tda

What is seen, cannot be unseen

I wonder if and how you can use this for steganography, hiding data in plain sight. I bet most automated image analysis programs would only consider the final image. I sure some highschooler can use this to bypass their schools contentfilter

I installed Claude Code on the Mac Studio, and gave it access over ssh to a Linux box. I instructed Claude to first compile on Linux, create reference artefacts, and only then start the work on mac. And then to make sure the Linux version still would compile with the changes made to the compiler toolchain, the Mac specific instructions etc.

I told Fable to make a nice plan, and fan out work to opus and sonnet agents. I then enabled remote connections and checked every few hours from my phone and nudged it in the right directions where needed. Have some push back once in a while. And once done I asked it to make a nice commit history and remove all the dead ends.

I ported a decades-old piece of scientific software (Delft3D — Deltares' Fortran hydrodynamics engine) to run natively on Apple Silicon. From my phone, mostly, because I could (and because AI could). Runs pretty decently on a Mac Studio M3 now.

It was officially Intel-compiler + Windows/Linux only; the port gets it building with gfortran on macOS/arm64, PETSc+MPI solver and all. Fable and GPT-5.6 chewed through the compiler warnings, dependency hell, and portability bugs largely on their own. 30 or so bugs uncovered and fixed along the way, several of them real and platform-independent.

I'm not a C or Fortran dev. Mostly I argued with the agents about benchmarking methodology and told them to re-measure when the numbers looked off. The rest was up to them. Still amazed about this AI thing going on

https://github.com/tdamsma/Delft3D

Nice use of civic tech! Have you ever showed this to open state foundation? https://openstate.eu/

Record the audio, transcribe it later. Your AI note taker will mangle some of the details in a subtle but wrong way. Or not, who knows? But if you have ever transcribed anything with whisper, it is both amazing how well it works, and how it can go completely off the rails if there is a prolonged pause or noise. It will just make up words, fill the gaps. I am assuming the transcription process for an AI note taker will not use anything fundamentally different than Whisper

... oxygen generation mechanism of diving suit for cyborg cockroaches

An actual quote from a Nature article

Though I fully agree a passive heat prevention solution is better, for my house the number were: nice shades (that open and close electrically): 7K, AC unit that also functions as heater in winter allowed the central heater to remain off foor all but 2 days this winter: 1800 (including installation). I used to hate AC's, and kinda still do hate the how they look. But I am pretty happy so sit in a 25 deg living room when it is 35 outside. And the AC runs on my own solar, I have plenty anyways

I was using fargate, route 53, managed postgres. Also the container registry i think. Not nothing, not too complicated

I literally migrated last week, and saved 8x by ditching AWS for Hetzner.

I run Nextcloud at home with 1.5TB of files and 2 users, on a reasonably sized server. It is painfully slow. Still better than OneDrive, but only just: synscing takes forever, never reaching a fraction of available bandwidth. Upload from my phone is flaky, often hangs and needs manual intervention. It is a battery drain. The whole experience with add-ons and the general UI feels like a 2010 PHP app.

I am grateful Nextcloud exists, but no app deserves a vibe coded Rust rewrite more than Nextcloud. Literally nothing to loose

You might want to add a filter to the colorings page. Under the "realistic" category, it is filled with images of Hitler:

https://www.coloringsai.com/en/coloring-page/hitler-doing-a-...

The filament handling is part of the printer to me. You wouldn't want a car without a gas tank (just strap a jerrycan on the roof), or a table saw without dust collection. Ok, bad examples but to me the filament handling is part of the problem I want solved. And I want it solved in a single machine/ setup. I have an old prima that I made en enclosure for and a bit Ikea tub with silica beads and 5 spools of filament routed through custom tubes to the MMU. That is a messy setup, takes a lot of space, and look like crap. If I spend north of 1k or maybe 2 for an upgrade, I want the entire problem solved. Not 80% with an implied hobby project included. No thanks

But what does Prusa offer for dust and moisture free filament storage? Why doesn't Prusa copy the "enclosed filament storage on top of printer' solution that is so much better then spools dangling on the side? I don;t want to but a BamuLabs I am willing to spend 2k+, but not on a printer that does not have a properly integrated filament solution.

https://archive.org/details/psychopathcode/psychopathcode deserves a mention here

Nice model, but also a great notebook. Really like this "literal coding" style for CAD in a notebook. I still sometimes prefer GUI cad for simple stuff, CadQuery is not always that easy to write. But when iterating over something complicated, the notebook style development is really nice. And being to compose a model using functions and iterators instead of the clumsy GUI workflows is a godsend. Also time I tried Opus was more helpful and capable than I would have thought. Not good enough to one-shot yet, but it is very helpful nonetheless

I think the comparisons were referring to land area, but I agree this is not that clear from from the comment

Ome advantage of python is that it is so slow that if you choose the wrong algorithm or data structure that soon gets obvious. And for complicated stuff this is exactly where I find the LLMs struggle. So I make a first version in Python, and only when I am happy with the results and the speed feels reasonable compared to the problem complexity, I ask Claude Code to port the critical parts to Rust.

