"humans will become so enamored with interacting with AI devices and ultimately, AI robots, we will lose interest in other humans and perhaps gradually die out."
Something like that seems to be happening with smartphones. But it's not losing interest in people. Quite the opposite. It's becoming so interested in unusual or entertaining people that boring ordinary people can't compete. And those entertaining people don't stick around and bother you if you don't want them to. They show you their best, that is, their best in the way of entertaining you--because that's why you've chosen to watch them.
This may eventually lead to loneliness, and a desire for actual people, but great difficulty in actually doing anything with other people because you never developed the knack, and, God, they are so borrrrring some times.
"In 2003, the typical female pet owner spent much more time socializing with humans than playing with her cat or dog. By 2022, this flipped, and the average woman with a pet now spends more time 'actively engaged' with her pet than she spends hanging out face-to-face with fellow humans on any given day....
"From "Why Americans Suddenly Stopped Hanging Out/Too much aloneness is creating a crisis of social fitness" (The Atlantic)."
I think you couldn't be more mistaken. Clearly you aren't an engineer even if your programmer expertise should give you some insight.
The tech for "b" requires far more mechanical components that tend to be expensive, never mind that more parts means more maintenance.
The potential of "a" is that AI software allows much reduced mechanical that is far more expensive on a per unit basis. Option "a" has much greater potential of being effective at a reasonable cost.
The statement: "But my guess is that the Fed is not eager to cut down on staff. Nor are firms in the securities business. Nor is the legal profession." is highly understated when referring to government monopolies.
Rao is being visionary, yet attempting to describe a current reality in flux, with noun labels that become better defined as they are used.
His 4 circle Venn diagram is really great: Embodiment, Boundary Intelligence, Temporality, Personhood.
Physical form, data flow-processing, time, I-you relations rather than I-it.
Including lots of new terms, like LMM-Large Multimodal Models as part of Embodiment, which I’ve maybe seen, but not yet used (thus not sure I know what they mean).
I never understand the all in one approach. at the very best you get average performance of everything included, at worst the car plane. Then if one part breaks, it is much more costly to fix or replace. But people keep trying and failing. Except for the smart phone. So who knows.
It’s a good way to think about the first point, in terms of the distinction between hardware and software. The question then becomes what is better for the domain of software and what is better for the domain of hardware?
Software’s main advantage is that it is relatively cheap to implement and highly flexible; Hardware is expensive to implement and relatively rigid, but, once implemented, is generally faster and more decisive/determinative in its action.
A humanoid robot is useful becuase the world is adapted to interface with a human. And so if you want a single robot that can do many tasks that humans do, a humanoid robot makes sense.
I do think, if you’re talking about a kitchen robot, you don’t want a humanoid robot. What you want is something that complements what humans can do, that offers us some improvements over the human design. Humans are extremely flexible and general purpose. But I want a machine in the kitchen that is extremely good at doing the task that I don’t particularly want to do or that I’m not well adapted to do. That’s the nature of kitchen appliances: I just bought a Japanese rice cooker. It is much better at executing a special-purpose activity – i.e., making rice – than I am using more general purpose tools – i.e., a gas stove and pot.
Now, if I wanted to replace people in a McDonald’s kitchen, maybe a humanoid robot would be good as a literal one-to-one replacement for people. Maybe it would be good if my purpose was to change menu items frequently. But in a McDonalds, currently, the priority is in producing a few items, quickly and efficiently. And it seems like a special purpose robot built for that task would be better.
I’m just kind of rambling, it’s an interesting question to think about.
> Software’s main advantage is that it is relatively cheap to implement and highly flexible; Hardware is expensive to implement and relatively rigid, but, once implemented, is generally faster and more decisive/determinative in its action.
> A humanoid robot is useful becuase the world is adapted to interface with a human. And so if you want a single robot that can do many tasks that humans do, a humanoid robot makes sense.
Consider this same idea in the context of Daniel Kahneman's System 1 vs System 2 cognition and you'll really be onto something!
I just finished re-reading "Caves of Steel" by Asimov (fun book gives a weird perspective on how the future will look). In the book the case for humanoid robots is that the cost of the positronic brain. The brain cost so much that it would be prohibitive to put on in each device. Today, microcontrollers that run small robots cost only dollars so obviate the need for expensive brains in each robot. Gates should consider that no one is developing a humanoid robot to drive a car. All the work on autonomous cars uses onboard computers. Adding a humanoid robot adds a significant level of difficulty.
