Tesla’s AI Day 2022 was a pretty awesome recruiting event for AI engineers. Utterly incomprehensible to most others, especially stock analysts and journalists. Just look around the web and see all the moronic takes on what Tesla presented. Hint: anybody drawing parallels to Boston Dynamics has no idea what they just saw from Tesla.
I watched it as it happened and found it fascinating, not that I could understand everything. My software engineering background was certainly not up to grasping the finer details of the hardware and mechanical engineering being talked about.
As an engineer I especially liked the bits where they talked about nasty bugs. And the fact that they had a ton of engineers there talking about what they were doing. Musk just stepped out of the picture for most of the presentation.
Musk was abundantly clear that Tesla’s goal is to produce a humanoid robot that is useful for something as quickly as possible, with the restrictions that it has to be affordable and mass manufacturable. These last two are not true of any robots today. Tesla intends to be using it themselves for something by next year, and possibly selling something in 3-5 years.
The robot prototypes were interesting, and the progress has been substantial. The work that remains is massive and should prove very attractive for new recruits.
IGU, could you maybe summarize for a lay person what was engaging about the robot presentation? I understand that it was pitched for the recruitment audience, not the public (or Wall Street) - but I didn’t quite get what it is that Tesla is saying that it brings to the table that the other robotics firms are not.
I think we know Tesla is a major user of robots in building its cars.
Elon seems to be saying that he plans to invest in building better robots. Maybe pushing the technology to new heights. As he has for EVs, SpaceX and others.
Technology leaders tend to attract the best talent. This is probably about hiring the best.
I think the most important things Tesla demonstrated are:
design for mass manufacturing (something at which they are experts)
design for affordability
vision, mobility, and navigation by way of machine learning
vertical integration
batteries
Those last three mean that they can iterate rapidly on both the hardware and software without waiting on outside vendors for custom parts. Plus they can use much of the tech developed for the cars for vision and navigation and power. That includes Dojo for training.
The point of “humanoid” that seems to be underappreciated is that the robot is designed to be generally useful. There are tons of specialized uses to which robots can be put, but a humanoid robot can do most anything a human can do if it can be taught to do it. That means it becomes largely a software problem. Much faster and cheaper to iterate.
There are lots of research projects out there. There are lots of industrial robots out there. But there are no mass market, cheap, generally useful robots. It’s a wide open field. And lots and lots of interesting problems to solve to make such a robot work well.
Thanks for the response, IGU. That helps me understand why Tesla’s AI day might have some appeal to potential employees.
As to the above quote, I’m not sure that I see it. Most of what humans do in the world involves tools - devices that exist to amplify or improve our ability to move and do things in the physical environment. To use a very Tesla example, we use cars so that we can move faster from place to place. We have a complicated interface that allows us to manipulate the car using various inputs (steering wheel, brake pedals). But in that case, it’s less efficient to build a robot to use the tool than to just make the tool smart enough to do what we want. The “driver” of an AV has direct control over brakes and tire angles and acceleration because it’s part of the tool, not using the tool. It’s not especially efficient to have a generalist robot that drives a car for us - it’s much more efficient to build a car that drives itself.
For many things, that same dynamic seems likely to apply. I don’t want a humanoid robot surgeon to pick up a laser cutter the same way a human would - I want a specialty surgical robot that is optimized for using a laser to cut, that isn’t limited by a human-style arm but has an armature that is designed especially for that purpose.
I guess he’s looking at the ‘home servant’ market? A personal assistant that can do lots of household chores in a way that you can’t replicate using the auto-assembly line robot model?
Among young engineers fresh out of college, Tesla is already about as high on their popularity list as a company can get. The nerdier the engineer, the more likely they will want to work at Tesla. At Tesla engineers get to really do stuff! Of course Tesla only wants the best, and they have an advantage because so many of the best want Tesla.
In the case of AI specialists, I suspect that the competition from other top AI companies is pretty fierce. I also suspect that in such a complex, cutting edge field the difference between the very best and the middle of the road people is profound. Some engineering fields are pretty stable. The AI field is still defining itself, and developing the knowledge of how to do it as they work. Tesla’s recruiting of these specialists with AI day tells them they will not just be at the cutting edge, they will be defining a new cutting edge. We have to best toys for you to play with!
When Tesla created their first assembly line for the Model 3 they got carried away with automation. Eventually they realized the problem and got rid of a LOT of robots, replacing them with people. There are some things that robots can’t handle well, but people are good at. Installing something floppy, like a wiring harness, for one example. I suspect a lot of the ideal behind the humanoid robot is for them to work in factories, particularly Tesla factories, doing jobs that dedicated robots are not good at.
