Today we're releasing Critic, a platform to review the code your agents write. Try it out at critic.run
Since agents started writing our code, it has become harder to understand the scope of a change. Your AI has all the context behind the code but you don't.
Critic lets the agent explain what changed and why, and answer questions from anyone reviewing its work.
Read more about MultiMatte xitter.gq/status/209779722711277… or build with it using NoBg.
MultiMatte is open source. Check it out on HuggingFace: hf.co/feyninc/multimatte
3
40
MultiMatte is built on Meta's SAM 3. Our biggest win came from converting binary masks that classify each pixel as being inside or outside an object, to alpha mattes that assign an opacity value to every pixel. This allows MultiMatte to better understand hair, motion blur, translucency, and fuzzy boundaries.
Across all our measured benchmarks, MultiMatte shows a step improvement over SAM 3. On DIS5K, S-measure rises from 0.674 to 0.908 (a 34.6% relative gain), and on DUT-OMRON from 0.792 to 0.901 (13.7%).
1
4
106
In training, a diverse data mix matters. Our first fine-tune used one dataset. It improved performance on camouflage images, but regressed elsewhere.
We took this as a data problem and built a 26.1K-image mix from 10 datasets covering camouflage, crowded scenes, high-res subjects, portraits, and anime.
1
1
130
We took an interpretability-first approach for building FeyNoBg
We studied the architecture of an existing model, BiRefNet. This led us to discover the importance of the third stage of its feature extractor. We expanded this stage from 18 to 24 blocks, giving it room to learn.
1
2
231
SQL queries can be syntactically sound, but still incorrect. A query can misread an ambiguous column or filter for a value that doesn’t exist, all without throwing an error.
SQRL handles this by inspecting the database before answering. It asks for the stored value instead of guessing.
1
2
96
The gains are architectural.
Today's best extractors are decoders limited by memory bandwidth. Pulpie is an encoder bound by compute.
GPUs are starved for bandwidth, not compute. As a result, Pulpie is performant everywhere. Our testing shows Pulpie to be 7x faster on A100, and 20x faster on L4 GPUs.
1
4
326
Introducing Pulpie, models for cleaning the web.
70% of a typical HTML page is ads, navigation, and sidebars. Pulpie removes this noise and returns a clean document you can use in pre-training or as context.
Previously, cleaning 1 billion pages with the best extractors cost $159,000. Pulpie brings that down to $7,900. A 20x decrease.
🔥🚀Chonkie 1.6.8 is out with PyEmscripten wheels!
🐍You can now run the full Python Chonkie library directly in your browser using Pyodide🦛
This is an early release that brings Chonkie’s core functionality to the web, but it's a big step towards 𝕓𝕣𝕠𝕨𝕤𝕖𝕣 𝕟𝕒𝕥𝕚𝕧𝕖 𝕡𝕪𝕥𝕙𝕠𝕟 𝕔𝕙𝕦𝕟𝕜𝕚𝕟𝕘.
We’d love to hear your feedback! try it out and let us know what works, what doesn't, and what you'd like to see next!
Happy coding! 🦛✨

















