IF THE MOON WERE ONLY 1 PIXEL 0 km- A TEDIOUSLY ACCURATE SCALE MODEL OF THE SOLAR SYSTEM
If the Moon were only 1 pixel - A tediously accurate scale model of the solar system.
One of the best sites ever.
Click. Scroll. Enjoy!
joshworth.com/dev/pixelspa...
Just be aware that the events are extracted from news sources in over 80 different languages.
At this moment, we can't include the URLs but we are currently looking if we can make this information available in a future release. Be assured that the process to extract events targeted only publicly available news sources, which you should be able to find via Google search.
If only a learning from the current situation would be to force push towards electrifying the grit and expanding renewables...something we could have learned already from the Russian invasion into Ukraine. But not in this country...
xLSTM Distillation: arxiv.org/abs/2603.15590
Near-lossless distillation of quadratic Transformer LLMs into linear-time xLSTM architectures enables cost- and energy-efficient alternatives without sacrificing performance.
Efficient xLSTM variants of instruction-tuned Llama, Qwen, and Olmo models.
As always, thank you @kinarnicholas.bsky.social for sharing our work with your network. 🙇
Seems like a pretty good fit actually.
People definitely prefer my dataset publication posts over anything else I post on this homepage^^
Obiously, there are tons of caveats to consider and remember, when working with news data (e.g. spatial bias, recency bias, and many more). We go on detail about this in the paper.
Nevertheless, I'm convinced that you can surface event information from news articles that would otherwise not exist.
Flash floods model
Preprint: doi.org/10.31223/X51...
Blog: research.google/blog/introdu...
Operational predictions at g.co/floodhub
Groundsource
Preprint: doi.org/10.31223/X5R...
Dataset: doi.org/10.5281/zeno...
Blog: research.google/blog/introdu...
It was a huge team effort from any people that aren’t on BlueSky. But big shout-outs to Oleg Zyldenko, Rotem Mayo and Deborah Cohen. 🙏
This however is just the beginning. The methodology can easily be extended. We're looking at extracting data beyond just verified news outlets, and expanding to other natural hazards that lack good datasets. 4/n
But We aren't just releasing data. We also shared details about our new flash flood forecasting model that was trained flood events data from Groundsource. This allows us to better predict flash floods up to 24 hours in advance. 3/n
We used Gemini to extract locations & times of past floods from unstructured global news reports in 80 langauges. The locations were further georeferenced. The result? 2.6 MILLION historical events across 150+ countries. 🌎 2/n
A worldmap of the spatial distribution of extracted flood events in the Groundsource dataset. The map displays the total number of flood events extracted by the LLM-based pipeline aggregated per grid cell. The data are visualized using a Robinson projection, with event counts represented by a logarithmic color scale. Red points indicate the spatial centroids of reference flood events from the GDACS database.
Excited to announce Groundsource - an open-source dataset of historic flood events! This has easily been one of the coolest projects I've worked on recently!
Thread 🧵 for details and all relevant links. 1/n
I present the most interesting graph ever made.
HUMAN ON BICYCLE beats every other living thing.
www.scientificamerican.com/article/a-hu...
Interesting read and something I have been thinking about recently. How fit is our education system for the new era and what does this mean for _learning_ (especially learning to learn), if students/pupils over rely on LLMs. What does this mean for my kids (2nd grade and kindergarten)...?
What are your thoughts on this article. It also has been a while since I came across Tim's name, but I remember from my early days of ML how he was already publishing hardware benchmarks for ML setups and that I read them carefully whenever I upgrade my home setup.
Some updates on NeuralHydrology and I'll also use this opportunity to give a shout-out to @gauchm.bsky.social for his continuous support and maintenance of NH
Are you working in large-sample hydrology?
If so, we invite you to submit an abstract to our session for EGU 2026:
HS2.4.2 "Large-Sample Hydrology: Advancing dataset developments, enhancing process understanding, and unifying insights through catchment modeling"
Dezember 8, 2025 :)
Or, if your model outputs uncertainty bounds, at least you learn/model the uncertainty of the task that you care about.
For those of you who are still undecided.
We (wife, 7 y/o, myself) spent New Years Eve playing "Stuffed Fables"
boardgamegeek.com/boardgame/23...
My wife's first roleplay game and also the most complex game we played so far with our oldest. Time was flying and we only made it to the start of the second chapter 😅
No, I was not thinking about FM models in this context. One easy example: if you want a model that forecasts flood warnings, why train the model to predict streamflow? Under ideal conditions, you could simply train the model to forecast "send alert/ no alert"
Hydrology Paper of the Day @kratzert.bsky.social on integrating water quality observations into the Caravan dataset: the need for proper metadata and processing of observations; a new river network with water quality stations; catchment delineation; and validation and flagging of suspect data.
Interesting point. I've been thinking about this as well in the past and I think, especially with ML, we have the opportunity to optimize for what we really want to know from the model, instead of training it on some proxy task
game screenshot showing an amsterdam-like city with the title "tramstertram" above it
since everyone in Brazil's lost Twitter, I'm gonna try to make sure anything I post there is posted here too 🥲
here's a free building game I shared there a few weeks ago! mattstark.itch.io/tramstertram
Very cool. Congratulations and surely deserved 💐
