Tesla AI day

[Buckminster]

https://twitter.com/elonmusk/status/1348043839022370817
Possible AI day later in 2021. Will almost certainly introduce Dojo which is scheduled for late 2021:
Project Dojo

 

[heltok]

My guess
1. Show current progress of FSD
2. Tell strategy for robotaxi rollout
3. Show Dojo
4. Release HW4
5. One more thing. Tesla Model Q aka robotaxi

Btw, I found this video to be informative:

 

[Yuha]

Don't forget this tweet by Elon

https://twitter.com/elonmusk/status/1307828265445400579

"Yeah, we will open Dojo for training as a web service once we work out the bugs"

 

[brianhn1]

Under appreciated feature already in every Tesla: Voice Command with AI

 

[Gilbert101]

My guess
1. Show current progress of FSD
2. Tell strategy for robotaxi rollout
3. Show Dojo
4. Release HW4
5. One more thing. Tesla Model Q aka robotaxi

Btw, I found this video to be informative:

Hmm. I keep my hands on the wheel all the time and get nagged anyway.

 

[adiggs]

Hmm. I keep my hands on the wheel all the time and get nagged anyway.

I realize it's OT, but I'm able to go for hours with no nags (and no defeat devices). The key is to keep one hand on the wheel rather than both, and to have that hand offset from the bottom of the wheel. I usually use my left arm resting on the arm rest on the door, with my hand hooked over the horizontal cross piece at the 9 o'clock position on the wheel.

The resistance in the wheel with AP active is high enough that I can hang my hand there with my hand/arm weight completely supported by the wheel. That's enough weight to let the car know, continuously, that I'm present and attentive (which in fact, I am ).


I've also discovered that on long drives, I can frequently rely on my hand's friction with the wheel to do the same thing but without a bit lower down the wheel - more like 7-8 o'clock, where it's just friction keeping my hand instead of sliding down the wheel to 6 o'clock. If your hand or hands are holding the wheel at 6 or 12 o'clock, then there's no twist on the wheel, and that's what is being monitored by the nag.

On long drives, I've also gotten in the habit of switching hands. It's a bit longer of a reach for the right arm, but same idea.

 

[Buckminster]

https://twitter.com/x/status/1382453719585808384

 

[Buckminster]

Could Tesla make their FSD 4.0 hardware into a stand alone product? People can develop their own NN - trained on Dojo. This would make Tesla the AWS of the future.

You buy the hardware from Tesla and integrate it into any number of products.

 

[Buckminster]

Still on target for end July:

https://twitter.com/x/status/1407050578149183488

 

[Buckminster]

This month! Maybe...

Can't wait. Connecting the Dojo chips in a novel 3D layout could be something special.

 

[new_sneakers]

Tesla AI Day will be August 19th - announced by Elon on Twitter just now - interesting to look in on this thread again.

Edit: "Convincing the best AI talent to join Tesla is the sole goal" - Elon

 

[ZachF]

Prediction: Tesla robotics division

 

[heltok]

Prediction: Hong and his team are doing snake charger for autonomous charging.

 

[MC3OZ]

Do we think Dojo is a ll software 1.0 human written code, or could some part of it be software 2.0 NN logic.
I think auto-labeling may have a fair chunk of soiftware 2.0 logic.
For NN training itself,I don't know enough about it to form a solid opinion.

 

[mongo]

Do we think Dojo is a ll software 1.0 human written code, or could some part of it be software 2.0 NN logic.
I think auto-labeling may have a fair chunk of soiftware 2.0 logic.
For NN training itself,I don't know enough about it to form a solid opinion.

I believe Dojo uses a NN to do self correlating temporal object recognition for auto labeling video data to then train the vehicle NN.
User interface to this subsystem would be classic code.

 

[new_sneakers]

Thinking back to Elon’s 25 April 2018 tweet, “Oh btw I’m building a cyborg dragon”

Surely this is the one-more-thing on AI day. “Dragon” gives them licence to make a cool mobile robot in a non-humanoid shape that works better for functionality. Am I crazy?

 

[jeewee3000]

Thinking back to Elon’s 25 April 2018 tweet, “Oh btw I’m building a cyborg dragon”

Surely this is the one-more-thing on AI day. “Dragon” gives them licence to make a cool mobile robot in a non-humanoid shape that works better for functionality. Am I crazy?

You are crazy.

In all seriousness, please PLEASE don't overhype AI day. There will not be a "one more thing", just a decent explanation of what Tesla has done in the AI field until now and what they are planning to achieve.

