Automated laundry folding is working just fine at industrial scale.[1][2]
Chicago Dryer, a century old company in Chicago in a very boring business, is doing quite well at automated folding. Their approach is about 20% AI, 80% clever machinery, and it gets the job done at scale.
Damp laundry fresh from the washer goes in, and folded stacks come out.
There's a vision system with some AI finding corners for a robotic grabber. Another vision system decides if items are worn or damaged and should be sent to the reject bin.
When I hear "folding laundry", I think "taking a load of laundry from the washing machine, including pants, shirts, underwear, bras, socks, sheets, towels, pillowcases with socks inside them and cycling pants, separating them, folding them, and putting them in stacks for each family member"
The automated solution in the video seems to work for loads of identical white towels.
Don't get me wrong, the automated laundry machines are impressive, but I don't think they accomplish what people thought of when they asked the question.
There was Foldimate, a startup which got publicity in 2019. They went bankrupt.[1]
Willow Garage demonstrated folding with their mobile robot. They shut down when the funding ran out. There was Laundroid, priced around US$ 15000, bankrupt 2019.[2]
Chicago Dryer's machines can fold shirts and pants. This is mostly used for commercial workwear, such as hospital uniforms. They have all the steps in the process automated.
If you buy all the options, you can start with a bin of damp laundry direct from an extractor and get folded items out. (Other companies make washing systems with automated load and unload to feed the "clean side" of the process.) There are vision systems, machine learning, and robotic grippers inside those big machines.
So home-sized is possible, but not cost-effective. Boring old Chicago Dryer is cost-effective compared to low-wage labor, and they have real products that sell.
The 2022 survey is interesting - he marks "Write python code to implement algos like quicksort" and "Answer easily googleable factual but open-ended questions better than an expert" as "already happened" and the experts estimate 3 and 5 years in the future - this points to his expectations for what "happened" different than the experts.
Like for the former "implement quicksort" they might be answering "write quicksort given an understanding of python and an understanding of the problem" rather than what happened which was "implement quicksort by reconstructing text from corpus".
And on the latter - I'm not sure that we've achieved the "answer easily googleable factual questions better than an expert", unless by better he means answer it with more confidence than an expert (sarcasm). LLMs will flat out make up facts for factual questions and answer with complete confidence and there is no easy way to extract structured data out of those answers to confirm them.
I think the idea of the marking these items as "these happened" is very handwavey and removes some of the interesting things about the survey. The most interesting thing in this article is where he compares the two articles and tracks deltas between the time estimates - I think that's a strong indicator of what we're finding is easier than we though in AI vs what is proving to be a bigger problem than we though in AI.
What would it mean for an AI to beat the best humans in a 5K race? Obviously robots can do that, is there some definition of "humanoid AI robot" which exists for such a task?
I think to have meaning the AI would have to work both as generally and usefully as a human, and fit entirely within a human size profile (not offloading compute remotely). What makes humans competing against each other even remotely worth paying attention to is the constraints, which is that we're all autonomous, we all have similar size, we all can function to a similar degree.
Plenty of animals can beat humans at all sorts of things, but they can't do so while also standing in as a human, so the comparisons aren't very interesting. Building a robot that has a human profile and can run and beat humans at a 5k race also isn't very interesting (at this point), because fine, you created a running robot, but that robot's not going to do it's own taxes at the end of the year, so in the sense of AI, who cares.
> beat humans at a 5k race also isn't very interesting (at this point), because fine, you created a running robot, but that robot's not going to do it's own taxes at the end of the year, so in the sense of AI, who cares.
It depends exactly on what we're saying. So far we have machines that can navigate language well, but aren't good at navigating real world environments. If we're talking about a road 5k race, with myriad ways that courses are marked and many, many distracting, this is at least an interesting problem (perhaps as difficult as self driving vehicles).
Being able to navigate the course is an interesting problem. I'm not sure beating a human at it is, unless we change the problem to be accomplishing something useful (lugging a certain amount of weight up a cliff face for rescue operations, for example), or make the contestant AI sufficiently human like.
A 5k race is about competition, not about doing useful work. For AI or a robot to compete and win and have it mean anything means it needs to be similar enough to the other contestants to make that competition matter, otherwise I could slap a picture of my face on an RC car with enough batteries and "win". We all realize intuitively that a "win" such as that tells us nothing useful, so would ignore it even if it somehow was allowed. AI is a bit less intuitive to many at this point, so it's not always as obvious.
Well, sure. IBM's Jeopardy PR stunt with Watson is an instance of that: it was definitely "buzzer doping."
At the same time, included in that "win" was a big technical feat. So it would be with winning a 5k race.
(Not to mention that right now nothing we have capable of bipedal or even quadripedal locomotion is capable of the feat, so if we rule out the RC car it's quite the mechanical accomplishment, too--- even if we're throwing a ridiculous amount of power density and big actuators at the problem).
Beyonh Watson buzzer doping, it's interesting but not as much as some people make it out to be. Since it isn't an autonomous package equivalent to a person. At the time, it was the size of a master bedroom. If we compared it to a room jull of smart people, does Watson still seem impressive in it's feat?