On the other hand, if I could get a humanoid robot to get out of bed for me...
Thanks for taking the time to comment. i think it comes down to what job you are doing. A dishwasher and a human both can wash dishes. Both do the same job very differently. Which is better? i don't know. Would a humanoid robot load the dishes into the dishwasher or do them by hand.
AS for driving a car. A robot could access all of the car's sensors through BlueTooth. This would give it an edge.
Not sure how it will play out. i expect that some tasks will be performed by a specialized robot and some will be performed by a humanoid robot. i hesitate to pick which tasks.
The L for language has two possible meanings which can both be considered parts of a more abstract concept. In terms of analyzing text as a training set, x means any kind of category of data of a type that tends to be structured in patterns. In terms of prompting or instructing for generation, action, or analysis, a human's text piece can be a step removed from the prompting stimulus, and one can anticipate that soon enough the AIs will self-refine, create, and interact with each other in ways that are too distinct from human communication modes to be termed "language", maybe something more like "protocol".
Do you think thinking of it in terms of ~"representative (of reality) tokens", and not getting overly concerned for now about what that exactly means, is the way to approach it? This feels about right to me.
"I remember in the 1990s that every time Freddie Mac issued a mortgage security, staffers would have to review the offering circular. That sounds like something to outsource to a large language model."
Some lawyers have already tried to outsource their legal research to large language models. I don't know whether it has ever worked well, but there have been two cases in the news where the AI cited legal cases that didn't actually exist, leading judges to sanction the lawyers.
Of course, humans can make errors as well. Do you think LLMs are already capable of doing this work as well as humans? If so, why? If not, will they be able to do so in the future? Will humans still need to check all of the AI's work? If so, what's gained by using the AI?
"humans will become so enamored with interacting with AI devices and ultimately, AI robots, we will lose interest in other humans and perhaps gradually die out."
Something like that seems to be happening with smartphones. But it's not losing interest in people. Quite the opposite. It's becoming so interested in unusual or entertaining people that boring ordinary people can't compete. And those entertaining people don't stick around and bother you if you don't want them to. They show you their best, that is, their best in the way of entertaining you--because that's why you've chosen to watch them.
This may eventually lead to loneliness, and a desire for actual people, but great difficulty in actually doing anything with other people because you never developed the knack, and, God, they are so borrrrring some times.
And there is this from Ann Althouse today:
"In 2003, the typical female pet owner spent much more time socializing with humans than playing with her cat or dog. By 2022, this flipped, and the average woman with a pet now spends more time 'actively engaged' with her pet than she spends hanging out face-to-face with fellow humans on any given day....
"From "Why Americans Suddenly Stopped Hanging Out/Too much aloneness is creating a crisis of social fitness" (The Atlantic)."
https://althouse.blogspot.com/2024/02/in-2003-typical-female-pet-owner-spent.html
It is astounding how many people will unself-consciously refer to themselves as "dogmom".
A vs B
I think you couldn't be more mistaken. Clearly you aren't an engineer even if your programmer expertise should give you some insight.
The tech for "b" requires far more mechanical components that tend to be expensive, never mind that more parts means more maintenance.
The potential of "a" is that AI software allows much reduced mechanical that is far more expensive on a per unit basis. Option "a" has much greater potential of being effective at a reasonable cost.
The statement: "But my guess is that the Fed is not eager to cut down on staff. Nor are firms in the securities business. Nor is the legal profession." is highly understated when referring to government monopolies.
Rao is being visionary, yet attempting to describe a current reality in flux, with noun labels that become better defined as they are used.
His 4 circle Venn diagram is really great: Embodiment, Boundary Intelligence, Temporality, Personhood.
Physical form, data flow-processing, time, I-you relations rather than I-it.
Including lots of new terms, like LMM-Large Multimodal Models as part of Embodiment, which I’ve maybe seen, but not yet used (thus not sure I know what they mean).
I never understand the all in one approach. at the very best you get average performance of everything included, at worst the car plane. Then if one part breaks, it is much more costly to fix or replace. But people keep trying and failing. Except for the smart phone. So who knows.
It’s a good way to think about the first point, in terms of the distinction between hardware and software. The question then becomes what is better for the domain of software and what is better for the domain of hardware?
