| | | Four months ago, laptop-runnable models couldn't complete a simple refactoring task. Now Qwen 3.5 and Gemma 4 succeed 9 out of 10 times on Simon Couch's coding benchmark. This post tracks the narrowing gap between frontier and local models, with concrete numbers on speed and capability.
Simon P. Couch |
| | |
| | Most databases assume clean, intentional usage but that breaks once AI agents start querying and writing on their own. This post shows how teams are adapting with safeguards to prevent things from quietly going wrong.
Arpit Bhayani |
| | |
| | Browser-based data tools have a packaging problem: lots of files, limited memory, and expensive extraction steps. In this post, Jeroen Ooms shows how webR avoids copying by mounting .tar.gz archives directly as a virtual filesystem. It's a neat systems-level trick behind faster R in the browser.
Jeroen Ooms |
| | |
| | | | The same data on a different map tells a different story. This post breaks down how projections distort the world in different ways, and why picking the wrong one can quietly skew your analysis.
Dominic Royé |
| | |
| | Datatype turns numbers into inline visuals that live directly in your text. It’s for sparklines, bars, even pie charts, rendered like typography so they scale and style just like a font.
Datatype |
| | |
| | | Everyone keeps reimplementing parametric curves. Torchcurves gives you a clean, differentiable version out of the box. Useful for anything from geometry to ad bidding to KAN-style models, without rebuilding the same components every time.
GitHub | Alex Shtoff |
| | |
| | ML Intern is an open-source project that autonomously reads papers, trains models, and ships ML related code using the Hugging Face ecosystem.
GitHub | huggingface |
| | |
| | | Order, then chaos. This post recreates a 1968 generative artwork where a perfect grid gradually breaks down. The interesting part isn’t the image, it’s how seeded randomness and language choices control that transition.
Jonathan Carroll |
| | |
| | |
|
