; a new valley fire started during the Aug. 17 lookout rotation.

↑ Back to top

THE LONG READ

A Proof Is Not a Proof: Terence Tao Builds the Infrastructure for the AI Math Era

The proliferation happened faster than anyone had planned for. AI systems began producing Lean code — formal, machine-checkable proofs — at a rate no human review process was designed to absorb. A repository appears on GitHub claiming to formalize some landmark result. The Lean file typechecks. But does it actually prove what the description says it proves? Those are two different questions, and for most mathematicians, answering the second one requires fluency in a language most of them don't have.

On Tuesday, Terence Tao announced a structural answer: Palomar, a registry of Lean-verified mathematics incubated by the Lean FRO and the International Center for AI and Research in Mathematics (ICARM). Tao — who won the Fields Medal in 2006 and has become one of the most visible figures in the AI-assisted mathematics movement — is serving on the registry's scientific advisory board alongside Jeremy Avigad (Carnegie Mellon), Ravi Vakil (Stanford), Akshay Venkatesh (IAS), Bryna Kra (Northwestern), and several others.¹

The analogy Tao reaches for is a preprint server, and it is apt. Palomar does not peer-review submissions for novelty or interest. It does something narrower and, for this particular moment, more urgently needed: it verifies that a repository actually delivers what it claims.

That verification runs in two layers. The first is purely mechanical. A tool called Comparator — built on the Lean prover itself — checks that the solution module actually typechecks and proves exactly the results specified in a separate "challenge file." No hand-waving, no clever axiom additions to smuggle in extra assumptions: the tool either passes the submission or it doesn't. The second check is non-deterministic and handled by a large language model: does the informal English description in the formalization.yaml file actually correspond, in semantic terms, to the formal Lean statement? This is where the mismatch problem lives — a Lean statement can be syntactically valid and semantically misleading, and no mechanical check catches that.¹

The three-artifact structure is worth understanding precisely. Every Palomar submission requires a challenge file (a short, human-readable Lean statement of the claimed results), a solution module (the full proof, arbitrarily long), and a formalization.yaml (metadata, informal description, relevant disclosures). The registry is a collection of GitHub repository snapshots — specific commits, not mutable repos — so what gets registered is a precise, reproducible artifact.¹

Tao ran himself through the process as a test. His formalization of the proof of Sendov's conjecture — a longstanding open problem in complex analysis — is now the first entry in the Palomar registry.¹ That he bothered to do this himself, rather than leaving it as infrastructure for others, is a signal worth noting. The bet implicit in Palomar is that formalized mathematics is about to become common enough that it needs the same basic institutional scaffolding that experimental science built around preprint servers and data repositories in the last two decades.

The timing is not accidental. The last eighteen months have produced a genuine acceleration in AI-generated Lean code, some of it impressive, some of it subtly wrong in ways that require expert reading to catch. The problem Palomar addresses is not that the proofs are bad — it is that the audience for evaluating them is too small and too busy. A two-layer automated check that any graduate student can interpret is a forcing function: it shifts the burden of basic verification from human experts to infrastructure.

Tao is explicit that Palomar is not a peer-reviewed journal. What it is, more precisely, is a trust layer. A submission that passes Palomar's checks carries a specific, bounded claim: the Lean typechecks cleanly, no hidden axioms were added, and a language model found the informal description consistent with the formal claim. That is not a proof that the mathematics is correct, interesting, or original. It is a proof that the file does what the label says — which, in an environment of automated proof generation at scale, turns out to be exactly the check that was missing.

The Palomar registry opened for submissions on Aug 18.¹ The submission documentation is detailed, and Tao notes that modern AI agents are useful for navigating the mechanical requirements — which has the slightly recursive quality that characterizes most of the AI-mathematics moment right now. Discussion is happening on the Lean Zulip.

↑ Back to top

FROM THE ARCHIVE

The Communists Were Never the Problem: August 19, 1953

The Shah had already fled the country "for medical reasons" when the CIA made its second attempt. The first had failed: Mohammad Mosaddegh got wind of the plot, called his supporters into the streets, and the Shah disappeared. British intelligence backed away from the mess. The Americans did not.

On August 19, 1953, the Iranian military — organized, cajoled, and in some cases bribed by CIA operatives — moved on Mosaddegh. Street protests, also CIA-financed, provided cover. The elected prime minister was arrested. The Shah returned within days. As a gesture of gratitude for American support, he signed over 40 percent of Iran's oil fields to U.S. companies.¹

The official rationale was communism. Mosaddegh had nationalized the British-owned Anglo-Iranian Oil Company in 1951, which was enough for British intelligence to conclude he harbored Soviet sympathies. The CIA accepted that framing. What Mosaddegh actually harbored was nationalism — a conviction that Iranian oil should benefit Iranians — which is a different thing entirely. He served three years in prison after the coup, then died under house arrest in 1967.¹ The Shah became one of America's most reliable Cold War clients, drawing U.S. economic and military aid for two and a half decades.

The revolution came anyway. Anti-Shah protests broke out in 1978; the Shah was toppled in 1979; Iranian militants seized the U.S. embassy and held its staff hostage until January 1981.¹ The threat the CIA had acted to prevent — a hostile Iran — arrived on schedule, produced not by communist infiltration but by the accumulated weight of a coup, a police state, and a population that remembered what had been taken from them. As a postmortem on the operation noted: nationalism, not communism, proved to be the most serious threat to U.S. power in Iran.

