Hacker News Reader

Comment by LarsDu88 | Hacker News Reader

I had the same thoughts as much as 2 years ago on how this will play out. Unlike with railroads and fiber optic cable however, the core infrastructural asset of GPUs tends to become rapidly obsolete after about 5 years.

Commoditization of this scale of compute is definitely going to be a boon for many fields of research. Unfortunately fundamental public research is exactly what is being cut right now in the US.

Long term, I think the real winners are going to be in robotics. Still an unsolved field, but Waymo proves that even a nearly 20 year slog to the finish line is viable. And robotics infrastructure may be more robust to obsolescence than the underlying compute. I find it odd so many companies are making humanoid robots though... Over engineering that reeks of bubble economics and possible fraud.

View on HN

I think the allure of humanoid robots is that they are drop in replacements for agents in a world desgined for humans.

If you want your robot to be a helper around the general populations houses for example, you would aim to make a general purpose bot capable of stairs, ladders, lying down, reaching high, stepping over things, holding awkward weights and loads while doing all of the above. Pinch, twist, push, pull, in all degrees of motion a human has etc.

Completely off-topic: I find it odd that we easily use this argument for humanoid robots (and also self-driving cars), but handicapped people are still bound to wheelchairs and have to constantly fight to change the environment and make it wheelchair-accessible.

If we applied the same logic, there should be a massive effort to ditch wheelchairs and build exoskeletons instead.

Much like wheelchairs vs exo skeletons, the simpler (and cheaper) tech tends to win. I'd imagine in a future where robots are everywhere, they'll use whatever cheap locomotion is appropriate for their tasks, probably predominantly wheels. It's a fun vision to imagine bipedal robots everywhere like in old sci movies, but I'm not convinced that's how it'll play out, the economics don't make that much sense.

Bipedal robots are more expensive to develop, build and maintain, more limited in their payloads, and because of the additional complexity, less reliable.

Cost will be the reason why these over engineered robots fail. Cost is the reason why indoor vertical farms fail, even though in their case, the environment is controlled and the tech is relatively simple compared to omnipotent field robots.

The most viable use case of AI is bullshiting humans, which is still a multi-billion market. Infrastructure hooray!

I look at this a bit more generously: We're pushing the boundaries of what's technically possible. At least for niche use cases, I'm convinced bipedal robots and GenAI will have lasting value. Are they the next automobile / electricity / smartphone? I'm sceptical.

Reminds me of some of the early designs imagined for automatic dishwashers which had articulated arms and hands. The dishwashers we have are nothing like what was envisioned by those early designs. It just takes some time to break out of the pattern of thinking that the recent past imposed due to the technology of that time.

> If we applied the same logic, there should be a massive effort to ditch wheelchairs and build exoskeletons instead.

Knowing the kind of markup on wheelchairs that means a YouTuber wheel chair look like a bargain (see Jerry rig everything wheelchair), I can't imagine how much the US healthcare "industry" would charge for a "medical grade" exoskeleton.

For disabled people who can use their arms and propel themselves, a wheelchair is low-maintenance, dependable, repairable, self-powered, and generic enough for there to be a healthy market and not too much vendor lock-in. Wheelchairs are not necessarily low-tech, but they are fully understandable. If the user of a wheelchair is unable to move themself for whatever reason, any able bodied person can step in and push. These things matter in any situation I think, but they matter even more in an emergency or a war zone.

Exoskeletons can't match that.

All of those are arguments for why robots should generally have wheels rather than legs, except for when legs are specifically needed.

> except for when legs are specifically needed.

Exactly, we need legs when they are specifically needed, and we already have wheeled robots so building legged robots that can move like a human will cover so many cases we currently cannot cover.

And even more important are arms and hands, and legs is a precursor to that, they are much simpler so its smart to start with legs to then try to make good arms and hands.

Give me the option of a humanoid looking across that takes care of all in house chores and one that's that utility based with wheels and arms, I will likely choose the humanoid one even with a 100% premium price.

I mean I wouldn't buy either unless I could be certain it's not uploading all data to the cloud and be fully controlled by a user hostile company, but if we're talking fantasy tech ala Detroit: become human... Yeah, it'd be willing to spend a lot of money to have all chores taken care of by a humanoid robot.

And in before someone talks nonsense again wrt "you already can, just pay someone to do it for you"... I do not want to have strangers in my home. This is also essentially why I wouldn't want any cloud connected bot anywhere innit.

The cloud situation is where it's probably going to fall down at first (haha). I don't see companies choosing to offer local model integration over the possibility of using the robot as a loss leader to a long term subscription model for access to the compute/inference.

But that's going to be hilarious. Imagine your internet goes down while the bot is half way down your stairs, or the in the middle of pouring a drink. Very fun.

