Top four highlights from Elon Musk’s Tesla AI Day – TechCrunch

Elon Musk wants Tesla to be seen as much more than an electric car company. On Thursday, Tesla AI Day, the CEO described Tesla as a company with deep AI activity in inference-level and training-level hardware that could later be used for applications beyond self-driving cars, including a humanoid robot that Tesla is apparently building.

Tesla AI Day, which started after a rousing 45 minutes of industrial music straight from the soundtrack of “The Matrix”, featured a series of Tesla engineers explaining various Tesla technologies with the clear goal of recruiting the best and brightest to join Tesla’s vision and AI team and help the company move toward autonomy and beyond.

“There’s a huge amount of work to make it work and that’s why we need talented people to get involved and solve the problem,” Musk said.

Like both Battery Day and Autonomy Day, the event was streamed live on Tesla’s YouTube channel on Thursday. There was a lot of super-tech jargon, but here are the top four highlights from the day.

Tesla Bot: An absolutely real humanoid robot

This piece of news was AI Day’s last update before the public inquiries kicked in, but it’s certainly the most interesting. After the Tesla engineers and managers talked about computer vision, the Dojo supercomputer, and the Tesla chip (which we’ll all come back to in a moment), there was a brief interlude in which what appeared to be an alien go-go dancer appeared on the scene. stage, dressed in a white suit with a shiny black mask for a face. Turns out this wasn’t just a Tesla stunt, but rather an introduction to the Tesla Bot, a humanoid robot that Tesla is actually building.

Image Credits: Tesla

When Tesla talked about using its advanced technology in applications outside of automobiles, we didn’t think he was talking about robot slaves. That’s no exaggeration. CEO Elon Musk envisions a world where the human toil like grocery shopping, “the work people least like to do,” could be taken over by humanoid robots like the Tesla Bot. The bot is 5’8″, 125 pounds, can deadlift 150 pounds, run at 5 miles per hour and has a head screen that displays important information.

“It’s obviously meant to be kind and navigate in a world built for people,” Musk said. “We set it up so that you can run away from it on a mechanical and physical level and most likely overpower it.”

Because everyone is definitely afraid of getting beat up by a robot that really has had enough, right?

The bot, expected to be prototyped next year, is being proposed as a non-automotive robotic application for the company’s work on neural networks and its advanced Dojo supercomputer. Musk did not share whether the Tesla Bot could dance.

Disclosure of the chip to Dojo. to train

Image Credits: Tesla

Tesla CEO Ganesh Venkataramanan unveiled Tesla’s computer chip, designed and built entirely in-house, which the company uses to run its supercomputer, Dojo. Much of Tesla’s AI architecture relies on Dojo, the neural network training computer that Musk says will be able to process massive amounts of camera image data four times faster than other computer systems. The idea is that the Dojo-trained AI software will be pushed to Tesla customers via over-the-air updates.

The chip that Tesla unveiled on Thursday is called D1 and contains 7nm technology. Venkataramanan proudly held up the chip that he said has GPU-level compute with CPU connectivity and twice the I/O bandwidth of the state-of-the-art network switch chips out there today that would be the gold standard. must be. He walked through the technical details of the chip and explained that Tesla wanted to own as much of its tech stack as possible to avoid bottlenecks. Tesla introduced a next-gen computer chip produced by Samsung last year, but it hasn’t quite managed to escape the global chip shortage that has been rocking the auto industry for months. To survive the shortfall, Musk said in an earnings call this summer that the company was forced to rewrite some vehicle software after it had to be replaced with alternative chips.

Aside from limited availability, the overall goal of taking chip manufacturing in-house is to increase bandwidth and reduce latencies for better AI performance.

“We can perform compute and data transfer at the same time, and our custom ISA, the instruction set architecture, is fully optimized for machine learning workloads,” says Venkataramanan on AI Day. “This is a pure machine learning machine.”

