Economics for the future – Beyond the superorganism
As USA and Brazil attest, one of the few remaining economic cans to kick is de-regulation and removal of environmental protection. As the economy gets worse, environmental initiatives (e.g. climate mitigation) will become less popular – not because people disbelieve or care less but because they’ll have less financial and emotional bandwidth to advocate for them.
A bleak look at the reality of human nature, we're not going to stop growing/extracting until physics puts a stop to it.
Removed the other electronics, now it's just an apple (pocket) watch. I love when people ask me about it.
Graph Learning
Aaand my notes from the graph learning working on Monday are live!
It's 9pm now. I've been sitting at Cafe Steam in downtown Rochester, MN trying to get this graph stuff wrapped up so I can think about the posters I saw today at ML for Healthcare.
Emily hosted lino cut art night at mine.
TUI time
Working on a little app to transcribe my voice memos, my first toy app in the vibe coding era. It's a going a lot better now that I've given up on the idea of a Tauri based mac app. Tauri is genuinely so fun, it's like Electron without the bloat - but designing the interface was starting to dwarf the relatively simple transcription functionality.
Now it's a stateless terminal user interface, inspired by the success of all these TUI code editors. This feels more maintainable? + text interfaces are having their moment, I half considered making an MCP server so it could be driven by an LLM interface. That might be where things are headed generally.
It's a rust app with whisper.cpp linked and compiled for Apple Silicon, fast and local. I hope some better open weight transcription models pop up soon, I don't want to give into the temptation to wire it up to a hosted transcription service.
Duumviri: Detecting Trackers and Mixed Trackers with a Breakage Detector (arxiv.org)
I've been thinking about how web fingerprinting, aided by deep-learning based classifiers, has likely gotten extremely good at tracking users across the web without the need of any cookie/supercookie.
Can we fight fire with fire? I think quick local models could be employed client side to detect and remove tracking behavior from websites. Dynamically rewriting the page source code.
I went looking around for work in this area and found this paper where am XGBoost models was trained to identify tracking code that was mixed in with legitimate website functionality.
Of course this only works on an open, DRM-free web. the game is already up on mobile devices where most people browse. How long until most day-to-day web browsing doesn't work outside of signed unmodifiable browser environments?
The Anti-Social Century (theatlantic.com)
Emily gave me a copy of the 2025 East Bay Bike Zine which lead me to this long form.
The essay hits on the social damage we've inflicted on ourselves by designing life around the car and the television. It's not a small wound, the effect has been so totalizing on our culture that it's difficult to frame a specific problem. Every aspect of the built environment has pushed us towards isolation and consumption.
He told me that the cardinal rule of contemporary apartment design is that every room is built to accommodate maximal screen time. “In design meetings with developers and architects, you have to assure everybody that there will be space for a wall-mounted flatscreen television in every room,”
I'm grateful I live in an old enough apartment that at least the architecture is not conspiring against me. My space is television free and people always comment on the fun layout even if they can't quite put their finger on what it is about it.
A little spot around the corner called Aurora Studio just opened up, it's not quite a bar or a cafe or a restaurant. It's a space to listen to music, look at art, hangout, eat w/ friends. Nanxi has desserts there on Saturdays. I think it's a template for a less lonely future.
Position: Graph Learning Will Lose Relevance Due To Poor Benchmarks (arxiv.org)
I heard an anecdote that graph modeling has been running into dead ends because you're better off smushing your data into a euclidean form, throwing away the graph structure and applying massive scaling/compute. This paper from the graph learning community discusses their strategy to reorient in the face of LLM success. We need GNNs to be as easy to run as ollama!
I think I'll attend the graph learning workshop at the Simons Institute in two weeks, need to do some homework though, what is a graphon?
Supervision (github.com)
Supervision is a library to make it /super/ easy to set up object tracking annotations using models like YOLO. This installed easily and ran on my mac without any fuss.
Hmm, if VGGT makes it dead simple to get camera poses from video frames and YOLO makes it this easy to get keypoints you could combine them to get fully 3D keypoints. I'll have to try this.
The library author has a video about using it for dwell time analysis, you could use this to recreate some of the results from The Lost Art of Lingering.
Tattoo idea, not gonna do it just thinking about it.
pocket computer update
Maybe I've lost the plot but I added a LTE -> Wifi hotspot for data access and a Bluetooth audio driver that lets me attach normal headphones. Since Google Fi doesn't support Apple Watch I needed this to go truly phoneless. Hiding my phone in my bag and telling myself I wouldn't touch it was not working. There are a bunch of issues with this setup but it's actually viable for leaving the house without a phone, let's see how this week goes.
VGGT: Visual Geometry Grounded Transformer
The best paper at CVPR this year was a 1.4B parameter (5GB) open-weight feed-forward neural network that takes in a collection of images and does all of 3D reconstruction in a single pass. It outputs depth maps, point clouds, camera poses and motion tracks all at once! You can download this and run it on a laptop right now.
As networks become ever more powerful, we ask if,
finally, 3D tasks can be solved directly by a neural net-
work, eschewing geometry post-processing almost entirely.
This is very different from gaussian splatting, though the visualizations look pretty similar. With splats/NeRF you're training a differentiable model to become an implicit representation the 3D scene – here you have a pre-trained model that outputs 3D information, not just synthetic views.
Free fast simple local computer vision on demand, incredible.
I've become obessed with the idea that maybe a pocketable Apple Watch could be the dumb phone we've been looking for. An actual flipphone is a non-starter for me but this thing has iMessage, email, maps, music, wallet etc. I love the idea of a semi-retro pocket computer that can keep me in the modern world but free me from algorithmic feeds.
I got this Rabbit R1 <> Apple Watch idea from @basvanderploeg but there were no files so please hold while I learn CAD.
Sunday morning, let myself get sniped into reinventing twitter (this) so I can post about my new CAD projects. It's also inspired by Simon Wilson's link blogging https://simonwillison.net/2024/Dec/22/link-blog.
How do we find things? (tadaima.bearblog.dev)
Timely post about non-algorithmic curation. Two years ago I took a year off all music recommendation algorithms to push myself to seek music from friends and aim to learn more about the music I'm consuming. That was before the AI slop machine went into full gear. Now Spotify has every incentive to juice your feed with cheap music from AI and musical ghost kitchens. There's never been a better time to ask your friends for some recs or tune into a local internet radio station.