Furthermore, mobile robots currently in home use--vacuum and mop robots--are all wheeled, of course. We've shown we can accommodate wheeled robots in the home if we feel like the payoff is worth it.

Well I think wheels match the use case there, a small bot close to the ground with just the one job. I think there will be many wheeled bots to begin with. But long term I don't see that form factor scaling to "able to do all tasks around the house".

It's super easy to come up with scenarios that a wheeled bot can't cope with, but again "good enough, cheap enough" will probably see lots of wheeled bots on the market. I am just trying to show why the pioneering companies would be interested in bipedal bots, it's a long term play.

Lastly, the elephant in the room is that basically all general purpose bots are a euphemism for military bots that will need to operate in unknowable conditions.

I'd say the same thing about mobile robots, especially since ADA means most of the public environment should be relatively accessible.

I hear you, I think there's some obvious points you've already considered on price and serviceability. But maybe we do see that coming soon? Exoskeletons currently require some mobility from the legs from the user in the products available today, but as automatic bipedal motion gets more reliable maybe that changes.

I have seen how robots currently behave when they lose their footing though, and I'd be bloody terrified to be strapped into one.

Maybe wheelchair users and robot manufacturers can share a force for getting wheeled locomotion into more spaces, but I think homes will always be a challenge as stairs are a requirement for denser living, and elevators are expensive.

There just aren't that many people in wheelchairs. There's not a lot of TAM in trying to build a better wheelchair that involves a huge capex.

So instead the government gets involved and demands a change to built environment instead of a speculative bet on the idea of a new technology.

I love how most people can't even afford a car without taking loans but we're talking about personal robot assistants as if it was right around the corner and the natural evolution of things

With the wealth gap increasing ever further, I think we will see technology in lock step with that divide.

I doubt robots will actually end up in every household, but a niche luxury product and utility for businesses makes some sense. Even if you think about it from that perspective, robot makers would still want them to be a universal robot not dozens of unique use case bots.

If a business can pay 30k for a general purpose extra set of hands I think that would be a no brainer, and I think the wealthy would see it similarly.

>I doubt robots will actually end up in every household, but a niche luxury product and utility for businesses makes some sense.

Sounds like a limited market/growth potential, hard to amortize the huge R&D etc. Could happen but will never justify the current levels of investment required.

As wealth becomes more skewed, it won't matter that most people can't afford a robot butler. The smaller proportion of people with the majority of the wealth will make it worthwhile.

Such robots aren't more expensive than cars though, you can buy a humanoid robot today for a fraction of the cost of a car. They lack intelligence but they can move around.

Increasing wealth inequality will increase this type of thing. William Gibson novels explain it.

The allure of humanoid robots is they don't have human rights while being able to replace humans. If it turns out that they end up being conscious, their lack of humanity will be all humans need to justify enslaving them. See: animals.

The half-life of iron is pretty low too, the advantage of the rail system is what it allowed us to do when it was cheap enough.

All the investment in AI should help bring infrastructure up to a higher level, power distribution and cooling for example are at a much higher level than would have otherwise been.

Who knows what use that might have if it suddenly becomes incredibly cheap.

(this is my silver lining thinking)

The half-life of metallic iron is apparently 2.6 million years, so I'm not sure what you mean there.

Rust.

Maybe some isotopes of iron have a half-life, stable isotopes don't decay (iron is the element where all decay chains end)

Hahahah,

fair point, maybe you could show me a 50 year old rail that is still worthy of being ridden. ;)

Even a 20 year old rail is problematic from what I understand (from a UK perspective).

It's not the tracks itself that need to be maintained first. When they are the issue the easiest fix is to swap the rails.

What need to be done first is the gravel and then also the ties. Expensive are also trackout/switches with motors, and of course the signal boxes. What is now the big deal is adoption to newer technologies like ETCS.

What needs the fastest maintenance nowadays, though, is software :-).

For some radioactive isotope probably. Uranium 235 half life is, what, > 500 million years? That would make iron significantly hotter. Normal Fe is effectively around forever.

On long enough timescales, the most stable thing in the universe is the iron isotope 56Fe. All heavier atoms will decay to 56Fe, and all lighter atoms will eventually combine to form 56Fe via quantum processes, even at zero temperature. 10^15000 years from now, there'll be iron stars comprised almost entirely of 56Fe.

The lasting infrastructure of railways was the rights of way and the stations, and I think you're hinting that building them in the 1840s allowed us to do that when no one cared very much about NIMBYs or bulldozing through the countryside.

What's the corresponding infrastructure of AI? The major cost - the GPUs - are effectively obsolete after 3-5 years. The physical location of the datacenters, power, cooling and fibre that connects them might be the lasting infrastructure. Is datacenter location important? Are we actually building up new power sources (apart from endless announcements about FANGs opening nuclear power stations, which as far as I'm aware have not happened yet)?