Is there a video somewhere of the one inch microdrive with acrylic display shown in the article?

One thing with python is that usually I will use one of the many c based libraries to get reasonable speed and well thought out abstractions from the start. I architect around numpy, scipy, shapely, pandas/polars or whatever. So my code runs at reasonable speed from the start. But transpiling to rust then effectively means a complete redesign of the code, data structures, algorithms etc. And I have seen the AI tools really struggle to get it right, as my intent gets lost somewhere.

So what I do now (since Claude Code) is write really bare bones (and slow) pure python implementation (like I used to do for numba, pypy or cython ready code), with minimal dependencies. Then I use the REPL, notebooks and nice plotting tools to get a real understanding of the problem space and the intricacies of my algorithm/problem at hand. When done, I let Claude add tests and I ask it to transpile to equivalent Rust and boom! a flawless 1000x speed upgrade in a minutes.

The great thing is I don't need to do the mental gymnastics to vectorize code in a write only mode like I've had to do since my Matlab days. Instead I can write simple to read for loops that follow my intent much better, and result in much more legible code. So refreshing!

And with pyO3 i can still expose the Rust lib to python, and continue to use Python for glue and plotting

Some operations (downsizing already heavily compressed video with the latest and greatest compression techniques) are CPU or GPU bound, but others, like bunching thousands of high res jopg into a lightly compressed timelapse are more likely to be IO bound. So how does this tool make the trade-off? I imagine some things can best be dealt with locally, whilest for ither operations offloading to a external GPU will be beneficial. Also makes a lot of difference if your bandwidth is megabits or gigabits.

The best thing about the average speed cameras is that between the checkpoints all cars drive at almost exactly the same speed. No one trying to overtake, just 5 lanes of traffic at 1km/h below the speed limit

I love these kind of stories, and I appreciate the effort to write these histories down.

I do think the interview could use a bit more editing, it now reads more like a literal transcript and that is somewhat exhausting to read. Take this excerpt:

> "Like it just fit my brain. It’s one of those where I don’t have to think about it, it just automatically makes sense to me. Much more than any other programming language. Even though I quite like C, I quite like all the pitfalls in C, I’m very comfortable in C and C++ and pretty comfortable in Java and PHP, but Python is just fun and it’s enjoyable and it makes sense. And this was the Usenet days, of course. I think the Python list, yeah, the email list gateway existed as well. So I think I just subscribed to Python-List and learned a lot about Python just from using it, following along."

I sometimes help out at a hobby vineyard of 0.7 hectare; weeding in the row is a lot of manual labour. Your platform seems like a good fit. I like tinkering and robotics, what kind of price point are you aiming for?

And for weeding?

I fondly remember finding and exploiting a buggy slot machine on the night the Euro got introduced. A classmate (I never played slot machines) made some money but didn't understand what was going on. I observed and it became apparent (in my slightly intoxicated state) the machine would pay out 2 Euro coins where is should pay out 20 cents. And when playing a 1 Euro game, you would often "win" 80 cents. Pay-out immediately and you got 8 Euro. Of course after a few rounds, the 2 Euro coins ran out and it would do some RNG to pay out 1 Euro with 80% chance. Don't know if I tried feeding it back the 2 Euro coins, I recall just made enough to have a free new years eve

The real problem is that BREP CAD kernels are hard. A few of proprietary kernels dominate the scene: Parasolid powers NX, SolidWorks, Fusion, and Onshape, while ACIS (owned by Dassault) is used by Inventor and BricsCAD. Catia uses Dassault's own CGM kernel. The open-source world relies mostly on OpenCASCADE, which is unfortunately a lot less capable than any of these.

Fillets and chamfers are a good example. They seem simple but are geometrically non-trivial, and OCC will fail on cases that Parasolid handles without complaint. You can push either kernel to its limits if you try hard enough, but OCC hits that ceiling much sooner. So any CAD tool built on top of it inherits that ceiling too.

I recently gave CadQuery (a Python wrapper around OpenCASCADE) and its Jupyter and VSCode integrations another try. Two years ago installation was a mess across conda, Docker, and pyenv, and the API itself felt like a dense, bespoke DSL you had to fight.

This time everything just installed, and Claude Code turned out to be pretty good. Designing with code is sometimes more work upfront, but iteration is so much better. You get proper abstractions: functions, encapsulation, loops. You can drop in a SAT solver to optimize part placement or grab data from an excel sheet. No more clicking through a GUI that crashes and loses your session. I've spent time with Fusion, SolidWorks, NX, OnShape, FreeCAD, and Rhino, and each has its merits, but none of them can benefit from the LLM revolution the way a code-first tool can.

I asked Claude Code to generate a set of Lego bricks in various sizes, apply a nice color palette, and pack them optimally into a grid. It needed some steering, but all in all I was impressed

Yes my router has open ports, but it does not do any port forwarding. So I can 'directly' connect any device behind my router without my router needing to know any specifics of which device that is. And I don't need to do any port forwarding of anything on my network and thus expose them to the whole internet; I just expose them to the users of my tailscale network (only me)

Nothing special, an edgerouter that allows installing tailscale