Tab Clearning (December 4 2024)
misc
Clearing out my open tabs
Tab Clearing
- QwQ announcement. Amusing tone - ‘What does it mean to think, to question, to understand?…’. Exciting model, by vibes at least. Playing with it via https://api.together.ai/playground/chat/Qwen/QwQ-32B-Preview until buddies get quants running locally with long context.
- https://gen-omnimatte.github.io/ Generative Omnimatte - Learning to Decompose Video into Layers, cool for VFX etc, good to see.
- https://cat-4d.github.io/ change view on any video, I want this in VR, even janky as it is.
- https://lilianweng.github.io/posts/2024-11-28-reward-hacking/ - fantastic post on reward hacking. Also yay, lilianweng is posting again after recently leaving openai. A boon for us all.
- https://transluce.org/neuron-descriptions describing all neurons in Llama 3.1 8B with a fine-tuned model. See also https://goodfire.ai/ which lets you steer generation with SAE features, looks very cool
- https://minihf.com/posts/2024-08-23-weave-agent-dev-log-1/ a representative sample of an interesting blog - the author is John David Pressman who I’ve been aware of since early latent diffusion stuff. Working on agents, and writing about LLMs and philosophy from a unique perspective. Also, and I mean this gently, very not normal.
- https://arxiv.org/abs/1906.09453 Image Synthesis with a Single (Robust) Classifier a random paper I want to re-visit with a new trick and modern classifiers maybe.
- https://www.interconnects.ai/p/openais-o1-using-search-was-a-psyop more speculations on o1. I liked Sasha’s talk Speculations on Test-Time Scaling which we did a watch party for.
- Streaming Deep Reinforcement Learning Finally Works Ever since a chat with @pcastr a while back I’ve been more bullish on RL. I just skimmed bits of this, but one day soon I’ll be ready to dive into RL I bet.
Twitter/X links
- Mistral canvas executing python in the browser. Gimmicky but I should try their canvas more.
- HunyuanVideo: A Systematic Framework For Large Video Generation Model Training 🔥 Jupyter Notebook 🥳 OS video models getting good, and big. Not including lots of other video gen and 3D gen tweets
- Sander Dieleman: Better VQ-VAEs with this one weird rotation trick! - paper seems a cool little trick.
- I see CPNNs, I click = imscore - a collection of SoTA image aesthetic/preference scorers.
- Structured outputs might not hurt after all why am I not suprised?
- Ensemble everything everywhere: Multi-scale aggregation for adversarial robustness - fun paper, hopefully colab at some point. imstack rediscovery.
Misc
- https://www.infinitescroll.us/p/why-you-should-read-web-fiction with some interesting-looking recs.
- https://www.marksaroufim.com/bookshelf Mark’s amazing book list
- https://nabeelqu.co/advice A nice advice post and a good idea for a personal site page.
- https://podcast.adobe.com/ New voice cleanup thing, might be amazing (and save me having to finish my idea of a hosted soln)
- https://www.youtube.com/watch?app=desktop&v=YwELr8ir9qM fun predictions of the future.
- https://huggingface.co/datasets/ailsntua/Chordonomicon A bunch (600k) of chord progressions, might be fun to play with. Not linked back to lyrics etc so not directly UG replacing though.