THE WORLD

Trump Blockades Hormuz; Iranian Missiles Kill Crew Member on UAE Tanker

↩ Developing story — first reported Jul 08 · previously Jul 10, Jul 12, Jul 13

• Trump Orders Hormuz Blockade — US struck Iranian coastal defenses, missile and drone sites in a third consecutive night of operations Tuesday. CNBC
• Iran Hits UAE Tankers — Cruise missiles struck the Mombasa and Al Bahiyah inside Omani waters, killing one Indian crew member and injuring eight. CNBC
• Hormuz Crossings Down 52% — Kpler data shows week-on-week traffic collapse through July 12; Brent crude rose 2% to $85 on renewed uncertainty. CNBC
• Trump Floats 20% Hormuz Toll — While ordering the blockade, Trump proposed charging ships 20% to transit what carries a fifth of the world's oil. CNBC

ON THE ROAD — AURILLAC, FRANCE

• Bastille Day Across Cantal — Fireworks, the bal des pompiers, and celebrations mark La Fête nationale tonight in towns across the region. Connexion France
• TDF Stage 10 Departs Aurillac — Today's Bastille Day queen stage runs to Le Lioran; expect center-city road closures and crowds through midday. Aurillac Tourisme
• Heavy Traffic July 12–15 — Bison Futé has flagged this stretch for high national congestion; allow extra time heading south toward the Pyrenees tomorrow. Connexion France
• Les Fêtes de Bayonne Open Tomorrow — The five-day festival in the Pyrenean gateway runs July 15–19; book restaurants in advance, it fills up fast. Connexion France
• Weather: Hot and Dry Today, Storms Thursday — 95.7°F, overcast, 0% rain in Aurillac today; thunderstorms enter the Pyrenees Thursday (58%) and Friday (63%).

↑ Back to top

THE PELOTON

Bastille Day in the Cantal: Stage 10 Climbs to Le Lioran on the Hottest Tour in Memory

↩ Developing story — first reported Jul 10 · previously Jul 11, Jul 12, Jul 13

AURILLAC, FRANCE — Stage 10 left here at 13:15 under 31°C — seven classified climbs, 3,791 metres of ascent, 166.6 kilometres of Cantal roads — and on Bastille Day, France is waiting at the summit.¹

The finish at Le Lioran is one Jonas Vingegaard knows well. Two years ago he ran down Tadej Pogačar on the same 1,240-metre road in the Cantal and beat him in a sprint on a day the Dane appeared finished. He arrives today 2:42 behind that same man in yellow, the gap built almost entirely in the Pyrenees on Stage 6.² Visma-Lease a Bike is not conceding. "I think Jonas is a leader by example," Matteo Jorgenson told Cyclingnews from the rest day. "He really believes that he can beat him, and we all do."

The jersey situation entering Stage 10: Pogačar carries yellow, Mads Pedersen the points jersey, Isaac Del Toro the white, Vingegaard the polka dot. Paul Seixas — 19 years old, in his first Tour de France — lies sixth, 28 seconds off a podium position.³ His directeur sportif Julien Jurdie told Cyclism'Actu on the rest day that the opening nine stages had been "extremely positive," crediting Olav Kooij's Stage 5 sprint in Pau with releasing pressure across the whole squad. "Winning a stage in the first week takes a huge amount of pressure off," Jurdie said. "Not many teams have won a stage so far, so it's a real relief."³ It is the first time in Seixas's career that he has ridden nine consecutive days; everything from Stage 11 onward is new territory.

The heat has accumulated into a formal dispute. The CPA negotiated a 2% time limit extension with ASO over the weekend — after Stage 9 was already shortened by 30km — then issued a statement calling for off-season talks with all stakeholders, saying it "urgently calls for discussions" before the 2027 season.⁴ Pogačar, at the Stage 9 finish in Ussel, went further: "If I could have the power to change it all, I would change all the calendar. I would not race in July and August in the hot places." Average temperatures at this Tour are running nine degrees Celsius above last year's race, according to VeloNews; riders have been losing up to 10 pounds of body weight per stage through dehydration.⁴ Jorgenson put the cumulative toll plainly: "I can't immediately remember a tougher edition. This is certainly the hottest."²

