dllu

Cool idea! I had lots of fun looking at stuff nearby.

Right now it seems that it simply removes links for stuff without a geotag that's nearby. But many articles are in fact nearby and are simply missing geotags, or perhaps encompass a slightly larger area where only a part of it is close by. Also, all the pictures don't have working links so I can't click on a picture to make it bigger.

Lack of anti cheat is rather outdated perception of Windows on Arm. Most popular competitive games like League of Legends have native arm versions now.

I recently had the displeasure of using Passkeys:

- trying to sign in to Okta on my personal Linux desktop computer. It has bluetooth and I had my phone right next to me, also with bluetooth turned on. But Firefox simply refuses to even show the option for "connect to your phone" or to give a QR code... it just said "tap your security key or biometric device". I had neither; my passkey is on my phone. Then it timed out and I gave up trying to log in.

- trying to sign in on my work Macbook using Google Chrome (enterprise managed one). For some reason, upon scanning the QR code, my phone was stuck on "Connecting..." and timed out. I had bluetooth turned on on both devices, they were five centimeters apart, both were connected to internet with a great gigabit connection. It turns out, I had to quit and restart Chrome and it magically started working.

- trying to sign in on my work desktop using a physical Yubikey remotely using VNC. It's not possible by design. I had to get my coworker to press my Yubikey in the office.

- trying to get Muse and OpenAI Dot to sign in using a passkey. It's not possible by design.

Generally, the user experience is really horrible with passkeys. It is also ironic that it trusts my phone (used for shitposting; easily lost or stolen) instantly, but simply refuses to trust my personal Linux desktop computer, which is fully encrypted, securely stored in my house, and contains all my life's work and everything that is dear to me.

I wonder if they were already training AI on decompiled games. It's plausible that the large labs are already using AI to decompile existing software way before random people on the internet found out this was possible. That would also explain why some AI are oddly good at generating video games.

My complaint applies specifically to the CLI. I have never used the gui app. Here are the various goal-breaking bugs I've run into:

- Goals get blocked due to .git and other stuff becoming randomly read-only when daemon auto-updates in the background (the linked issue in my comment above)

- Goals get blocked due to transient "Selected model is at capacity"

- Goals get blocked due to "Codex ran out of room in the model's context window". Somehow, both auto-compaction and manually running /compact don't work. https://github.com/openai/codex/issues/52567#issuecomment-61...

- Goals get blocked due to bwrap randomly failing with `bwrap: loopback: Failed RTM_NEWADDR: Operation not permitted`. I traced this one to `apparmor` on Ubuntu 24.04 interfering with bwrap and found a local workaround.

The daemon has many bugs. For example, when auto-updating, it would randomly ruin the permissions and block the goal as .git became read-only. Generally, it is extremely annoying how many bugs Codex has. I basically can't trust it to execute on any goal since it would often get blocked every few hours due to random bugs.

https://github.com/openai/codex/issues/52251#issuecomment-60...

new one is actually smaller than the original (3760 mm long vs 3830 mm). but batteries and modern safety makes it heavier

He has a writeup about the office and desk setup too: https://arslan.io/2025/11/18/my-two-part-desk-setup/

Chess is an interesting case. I remember in 2023, GPT 3.5 or something used to be surprisingly good at chess. There was even a "stochastic parrot chess" website [1]. I recall it was playing decently at around a 1800 level. Even as a fairly okay player myself (2100 bullet on lichess), I struggled to beat it. However, modern LLMs are a lot worse at chess. I guess having too much chess data in the training set probably regressed performance on stuff that actually matters, like coding.

[1] parrotchess.com, no longer available. Previous discussions: https://hn.algolia.com/?q=parrotchess.com

Yeah but on the US government website [1] it says

> Submit the original, unchanged photo.

> Do not change your photo using computer software, phone apps or filters, or artificial intelligence.

and many other countries have similar requirements.

EDIT: I just read the "Common Questions" that says: "This editor crops and centers the photo you provide and can replace the background. It does not invent or beautify a face." I guess that's not as bad then. Still mildly iffy and weird marketing though.