Venkataramanan also unveiled a training tile that integrates multiple chips to get higher bandwidth and incredible computing power of 9 petaflops per tile and 36 terabytes per second of bandwidth. Together, the training tiles make up the Dojo supercomputer.

Towards fully self-driving and beyond

Many of the speakers at the AI ​​Day event noted that Dojo won’t just be a technology for Teslas Fully self-propelled (FSD) system, it is definitely an impressively advanced driver assistance system that is certainly not yet fully self-driving or autonomous. The powerful supercomputer was built with multiple aspects, such as the simulation architecture, which the company hopes to expand to be universal and even open up to other automakers and technology companies.

“This isn’t meant to be limited to just Tesla cars,” Musk said. “Those of you who have seen the full self-driving beta can appreciate the speed at which Tesla’s neural network learns to drive. And this is a specific application of AI, but I think there are more applications down the road that will make sense.”

Musk said Dojo is expected to be operational next year, after which we can talk about how this technology could be applied to many other use cases.

Troubleshoot computer vision

During AI Day, Tesla again supported its vision-based approach to autonomy, an approach that uses neural networks to make the car function ideally anywhere on Earth through its “Autopilot” system. Tesla’s head of AI, Andrej Karpathy, described Tesla’s architecture as building an animal from scratch that moves around, perceives its environment and acts intelligently and autonomously based on what it sees.

Andrej Karpathy, head of AI at Tesla, explains how Tesla is managing data to realize semi-autonomous driving based on computer vision. Image Credits: Tesla

“So of course we’re building all the mechanical components of the body, the nervous system, which has all the electrical components, and for our purposes, the autopilot brain, and specifically for this part the synthetic visual cortex,” he said.

Karpathy illustrated how Tesla’s neural networks have evolved over time and how now the car’s visual cortex, which is essentially the first part of the car’s “brain” that processes visual information, is designed in combination with the broader neural network architecture, so that information flows more intelligently into the system.

The two main problems Tesla is trying to solve with its computer vision architecture are temporary closures (such as cars at a busy intersection blocking Autopilot’s view of the road behind) and signs or markings that appear earlier on the road (such as if a sign is 100 meters away). back says the lanes will merge, the computer once struggled to remember that by the time it reached the merge lanes).

To fix this, Tesla engineers fell back on a spatially returning network video module, where different aspects of the module track different aspects of the road and form a space-based and time-based queue, both of which are a cache of data that the model can reference when making predictions about the road.

The company harnessed its more than 1,000-person manual data labeling team and showed the public how Tesla automatically labels certain clips, many of which are pulled from Tesla’s fleet along the way, for labeling to scale. With all this real-world information, the AI ​​team then uses incredible simulation, creating “a video game with Autopilot as the player.” The simulations are particularly helpful with data that is difficult to find, label, or in a closed loop.

Background information on Tesla’s FSD

At about forty minutes in the waiting room, the dubstep music was accompanied by a video loop showing Tesla’s FSD system with the hand of an apparently alert driver simply grazing at the wheel, no doubt a legal requirement for the video after investigation. to Tesla’s claims about the capabilities of its certainly non-autonomous advanced driver assistance system, Autopilot. The National Highway Transportation and Safety Administration earlier this week said: they would open a preliminary investigation into Autopilot after 11 incidents in which a Tesla collided with parked emergency services.

A few days later, two US Democratic senators appealed to the Federal Trade Commissionn to investigate Tesla’s marketing and communications claims around Autopilot and its Full Self-Driving capabilities.

Tesla released the beta 9 version of Full Self-Driving in July with great fanfare and rolled out the full suite of features to a few thousand drivers. But if Tesla wants to keep this feature in its cars, it needs to take its technology to the next level. That’s where Tesla AI Day comes in.

“Basically, we want to encourage anyone interested in solving real-world AI problems at the hardware or software level to join Tesla, or consider joining Tesla,” Musk said.

And with technical nuggets as deep as Thursday’s plus a beating electronic soundtrack, what red-blooded AI engineer wouldn’t frown to join the Tesla crew?

You can check out the whole thing here: