Tesla AI5 chips will be volume production in 2026.
Production version 1 of Teslabot will be produced in 2025. Currently Teslabot actuators are in the hand but the new version has actuators in the forearm. The cost of a super useful humanoid robot will be about $20,000.
Tesla FSD for Cybertruck will be in mid-August.
The GPU and Dojo AI training clusters have big electrical demand issues. The power can drop 50 megawatts in milliseconds.
AI training compresses reality. The weaker inference chip HW3 and HW4 in the cars have about 15% of the power of some GPUs. This means more compute on the training must be used to make up for the lower compute in the car.
When Tesla FSD enables unsupervised driving then it will enable cars to go from about 10 hours of use per week to 50-100 hours per week.
10 billion bots times $20,000 would be $200 trillion in value.
Brian Wang is a Futurist Thought Leader and a popular Science blogger with 1 million readers per month. His blog Nextbigfuture.com is ranked #1 Science News Blog. It covers many disruptive technology and trends including Space, Robotics, Artificial Intelligence, Medicine, Anti-aging Biotechnology, and Nanotechnology.
Known for identifying cutting edge technologies, he is currently a Co-Founder of a startup and fundraiser for high potential early-stage companies. He is the Head of Research for Allocations for deep technology investments and an Angel Investor at Space Angels.
A frequent speaker at corporations, he has been a TEDx speaker, a Singularity University speaker and guest at numerous interviews for radio and podcasts. He is open to public speaking and advising engagements.
Where will the data to train AI robots come from?
Tesla has attached many sensors to their vehicles. Tesla requires millions of cars to train their version of AI for cars/trucks – FSD.
Waymo/Cruise are losing the battle against FSD because of all the training data available to Tesla.
But unless Tesla uses sensors attached to humans to train their robots EN MASSE (numbering at least in the millions), then Tesla will lose the battle to someone willing to train their robots the same way Tesla trained FSD.
Please get the message out to Elon and the head of Optimus/Teslabot.
I am serious. They need to be focusing on this. Either through volunteers or paid trainers.
Apparently they can train the robots using 3rd person perspective videos, they don’t need to instrument the people. Though there is talk of using the robots in a teleoperated mode for training purposes.
But, yes, the current approach to AI training, using insanely massive databases of example behavior, really is not terribly useful for robots. They really need to up the AI game to be able to learn in a more human fashion.
While Tesla vehicle robots are the foundation of being able to build humanoid robots, the situations aren’t perfectly analogous. Tesla isn’t in the business of selling mass market non-robotic humanoid figures, so there is no good analogy to including the hardware while developing the software. When Tesla started putting HW3 in every vehicle, it had limited use for the data, it just needed to grow the fleet for future use. Even now, it uses it to pull in a small number of edge case examples. That isn’t similar to training needs for humanoids which will mostly be trained on repetitive tasks in a very concentrated way – at least for many years.
Vehicles can’t just stop if they don’t know what to do and they move at high speeds with a lot of mass. Humanoids are not so dangerous, move much more slowly, can afford to just stop if they are confused.
Vehicles need to be capable of handling nearly every possible scenario to be safe. Humanoids don’t. They don’t even need to individually know most tasks that have been trained. They can simply download a skill set if a need arises for it.
Most humanoid training will be dozens of humans teleoperating bots performing a task over and over all day long. Like the example of a bot folding clothes or packing objects while teleoperated that Tesla showed. They aren’t being trained for hard to anticipate risky situation in the wild like vehicles on the road.
Will all these Tesla Bots drive home with a Tesla, after work?
I ask for a friend.
They will make a hell of a team for sure.
The car takes the bot anywhere a person can/needs to go, and the bot does stuff people does but where the car can’t go.
Doing chores, buying groceries, taking or picking stuff up.
And the car can act as the Starlink hub for giving the bot full connectivity.
The only thing stopping this from being a revolution in productivity, is the fact many formerly developed places are transitioning into low-trust societies, and that the cars and bots will be targets of crime.
There’s probably not much money in the bot hardware. Dozens of Chinese companies are going to race to the bottom with pricing. Similar to how they beat Tesla EVs in China.
The key to Tesla valuation is either they are better at software, or at least better than non-Chinese companies at software – and Chinese software is kept out of Western markets.
Tesla is pretty good at low-cost manufacturing. China’s main advantage is cheap labor, which doesn’t matter so much when the labor is robots. As for software, I don’t think they have any advantage today.
Tesla like Apple uses proprietary hardware and software designed together. In the case of Optimus, it uses the same Inference engine that it designed and mass produces for Tesla robotic vehicles and the same NN training supercomputers used to turn big data into NN.
Tesla will beat EV competition in China by licensing them FSD and making most of the profits from robotaxis they build.
I don’t think Musk’s comments about selling Optimus reflect what Tesla will actually do in the foreseeable future.
Tesla will lease Optimus to users in an all-in package including training, maintenance and all operating expenses. This will be priced to produce major cost savings for the user vs current all-in labor costs and be very easy to justify based on the accounting with no upfront capital costs or depreciation and little risk.
EG A bot that replaces 4 workers who each cost $75k per year all-in each might lease for $10k per month or $120k per year all-in. So $300k of labor expense replaced for a savings to the user of $180k per year. Tesla’s marginal cost to build might start off at $40k so it earns 3X the cost in the first year – and its main on going expense is training on its Supercomputer clusters.
Tesla wouldn’t sell robot hardware they’d sell labor replacement. Competitors would be selling robot hardware and leaving customers to try to figure out how to use it.
Tesla will only prosper outside of China if they pivot to releasing some hard hybrids – PHEVs with similar gas range to ICEs with 100 miles or so of pure electric. Hard hybrids are the only path forward to moving the market and ensuring a better electrical future with choice for All. Pure BEVs and the regulations that ‘mandate’ them will lead to local and possible national BEV collapse or even vast reductions in personal vehicular traffic -a national tragedy. Elon collaborating or compromising? – crazier things have happened.