While having a built for purpose machine for training is a huge advantage, for me another major advantage of dojo is owning the machine as capex as opposed to renting it via oppex. Getting FSD working is going to take more than a big machine and a bunch of data. They’re going to need countless iterations tweaking the training algorithms and then subsequent also computationally expensive simulations to test the viability of the trained models on a plethora of situations. If I’m having to shell out millions for each training and testing cycle I’m going to be a whole lot less iterative/agile than if it’s just the cost of electricity. In an unknown space like this I’d way rather take a try out as many things as possible approach than a justify to the board my ongoing 400m oppex spend world.
My bullshit alarm went off when I saw the first Cerebras slide...it is basic knowledge that you can run compute inside specific regions on the major clouds. No one is running half their job in Austin and half their job in Seattle. Maybe the point about the big players using dedicated infra still stands, but this is a very elementary mistake for them to call out. Weird
I'm also very skeptical about the Tesla humanoid bot program. The hardware costs alone make the economics very questionable / impossible today. I say this having had experience making angel investments in the robotics space since 2017. Even if hardware costs could be improved by an order of magnitude, a general purpose humanoid bot is extremely challenging and not even Boston Dynamics is close to producing one. I would ascribe basically zero value to Tesla's humanoid robot program today.
I don't doubt that it's possible to run a cloud job in a single location. But that is not the central point here in my opinion. It is that a dedicated deep learning computer system will have superior performance and/or lower cost vs. a general purpose data center approach. It is therefore a likely scenario that the most profound models will be trained on those and demand for this type of computing will rise disproportionately. It would then be where most of the AI value is created (and perhaps captured).
With respect to the bot program, even if that does not come to fruition, Dojo will likely cement Tesla's FSD lead. I don't doubt that markets are ascribing zero value to the bot, so there is zero downside. What's more relevant to me is that the ambition of Tesla's FSD program is to come up with a generalized vision-only real world AI solution. If/Once successful, there will almost certainly be further commercialization potential.
It is also important to consider that Boston Dynamics does not employ NN training like Tesla does. Their bot performances are great on video, but we have no idea about the underlying software. Might be all pre-programmed. I don't consider them a relevant peer/benchmark. Closest competitor is Google in my view.
From what I remember DOJO was supposed to be online last July. Their chart indicates it has been online since January this year.
As w all things Tesla, I think DOJO will run late and underwhelm for quite a while. Their 300k GPU equivalent expectation for October 2024 will not be reached. I suspect they will reach October 2023 targets only on October 2024.
Tesla always delivers. But never on time. They simply tackle super ambitious issues w unrealist timelines driven by Elon. Elon is too over optimistic but I understand he has to lead and be bold.
To start at zero and 100x the performance of the worlds fastest supercomputer in 3 years is simply not going to happen. I wish Tesla could so it. But there is enough evidence to show they never meet these ridiculous timelines. They WILL deliver. But not 100x by next October and probably 50x or less in 3-4 more years.
Yeah the Cerebras comment was more of an aside and a red flag about Cerebras
I disagree with the idea that Tesla leads in FSD. Cruise and Waymo are on the road doing driverless rides every day now with passengers, and Waymo in particular has long deep rooted expertise in talent, compute, and self driving software. Tesla has made strange development choices which have not always been correct, and there are many public disengagement examples showing fundamental errors in the existing software. That's not to say they won't be one of several main players, but the reality today does not align with the megabull thesis that Tesla will be the lone winner (or #1) in autonomous vehicles.
Comparing Boston Dynamics to Teslas ability to make humanoid robots is people comparing Teslas ability to make an EV vs legacy auto in 2010. We all know how that turned out.
As for cost. Tesla is amongst the best and innovative manufacturers in the world. And Wrights Law is applicable to any manufactured goods. Cost will come down as they scale production of robot parts and robots. People made the same argument when the 240 mi $120k+ Tesla Roadster was announced/revealed in the early 2000's. Tesla has all the key ingredients to train a humanoid AI and develop and manufacturer it at scale.
In the 1990's cell phones were also considered to be way too expensive to ever go mainstream. To say nothing about the limited coverage area and weak reception and battery life. As well as the internet that was an impossible dream in the mid-90's.