[1] https://travel.state.gov/en/passports/apply/help/photos.html

Also, I don't have a monopoly on "pictures that look like this".

Adam Magyar and Jay Mark Johnson have line scan photos in fine arts as well, and they were in fact a big source of inspiration when I started out on my own line scan journey in 2017.

- https://www.magyaradam.com/wp/?page_id=368

- https://www.jaymarkjohnson.com/works

Seth Itow, Philomena, and Elm-chan both also have hobbyist line scan camera setups:

- https://www.sethitow.com/posts/2026-linescan/

- https://philo.gay/linecam/

- https://elm-chan.org/works/lcam/report.html (be sure to scroll down and click on the train pics)

There are also Japanese rail fans:

- https://kr64.seesaa.net/ with a bewilderingly huge collection of trains all over Japan,

And faux-line scan cameras using regular video cameras, stitching together thin chunks rather than 1 px columns:

- https://railfile.jp/

- https://x.com/yamakox/status/2089934720759144653

- https://kakeyama.image.coocan.jp/index.htm

- https://trains.jo-m.ch/#/trains/list

Line scan cameras are also commonly used for machine vision in the rail industry to inspect trains for damage and detect graffiti and stuff. And conversely, line scan cameras mounted on trains are used to inspect rails and tunnels.

I actually reached out to Philo on GitHub Issues but she hasn't responded yet [1]. I also mentioned her cool camera on my blog [2].

By the way, I recognize your username as the person that Philo credits with "rewriting grindstone in idiomatic NumPy". That's super cool!

[1] https://github.com/FlamingSpork/grindstone/issues/1

[2] https://daniel.lawrence.lu/blog/2026-06-12-standalone-cnc-li...

I just set up a shop at https://shop.lawrence.lu/. I also emailed Rick Laubscher and Alison Cant at streetcar.org [1] yesterday but they haven't gotten back to me yet.

[1] https://www.streetcar.org/about-sfmsr/

If you want to use a regular phone or camera, check out TrainScanner!

https://github.com/vitroid/TrainScanner

It converts a regular video into a line scan style photo. It works by stitching together small chunks rather than 1px thin slices, so the background has a repetitive texture rather than being smooth horizontal stripes. But at lower resolutions, it looks similar, especially if you pick a spot with a clean background. Also the trains turn out very detailed.

Example work by Kakeyama from Japan using this: https://kakeyama.fan.coocan.jp/tec/e5e8r.htm

If people are moving, they may get stretched or distorted. But they won't ever get blurry, since the capture rate is like 10,000 lines per second.

Muni heritage weekend was shot on my new Voigtlander Apo Ultron 90mm f/2. The line scan camera's sensor is 28mm long so any APS-C lens will work, not necessarily full frame. I modded my line scan camera from Nikon F mount originally to Sony E mount so that I can use more compact high quality lenses like M mount ones.

Previously I've also used these full frame lenses:

- Myutron 50mm f/2.6 (Shinkansen N700A, SF Cable Car 2017)

- Voigtlander Nokton 58mm f/1.4 (Brisbane Queensland EMU)

- Samyang 135mm f/2 (Madrid AVE Class 102 and Shanghai maglev)

- Voigtlander Apo Lanthar 125mm f2.5 (Yamanote Line, CR400AF)

- Voigtlander Apo Skopar 90mm f/2.8 (London Tube photos)

- Voigtlander Ultron 35mm f/1.7 (Hong Kong trams)

and APS-C lenses:

- Cheecar 50mm f/1.2 (Osaka trains)

- Samyang 85mm f/1.8 CS (Caltrain)

I took one of my baby on the swing. It needed custom polar coordinate dewarping:

https://i.dllu.net/swing_07_7668f7f16ff7ac2a.png

Also motorcycle cops:

https://commons.wikimedia.org/wiki/File:San_Francisco_Police...

