The third question from this post: https://garymarcus.substack.com/p/dear-elon-musk-here-are-five-things
The full text is: "In 2029, AI will not be able to work as a competent cook in an arbitrary kitchen (extending Steve Wozniak’s cup of coffee benchmark)."
Judgment will be by me, not Gary Marcus.
Ambiguous whether this means start or end of 2029, so I have set it for the end.
A humanoid robot using a simple coffee machine to make coffee, not huge evidence but still notable https://twitter.com/adcock_brett/status/1743987597301399852
@MartinRandall If this is intended as a criticism of the criterion, it is a bad one.
He could do that with a few weeks of training. I think if an AI can do it with a few weeks of training, that is pretty much as good as being able to do it without the training.
@DavidBolin On the object level, I doubt that Gary Marcus could be a competent cook in an arbitrary kitchen with a few weeks training in that kitchen, but of course that depends on "arbitrary" and "competent" which are ill-defined. Most of my evidence on rapid career switches comes from reality TV shows, which imply greater than a few weeks retraining.
I also think that being able to work as a competent chef in an arbitrary kitchen with no additional training is much more impressive. In particular it implies adversarial robustness that is hard to achieve in AI with current techniques. In humans it implies a depth of experience in cooking, linguistics, and world culture that is rarely achieved.
We're both betting NO so I think we both think this milestone is unlikely by the date. I'm not really criticizing the criteria, just pointing out that it's superhuman to justify my bets.
There's no economic reason to develop this rather than adapting kitchens for the needs of AI cooks.
My personal impression: of the 5 Gary Marcus questions this is the hardest. The other four are "shovel-ready" in the sense that I think the other four may not require additional breakthroughs, but AFAIK solving the robotics problem here would require at least one major advance (I do not follow the robotics literature very closely, however)
@vluzko This would need AI to be applied to robotic control, much better object recognition, the ability to understand a wide variety of non-objects (e.g. soup, mashed potatoes), the ability to navigate an arbitrary tight space, the ability to safely navigate around humans, and major advances in robotic limb design.