Weeknotes 2026 W09: Optimized away

I’m still not quite fully back to health. Low-key misery. At least Berlin is sunny and reasonably warm now.

Quick bits

• I don’t think I’ll ever get used to the random explosions in Berlin. I’ve still got no idea what these are.

• A belated happy Twin Peaks day! Diane, 11:30 AM, February 24th. Entering the town of Twin Peaks.

• Yesterday had the clearest moon ring I’ve ever seen.

• I’m getting quite comfortable with Colemak. I still occasionally fall back to QWERTY by accident. Old habits die hard, I suppose.

• My phone corrected “pipes bursting” to “puppies bursting.” Eww. iOS autocorrect sure has been getting worse.

Nanoc’s getting faster

It took me a while — quite some thinking and refactoring was needed — but I’ve finally landed a changeset to switch to a push strategy in the outdatedness checker, which brings a nice speedup: my own web site compiles about 25% faster. The algorithmic complexity of the dependency checker is now O(#changes) rather than O(#items). The newly optimized code is so fast that it doesn’t show up in my CPU profile anymore — it’s been optimized away!

There are also some other improvements like using a bitset rather than a Set, which I can’t believe I didn’t think of earlier.

There is more work necessary before a release, though. There is a specific performance regression that I’ll need to fix first, but it shouldn’t take too much effort.

I would love for the compiler’s algorithmic complexity to be O(#changes) overall, rather than O(#items). That’s far from easy — unless Nanoc would start relying of a stream of filesystem change notifications. But that would be a pain to work with, and on top of the complexity of filesystem change notifications, there is now the challenge of managing continuously evolving in-memory state.

There are more optimizations still pending. Roughly 30%–40% of the compilation time is spent loading and storing temporary data needed for incremental compilation. Most of that time is spent in zlib and Marshal.dump and Marshal.load. It would help to only read/write what’s necessary, but that could also backfire: I don’t want to kill performance by turning a single slow read into a thousand slightly-less-slow reads.

AI for codegen: still terrible

Every time I talk about how bad AI code generation is, people will chime in to tell me that I simply haven’t used in in the last few months, and that AI has gotten so much better in the last 4–8 weeks. So, once again I am indulging people, and I gave Claude Opus 4.6 a try. Alas:

• It still often fails to generate code that compiles.

• When asking to fix a failing build, it’ll occasionally claim that there are no failures and thus not do anything.

• It will, for no reason, rewrite existing, perfectly valid code.

• It will, for no reason, delete existing, perfectly valid code.

• When I ask it to debug build failures, it will come up with explanations for build errors that don’t actually occur — and the explanations are wrong, too.

• When asked to parse JSON, it will write a broken JSON parser from scratch, instead of using an already existing JSON parser.

• It occasionally will start outputting the raw Claude protocol interaction (<tool_call>) and outputting code where the newlines are replaced with backslash-n sequences.

• It will hallucinate files that don’t exist in my project.

It’s like babysitting a delusional nightmare. I am non-stop facepalming.

AI code generation is so slow, and waiting for code output is infuriating when the output is just garbage 90% of the time. I could have been writing the proper thing by hand; it would be faster and way more correct.

This is especially true in “agentic” mode, where it confidently goes off on a long and winding trail that is utter nonsense, generating thousands of lines of garbage that I have no choice but to throw away.

Again, this is with Claude Opus 4.6, which at the time of writing is — supposedly — the cream of the crop¹¹1 More like the cream of the crap, am I right?!  for code generation.

I’ve managed to get it to work properly for simple changes akin to global search and replace, but it’s never 100% correct, and I could’ve used a regex instead for a faster and better result.

I’m baffled. Not because I expected it to be great (it’s a token predictor after all — a stochastic parrot), but because I fail to understand why people believe this technology is revolutionary.

Entertainment

• After Apple gave me (for no reason that I can think of) three months of free Apple TV+, I’ve been binging Slow Horses.²²2 Slow Horses (See-Saw Films, Apple TV+, 2022).  It’s full of tropes and doesn’t tread any new ground.

Toots and skeets

Before you ask a diffusion model to barf up a banner image for your article, please consider searching one of the many existing sources of fine & free media:

• Public Domain Image Archive
• Openverse
• Wellcome Collection
• Digital Bodleian
• Getty Open Content Program
• Wikimedia Commons

 — Pavel A. Samsonov

Links

• Times New Resistance: Delightful.

• Scientists crack the case of "screeching" Scotch tape (Ars Technica): Scotch Tape screams because it is painfully being torn apart at supersonic speeds. Got it.

• H-Bomb: A Frank Lloyd Wright Typographic Mystery (Paul Lukas)

• If This Dam Fails, It Pollutes Half of Europe. (What On Earth Is This?): What the actual fuck.

Tech links:

• Using git while trans (Brooke Hart): Might be useful to some!

• I Verified My LinkedIn Identity. Here's What I Actually Handed Over. (rogi): Yikes.

• On Alliances (tante)

• What Are Capabilities? (Habitat Chronicles): I’m fond of capabilities and this is a good overview.

• Tests Are The New Moat (Daniel Saewitz): This is a technique that I could be interested in using!

• The Death of the Software Craftsman (David Bryant Copeland)

• Model Collapse Ends AI Hype