I think the missing piece is the generative AI being tied to a physical embodiment. My point is, we need the physical machine but we also need to work on this generative AI aspect to manipulate the machine. Unfortunately, making pictures and writing text seems to be the lowest hanging fruit of the Transformer AI era since we have significant datasets already available. Several companies are working on creating datasets to train transformers to apply to other tasks and I suspect we'll start seeing the results in a couple to a few years. Not sure how long before we have a consumer bot doing the laundry and dishes though.
Correct me if I'm wrong here, but it seems that you're taking the wrong lesson from the above quote.
It's not an instruction manual on what parts of the human experience are not yet fully captured by AI.
It's a condemnation on the entire purpose of AI as actualized in our economy, re: it's presumed lofty sci-fi-inspired aspirations (giving humans more creative liberty and freedom) versus how it's actually being operationalized (replacing their creative liberty and freedom).
Unironically, yes. Detailed, nuanced, human level modeling of the world provides an interface with automation that allows for complex tasks and workflows to be built and trusted. "Go clean up the kitchen" might involve something where the question of "is this trash, or is this an adorable drawing made by a child that should be saved" or other such collisions between seemingly clear instructions and the vast context baked into the world requiring interpretation at a more-or-less human level. The system could even frame the above decision as "get clarification, or risk being thrown in the trash."
One of the huge advances in the last 5 years has been categorization of images, alongside segmentation, in conjunction with reasoning about categories. LLM chatbots and generative images and audios are the simplest and least nuanced uses of transformers, with the least friction between training and deployment of a product. Generative models tend to allow perception as detailed, varied, and nuanced as their output produces, and integration of these software capabilities will happen in new and surprising ways for many years to come.
TLDR; what they can create, they can see - advanced image models enable advanced computer vision.
Thanks for the response; I see your point that housework can have pretty significant cognitive requirements. I'm not sure I understood sufficiently the argument for why LLMs are well-placed to attack that sort of categorisation, but it's something for me to mull over.
My favorite version of this sentiment is also the earliest one I know of -- even beating the SMBC comic linked by others here by an additional half a year -- by Alex Krokus, in no small part as it doesn't fall into this trap of talking about physical tasks like laundry/dishes or plumbing: "but i like making art | can it do my taxes? -- no it can't".
