• The secret to Tesla’s Optimus robot
  • Tesla’s full self-driving technology is reaching critical mass
  • Is Musk coming for Amazon’s business?

Dear Reader,

As I promised yesterday, we’re going to spend today’s entire issue digging deeper into the announcements at Tesla’s AI Day from last Friday and understanding why they are so significant.

And we’ll start with…

The rise of Tesla’s Optimus…

At its event last year, Tesla revealed plans to develop and manufacture a humanoid robot called “Optimus”. It was just an idea at the time.

And Tesla announced its plans by having a human dress up as a robot and prance around the stage. Musk of course was having fun with the event. I found it hilarious only because I knew that there was substance behind the announcement.

Elon Musk said that Tesla would have a commercial prototype within 12 months’ time. Most people scoffed at this. It didn’t sound feasible to go from a human in a robot costume to a prototype so quickly. But I knew Tesla would get it done (for reasons I’ll explain in a moment).

That’s why everyone was curious to get an update on Optimus on Friday. And Tesla did not disappoint…

First, Tesla brought out its development platform for attendees to see. That’s the robot below in the middle. Tesla was able to whip that together in a matter of just a few months as a way to “fail fast” and iterate quickly.

Source: Tesla

Obviously, it looks very mechanical and nothing like the concept image on the left. Tesla engineers brough it out on stage and it was capable of unaided movement and walking. 

The team said it was the first time they had “released” it untethered. I’m sure some competing companies chortled when they saw this, forgetting that Tesla went from scratch to a functional robot in just a few months.

After Tesla showed off the development platform, it brought its latest generation of Optimus. That’s the image we see on the right.

This is an early commercial prototype that Tesla is actively improving. And as we can see in the image, the aesthetics are greatly improved.

We should keep in mind that Tesla’s goal is to produce a robot assistant who can help with tasks in the office or in the home or in an industrial setting. Tesla’s goal isn’t to make something that’s capable of acrobatic movement like Boston Dynamics did last year.

Optimus is designed to climb and descend stairs. It can walk forward, backward, or sideways. It can squat down to pick up an item. And the robot can even twist while walking with an object in hand.

Optimus is also capable of carrying around boxes and watering plants. It can even use tools like a screwdriver or a drill. These are things that require fine motor skills.

So how did Tesla go from an idea to a prototype in such a short period of time? How did Musk and his impressive team pull it off so quickly when virtually everybody thought it was impossible?

The answer has to do with its advanced artificial intelligence (AI). The hardware is comparatively easy compared to the software; it’s the “intelligence” that is much more difficult. 

And for that, Tesla didn’t need to start from scratch. Tesla was able to take its AI software for self-driving cars and use it as the foundation for Optimus’ “brain”.

In other words, the same AI that can drive a car can help a robot navigate its environments safely and with purpose. Tesla had a major leg-up right from the start.

This has been one of the biggest misunderstandings about Tesla. It is not a traditional automotive company. It is a tech company. More specifically, it’s an artificial intelligence company.

And Tesla has already proven that it can manufacture advanced electric vehicles (EVs)—the hardware—at scale. It can clearly do the same for robotics.

That’s important because scale equates to lower costs. And Tesla wants Optimus to cost less than $20,000. At that price point, Optimus would be realistic for even small- and medium-sized businesses, as well as the consumer market.

Tesla proved the critics wrong when it built the most technologically advanced electric vehicles on the planet at scale. It is going to do exactly the same thing with Optimus for robotics. 

Tesla’s self-driving cars now react faster than humans…

Tesla also revealed just how much progress it has made with its full self-driving (FSD) technology at this year’s AI Day.

And Tesla is making phenomenal progress. This image tells the story:

Source: Tesla

As we can see, Tesla only had 2,000 customers using its FSD software last year. Fast forward to today and 160,000 people are using it. That’s a huge jump.

For any readers who are interested, I tested Tesla’s FSD technology last year. We can see the results for ourselves below.

The technology was impressive at the time. And it’s dramatically improved since then.