The riders assigned to Le Lioran for the rest day found the conditions unsupported. Uno-X Mobility and Alpecin-Premier Tech shared an apartment hotel that Magnus Cort subsequently reviewed on Instagram, rating it one out of seven stars: "only a small smell of rot... dirty, no aircondition and no WIFI."⁵ Brothers Anders and Tobias Halland Johannessen posted photos of their solution — beds moved onto the balcony, sleeping outside rather than in their rooms. Alpecin's Silvan Dillier posted a room-tour reel and did the same; his roommate Tim Marsman stayed inside with a portable AC unit. UAE Team Emirates-XRG and Lidl-Trek also reportedly had no air conditioning at a separate hotel. ASO books and assigns all Tour accommodation; teams are not permitted to pay for alternatives, in the name of competitive fairness.

The team that races next year as Soudal Safety Jogger will still have Quick-Step on its jersey. Soudal's announcement of the Safety Jogger co-title partnership — confirming the new name and the Belgian PPE manufacturer's elevation from partner since 2022 to co-title sponsor from 2027 — included a detail that changes the read on the transition: the Quick-Step flooring brand has committed to supporting the team through 2030, with its name remaining prominent on the race kit.⁶ Quick-Step's 24-year run as title sponsor ends; its presence on the team does not.

FDJ United-SUEZ signed Kate Courtney to a 30-month contract effective immediately and running through the end of 2028.⁷ Courtney, 30, is the reigning MTB cross-country marathon World Champion and the current US Pro Road Champion — she outsprinted Lauren Stephens at road nationals. General manager Stephen Delcourt cited her "great potential on short to mid-range climbs and demanding efforts."⁷ She makes her Women's WorldTour debut at La Périgord Ladies on July 18, and is confirmed for Tour of Britain Women, August 19-23.

↑ Back to top

THE LAB

What Grok's Coding Agent Was Sending Home

A security researcher at Cereblab ran xAI's Grok Build CLI through a proxy and captured every byte it sent home. The findings are specific: version 0.2.93 was uploading entire repository snapshots to a GCS bucket called grok-code-session-traces, independent of which files the agent actually opened. A 12 GB test repository generated 5.10 GiB of uploads across 82 storage requests. The model-processing channel consumed 192 KB; the storage channel consumed 5.10 GiB — a ratio of roughly 27,800 to one.¹ The researcher proved the upload was complete by cloning git bundles recovered from captured traffic and recovering files the agent never touched during the session.

The config option disable_codebase_upload=true exists in the [harness] section. It does not work the way the name implies: even with the "Improve the model" toggle disabled in the Grok UI, the server returned trace_upload_enabled: true.¹ The relationship between client-side settings and server-side behavior is, per the analysis, undocumented. The wire-level research was published as a GitHub Gist and discussed on Hacker News; some commenters proposed a benign interpretation — uploading the full codebase lets the model inspect it without client round-trips during "thinking." Others noted this is a weak justification given the privacy implications, particularly for repositories containing .env files with production credentials.

The broader takeaway the researcher draws is accurate and not specific to Grok: choosing a coding agent harness is also choosing what leaves your machine and under what terms. The gap between "context for this task" and "persistent data collection" is frequently unspecified in documentation, and the only reliable verification method is running the tool through a proxy yourself.

On the more entertaining end of the week: Peter Gostev built DOOMQL, a Doom-like game in which SQLite is the actual game engine — not just a data store, but the engine for movement, collision detection, enemies, combat, progression, and every RGB pixel rendered to the terminal. The ray tracer is a single large recursive CTE. Gostev built the project using GPT-5.6 Sol.² Simon Willison picked it up on July 13, cloned the repo, and used Datasette Apps to mirror the live SQLite game state into a browser view while playing in the terminal. A single prompt to Claude Fable 5 — "Build an app that displays the current state of the screen using the frame_pixels view with its x, y, r, g, b columns. have it refresh once a second." — produced a working live display; adding "add a minimap" completed it. The full render SQL and the Datasette App source are both linked from Willison's post.

Chips and Cheese published a detailed analysis of ACE, the x86 Ecosystem Advisory Group's forthcoming AMX accelerator that switches from inner products to outer products — the same architectural choice Arm made with SME2. Where AMX's original TMUL accelerator takes all operands from tile registers, ACE takes A and B inputs from AVX-512 vector registers and accumulates into tile registers. That frees the full tile register file for output accumulation, enabling a 32×64 element output tile and avoiding the register pressure that forces TMUL code to spill inputs through L1D.

