The key here is that it’s depending on the human inability to connect the sum of relevant knowledge, but said knowledge comes from humans.
Theres going to be this field day of low-hanging fruit that ML can round up, but after that I suspect it will be in fits and starts as a “connection maker” rather than some proof producer.
I think we're underestimating just how much low hanging fruit there is. I've been trying to apply this LLM research process to physics (QM and solid state) and there is so much missing in Physlib and the rest of the Lean ecosystem that most of my work has been trying to formalize the theories and validating them against the specification problem (and mostly failing badly).
It's not only going to be "connection maker". If and when robotics advance to a point where the LLMs are embodied, they can run experiments in the physical world and find new knowledge.
So applied math, math, applied math, applied math, and maybe some others.
That makes sense, although I'd argue that at least in the realm of HEP theoretical physics has extraordinarily expensive kit compared to what scientists make. See: The LHC.
I think it's telling that "Researchers cost more than any other aspect of research" seems to involve primarily math as inputs and outputs. Like I said the obvious outlier is theoretical physics because that tends to involve significant resources both to explore (supercomputers) and verify (colliders, telescopes, interferometers, etc). Most people involved in research in any number of fields do not find that their collective salaries are the majority of the cost required; math is an outlier in many cases.
Robots in labs already exist, but mercifully they're not hooked up to anything as unpredictable as an LLM. Robots tend to work best as specialists doing high-throughput, extremely repetitive tasks which nonetheless require a degree of precision. Giving a robot a "human" body makes very little sense if we're talking about the needs and productivity of a non-human; humanoid robots are marketing for humans.
Humanoid robots are obviously more than marketing. The entirety of human civilization is human shaped. Making robots that are human shaped is easier and more efficient than redesigning and rebuilding everything that exists.
> The entirety of human civilization is human shaped.
That's the marketing pitch.
A plumbing robot doesn't need to be humanoid, an octopus shape may well be better for all the awkward corners. A robot police officer could be the municipality itself for sensory nodes (essentially the sales pitch of Flock etc.), plus some drones or robot dogs to perform arrests*.
The robot vacuum cleaners and lawnmowers we already have are nothing like a human. A robot taxi driver can be just the car. Robot dogs are already used for maintenance and security sweeps.
If you've got wheelchair access, you've got wheeled robot access. If you've got guide dog access, you've got access for Boston Dynamics' Spot.
* this may be a bad idea with current robotics, but I aver it's not improved by making those robotics humanoid.
You are not understanding what humanoid robots are about. Those are specialist robots you are describing. The promise is of course one robot that can do the plumbing, clean your house, do the dishes, build a house, and basically every physical job a human can do. It's extreme lly likely that at least a somewhat humanoid shape is required for that.
Humans are really bad at everything we do compared to the specialized creatures in nature, it’s just that we can do many things that sets us apart.
A humanoid robot is incredibly dumb for that reason. At least add a set of arms and legs and 360 vision. And obviously dislocated joints than can move freely and hands should have two thumbs and more fingers.
It’s not hard to come up with a bunch of improvements for humans, it’s just that making robots in our image is a lot more trivial because you only have to solve for those same averages attributes that we have.
Adding arms and 360 vision is not free. Look at octopi with their insane nervous system required to support their tentacle. The one who wins is not the one that makes the best robot ever. The one who wins is the one who builds a robot that can do the most, while being the cheapest.
Vision is hard regardless of the angle, but scaling it from a normal camera to 360° doesn't add much cost or difficulty once you've done the hard part of turning pixels into a suitable latent space.
> Look at octopi with their insane nervous system required to support their tentacle.
500 million neurons across the whole animal, brain included. We'll only know the synapse count when someone does a full connectome scan of one, but based on the vague estimates I see with a quick search, their whole mind is less complex than a SotA LLM today.
> The one who wins is the one who builds a robot that can do the most, while being the cheapest.
This is a reason for specialists, not generalists.