I think that’s the part that’s tripping me up. If you have a problem that robots can’t handle well (installing something floppy), and you solve that problem for the humanoid robot, haven’t you also solved it for dedicated robots as well? Whatever you’ve done to the AI, or the physical apparatus, of your humanoid robot can also be done with a dedicated robot - right?
Not necessarily, I think. Humans, and (presumably) humanoids, have general purpose manipulators called hands. Likewise, general purpose orienting tools such as two eyes far enough apart for depth perception. An AI that knows how to use such general purpose capabilities might not adapt well to controlling specialized hardware. Also, the same humanoid could be used at any of the tasks for which the systems have learned to get the job done. Humanoid workers would have more parts than dedicated robotic devices, but maintaining a work force of identical humanoid workers may in the end require fewer spare parts and reduce delays when one malfunctions. A specialized unit requires specialized parts. A hundred different devices with different specialties… a lot of parts, a lot of spares.
I mean, maybe? I can certainly see that being the case in lots of other circumstances - but on an auto assembly line, it seems pretty unlikely. For our mesh installer, they certainly need eyes and hands - but do they really need legs? Is using a torso and legs to move the hands into the right place the best and most efficient design, or would something more bespoke to that assembly line work better? Might something with three hands work better? Certainly if you have a prototype “hand” you might want to generalize that “hand” so that it’s understandable by AI and has standardized spare parts - but having hands specifically (and only) in the configuration that corresponds to a humanoid robot seems very limiting.
In an assembly line, you’re really tailoring each step and each component to work as efficiently as possible with the other parts of the line. Given how repetitive and constant the tasks are, it’s hard to see why having a bog-standard humanoid robot would be better than having something engineered to do that job better.
I think what’s tripping you up is that it’s not about efficiency of the individual unit at a particular task, it’s all about mass production. Unless you can manufacture millions, you can’t get the costs down. So a general purpose, “not exactly right for anything but good enough for many things” type of solution is the target.
It will always be possible to design bespoke better hardware. But it is almost never worth the cost. Especially since the tasks change over time. Tesla is all too aware of the stultification that results from a mindset of “got me a hammer, so all problems look like a nail”.
Once you have something generally useful that you can mass produce, then you can differentiate mostly through software. And because it’s humanoid, it can use all those tools we have already lying around for human use. Like hammers.
But you have to start somewhere and get your product out into the world. So that’s what’s key here. You can do that for some people at $20,000 a pop, and drive the cost down from there. You can’t do that for enough people if your robot is $200,000 a pop. Most are currently more expensive than that, and have no viable path to becoming affordable.
Oh, I agree - if they could make an AI humanoid robot for only $20K, that would be just a ridiculously valuable thing. Is that the price point they’re talking about? That wouldn’t just be a factory product - that would literally change the world. Color me skeptical that they could ever get close to that type of price for a product safe enough to sell to ordinary folks…but that’s quite a lofty aspiration.
BTW, looks like Musk has decided to stop fighting the TWTR acquisition - so that will be interesting as well!
Possibly more than any other car company, Tesla is constantly updating their designs. This is one of the recurring themes in Sandy Munro’s videos on their cars.
Musk/Tesla is attempting to define Tesla’s AI as General AI (GAI)
This is an important concept, IMO, in order to follow the megatrend that is Musk’s vision.
The term AI has been criticized as “not really intelligence”.
ML may or may not be AI, depending on who is speaking.
The automated phone responses have been called AI, but if so, are really DUMB AI.
IMO, most bespoke robots are not AI. They are similar to the automated phone responses.
Call these “bespoke AI”? Cause the decisions are linear, algorithmic “by the book”.
The recent AI day presentations stated many times that the “physical machine that achieves some action” is JUST THE PACKAGE.
The GAI for the car is the same GAI as for the bot, and a strong implication that that same GAI could operate a toaster, or a “bespoke” robotic machine, or a space ship.
THE thing that separates TeslaBot from all the others is the “humanoid hand with an OPPOSABLE THUMB”. This was pointedly stated several times.
IMO the GAI that Musk envisions will be able to make “humanoid thinking-outside-the-box decisions” rather than the linear, iterative, algorithmic “by the book” decisions of the bespoke AI.
And THIS GAI is the mulri-trillion dollar difference that separates Tesla (and SpaceX, Neuralink, Boring Co) from the pack.
The pack wants a piece of the pie.