The cache bandwidth numbers are the most consequential part of the analysis. For a 32K×32K INT8 matrix multiply at optimal tiling, AMX TMUL requires roughly 1.7 TB of cache traffic; ACE drops that to approximately 1.5 TB.³ A 512-bit SME implementation sits at around 2.0 TB, while a 1024-bit SME implementation reaches 1.0 TB. Per MAC operation, ACE comes in at roughly 0.046 bytes versus AVX-512-VNNI's 0.10 bytes, meaning an L1D capable of 128 bytes per cycle could sustain approximately 2,785 MAC operations per cycle with ACE versus 1,228 with VNNI. ACE also adds FP8 support via a 1024-bit Block Scale Register (BSR0) implementing OCP Microscaling Format scaling, and a new VUNPACKB instruction that, combined with AVX-512 permute instructions, handles quantized weight dequantization for any 2–7 bit format in three instructions.

No ACE hardware exists yet. Intel and AMD may be targeting next-generation parts, but as Chips and Cheese notes, ISA design and hardware implementation are different questions — how those implementations land will matter more than the spec. One interesting detail: the current AMX VEX encoding supports up to 16 tile register slots; AMX only used 8. If Intel expands to the full 16 slots in an ACE implementation, the cache efficiency would reach parity with a hypothetical 1024-bit SME implementation.

Trending today: GitHub is saturated with AI agent wrappers, CLAUDE.md skills collections, and coding-assistant proxy forks — no technical novelty cleared the bar.

↑ Back to top

THE LONG READ

Eight Kilobytes Per Frame: How Game Arts Made the Sega CD Do the Impossible

The Sega CD's single-speed CD-ROM drive delivered 150 KiB per second. Its main processor ran at 12.5 MHz. A full-screen image at 15 frames per second implied something like 600 KiB per second just for raw pixel data — four times what the drive could physically deliver. Game Arts shipped Silpheed anyway, with near-fullscreen cutscenes that made the press question whether they were looking at pre-rendered video or real-time 3D. They were looking at pre-rendered video. What they couldn't see was what it took to make that video exist: a compression system that squeezed every frame into 8 KiB, built entirely from constraints Game Arts refused to compromise on.¹

Fabien Sanglard — the engineer and writer known for dissecting the source code of Quake and Doom — spent two weeks reverse-engineering Silpheed's FMV format in his spare time.¹ His writeup is the kind of systems archaeology that makes you realize how much engineering ingenuity was buried inside consumer hardware that most people treated as a novelty.

The core problem: the Sega CD had no framebuffer. Rendering went through a tile system — 768 tiles of 8×8 pixels, arranged on a 32×28 grid to form a 256×224 image. This looked like a liability. Sanglard's analysis reveals it was actually the foundation of the entire compression scheme.¹

Most developers making FMV games at the time tried to cram real video into the constraints — small windows, compression artifacts, dithering that made images look like they had been printed on wet newsprint. Game Arts started from the other direction. They chose flat-shaded polygons, 16 colors total, no Gouraud shading, and minimal dithering. Choices that looked like concessions but were actually load-bearing decisions that made everything else possible.

The first compression layer came from how tiles repeat. In the example frame Sanglard analyzes in detail, 456 of 896 tiles are single-color fills. Rather than store each one individually, Silpheed's encoder issues the 15 plain-color tiles once per frame and references them over and over in the tilemap.¹ Half the bandwidth problem, gone.

The second layer is where the engineering gets genuinely surprising. The Mega-CD ASIC included a feature called "Font bit" — intended to render two-color kanji characters quickly. You write two palette indices and a 16-bit bitmap to three registers; the hardware expands the 1-bit pattern into 4-bit pixel data representing two rows of an 8×8 tile. A font feature. Silpheed's engineers repurposed it as a tile decompressor. For the tiles made of exactly two colors — which turned out to be most of the remaining tiles — they used the font registers instead of storing raw nibble data. Three bytes in, eight bytes out. A 37% bandwidth reduction on the tiles that survived the first pass.¹

After those two optimizations, some tiles still needed to be stored raw. But even the tilemap itself got compressed: rather than storing all 768 tile indices explicitly, Silpheed stored a 768-bit bitmap where a 0 means "auto-increment from the previous index" and a 1 means "read an explicit value." In practice this cut tilemap overhead by around 30%.¹

The fractal levels — two stages with complex textures the developers called "fractals" — defeated every trick. No solid tiles, no two-color tiles, nothing that compressed well. The developers' only option was to drop to 7.5 frames per second and store everything raw. You can hear the concession in how Sanglard describes it: the fractals just didn't yield.

