Comment by jubilanti
5 days ago
The world of computer vision is much bigger than multimodal LLMs. You'd run an ensemble of specialized models for 3d mapping, object classification, path validation, and so on. On a raspberry pi 5 8gb you can run what you need to self drive an RC car on an obstacle track at 10 FPS.
Today, a lot of it is just integrated VLAs. End to end backprop covers a multitude of sins - a single integrated model stack in cross-attention beats "an ensemble of specialized models".