The reason for that is that Tesla’s “fleet” of self-driving cars (its customers’ cars) act as real-world data collection for self-driving. As Tesla increases the number of beta users, it multiplies the data set from which its self-driving AI can learn.

Tesla has run more than 75,000 training models over the last year in an effort to optimize and improve its self-driving software. As a result, it released 35 distinct software updates for FSD during that timeframe. Incredible.

Compare that to what we normally see with computing systems like our laptops or smartphones. For iPhone users, Apple typically releases a new software update every one or two months.

That equates to six to 12 updates every year. By comparison, Tesla is doing new updates at least every other week. It is iterating with its software and artificial intelligence at a pace never seen before.

As a result, Tesla cars operating in FSD mode can now make decisions in less than 10 milliseconds. Compare that to humans which take somewhere between 390 and 600 milliseconds to react to a traffic situation or dangerous incident.

And that means Tesla has basically addressed all the big challenges. The company is now focused on solving for those last 1% of complex cases that the FSD software still hasn’t fully refined.

But given all the data Tesla has coming in to train its AI from those 160,000 plus vehicles, I’m confident it will figure out these last challenges very quickly. I wouldn’t be surprised if Tesla announces that it has achieved Level 5 FSD before 2023 is out.

At that point, Tesla’s cars will be 100% autonomous. That’s the holy grail of autonomous driving.

There’s a bigger motive behind this announcement…

As we just discussed, most people think of Tesla as an EV company. Very few see Tesla for what it is—an AI company. And even fewer realize that Tesla is now also a high-performance computing company. And this may very well signal that Tesla has ambitions to enter a new market.

We all know that Tesla’s artificial intelligence is some of the most advanced in the world. What it is doing is unique. And to accomplish its goals, Tesla went so far as to design its own semiconductors and its own supercomputer for training its AIs. It’s called Dojo. We talked about it last summer.

At the time, Dojo was the world’s fifth most powerful supercomputer in the world. And Tesla uses it exclusively to train its AI. That’s how it makes so much progress so quickly.

And at the event, we got some interesting news for Dojo…

Tesla has been working hard on what it calls an “ExaPOD”. This is a method for scaling Dojo so that it becomes more and more powerful. Here’s a look:

Source: Tesla

The text on the image might be difficult to read. But it says that Tesla’s ExaPOD will enable 1.1 exaflops of computing power. That equates to one quintillion (1,000,000,000,000,000,000) floating-point operations per second (FLOPS).

It’s hard to get our mind around processing power like that. To give us some idea, we can consider that an average laptop might be able to do 7,000,000,000 FLOPS. Our laptop would be like an abacus compared to an ExaPOD.

This alone would make Dojo one of the top two or three most powerful supercomputers in the world. And Tesla announced that this scaling method will enable computation at one-sixth the cost of the world’s other GPU-based supercomputers.

Tesla plans to build seven of these ExaPODs in Palo Alto, CA. Once all seven are running, Dojo will easily be the most powerful supercomputer on the planet. And again, it will be 100% focused on AI and machine learning (ML).

But here’s the thing—a supercomputer running at just one exaflop would be more than enough to develop and improve the AIs that Tesla is developing for vehicles and Optimus. It doesn’t really need any more “muscle.”

So why does Tesla need seven exaflops?

I can’t help but think there’s a bigger play at work here.

I believe Tesla may be quietly positioning itself to become a cloud service provider. With seven exaflops of computing power, it could easily compete with top cloud offerings like Amazon Web Services (AWS) and Google Cloud.

Or perhaps Tesla is working on a completely new AI-enabled product that it hasn’t talked about yet. That’s also a distinct possibility.

Perhaps Tesla is looking at developing swarms of industrial robots capable of taking on challenging construction projects in short periods of time. Musk has already tackled tunneling technology with his Boring Company. But perhaps there’s something even bigger that leverages the autonomous technology that Tesla’s EVs and humanoid robots take advantage of….

There’s something big going on behind the scenes at Tesla.

I am going to be watching closely for clues as to what it may be. There’s always breadcrumbs for those of us who are paying attention…


Jeff Brown
Editor, The Bleeding Edge