How Tesla Is Building an AI Powerhouse
Tesla (Nasdaq: TSLA) is on the verge of a monumental breakthrough in machine learning and the only thing people are talking about is its stupid humanoid robot.
On Thursday, executives at the electric vehicle (EV) company held its AI Day. The two-pronged goal of the event was to showcase its artificial intelligence (AI) progress and recruitment of new engineers.
Plans went awry when Tesla Bot — a 5’8” humanoid robot — appeared on stage.
Humanoid distraction aside, investors should look past the bot and strongly consider buying buy Tesla shares.
Let’s be clear: The AI day presentation was mind-blowing. Executives at Tesla are aiming so high, it’s hard to put the scale of innovation in perspective.
Related Post: Tesla’s Recall Bodes Well for Tech Investors
Lex Fridman, an acclaimed AI researcher working at Massachusetts Institute of Technology (MIT) and frequent Tesla critic, characterized the event succinctly: “Tesla AI Day presented the most amazing real-world AI and engineering effort I have ever seen in my life.”
For Fridman, the event represented a fresh and ambitious new course for AI research. He has harshly criticized CEO Elon Musk in the past for downplaying the difficulty of the full self-driving (FSD) problem.
In Fridman’s view, the obstacles to FSD were so daunting, he didn’t think any firm could successfully navigate the landscape any time soon.
Tesla AI Day changed his mind.
What engineers did was completely reimagine computer vision by thinking exponentially bigger. Then, they built models — to collect and label the data — and a new processor to make sense of it all.
On his YouTube channel, Fridman discussed how Tesla researchers are now collecting data in 3D, as opposed to the 2D industry standard.
They’re also fusing all of the data inputs and doing machine learning on the combined dataset rather than piecing the information together later. New algorithms model time and use heuristics to plan possible routes.
The idea of AI conjures up rooms full of computers deciphering data and making choices on the fly. In reality, Tesla still employs 1,000 engineers who manually label street signs and orange road cones.
Currently, its AI system uses 10,000 GPUs to auto-label a lot of this data. On Thursday, Musk said the software is improving daily and is being completely retrained about every two weeks … and that leads to the third big breakthrough at Tesla AI Day: processing.
The Dojo computer is Tesla’s supercomputer for model training. The company is upgrading its data centers with Dojo processors specifically designed to optimize neural networks. Engineers claim these breakthrough megachips offer four times the performance of off-the-shelf GPUs while consuming 20% of the footprint.
It’s the kind of processing power required to do all the things Tesla engineers want to do with 3D-data collection, labeling, neural network model training and redeployment. That loop — data collection, labelling, training, redeployment and repeat — gives Tesla a fighting chance to finally solve FSD, according to Fridman.
Tesla is building a fully integrated AI stack that isn’t being duplicated anywhere in the industry. And most of the vehicles in its near 800,000-EV fleet are sending data back to Tesla every day.
The greater the amount of data in the pipeline, the faster the algorithms learn to solve FSD, including rare edge cases like packages falling off trucks and sudden bad weather.
That story should have been front and center Friday morning … yet most of the post-event coverage was about Tesla Bot, which may be available as a prototype next year.
I get that the event was as much about recruitment as it was self-driving cars. However, the narrative got lost with the robot. It’s cool, but also a distraction.
Investors should focus.
Tesla is building a fully integrated AI powerhouse. Fridman even suggests the scalable Dojo supercomputer might later challenge Amazon Web Services (AWS) as an AI-as-a-Service (AIaas) platform.
Related Post: Tesla Is Still Winning the Race for Autonomous Cars
Tesla shares ran up going into the event, then dropped to previous levels.
I believe shares could run up to $900, so even if you were to buy at around $700, that’s a 28.5% gain.
Other disruptive technology investors, like Cathie Wood, are still very bullish on Tesla. Her price target is $3,000. That’d be good for a cool 328.5% gain from current levels. She doesn’t care if a company is profitable now … she wants to be on board when the money starts printing, and so do I.
Longer-term investors should strongly consider taking advantage of any weakness to add Tesla shares.
Jon D. Markman