Video of the Week: 1X World Model

OK robot nerds - here's a big one.

NEO from 1X has started generating its own videos of human-completed tasks so that it can learn how to do them for itself. This week's Video of the Week shows one of them educating itself to perform various tasks, whilst staring into space like 15-year old me during a French exam.

Unlike 15-year old me during a French exam, however, NEO is running multiple scenarios through its head in order to get the answer right - as opposed to wondering whether I could win a fight with an owl or who invented kerbs.

NEO's prowess comes from its new AI 'world model', which allows it to derive logical conclusions from things that are +like+ what it sees rather than having to see the exact thing that it's being asked to do. These world models are going to be increasingly talked-about this year, with everyone from Ilya Sutskever (legend that co-created the AlexNet neural net which rebooted AI in 2012) to Yann LeCun (one of the 'Godfathers of AI' who built Meta's LLaMA open-source models) creating labs to build them.

The sheer volume of data that robots have to ingest to become broadly useful has always been the biggest barrier to their adoption. Slowly - so, so slowly - putting things in a dishwasher and/or gently dusting a table are not the chores that make them essential. However, if they can access videos of a human making a roast dinner, then make the leap to creating a video internally of its own hands completing the task in the most likely way, then we're into 'recursive learning' - it trying things then remembering what works for next time.

So this is big - and will be copied and built upon by every other robot lab out there. I still don't think that they'll be in my mum's house before 2035 but this is another great step in learning.

(Which is also very unlike15-year old Mal's French exam ...)