OP here, ask me anything! I've also written about the process previously:

- Line scan camera image processing: https://daniel.lawrence.lu/blog/2025-09-21-line-scan-camera-... Previous discussions:

  - https://news.ycombinator.com/item?id=44996938

  - https://news.ycombinator.com/item?id=46383593

  - https://news.ycombinator.com/item?id=46388084

- Standalone line scan camera hardware: https://daniel.lawrence.lu/blog/2026-06-12-standalone-cnc-li...

- Line scan camera image processing 2: https://daniel.lawrence.lu/blog/2026-08-02-line-scan-camera-...

Rhino is unusually suitable for AI agentic generation since it is very command line driven and there are a lot of example RhinoScript projects. McNeel, the Rhino developers, also have a lot of open source stuff, Rhino MCP, and explicit support for headless Rhino [1], which makes it even more amenable.

[1] https://github.com/mcneel

Since NVIDIA owns huggingface now and huggingface has the excellent Candle [1] crate for inference on Rust, this seems like a good step towards nice native Rust kernels.

[1] https://github.com/huggingface/candle

haha, I see you recognized me from my line scan photography work. But unfortunately that only gets rid of perspective in one axis but not the other.

Probably just taking photos with a really long telephoto lens would be good.

In theory, you'd use a telecentric lens [1], but then the lens would need to be bigger than the subject, which is not super feasible for a car-sized subject. I guess for line scan, the telecentric lens just needs to cover a small vertical strip of the car so it just needs to be really tall.

[1] https://en.wikipedia.org/wiki/Telecentric_lens

Extreme vibe coding!!!

* better squares-in-circle packing for n = 31 and n = 34: https://github.com/dllu/packing and https://erich-friedman.github.io/packing/squincir/

* vibe coded CAD application (paused due to lack of quota on my $200/mo plan). Rather doubtful if GPT-6 Astra can one shot a correct B-rep kernel but whatever: https://github.com/dllu/viboceros

* ray traced chess demo: https://github.com/dllu/chess-rtx

That's not the same thing as those are photos with perspective distortion.

I found it interesting that Thorsten mentions the importance of attaching screenshots.

In my experience, it varies widely by workload. Vision is great for some UI stuff but extremely terrible for my work, which involves visualizing connectivity and geometric stuff for road structure for autonomous vehicles. It can get super frustrating when I paste a screenshot and go "this polyline isn't connected when it should be" or "those polylines are self intersecting in an invalid way" and then the AI literally can't see what I'm talking about. It would have to then write debug code to dump out the coordinates of the polylines to convince itself that the bug even exists. Then it would often claim that it's fixed when it is clearly not.

I previously complained about GPT 5.6 Sol not being able to solve a simple "follow the lines" puzzle [1] without having to write custom Python code for it. GPT-6 Astra, however, is vastly improved, and could do it without writing Python. Still, for my work, it can occasionally struggle.

[1] https://news.ycombinator.com/item?id=49334444

I stand corrected. However, fundamentally, the idea of "tiny CPU-only neural network" combined with traditional alpha-beta search is substantially different from "big GPU network" combined with Monte Carlo Tree Search. And historically the NNUE came from a 2018 idea for shogi engines rather than from AlphaZero.

The Stockfish NNUE is completely unrelated to AlphaZero.

I recently scored an amazing deal on the Samsung Odyssey G8 G80HS, which is 32", 6144 x 3456 at 165 Hz for only $700 via perks at work [1]!!! Even if computers typically cannot game at 6144 x 3456 px, it is incredibly nice to simultaneously have the fast refresh rate to glide your mouse cursor around smoothly, and have the super high image resolution to read super crisp text. The monitor can also be run at half resolution but at twice the framerate of 330 Hz, which is amazing for older or competitive titles. The only downside is that the quality control is not too great (my friend ordered the same one and it was defective) and the plastic bezel of the monitor seems very generic.

[1] https://i.dllu.net/2026-09-09-20-40-23_b59cb931ec7bd913.png

The windshield wiper controls are a physical button plus physical scroll wheel on the steering wheel. You don't have to touch the touchscreen.

I feel like taking pictures is a fun experience in itself. For me, it isn't really "missing the experience". If anything, it makes me pay more attention to the amazing view and details that I'd have otherwise missed as I try to line up the shot perfectly.