Boston Dynamics is a horse and pony show. Their robots use no AI and are codes by hand to do impressive looking demonstrations only. What useful has come out of anything BD has done sonce they first showed walking robots about 15 years ago? Nothing. Boston Dynamics (as far as robots go) will be the last company to produce himanoid robots. It's academic institution and not a manufacturer.
I suggest you go listen to Lex Fridman's recent interview with Robert Playter from Boston Dynamics to get a better understanding of their roadmap and capabilities. It may also open up your eyes to the difficulties Tesla will face in developing the physical capabilities necessary for humanoid robots.
This will take many more years at a minimum, and even then you have to think about what the robot cost will be versus the tasks it will be displacing. Your average citizen is unlikely to drop $50K+ to replace vacuuming, dishwashing, and housekeeping.
Robots will be built for industrial purposes first and then commercial. Consumer robots is bot something I see until those markets are scaled to meet demand. Even at the consumwr level I see a humanoid robot being rented by the hour or for a specific taks or project. Outright ownership won't be economically feasible for most bcz what will you do with it 24/7? I don't even think of consumer level robots. That will come. But industrial and commercial humanoid robot demand will be almost unlimited.
I have listened to parts of the interview and there is nothing earth shattering IMO. BD has no manufacturing capabilities. BD was founded in 1992. Their first robot was operational (a demo as always the case) in 2004. It's been 20 years and they are still just doing the same thing. Their history says it is highly unlikely that they will be able to scale a humanoid robot. At 31 years old I think the culture is to do interesting research to enable impressive looking demos for the sake of research.
They don't have a working prototype (that is trained by an AI and not human coded). And they are not a manufacturer and has done nothing to scale the production of components. To be successful you need a working prototype as well as a clear path to scale (which requires complicated manufacturing in this case). BD has none of this.
Tesla is making all the right moves to enable mass manufacturing of a robot. They will be later than any of their targets they set. I don't expect an operational factory USEABLE robot for at least 5 years. And in 10 years I can see them in factories all over the world as well as in many commercial applications (like delivery robots in autonomous vehicles). Autonomous EVs and humanoid delivery robots are a must for internet commerce to expand (just one massive commercial use for robots).
While having a built for purpose machine for training is a huge advantage, for me another major advantage of dojo is owning the machine as capex as opposed to renting it via oppex. Getting FSD working is going to take more than a big machine and a bunch of data. They’re going to need countless iterations tweaking the training algorithms and then subsequent also computationally expensive simulations to test the viability of the trained models on a plethora of situations. If I’m having to shell out millions for each training and testing cycle I’m going to be a whole lot less iterative/agile than if it’s just the cost of electricity. In an unknown space like this I’d way rather take a try out as many things as possible approach than a justify to the board my ongoing 400m oppex spend world.
I agree with you wholeheartedly! This is early stage of a new industry. best to figure it out at the forefront!
great take, thank you!
My bullshit alarm went off when I saw the first Cerebras slide...it is basic knowledge that you can run compute inside specific regions on the major clouds. No one is running half their job in Austin and half their job in Seattle. Maybe the point about the big players using dedicated infra still stands, but this is a very elementary mistake for them to call out. Weird
I'm also very skeptical about the Tesla humanoid bot program. The hardware costs alone make the economics very questionable / impossible today. I say this having had experience making angel investments in the robotics space since 2017. Even if hardware costs could be improved by an order of magnitude, a general purpose humanoid bot is extremely challenging and not even Boston Dynamics is close to producing one. I would ascribe basically zero value to Tesla's humanoid robot program today.
I don't doubt that it's possible to run a cloud job in a single location. But that is not the central point here in my opinion. It is that a dedicated deep learning computer system will have superior performance and/or lower cost vs. a general purpose data center approach. It is therefore a likely scenario that the most profound models will be trained on those and demand for this type of computing will rise disproportionately. It would then be where most of the AI value is created (and perhaps captured).
With respect to the bot program, even if that does not come to fruition, Dojo will likely cement Tesla's FSD lead. I don't doubt that markets are ascribing zero value to the bot, so there is zero downside. What's more relevant to me is that the ambition of Tesla's FSD program is to come up with a generalized vision-only real world AI solution. If/Once successful, there will almost certainly be further commercialization potential.
It is also important to consider that Boston Dynamics does not employ NN training like Tesla does. Their bot performances are great on video, but we have no idea about the underlying software. Might be all pre-programmed. I don't consider them a relevant peer/benchmark. Closest competitor is Google in my view.