Tesla GAI wants it all.
But surely other robotics companies are working on creating robotic “hands?” It took barely an instant on google to find some examples.
Is that something you really need - or even want - in a robot to install a wiring harness in a car? Or to work in a warehouse?
I can understand the value of such an AI in making decisions - but I can’t see why you would want or need those decisions to be made in the robot that’s doing the physical work. I might want that GAI to be my factory foreman - or my plant manager. That GAI can supervise the robots - give it directions and input. But I don’t need that GAI to be in a humanoid container with a hand and opposable thumbs - it can be in a box with a good bluetooth connection.
Oh, I agree - if they could make an AI humanoid robot for only $20K, that would be just a ridiculously valuable thing.
That’s what Musk said is what they expect it to sell for. So they’ll have to make for for closer to $12K. And yes, that’s possible because of the things I enumerated earlier: mass manufacturing, vertical integration, basic operational pieces are already developed software (vision, navigation, battery pack). Whether it will work out remains to be seen.
So, yes, they fully intend to profoundly change the world. For everybody. This is, of course, immensely attractive to the best engineers. And the fact that it’s going to take a few years makes it even more attractive.
I spent most of my career building really cool stuff that never shipped, or at least never shipped in quantity. Then I went to Apple. It was to me a surprisingly satisfying thing to see my work product being used by huge numbers of people. It was very different from building something great, but too early, so the only people who ever appreciate it are the ones on your team.
I can see that. It must be really nice to work someplace that’s willing to devote large amounts of resources to trying to actually make and sell a moonshot product - a straight “skunkworks to market” pipeline. Didn’t Google built some of their industry rep on that? They’ve had all kinds of additional stuff going on - urban design and planning (Sidewalk), self-driving cars (Waymo), wearable VR (oh, Glass!), Deep Mind, and a host of others.
And of course, their own AI and robot efforts as well:
Much more fun that working on incremental improvements for existing products and just doing new stuff as a side project.
I can see that. It must be really nice to work someplace that’s willing to devote large amounts of resources to trying to actually make and sell a moonshot product
Well, it should be quite clear that it’s not intended as a moonshot product. While there’s a non-zero chance it will fail, it’s an obvious variation on the main path of achieving full self driving, which is a very high priority within Tesla. Nobody knows exactly how long FSD will take, but everybody knows it will happen. Right now, people are holding their collective breath to see if the deployment of Dojo for neural net training is the magic that’s needed to push FSD significantly forward.
An autonomous robot on wheels is not conceptually different from an autonomous robot on legs. The hardware details have mostly been done before in a high cost, hard to manufacture way, but Tesla knows how to work that. Meanwhile, selling into the workplace before the home will simplify requirements and deployment, in the same way that it’s easier to do autonomous driving on limited access highways as compared to city streets.
And fundamentally, batteries are the long term bottleneck. So since a humanoid robot uses much less battery power than a car, the profit margin per battery should be much higher on the humanoid robot. So getting the price for a robot low enough may be easier than it appears.
The biggest question for me is how it affects Tesla’s mission: “to accelerate the world’s transition to sustainable energy.” I don’t think a humanoid robot has much to do with that.
In a word: Yes!
but, really it’s not what I or you or any other reader ‘wants’… the message I got from the AI day was “That’s what Musk/Tesla wants”… Musk/Tesla is building ‘it’ and hoping ‘they will come’ with uses for ‘it’.
Is it? To this layperson, it seems like it would involve qualitatively different types of skills that the “driver” of a car would never have to learn.
If I ask a robot, “Go upstairs and bring me down the blue box labelled ‘Widget Parts,’” that will involve some of the basic skills a car driver would know - understanding a destination input, planning a path through physical space to avoid obstacles, identifying an object by its appearance, etc. But it also involves skills a car driver would never have learned - how to balance while going upstairs, how to pick up an object with enough force to hold it securely but without damaging it, etc. Those seem like fundamentally different - and very difficult - engineering problems. Especially if you’re trying to do it at a very low price point.
Again, I am a total layman in these areas. It seems like it’s much a much different technical challenge to teach a robot to walk up stairs on two legs than to move around on four wheels - but maybe that’s wrong?
Doors and stairs are designed for use by humans that stand upright and can walk. Ditto the height of desks and work tables.
Robots that can navigate like humans are not stopped by these human adaptations. They are theoretically more versatile. Of course re-designing for wheeled robots with wider base for stability and better balance, less top-heavy is a logical next step. Can humans adapt to those changes. Low ceiling. Have to duck? Or squat?
