Technology

LLMs and music generation.

I'll head this off at the start: The comparison is already extremely skewed. There's no hard "right" or "wrong" with creating text or images. Fuck up a sentence? Eh, whatever. Miss a note? Call it Jazz. Extra finger? ...okay, maybe there is a "wrong" but it's not a costly mistake.

Generate a dish-cleaning procedure that ends up trying to place down the cup where it is typically placed down, except the randomly deviating space is now partially occupied by a plate also placed within typical deviation of where plates typically are? Cup tips, falls, crashes and now the floor has irregularities not accounted for in the original procedure generation. Do I abort, try to generate a new procedure or carry out a floor-cleaning routine? How do I identify which option to go for, assuming I detect the irregularities at all?

The pattern-generation/-imitation approach used in GenAI is ill-equipped for systems requiring strict conditions and semantic requirements attached to parts of the pattern, for the same reason we've had incidents of agentic GenAI breaking production systems and their general tendency to produce shitty code that needs human supervision: There is no meaning to the tokens. That's good and well when you can kimit permissions and double-check before applying their suggested changes, but it works poorly with physical tasks.

Which is why

Automated dish cleaning machines exist that don't use AI, but they're industrial and expensive

They're the physical equivalent of specialist systems, built for a specific task and very good at it, better than humans even. I would hope they'll become more affordable at smaller scales, but I think the current hype about universal systems isn't gonna deliver the results we're hoping for. Humans are sorta good at many things, but I think imitating them shouldn't be aiming for imitating sorta good utility.

I'd love to see the development of human-like machines be treated and funded as a fancy toy, maybe a platform to explore the nature of sentience and sapience, a vehicle for science. I don't need them to have practical applications any more than I need other research to start out with a practical application: maybe an actual use will emerge down the line, maybe the lessons learned will contribute to other research, who knows. Just let the nerds go wild with the nerdery and keep business out of it.

The universal household robot would probably be a natural language interface for a dispatching system instructing and coordinating specialists. And if manually placing my dishes in a receptacle or sorting my laundry is all I have to do, I'll be happy.