MC3OZ
Active Member
I fully expect Tesla to make a bot prototype this year, and perhaps set up small scale production in 2023 or 2024.I think Elon has misjudged and underestimated something about their humanoid robot. I believe TESLA should be able to begin selling them profitably this year and add a significant income stream from them about that soon.
The challenging part of making a useful robot is solving the general intelligence problem. We see that this problem has largely been solved in developing FSD. The robot only has to be smart enough to work in a warehouse or supermarket, moving goods around, where there is no risk of killing anyone through a road driving error.
How much would such a robot cost if they were manufactured at a small scale? I don't see any expensive component in them. Even at the price of a Model 3 they would earn their cost back through wage-substitution in a year.
(I know I am not the first to make this point but I don't want to go back through the thread to find an earlier post.
Random addenda: Wireless charging might be best for them. Small independent robot-training companies could expand their functionality. Dojo could train them to do the complex task of packing groceries.)
I expect the bot (at scale) to cost less than a model 3 mainly because it needs fewer batteries.
At small scale they might be more expensive.
IMO the reason to start ASAP is because like FSD this is an iterative "fleet learning" exercise. So the first bots deployed in factories might be more of a hindrance than a help initially. But the bot becomes useful as soon as it can repeat any task in a well defined environment consistently with a low error rate.
So I think that the bot will start out like the early days of AP 2.0, when most thought it was a regression from AP 1.0.
But it is likely that the bot design itself will rapidly improve and high volume mass production will probably happen 2025 or 2026. A factory churning out up to 1 million bots per year would not be science fiction. It is a compact product with no need for paint.
All those bots learning on the job is a lot of neutral net training, possibly training a different set of nets for different robot tasks. So a bot may run a set of nets suited to what it is doing with each net covering a set of closely related tasks.
That is a lot of training hence Dojo. Using a training the bots in house has the advantage of gaining productivity benefits. But the additional advantage is there is no outside commenary on the error rate and the rate of improvement, the customer will be patient.
Last edited: