The category is: p(Doom) Generation
06.03.2026 17:21
👍 0
🔁 0
💬 0
📌 0
The category is: p(Doom) Generation
The situation monitor urge to shop at the military surplus
ohh! 👀
Can you run a single function remotely ?
Here are my findings 🔎 ocramz.github.io/posts/2026-0...
and a new library 'ascend' ⬆️☁️ : github.com/ocramz/ascend
Leftwing PM Pedro Sanchez bars US planes from Spanish air bases podcasts.apple.com/gb/podcast/w...
Calling your crowdfunding equity "for punks" and liquidating with them last in line was a little on the nose huh
A can of Vault City winter berry crumble pastry sour, which literally tastes like a blackberry pear and strawberry crumble in a beer
Vault City mon amour 🏰💜🍐🍇🍓
Cover art for Batu — Frostbite
V cool cover art timedance.bandcamp.com/track/batu-f...
Kiss Facility debut album reminiscent of early Grimes, Cocteau Twins, Slowdive 🙂↕️🙂↕️🙂↕️
Another ex: when I found I couldn't load TfL's official GTFS file with city2graph to make node/edge graph data structures of transit routes, Claude suggested registering for a 3rd party data provider to use their readymade one & I had to push to generate one from open data myself
People talk about vibe coding offering agency (in the sense of lowering the barrier to mastery experiences) but LLMs also structure said agency, e.g. I recall when I suggested some ideas around transit modelling/prediction Claude tried to dissuade me as it'd all already been done
Took me a minute but just realised the affordances a tube network distance field provides
originally wanted something like this so I could objectively order prospective gyms by travel time cost, but with OpenStreetMap data it can really be any category (and no, agents don't solve it)
Claude stop putting fallbacks in every single part of the code challenge !! unhinged
Sega Bodega goes Arabian shoegaze o'er Westfield 🙂↔️ youtu.be/PB3NnRYHmgc
First time using graph nets but these came out p nice:
💽 Policy network: routes through TfL tube stations (Underground only) huggingface.co/permutans/tu...
💽 Value network (distance field): predicts travel time between stations without actually doing a rollout huggingface.co/permutans/tu...
A simplified illustration of Hal 9000 from 2001. With a speech bubble saying Thanks for calling that out. That was bad advice on my part.
Techno industry MeToo moment this week is the real Dark Woke
Reprocessed a TfL platform interchange dataset that was FOI'd in 2015 and put it up on the 🤗 Hub huggingface.co/datasets/per...
🎉 TfL Underground inter-platform data acquired www.whatdotheyknow.com/request/inte... (download: www.whatdotheyknow.com/request/inte...)
via FOI, 2015 www.whatdotheyknow.com/request/inte...
Journey routes from West Ham to Bond Street either directly via Jubilee or going further East to Stratford then travelling back West on the Central line. The timings shown are purely inter-station travel time (i.e. assume instant transfer between platforms with no waiting for connections)
It'd appear I overestimated the speed of the Jubilee… maybe I just prefer it as it's a nicer line (this model doesn't include transfer timings)
it can actually! but apparently not many people use it, it's called "sampling": servers can request completions from clients x.com/permutans/st...
Yes! I'd made a TfL API interface lib but the route planner needs a model, so I converted API to GTFS then converted that to node/edge parquet + loaded in PyG to train policy networks, it learnt Dijkstra optimal shortest paths perfectly! Now just gotta add more lines & per station transfer penalties
The policy network from the tubeulator-models repo showing inference of a route from Liverpool Street to Westminster The route takes 3 hops, in 3 lines, with 2 transfers, and may not be the fastest once transfer timings are accounted for (etc etc…)
TUBEULATOR REAL 2026 🚂
whyyyyy is there a log me out immediately keyboard shortcut
I took the point of it as to allow people using it to *not* have to be at their computer (babysitting a long running process)
lol my model is teleporting
Remember the use-mention distinction?
A rollout with 90% step accuracy (as before) but exploring 5 routes in parallel and storing only the one that reaches the destination, or the fastest if multiple do. The stratified scores show that there is no longer a different outcome depending on the length: all the buckets get 100% success
Routing achieved with beam size 5! 😁💯💯💯💯💯
With greedy rollout one wrong hop spoils the whole trajectory, this model is exploring multiple in parallel (and keeping the best scoring one when multiple reach the destination)
Routes 7% too long and step accuracy still at 90%…
Training a GATv2 encoder + Transformer decoder learning a routing policy over tube routes. The results are presented as a point statistic (89% success) as well as stratified by route length, and the first 2 buckets (up to 10 stations) reach 100% accuracy, indicating the model has fully learnt them
All aboard 💯🚆💯
(Trained a routing policy instead of an autoregressive sequence learner and it's way more effective)