If robotics will be mentioned, it'll be along the lines of: "potential future applications" of the shown tech/AI. Very similar to what they did at the last Neuralink update.

They will NOT say: "here's a robot functioning today that does such and such".

This hyping ahead of Tesla events always leads to market disappointment and SP drop after the event. If I see more hype (I'm looking at you Solving the Money Problem) I think I'll sell some calls the day before the event.

 

[new_sneakers]

Thank you - you have talked sense into me.

 

[Buckminster]

 

[Buckminster]

The new D1 chip *and* the training tile incorporating 25 D1 chips on one big tile - which can be grouped in cabinets of 120 tiles, all of which communicating on all channels at twice the speed of the current fastest comm chips, plus their software and data management/ store advances, are a huge advance over existing tech.

As usual I don't think investors will understand the impact this has both for FSD and for computing ( Dojo as a service) or even longer term for TeslaBot.
So looking forward to Rob Maurer's take tomorrow (he was one of the few non engineers/ potential recruits invited)

And for the record, Tesla is becoming an AI (+robotics) company - from their new AI page AI page

" .. We’re seeking mechanical, electrical, controls and software engineers to help us leverage our AI expertise beyond our vehicle fleet. "
View attachment 698944

The Tesla Day presentation highlighted all the impressive advances being made right now, in all but the TeslaBot area.

View attachment 698946

I think they can be priced much lower than that. The most expensive part of a tesla vehicle is the battery. At 125 pounds, there aren't many batteries in the robot. Vision can use same cameras as the vehicles. Other parts can be common mass produced components. Then there is the FSD computer. I think Tesla could get a basic robot to under $10,000 with mass production.

The software will be where they make money. Similar to video game console makers, Tesla could sell the robots at cost, or even at a loss, and make money selling software packages. Pick and choose what software licenses you want and it may all or most of it be subscription based.

$9999 for your robot, $200 a month in subscription fees. Could be a wide range depending on the software. Want a general laborer for doing landscaping? If you were paying a human $15 an hour, 40 hours a week, 4 weeks a month that's $2400 a month, not counting overhead.

This will bring to discussion the idea of universal basic income. I realize that is political, but Elon is right. That discussion is going to be needed in the next few years.

Anyone that was wondering if Tesla will monetize Dojo to create AWS for AI, I believe your answer is right here:

View attachment 698998

I'm pretty sure they wouldn't have made dojo so easy to drop into PyTorch coding unless they had plans to monetize Dojo usage for AI. They also mentioned during the talk how they could control how many D1 chips were used for a processing job, which would imply to me that they are already considering how they can handle many different jobs of many sizes simultaneously.

Yes a disagree from me. The TeslaBot is simply the natural progression. Let me give just one example;

GigaNevada where warm bodies are in short supply because not enough talented people want to live in the middle of nowhere.

Enter TeslaBots to work the lines and to handle logistics inside the factory (this is the natural progression of FSD). No more labor issues or housing issues at GigaNevada, free to grow to full dreamed size.

I think where people are getting stuck is imagining a humanoid robot in every home ala some sci-fi movie. Slow your roll. Oh, that’s totally happening, but you can breathe easy knowing that many of us here will be dead when that happens. It’s going to be your children and grandchildren who’ll need to worry about the robots taking over. If you’re not old like many of us - you’ve at least got some time to get used to the idea and to buy more TSLA.

Other applications:

Robot mining
Robot firefighting
Robot medical staff to help during pandemics
Robot in home care for the elderly
Robot anything you can imagine
Robot surgeons so there isn’t a 6 month wait list for elective surgeries
Robot jail guards
Robocops (yes, I went there)
Robot teachers
Robot dog walkers

I’m sad I’m not 20.

I believe the answer is yes. Even low cost employees can be costly when you add in health benefits, 401k, etc.
In addition, robots:
- can work 24/7
- don't take vacation
- don't have sick days
- will work on Holidays
- won't take off 4 months on disability
- won't sue you
- won't demand annual salary increases
- won't be unionized
- won't work from home spending all day on TMC
- will provide consistent quality
- etc

Yes - there is maintenance and investments to upgrade but I can see the cost/benefit analysis working out just as we see automation in factories today taking work away from the low wage worker.

You refer to a high cost robot....so your point is valid; the cost of the robot can't be $1m. I think if would need to be $200k or less at scale.

Boston Dynamic robots run through obstacle course. Teslabot looks at the commotion, sits down on a lounge chair, sighs and reads the newspaper.