Tesla Optimus Sub-Prime Robot

[Buckminster]

I am very pleasantly surprised by the new optimus video. Especially as its just tweeted as 'no big deal' rather than building a whole AI day around it. That suggests that maybe progress is accelerating, and when there is another AI day, there will be bigger leaps to show.

I've watched boston dynamics closely for years. It seems they are... stuck. The company has changed hands a few times, and regardless who owns them, it seems that their only useful product is parkour videos on youtube. I know they sell spot, but not in vast numbers. I still have never seen one in real life.

That tells me that there is something fundamentally limiting about the approach boston dynamics are talking. I suspect people buy the company, have grand plans, and then realise...no. This seems to be the opposite of Tesla bot.

Elon has spoken a lot about the many 'local maxima' encountered by the FSD team, and is rightly focused on avoiding that issue with the bot. I still think that he will be keen to get the bot in the factory, working alongside humans in large numbers ASAP, even if it just means they carry empty boxes and stack them back on a truck all day. They will want those tens of thousands of hours of video, and data about when they screw up, to further train them. They are also building an actual robot, not just a chip inside a robot, so they want tens of thousands of hours of wear and tear on every single component.

I used to think we were years out from bot doing anything economically useful for Tesla, but maybe not. Maybe some time in 2024 we see bots carrying out some basic tasks, en masse on the production line. When we do, the UAW is going to have a heart attack...

 

[SOULPEDL]

Was I dreaming???

 

[Knightshade]

We finally know how Tesla will beat the comptetition!

If do right, no can defend!

 

[Andy O]

It looks to me like Optimus is ready to be used ("employed") to sort recycling.

 

[DanCar]

Long read if you want to understand more about training the bot.

https://twitter.com/x/status/1705982525825503282

 

[Buckminster]

https://twitter.com/x/status/1706142947769561530

Does this mean a simplified bot?

 

[Buckminster]

https://twitter.com/x/status/1706153863491830227

 

[Buckminster]

27. Manufacturing Engineer, Process Development, Tesla Bot (Sparks, Nevada)

52 mostly non manufacturing positions:

https://twitter.com/x/status/1706062018766332056

 

[Buckminster]

https://twitter.com/x/status/1706285260928061775

 

[Knightshade]

https://twitter.com/x/status/1706285260928061775

Would like to think this'd quell the junk we got back in July of people claiming they were already in use on the factory floor, given it's saying they're aspirationally hoping to do that with prototypes next year (and given its an Elon aspiration who knows when it'll actually happen)

 

[heltok]

The bad: More expensive, will be able to do fewer things, will perform them worse, less robust
The good: Available for presale

Competition will be coming for humanoid robots also.

 

[heltok]

https://twitter.com/x/status/1707812100943524220

 

[Christine69420]

Randy Kirk argues why Tesla will achieve an Optimus runrate of 1 mill in june 2024. And why customers will be able to buy or rent Optimi as early as Q2 in 2024.

TL;DW: Tesla hiring lots of people costing a lot in wages. If they only make a few 1000 robots in 2024 each Optimus will cost $12.000 just for the fixed costs of overhead. Tesla is excellent at achieving return on investment. They will want that for the robots too.

 

[Knightshade]

Randy Kirk argues why Tesla will achieve an Optimus runrate of 1 mill in june 2024. And why customers will be able to buy or rent Optimi as early as Q2 in 2024.

Is the "why" the author smoking crack?

I think we'll be able to have it do something useful in our factories sometime next year.

And that's Elon couching it (if you look at the whole paragraph) in exactly the "I think" and "reasonably confident" language he's been predicting FSD NEXT YEAR for 7 years with too.


But sure, there'll be a million run rate and public sale in June, that totally isn't insanely far from reality.

 

[Rarity]

27. Manufacturing Engineer, Process Development, Tesla Bot (Sparks, Nevada)

52 mostly non manufacturing positions:

https://twitter.com/x/status/1706062018766332056

Just by looking at the roles listed, it seems like they are about one year out from initial production for factory use. Maybe a bit more.

Harmonizing the timing of the ramp with the available compute will be important.

