Louis Maddox
@permutans
Combinatorially curious
Pushed the initial subsets of a BBC News curated @huggingface FineWeb dataset overnight, expecting ~400M entries total ⏳⏳⏳ 💾 github.com/lmmx/bbcfw/ 🤗 permutans/fineweb-bbc-news huggingface.co/datasets/permu…
<1B SmolLM4 🔜👀🍿
We did not start the training yet and we want to try different approach for this size so it's gonna take a bit more time!
Who said that SmolLM2-135M can't generate images ?
A new world wodel from Meta FAIR after all the chaos. 🌍 Meet DINO-world: a generalist video world model that predicts the future—in latent space. Trained on uncurated videos with DINOv2, it learns diverse temporal dynamics (driving, indoors, sims), beats prior models on…
2025 model training vibes: arch people losing themselves in muon fantasy loss valley while synth people commit borgesian monstruosities.
Dreamt that YouTube put out an announcement that users could upload simulations now (probably from world models), and I visited the homepage and there was in-dream AI slop
So an art school kid did a grad show with a robot dog and the backlash is how I find out these are now used in warzones…
Birb Storage
Bird saves and reproduces data: A PNG image of a bird ( photo of a bird -> spectral synthesizer ) was reproduced by an adult Starling bird youtu.be/hCQCP-5g5bo?si…. It seems to have reproduced the sound in conjunction with some additional notes which made it not detectable…
They did surgery on a quail egg ❕
At WAIC 2025 in Shanghai, we witnessed a surgical robot delicately peeling a quail egg. Truly impressive!
From March, sharing information about (“popularisation of”) VPNs has been banned in Russia techradar.com/pro/vpn/russia… Apple has apparently been ‘the most compliant’ in removing VPNs on request of the Russian gov (and Meta has been declared an extremist org…) techradar.com/vpn/vpn-privac…
Encountered this one yday (vaguely recall it impressed me), didn’t know it was one of the codenames people are watching (“lobster”, “nectarine”, “starfish” are rumoured GPT-5 pre-launch testers). Kimi K2 wasn’t so good at code, Qwen 3 Coder was decent, Gemini Flash also v good
Kinda amazing: the mystery model "summit" with the prompt "create something I can paste into p5js that will startle me with its cleverness in creating something that invokes the control panel of a starship in the distant future" & "make it better" 2,351 lines of code. First time
the site’s lawyers floated the idea of a monthly quota to keep it below the Category 1 threshold, meaning that UK access to Wikipedia could become less like browsing the web and more like trying to buy a concert ticket, with a cap on how many people get in biometricupdate.com/202507/uk-high…
Wikipedia is currently in a legal battle with the UK government to try and stop the platform being censored in the UK - or even completely blocked - thanks to the Online Safety Act.