One more layer: palette cycling. The laser fire and explosions in gameplay appear multi-colored, which seems impossible given the 16-color constraint. During gameplay sequences, Game Arts reserved four palette slots at the tail of the palette and rotated them every frame via the main CPU. The result is a shimmer effect across every pixel using those palette indices. It also meant artists designing gameplay sequences had only 12 usable colors — the other four were always cycling. Cutscenes got the full 16 because there was no laser to render.¹

The writeup includes an ASCII diagram of the full Genesis/Mega-CD dual-CPU architecture, per-frame tile breakdowns with grids overlaid, and a complete storyboard Sanglard generated with tools he designed during the reverse-engineering process. It rewards careful reading. What it amounts to is a case study in constraint-driven engineering: not the story of people finding workarounds for bad hardware, but of people who understood the hardware well enough to turn its oddities into a compression pipeline. The tilemap that looked like a limitation was the mechanism. The font register that was meant for kanji was the decompressor. The 16-color palette that seemed like a creative handicap was a performance budget.

The piece surfaced on Hacker News on July 13. It's been there before in various forms — Sanglard's site has a long track record of this kind of work — but this one is worth your time.

↑ Back to top

FROM THE ARCHIVE

The Dream of Mars: July 14, 1965

The signal left Mariner 4 at 10 watts. By the time it reached the 85-foot antenna dishes in South Africa, Australia, and Goldstone, California — having traveled roughly 191 million miles — it had dwindled to one billionth of one billionth of a watt.¹ Barely perceptible, NASA's engineers noted. What it carried was the end of a very old dream.

On the evening of July 14, 1965, Mariner 4 made its closest approach to Mars: 6,118 miles from the surface, at 8:00 p.m. Eastern time.¹ The TV camera had started shooting roughly 40 minutes before closest approach, capturing 21 images through red and green filters.¹ The following day, those images began the long transmission home. They showed a heavily cratered, ancient surface — barren, scarred, unmistakably dead — with daytime temperatures reaching minus 148 degrees Fahrenheit.¹ A faint radiation belt, barely one-tenth of one percent of Earth's, was the planet's only active signature.

That was it. Decades of speculation about Martian canals, civilizations, anything living — dispatched by 21 photographs in a single pass.

Mariner 4 was NASA's second attempt. Its twin, Mariner 3, had launched three weeks earlier and failed when its payload shroud refused to jettison; the solar panels never deployed and the spacecraft drifted off into solar orbit, inert. Engineers built a new all-metal shroud for Mariner 4. It worked. The spacecraft launched November 28, 1964, made one mid-course correction in December, and flew a seven-month arc to its appointment with Mars.

The occultation data from the flyby told another story equally as deflating: the Martian atmosphere was so thin that any future lander would need retro-rocket engines, not just parachutes, to slow down in time. There was no cushion to work with. The place was essentially naked to space.

NASA maintained contact with Mariner 4 until October 1, 1965.¹ The spacecraft was reactivated two years later for attitude control tests and kept talking to Earth until December 31, 1967 — more than three years after launch. By then the conversation had long since moved on. The cratered desert was the baseline now. Everything that followed — the orbiters, landers, rovers — was built on what those 21 photographs confirmed: Mars is a harder problem than anyone had hoped, and the fantasy was finally over.