 

[Buckminster]

Just by looking at the roles listed, it seems like they are about one year out from initial production for factory use. Maybe a bit more.

Harmonizing the timing with the available compute will be important.

I think you are looking at what will probably be v2 or v3. V1? prod is due to start ramping in November. These will all go to Tesla gigafactories most likely.

 

[Rarity]

I think you are looking at what will probably be v2 or v3. V1? prod is due to start ramping in November. These will all go to Tesla gigafactories most likely.

What order of magnitude of v1s do you expect to be produced?

Your timeline seems a bit ambitious for getting these bots into the factory. They are still at the foundational level on movement and control.

Robotics has been an industry moving at glacial speeds of innovation. I feel like Tesla Bot is different because of the embrace of end-to-end AI, but the hiring and compute just seem insufficient to support much actual factory work in the near term.

 

[Buckminster]

What order of magnitude of v1s do you expect to be produced?

Your timeline seems a bit ambitious for getting these bots into the factory. They are still at the foundational level on movement and control.

Robotics has been an industry moving at glacial speeds of innovation. I feel like Tesla Bot is different because of the embrace of end-to-end AI, but the hiring and compute just seem insufficient to support much actual factory work in the near term.

V1 bots might be employed building V2 bots. That would solve both of your problems. Even if they only make 1000 bots by March, where are they going to go? At the very least they will have them doing very small specific jobs in the factory. This is important for different reasons than the s/w such as how they get on in a factory with humans around, reliability, battery perf etc. The slides suggested that the unboxed methodology will be designed around bot capabilities. Not needing to lean into vehicle etc.

I think the best comparison here is raptor engines. Elon was worried about speed of production two years ago even though none are yet to go orbital. When the stars align, raptor / starship / bot production will be stellar.

 

[Rarity]

V1 bots might be employed building V2 bots. That would solve both of your problems. Even if they only make 1000 bots by March, where are they going to go? At the very least they will have them doing very small specific jobs in the factory. This is important for different reasons than the s/w such as how they get on in a factory with humans around, reliability, battery perf etc. The slides suggested that the unboxed methodology will be designed around bot capabilities. Not needing to lean into vehicle etc.

I think the best comparison here is raptor engines. Elon was worried about speed of production two years ago even though none are yet to go orbital. When the stars align, raptor / starship / bot production will be stellar.

One thing that I like about Optimus is that it doesn't seem all that complicated from the hardware side, once you factor in the existing Tesla bin of mass manufactured sophisticated parts. We are getting a lot of cutting edge stuff for free. A production rate of a thousand a year (or two) in March seems a bit ambitious, but perhaps not by much. When you stated that production will be "ramping" in November, I had another scale of operations in mind.

Lots to ponder. I suspect that the constraint to the first 1,000 is almost entirely training compute and engineering the logistics of learning. On the FSD side, it started to get good once they hit one million videos learning material. Assuming that the scale for the bot is similar or more demanding, that all takes logistics and compute. Then there is a mass manufacturing process overlay of two or three years that constrains going to 50,000 or 100,000 of these. My mental model for the mass manufacturing of the bot is more like the Starlink terminal rather than Raptor.

 

[Buckminster]

One thing that I like about Optimus is that it doesn't seem all that complicated from the hardware side, once you factor in the existing Tesla bin of mass manufactured sophisticated parts. We are getting a lot of cutting edge stuff for free. A production rate of a thousand a year (or two) in March seems a bit ambitious, but perhaps not by much. When you stated that production will be "ramping" in November, I had another scale of operations in mind.

Lots to ponder. I suspect that the constraint to the first 1,000 is almost entirely training compute and engineering the logistics of learning. On the FSD side, it started to get good once they hit one million videos learning material. Assuming that the scale for the bot is similar or more demanding, that all takes logistics and compute. Then there is a mass manufacturing process overlay of two or three years that constrains going to 50,000 or 100,000 of these. My mental model for the mass manufacturing of the bot is more like the Starlink terminal rather than Raptor.

My thinking was a thousand in total by March/April. They won't want to make more than that for V1. This is much more complex than Starlink. They won't ramp until the design is right. Maybe Solar Roof is a better example.