I’d look to the lasting infrastructure improvement from the dotcom era. Such as laying fiber.

A big ones here may also be increased technological literacy, the rise of a new UI paradigm (chat with a non-human) and the structuring so much data in the world that while it previously existed was hard to meaningfully leverage because it was unstructured.

And, last but not least, lowering the barrier to entry to starting tech companies by eliminating and launching a new generation of SMB-like tech startups that don’t need to take VC-money and scale to survive. And as a result can can solve problems facing niche industries (not to be confused with things like Wix or Etsy that lowered the barriers business selling real world products to create an online presence)

If nothing else, mainstreaming AI will have the same impact mainstreaming spreadsheets did.

> however, the core infrastructural asset of GPUs tends to become rapidly obsolete after about 5 years.

Is it all about the actual GPUs though, is that the only "infrastructure" being built? A list from the top of my head of things that I'd say do last:

1. Data center buildings (take a while to build, contents completely aside).

2. Organisations and processes for running operations and procurement in said data centers - doesn't take decades to build for sure, but it's something worthwhile to already have.

3. Advances in the actual chips, i.e. more powerful processing units.

4. Advances in chip fabrication.

5. Chip fabrication facilities and organisations (similar to #1 and #2).

So sure, GPUs are highly temporary. But a lot of the things being developed and built around them much less so.

I do think one possible bubble burst scenario is that we'll have cheap compute available for decades but not a lot of great ideas of what to do with it. That is not unlike the 2000s I suppose.

Consider the second order effects of building all those data centers.

The GPU hardware rots and becomes obsolete in a matter of years, but the national infrastructure required to support the physical sites isn't going away. Things such as...

- improved power distribution networks

- logistics arrangements to build and support the DC sites

- lots and lots of new fibre interconnects to support the massive bandwidth needs

- hopefully: better power delivery planning laws

- plumbing infrastructure, because all that hardware requires cooling

Some of the DC sites will be decommissioned from their initial use, but given the physical security requirements, might morph into handy higher-security industrial facilities with only small repurposing. Such reuse cases would especially benefit from improved logistics (see above).

I'll be honest: this list sounds like grasping at straws to find something positive in this fiasco and doesn't come close to building something lasting and valuable for USA.

That's fine, but the non-GPU infrastructure represents half of the cost of the datacenter. The physical GPUs and compute represents the other half and will rather rapidly depreciate.

This is not as hardy as fiberoptic communication lines used to build the internet, or railroad lines used to build transportation infrastructure.

Yes the thing that might be missed in the point about how the big buildout of GPU compute is going to be the backbone of the future etc is that, unlike railroads and dark fibre, the GPU compute gets obsolete really really quickly. So it's not the same.

I had a friend who got a Sun cluster for basically free when the 2000 dot com bubble burst. And when we were doing recreational math contests a couple of years later it was slower than our laptops.

So it is very likely that a load of today's GPU compute is very competitive next year or the year after?

The AI bubble bursting will kill investment in the next gen hardware in the west.

But china will come to market with its first gen that it is currently building to replace its dependency on the west and will leapfrog the west etc. China isn't really completely dependent on competing in our AI bubble, its using AI for its own things and will plough on even when the west bubble bursts. Seems obvious?

Still, there has been so much talk about the AI bubble bursting last week and this is the the best writeup.

I do agree with you, but I think there a non-zero chance the situation might be different now.

We are not getting the same insane gains from node shrinks anymore.

Imagine the bubble pops tomorrow. You would have an excess of compute using current gen tech, and the insane investments required to get to the next node shrink using our current path might no longer be economically justifiable while such an excess of compute exists.

It might be that you need to have a much bigger gap than what we are currently seeing in order to actually get enough of a boost to make it worthwhile.

Not saying that is what would happen, I'm just saying it's not impossible either.

I suspect that if AI doesn't use the compute then something else will happily fill the gap - like building more lanes on a highway. It will be interesting to see what fills the GPU compute glut though.

Regards robot form factor; I'd rather R2D2 than C3PO. I don't want anything approaching the Uncanny Valley; I want a machine that does handy things!

What’s the intersection between the current LLM driven bubble and robotics? Apart from Tesla’s PR?

Just look at every Nvidia GTC presentation for the past 3 years and Chinese investments into companies like Unitree. The plethora of robotics startups that all build humanoid machines (rather than wheeled arms)

On the other hand, since Moore's Law has flattened a lot in general, does this apply anymore?

I think at least with CPU:s the depreciation has slowed down a lot compared to 15 years ago.

Reply on news.ycombinator.com