Building actors is much harder. All these models are really good in interpolating between training data points. It’s much harder to do that for a set of actions, because the space of valid action sequences is far smaller and sparse.

Right now what we see is the GPT-2 of Video Generation. It's truly impressive, but just wait for GPT-3 and GPT-4

I was able to create a high quality animated video from a pencil sketch under 5 minutes . 🤯🤯🤯 First turn your sketch to a Pixar style image using ControlNet in DiffusionBee Then use Luma AI to turn it into a video. Wow 😮

Right now a lot of LLM usage is for very basic tasks. eg. Extracting info from text We will soon have several small task specific LMs which are faster, cheaper and more accurate!

MLX can now run on CUDA devices! 🔥 A game changer for cross platform local AI

ARC-AGI-3 preview is out It consists of a few games which AI agents can play. It's a good test of few shot learning, generalization and adaptation to new scenarios. > No instructions given to AI > Easy for humans, but hard for AI For perspective: most RL algos have to play

I've left Meta Reality Labs. Miss working with the best team for realistic avatars. Damn, I definitely miss using thousands of GPUs and running super large-scale training experiments. Now back to being GPU poor, but excited to start something new: building tiny AI models for