Software’s main advantage is that it is relatively cheap to implement and highly flexible; Hardware is expensive to implement and relatively rigid, but, once implemented, is generally faster and more decisive/determinative in its action.
A humanoid robot is useful becuase the world is adapted to interface with a human. And so if you want a single robot that can do many tasks that humans do, a humanoid robot makes sense.
I do think, if you’re talking about a kitchen robot, you don’t want a humanoid robot. What you want is something that complements what humans can do, that offers us some improvements over the human design. Humans are extremely flexible and general purpose. But I want a machine in the kitchen that is extremely good at doing the task that I don’t particularly want to do or that I’m not well adapted to do. That’s the nature of kitchen appliances: I just bought a Japanese rice cooker. It is much better at executing a special-purpose activity – i.e., making rice – than I am using more general purpose tools – i.e., a gas stove and pot.
Now, if I wanted to replace people in a McDonald’s kitchen, maybe a humanoid robot would be good as a literal one-to-one replacement for people. Maybe it would be good if my purpose was to change menu items frequently. But in a McDonalds, currently, the priority is in producing a few items, quickly and efficiently. And it seems like a special purpose robot built for that task would be better.
I’m just kind of rambling, it’s an interesting question to think about.
> Software’s main advantage is that it is relatively cheap to implement and highly flexible; Hardware is expensive to implement and relatively rigid, but, once implemented, is generally faster and more decisive/determinative in its action.
> A humanoid robot is useful becuase the world is adapted to interface with a human. And so if you want a single robot that can do many tasks that humans do, a humanoid robot makes sense.
Consider this same idea in the context of Daniel Kahneman's System 1 vs System 2 cognition and you'll really be onto something!
I just finished re-reading "Caves of Steel" by Asimov (fun book gives a weird perspective on how the future will look). In the book the case for humanoid robots is that the cost of the positronic brain. The brain cost so much that it would be prohibitive to put on in each device. Today, microcontrollers that run small robots cost only dollars so obviate the need for expensive brains in each robot. Gates should consider that no one is developing a humanoid robot to drive a car. All the work on autonomous cars uses onboard computers. Adding a humanoid robot adds a significant level of difficulty.
On the other hand, if I could get a humanoid robot to get out of bed for me...
How is a kitchen helper comparable to a driving helper? I don't see the similarity regarding devices needed to do the jobs.
Just the same, if one had a humanoid robot, why couldn't it drive the car?
Thanks for taking the time to comment. i think it comes down to what job you are doing. A dishwasher and a human both can wash dishes. Both do the same job very differently. Which is better? i don't know. Would a humanoid robot load the dishes into the dishwasher or do them by hand.
AS for driving a car. A robot could access all of the car's sensors through BlueTooth. This would give it an edge.
Not sure how it will play out. i expect that some tasks will be performed by a specialized robot and some will be performed by a humanoid robot. i hesitate to pick which tasks.
I agree with two caveats.
A specialized robot has far less need for AI.
I think Kling's kitchen helper leans far more towards humanoid than specialized.
You should follow Rao and start using LxM instead of LLM because things have already moved beyond text and that's where all the future action is.
What things other than "L" do we find under "x", that aren't translated to "L" (tokens) as an intermediate step?
The L for language has two possible meanings which can both be considered parts of a more abstract concept. In terms of analyzing text as a training set, x means any kind of category of data of a type that tends to be structured in patterns. In terms of prompting or instructing for generation, action, or analysis, a human's text piece can be a step removed from the prompting stimulus, and one can anticipate that soon enough the AIs will self-refine, create, and interact with each other in ways that are too distinct from human communication modes to be termed "language", maybe something more like "protocol".
Do you think thinking of it in terms of ~"representative (of reality) tokens", and not getting overly concerned for now about what that exactly means, is the way to approach it? This feels about right to me.
"I remember in the 1990s that every time Freddie Mac issued a mortgage security, staffers would have to review the offering circular. That sounds like something to outsource to a large language model."
Some lawyers have already tried to outsource their legal research to large language models. I don't know whether it has ever worked well, but there have been two cases in the news where the AI cited legal cases that didn't actually exist, leading judges to sanction the lawyers.
Of course, humans can make errors as well. Do you think LLMs are already capable of doing this work as well as humans? If so, why? If not, will they be able to do so in the future? Will humans still need to check all of the AI's work? If so, what's gained by using the AI?
reading an offering circular is an exercise in proofreading, not research