↑ Back to top

THE FUNNIES

Hot Bikes, Leaky Agents

*After XKCD — on the gap between a coding agent's "disable upload" setting and what actually crosses the wire. After Garfield — on what pro cyclists' hotel accommodations look like during the hottest Tour on record.*

↑ Back to top

ALSO NOTED

Also Noted

• LAPD Drops Flock Safety Contract Over Privacy Concerns — The LAPD let its three-year license-plate camera contract with Flock Safety expire Saturday, citing "serious concerns around civil liberties and civil rights issues, particularly around privacy," making the nation's third-largest police department one of the biggest agencies to exit a network that comprises at least 80,000 cameras across the US. techcrunch.com Jul 13, 2026
• Bangkok Bar Fire Kills 27 — A fire at a live-entertainment venue in Bangkok's Chatuchak district killed 27 people and injured 63 others Sunday night; investigators initially focused on an electrical fault near the stage or ceiling as the likely cause, with the precise origin still under investigation — and found objects obstructing both emergency exits, including beer crates near the kitchen exit and a table used to sell sweets blocking a second route. khaosodenglish.com Jul 13, 2026
• Irish Data Centers Now Consume 23% of the Country's Electricity — Ireland's Central Statistics Office reported that server farms consumed 7,663 GWh in 2025 — up 10% year-on-year, exceeding urban household consumption, and more than four times their 2015 share — even as a Dublin-area moratorium on new grid connections held through nearly all of the year. theregister.com Jul 11, 2026
• Claude Code Uses 4.7× More Tokens Than OpenCode on the Same Task — A wire-level study by Systima AI found Claude Code sends roughly 33,000 tokens before a single user prompt arrives (versus OpenCode's ~7,000), re-writes its cache prefix mid-session at up to 54× the rate, and on a fanned-out two-subagent task consumed 513,000 tokens against 121,000 for the identical work done directly — with the token gap persisting across both Sonnet and Fable model families. systima.ai Jul 12, 2026
• DOGE Used AI at HUD to Flag Regulations for Rescission, Then Withheld the Prompts Under Novel Privilege Claims — Documents obtained by Democracy Forward via FOIA show that DOGE staff at the Department of Housing and Urban Development used AI tools to identify agency rules for potential rescission; more than 100 requested documents were withheld, including one titled "GPT defined Econ Analysis approach," under claims including a novel "deliberative AI input" privilege that legal experts say has no established basis. wired.com Jul 14, 2026
• Seahawks Sale Seen as "Ripe" for a Public Stadium Funding Ask — Sports business analysts say Vinod Khosla's $9.6B purchase sets up conditions for seeking public money to upgrade or replace Lumen Field, as the team's five-season remaining lease parallels the Denver Broncos' stadium-replacement timeline, though experts consider city relocation unlikely given the absence of viable NFL expansion markets. kiro7.com Jul 14, 2026
• Cloudflare Precursor Analyzes Mouse-Physics Across Full Sessions to Catch Bots — Cloudflare's new Precursor system injects lightweight obfuscated JavaScript to capture wrist-pivot arc geometry, physiological hand-tremor frequency, and cognitive-click delay throughout an entire visitor session rather than at a single checkpoint, making behavioral fingerprints session-scoped so they cannot be reset on page refresh — it feeds directly into bot score and runs free until GA later this year. blog.cloudflare.com Jul 13, 2026

↑ Back to top

THE QUESTION

The Spec Is Not the Implementation

The documentation said the upload was disabled. The wire said otherwise.

As THE LAB reports today, a security researcher ran a popular AI coding agent through a proxy and found that a client-side toggle to disable uploads had no apparent effect on server-side behavior: 5 GiB of repository contents left the machine regardless, including files the agent never opened during the session.¹ Today's LONG READ traces a structurally identical gap across three decades: Fabien Sanglard's reverse engineering of a 1993 Sega CD game found that its entire compression pipeline — the mechanism that made near-fullscreen video possible on deeply constrained hardware — was built on registers documented for kanji font rendering. Game Arts knew what it was doing. Nobody else did until Sanglard spent two weeks looking.²

The audit mechanism in both cases is the same: someone willing to check what the system actually does, not what the label says. A proxy session. Two weeks of reverse engineering. Neither method scales to routine practice. For a 1993 video game, the gap between documentation and implementation is charming retrospective archaeology — engineering ingenuity buried in hardware most people long since discarded. For a coding agent with access to a production codebase that may contain live credentials, the same kind of gap is a different order of problem. The question today's edition raises isn't whether any particular tool is trustworthy. It's whether documentation can still function as a contract when systems are this opaque — and if it can't, what fills that role.

↑ Back to top