The Playground from Richard Powers has been an entertaining and mysterious read so far, no idea where it’s going but like where it’s been.

And of course next up in the queue is @chiphuyen.bsky.social’s new book on AI engineering 🤓

"Robot, remove the stain"

The robot then grab some scissors ✂️ and cuts it out of the t-shirt 👚.

This is a great example of why we need more data for robotics from Danica Kragic at @neuripsconf.bsky.social

Ask and you shall receive @scienceartmagic.bsky.social ! Qwen Vision Language Models now live on Oxen.ai

Thanks FireworksAI for doing the heavy lifting 🎆

Tomorrow we'll be diving into how you can use synthetic data, chains of thought, and inference time scaling to add reasoning to Visual LLMs.

TLDR ~ LLaVA-CoT beats out some of the closed source models on many benchmarks!

Thanksgiving reminder, enjoy what you have today

@oxen-ai.bsky.social now supports Groq for Visual LLMs including Llama 3.2 Vision 90B, 11B and LLaVA 7B

Can confirm, it's fast 🔥

First test processed 125 receipts in 24 seconds for $0.02

Posting on Reddit is awesome because you get random experts telling you about models and techniques you’ve never heard of before

Oxen.ai can now see! 🖼️ 👀

Simply select a dataset with images, type in your prompt, and let GPT-4o or GPT-4o-mini do the work for you.

Let us know if there are other vision models you'd like to see us integrate.

Talking Upcycling MoE with Ethan He from Nvidia, and learned a lot. It's a pretty simple approach that costs 1/8 the compute to get performance 2x the model size (according to scaling laws). Highly recommend listening to the whole presentation.

youtu.be/Cepja572USI