↑ Back to top

THE FUNNIES

Paused for Scoring Too Well / The New Objectives

*After XKCD — on OpenAI pausing Astra's training after the model aced the cybersecurity benchmarks it was supposed to fail. After Garfield — on Primoz Roglic's team director officially announcing that winning a stage is now the primary objective, GC is secondary, following the training-ride collision.*

↑ Back to top

ALSO NOTED

Also Noted

• Elisa Balsamo wins the Egmont Cycling Race — The former world champion sprinted off Lucinda Brand's wheel in the 130.2km Belgian race in Zottegem for her first victory since four Giro d'Italia Women stages in June, Lidl-Trek's first team win since the same. cyclingnews.com Aug 18, 2026
• Arnaud De Lie to Tudor on a four-year deal — The 24-year-old Belgian sprinter-classics specialist (35 pro wins, GP Québec, Belgian national title) leaves Lotto-Intermarché after six seasons for the Swiss ProTeam through 2030. cyclingnews.com Aug 17, 2026
• Frank Beard, ZZ Top drummer, dies at 77 — Beard — the sole clean-shaven member of the Texas trio, the backbeat behind "Sharp Dressed Man," "Legs," and "La Grange," and a member of the Rock & Roll Hall of Fame since 2004 — died Monday in hospice care in Richmond, Texas; the band canceled two tour dates and said it will resume in Austin this weekend. kiro7.com Aug 19, 2026
• Cerebras ships the CS-4 with a 2× overclocked WSE-3 Turbo — The new machine doubles the original wafer clock speed (to approximately 2.8 GHz) with the same 900,000 cores and 44 GB on-wafer SRAM, while the modular Nexus rack separates compute, power, and NICs into independently upgradeable backpacks and delivers three times the compute density of prior Cerebras systems; select customers get early access now, general availability later in Q3. nextplatform.com Aug 19, 2026
• Wired reconstructed Flock Safety's AI police tool from publicly served code — Flock's "Investigate" interface — which logs 20 billion plate scans a month and lets officers query vehicle locations, registered owners, and arrest records in natural language — served its full authenticated UI to unauthenticated visitors; officer search-justification fields require only three characters and carry no substantive review requirements. wired.com Aug 19, 2026
• A bookseller's AirTag traced a bulk order to an Amazon book-scanning facility — 404 Media (via Simon Willison) placed a tracker in one of roughly 1,000 books ordered through Biblio and watched it arrive at the VGT3 section of Amazon's LAS8 Las Vegas warehouse, where worker forums confirm books are destructively scanned for AI training data. simonwillison.net Aug 17, 2026
• Godot 4.8 dev 3 lands with 176 fixes from 91 contributors — Highlights include native admonition support in the documentation system, per-token GDScript error underlines replacing full-line highlighting, glob-based FileSystem search, and a new visionOS module that lays groundwork for future tvOS support. godotengine.org Aug 2026
• Eight countries now list the demoscene as intangible cultural heritage — Norway's April recognition brought the count to eight (Finland, Germany, Poland, Netherlands, Switzerland, Sweden, France, Norway), while Art of Coding organizers collect community consent for a UNESCO multinational nomination — a bureaucratic compression problem the Hyperlific essay frames as "the scene's newest size limit." hyperlific.com Aug 3, 2026
• Graphics Programming Weekly 451 covers adaptive ray reallocation and PlayStation Spectral Super Resolution — The issue also includes a walkthrough tracing a single global-load instruction through the full RTX 4090 memory hierarchy from warp registers to RAM, AMD's migration of its open-source MLP inference library from Cooperative Vectors to the D3D12 Linear Algebra Matrix API (Shader Model 6.10), and a procedural 2D grass renderer with a live browser demo and full shader source. jendrikillner.com Aug 16, 2026

↑ Back to top

THE QUESTION

Pause the System, or Build the Institution Around It?

The choice between pausing a system and building the infrastructure to evaluate it has no obvious right answer in the moment — only in retrospect. Today's paper carries one example of each, from domains that don't usually speak to each other.

As THE LAB reports today, OpenAI halted training on Astra after internal benchmarks showed the model had crossed a threshold for cybersecurity capability that existing containment infrastructure couldn't handle. The response was to stop: no more training runs until sandboxes are hardened, chain-of-thought monitoring is deployed, and human review can operate within 30 minutes of a concerning signal.¹ As THE LONG READ reports, Tao's response to AI-generated formal mathematics was the opposite — build a registry before the problem becomes unmanageable, because AI systems are producing Lean proofs faster than any human expert pool can absorb and verify. Palomar doesn't slow the AI down; it builds the evaluation layer that was missing.

Both choices are defensible. OpenAI's argument is that running an inadequately contained system is worse than the cost of delay. Tao's implicit argument is that the output won't slow down regardless, so the institution has to run to catch it. The question worth carrying today is whether "pause until safe" scales as a strategy — or whether, at sufficient velocity, the only move is to build faster than the capability is arriving.

↑ Back to top