Interesting that AMI is betting on video-first world models. A 4-year-old learns physics mostly through interaction, pushing, dropping, breaking things, not just watching. Vision helps but the feedback loop from acting in the world seems at least as important. Still, glad someone is putting $1B on a fundamentally different bet than "more text, bigger model."
The energy numbers are the real story here, 70-82% reduction on CPU inference. If 1-bit models ever get good enough, running them on commodity hardware with no GPU budget changes who can deploy LLMs. That's more interesting than the speed benchmarks imo.