>If you can train a machine using video, and get it to repeat what you have taught it

I might be wrong, but from what I know about Transformer models and neural nets in general, what you are describing sounds a lot closer to something like an AGI.

When you say “train a machine” here it sounds a lot closer to “teach a machine” rather than “finding weights and baises for an existing function”.

I think we are still a ways off from AGI.

I dont think AGI is required. But multimodality for sure and some piece to plan/execute actions. Prompting techniques now, would land you clear tasks, validations and such, given some objective "replace table leg". A unit to process this would be required.