The best container ship is a terrible pleasure yacht, and vice versa. You use container ships pretty much constantly, even though you (almost certainly) don't own one, by the power of indirection. You can rent a pleasure yacht as desired without owning own.
I will ignore the insane comparison between neurons and LLMs. But
> This is a reason for specialists, not generalists.
Precisely not. The amount of work a human can do is basically unenumerable. Requiring a specialist for each task will quickly run up costs to be completely unreasonable.
> I will ignore the insane comparison between neurons and LLMs. But
Please don't.
I mean, you yourself specifically said to "Look at octopi with their insane nervous system", and that's exactly what I'm doing and my conclusion is "doesn't look that bad compared to what else we do now".
Also note that I expressly noted that synapses, not neurons, are the important factor in parameter count.
> The amount of work a human can do is basically unenumerable.
> Requiring a specialist for each task will quickly run up costs to be completely unreasonable.
The existence of specialists in professional settings proves this claim false.
As does the existence of special-purpose robots, all the way down to injection moulds designed for one specific part, which overall make the end product cheaper, a process which probably started with whoever invented whichever came first of the wind or water powered grain mill.
Much of the economic gain of the industrial revolution was replacing the very expensive salaries of general-purpose humans with machines that, despite being extremely single-purpose and having high purchase and maintenance costs, were still cheaper than the wages of the huge number of humans needed for equal output.
Since the industrial revolution we have been the species that makes tools even though it sometimes also uses slavery, not the species that relies on slaves and has tools for decorative reasons.
Yeah. Can't wait to see three or four armed robots. Or five or six!
How do you train a robot to use three hands effectively when we only have two? Then again, why is the robot limited to being one robot? If two humanoid robots are in the same area, they don't have to be distinctly controlled. If they're both controlled by the same AI, a third arm on one body is the same as that arm being attached to another body.
Things you need to take into account that you've almost certainly glossed over:
Safety. Human-robot interactions are generally dangerous and avoided, unless the robot is specifically designed to interact with people. In those cases you often sacrifice speed, strength, and flexibility for safety and softness. Having someone come in with a specialized plumbing robot makes sense, you owning one probably doesn't, and you owning a generalize android capable of plumbing makes less sense still.
Cost. The more compact, complex, and interactive your robot is the more it costs. Make a strong, compact, complex robot safe for interactions with people in the wild is non-trivial and adds costs. The software required to do all of this is hypothetical, but obviously also costly.
Need. I understand the dream of a robot to do whatever you want is very much part of our culture, but when you consider the downsides do you really need it? I don't need a plumber living in my house any more than I need a carpenter or a landscaper to live on premises. At most these are services I would need occasionally or on a schedule. I also doubt my need for them will overlap much, unless we're talking about building a new dwelling.
So why do I need a generalist in my life that's going to cost more than you can imagine, when the means to hire existing human generalists is cheap, quick, and frankly less likely to accidentally punch a hole in you.
Generalist humanoid robots that are able to operate in unstructured environments weren't an option (they still aren't an option for the majority of operations). The humanity had no ways of building them.
Anyway, it's true that replacing an automated production line with a crowd of generalist robots doesn't make sense. Generalist humanoid robots are intended to replace the remaining human workers.
They weren't an option because they are much harder to build, that's my point. Fat from being the easy option, they are still the unlikely option, compared to purpose designed machines.
Historically this has never been correct. Turns out you get more efficient systems when designing them without how a human would accomplish a task in mind.
Outside of sci-fi, marketing proposals, and niches like "Elder care in Japan" there are very few humanoid robots. By contrast non-humanoid robots have been mass produced and used in industry for decades. Arms. Carts. Trollies.
No people. If you want something with fine motor control and dexterity, it's easier to make that the robot and then have another robot bring the workpiece to the arm than it is to build a single robot that can walk around and do it. There are compromises in human features because we're generalists.
Theres going to be this field day of low-hanging fruit that ML can round up, but after that I suspect it will be in fits and starts as a “connection maker” rather than some proof producer.