Skip to content
self.md

■ SIGNALS // RADAR SIGNAL

the AI debugger that wanted to quit

Torvalds pushed past an AI debugger’s false impossibility; HDR JPEGs add an invisible second layer; hard systems become easier to enter but not safer to misjudge.

Published Updated Read 4 min
■ [!] ON THIS PAGE ▼

self.md radar — 2026-08-23

A Linux bug hid in the last 2 KiB of a page. A logo can now carry a second image that only HDR screens see. And a few people are using agents to touch the kind of code they used to leave behind glass. The common problem is not whether the machine can move fast; it is whether somebody can still tell it when the evidence says it is wrong.

1. Linus Torvalds used an AI debugger that wanted to quit

A Linux kernel fix changes a round_up() to round_down(). That tiny change stops an Intel Xe driver from handing compression-hardware storage to the VRAM allocator as usable memory. On Torvalds’s Battlemage test machine, the mistake exposed the last 2 KiB of a page; a Mesa VM kept landing page tables there, and the compositor would restart into a black screen.

The route to that line was not tiny. Torvalds says the investigation took 24 debug patches and 18 kernel boots. An AI did much of the instrumentation and analysis, but repeatedly called the bug impossible and suggested writing a report instead. He kept pushing it. The agent kept doing the work.

That is a decent description of the useful division of labour. The model can shovel logs and propose paths through a mess. It does not get to decide when the anomaly has been explained away. The person who owns the machine needs enough stubbornness, and enough access to the traces, to keep going after the assistant has become politely certain.

reading: the kernel commit · Torvalds’s note, collected by Simon Willison

2. HDR images give a logo a private second layer

So Very Bright turns an ordinary logo into an HDR JPEG. On a compatible display, selected light pixels can appear up to 7.5 times brighter than SDR white; on most screens, the same file remains a normal JPEG. It uses an ISO 21496-1 gain map for the web and a BT.2100 PQ version for LinkedIn, which currently preserves it.

This is not a revolution in branding. It is a small, strange reminder that a file is no longer one thing for every reader. The base image stays intact, while a small grayscale companion tells capable software how much farther a pixel may go. Dark colours do not magically glow; the site is unusually clear about that, along with the short list of browsers and apps that honour the extra data.

The more interesting consequence is visual rather than technical. A public object can now carry a detail for people with the right screen without becoming broken or exclusionary for everybody else. Most interface work will use that power badly for about six months. Still: it is a real new material for people who care how a thing feels, not only whether its upload completed.

reading: the HDR JPEG tool and technical notes · the HDR gain-map image format

3. agents are making difficult code less socially expensive

In a new essay , Armin Ronacher notices a shift beneath the louder claim that programming is “solved.” Agents make it easier to pick a language the programmer does not already know, to rewrite code into another one, and to work on performance problems that used to demand a narrow apprenticeship before anyone dared touch them.

His examples are concrete: Cloudflare’s Artifacts service uses a pure-Zig Git-protocol engine compiled to roughly 100 KB of WebAssembly; Vercel’s fx is a small, fast Zig coding agent. Ronacher’s point is not that expertise evaporates. It is that Rust, Zig, DWARF, eBPF, custom drivers and old hardware are becoming available to more people as targets for assisted investigation.

That is good news with an asterisk the size of a server rack. Lowering the social cost of entering a difficult domain does not lower the cost of being wrong there. If agents make the hard parts reachable, then test rigs, source trails and people who can read a failure have to become reachable too. Otherwise we have only made the blast radius more inclusive.

reading: Ronacher’s essay · Cloudflare’s Artifacts documentation

more to read

  • More than just code review — a compact argument for verifying an agent’s work through more than line-by-line inspection.
  • Stop Making TUIs — a useful provocation about native personal software now that interface code is cheaper to produce.
  • LLM 0.33 — a small release with a very ordinary lesson: transitive dependencies still choose when your tool stops installing.
  • Personal AI OS tools — the control-plane map for personal agents, receipts, memory, and tools
  • AI coding assistants — compare coding workbenches by review surface, permissions, cost, logs, and escape hatches
  • Best MCP servers — connect files, browsers, memory, search, and workflow tools without turning the stack into soup
ai-assisted-programming software-reliability interface-design