Thanks for the article and insights.
From what I remember DOJO was supposed to be online last July. Their chart indicates it has been online since January this year.
As w all things Tesla, I think DOJO will run late and underwhelm for quite a while. Their 300k GPU equivalent expectation for October 2024 will not be reached. I suspect they will reach October 2023 targets only on October 2024.
Tesla always delivers. But never on time. They simply tackle super ambitious issues w unrealist timelines driven by Elon. Elon is too over optimistic but I understand he has to lead and be bold.
To start at zero and 100x the performance of the worlds fastest supercomputer in 3 years is simply not going to happen. I wish Tesla could so it. But there is enough evidence to show they never meet these ridiculous timelines. They WILL deliver. But not 100x by next October and probably 50x or less in 3-4 more years.
Yeah the Cerebras comment was more of an aside and a red flag about Cerebras
I disagree with the idea that Tesla leads in FSD. Cruise and Waymo are on the road doing driverless rides every day now with passengers, and Waymo in particular has long deep rooted expertise in talent, compute, and self driving software. Tesla has made strange development choices which have not always been correct, and there are many public disengagement examples showing fundamental errors in the existing software. That's not to say they won't be one of several main players, but the reality today does not align with the megabull thesis that Tesla will be the lone winner (or #1) in autonomous vehicles.
I don't think Waymo and Cruise have a scalable solution. But I guess that is what makes a market and we will have to wait and find out.
Comparing Boston Dynamics to Teslas ability to make humanoid robots is people comparing Teslas ability to make an EV vs legacy auto in 2010. We all know how that turned out.
As for cost. Tesla is amongst the best and innovative manufacturers in the world. And Wrights Law is applicable to any manufactured goods. Cost will come down as they scale production of robot parts and robots. People made the same argument when the 240 mi $120k+ Tesla Roadster was announced/revealed in the early 2000's. Tesla has all the key ingredients to train a humanoid AI and develop and manufacturer it at scale.
In the 1990's cell phones were also considered to be way too expensive to ever go mainstream. To say nothing about the limited coverage area and weak reception and battery life. As well as the internet that was an impossible dream in the mid-90's.
Boston Dynamics is a horse and pony show. Their robots use no AI and are codes by hand to do impressive looking demonstrations only. What useful has come out of anything BD has done sonce they first showed walking robots about 15 years ago? Nothing. Boston Dynamics (as far as robots go) will be the last company to produce himanoid robots. It's academic institution and not a manufacturer.
I suggest you go listen to Lex Fridman's recent interview with Robert Playter from Boston Dynamics to get a better understanding of their roadmap and capabilities. It may also open up your eyes to the difficulties Tesla will face in developing the physical capabilities necessary for humanoid robots.
This will take many more years at a minimum, and even then you have to think about what the robot cost will be versus the tasks it will be displacing. Your average citizen is unlikely to drop $50K+ to replace vacuuming, dishwashing, and housekeeping.
Robots will be built for industrial purposes first and then commercial. Consumer robots is bot something I see until those markets are scaled to meet demand. Even at the consumwr level I see a humanoid robot being rented by the hour or for a specific taks or project. Outright ownership won't be economically feasible for most bcz what will you do with it 24/7? I don't even think of consumer level robots. That will come. But industrial and commercial humanoid robot demand will be almost unlimited.
I have listened to parts of the interview and there is nothing earth shattering IMO. BD has no manufacturing capabilities. BD was founded in 1992. Their first robot was operational (a demo as always the case) in 2004. It's been 20 years and they are still just doing the same thing. Their history says it is highly unlikely that they will be able to scale a humanoid robot. At 31 years old I think the culture is to do interesting research to enable impressive looking demos for the sake of research.
They don't have a working prototype (that is trained by an AI and not human coded). And they are not a manufacturer and has done nothing to scale the production of components. To be successful you need a working prototype as well as a clear path to scale (which requires complicated manufacturing in this case). BD has none of this.
Tesla is making all the right moves to enable mass manufacturing of a robot. They will be later than any of their targets they set. I don't expect an operational factory USEABLE robot for at least 5 years. And in 10 years I can see them in factories all over the world as well as in many commercial applications (like delivery robots in autonomous vehicles). Autonomous EVs and humanoid delivery robots are a must for internet commerce to expand (just one massive commercial use for robots).