Similarly, getting a mechanical package that can traverse a course like a human using locomotion like a human, which would be much more impressive if it's not running of flat asphalt. But like you note, we can't do that even entirely abstracting away the AI portion, so the question of what will happen when an AI wins a 5k is mostly moot, there's many steps to get there that we haven't gotten close to, and even when we've solved all the aspects separately (a general AI, locomotion, power density) it will likely be a while after that (if ever) before they are solved together in a package that compete.
Im not saying humans are the epitome of these systems come together, but I doubt evolution has left us with a completely horrible design, especially if were talking about thinking and running, two things humans are known for being quite good at (to our own knowledge).
Eh, picking arbitrary metrics like that -- the size of Watson -- isn't too useful. For most of these things, like Watson-- we're concerned about whether it could do it more cheaply (less resource intensively) than a human. Watson also does the job on much less sleep and doesn't get bored.
(And a room of smart people is not likely to be much better at this kind of quick decision task than a single smart person).
And now, you can fit 16TB of ram in a couple rack units with boring hardware, though you'd need perhaps a quarter rack to get equivalent memory bandwidth.
> but I doubt evolution has left us with a completely horrible design, especially if were talking about thinking and running, two things humans are known for being quite good at (to our own knowledge).
LLM's are making me less sure that we're so good at thinking. We may be good at making quick decisions and navigating social hierarchies in a relatively low power budget, but that's a different thing...
> we're concerned about whether it could do it more cheaply (less resource intensively) than a human.
That's what I was getting at before. For a specific work based competition about who can do a specific job better, I agree. For a general "run five kilometers" case without any other explicit constraints, those constraints are basically "be human", because it's about competition.
Sometimes we even break it down more than that when "being human" isn't enough to provide useful information across the population of contestants. There's a reason some competitions are broken into separate categories for the sexes. Putting biological males and females together in power lifting competitions is less useful from a standpoint of determining how like organisms compete than not (which is one of the reasons, along with using substances to move someone from one category closer to another, people get upset about gender in sports competitions now).
> (And a room of smart people is not likely to be much better at this kind of quick decision task than a single smart person).
To a specific degree it certainly does. As many people as you can fit around the buzzer, or giving each person their own buzzer that connects to the central "real" buzzer seems like a simple enough solution to yield good results - as long as false buzzers are low enough.
> And now, you can fit 16TB of ram in a couple rack units with boring hardware, though you'd need perhaps a quarter rack to get equivalent memory bandwidth.
Don't forget power. If it's not solar powered, you're forgetting a very large aspect of what it means to be a human. The equivalent "human" would likely just be a brain with some electrodes for input and output, so would still be much smaller, if it was possible. Until Watson is able to move itself to a power source, sip from that power to refill some battering, and them be autonomous for a few hours, even it it's presented as the same "size" as a human I'll not be convinced.
Hell, they can start by ignoring the movement and just giving it a battery pack that can actually sustain it for a few hours of operation in a similar size package. We're closer to that, but I'm not sure where there yet (it probably depends on how much we care about weight or mass compared to volume).
> LLM's are making me less sure that we're so good at thinking.
That' funny, LLM's are doing pretty good and convincing me most the time we're thinking in a similar manner, with the same problems. ;)
What an insightful article. I vaguely remembered the questionnaire mentioned in the beginning paragraph, and how there were some inconsistencies. Glad I read the rest of the article before jumping in here to comment, because the author goes into detail about exactly those issues, where they might stem from, and what that says about the prediction domain.
It's really well written - I haven't heard of asteriskmag before, glad I did.
"Provide phone banking service as well as a human".
We got to that point, but mostly because of how much human provided service has degraded. I don't blame the humans in question, but how much they are trained to behave like robots, and the little they are trusted and allowed to do.
> We should, as a civilization, operate under the assumption that transformative AI will arrive 10-40 years from now, with a wide range for error in either direction.
So, this is not actually saying much. Anything that far out may well also just end up not happening, and it’s too vague an assumption to usefully operate on.
I see your point, but sometimes it's good for HN readers to work a little (just a little!), and I feel like this is one of those cases. Not everything should be explicitly spelled out - especially when the title isn't obviously baity and the article is substantive.
With the link to Algolia, I feel like "work a little" is a tagline you are trying to promote.
But look at it this way. Each title on HN is an advertisement to learn about an interesting...something.
Does the title interest you enough to read the article? Because if it doesn't accomplish that task, it doesn't get clicked on, or potentially not clicked on by the correct people.
Also, copying the first few lines: "Through a Glass Darkly Scott Alexander Nobody predicted the AI revolution except"
For the future: @dang, is that ok?
Damp laundry fresh from the washer goes in, and folded stacks come out. There's a vision system with some AI finding corners for a robotic grabber. Another vision system decides if items are worn or damaged and should be sent to the reject bin.
[1] https://www.chidry.com/products/genesis
[2] https://www.youtube.com/watch?v=7bd900ehE9M