Tesla's AI Super Weapon Is Officially Dead
Dojo's disbanding is so much worse for Musk than you think.
Tesla is up a very shitty creek with no paddle in sight. We all know that. In fact, the writing has been so glaringly on the wall for so long that we have become completely desensitised to Musk’s and Tesla’s downward spiral. Now, it just feels like water off a duck’s back. But the final nail is currently being driven into Tesla’s coffin, and nothing exemplifies this downfall more than the recent disbanding of Tesla’s Dojo AI supercomputer team.
Tesla’s in-house Dojo supercomputer was possibly the only advantage Tesla had in the automation race. The original idea was to have total vertical integration.
AI models require a significant amount of “training”, where supercomputers process a huge amount of data, recognise patterns in that data, and create a neural network (AI) that can identify and replicate those patterns. AI startups like OpenAI and Waymo use supercomputers built on third-party chips from the likes of Nvidia to achieve this. These off-the-shelf computers are cheap to buy, and while they are optimised to train AI, they are still general-purpose, making them less efficient.
Tesla took a different approach with their Dojo supercomputer. By designing the chips in-house, they could optimise the supercomputer for training their self-driving AIs. As such, they would use far less computing power and energy to optimise an AI model than these off-the-shelf chips. Given that energy costs are the primary expense for AI development, rather than computer hardware, this approach is significantly more cost-effective, provided Tesla can deliver the chips on time.
And this wouldn’t just give Tesla a cost advantage. Dojo was a key factor in Tesla overcoming one of the biggest limitations in the AI industry, the efficient compute frontier, and leapfrogging the AI automation industry in one fell swoop.
I have covered the efficient compute frontier many times (read more here), but effectively, this concept explains how AI training has diminishing returns. In order to improve your AI, you need to train it on more data, which requires more computing power and energy, which is wildly expensive. However, the relationship between the amount of training data and the performance of the AI isn’t linear. There are diminishing returns. So, if you double the training data of a small AI, you might see a performance increase of 5%, but if you then double the training data again, you might only see a performance increase of 1%, and so on. It is a Sisyphean task. At a certain point, the cost outweighs the performance gain, and creating a bigger, better AI becomes unfeasible.
Now, Dojo wouldn’t have negated the efficient compute frontier, but it would have enabled Tesla to dig deeper into these diminishing returns by theoretically making AI training far cheaper and pushing the point at which it becomes unfeasible further down the line. In short, Dojo was supposed to be Tesla’s “secret sauce” to develop AIs that are substantially more accurate than those of its competitors.
And Tesla needs significantly better AI than its competitors.
Again, I have covered this subject several times before (read more here), but other self-driving companies use a plethora of sensors, from lidar to radar and ultrasonics, as well as cameras, to understand the world around them. They also use highly detailed 3D maps of their operational areas as a reference and run multiple AIs using these different sensor types. This enables them to identify and negate errors in the AI, allowing them to drive safely without requiring a painfully precise AI. This is known as redundancy.
Tesla doesn’t have redundancy. Musk overruled his engineers and forced FSD (Full Self-Driving) to use a vision-only system, which exclusively involves computer vision AI and cameras. They also don’t use detailed maps as a reference, and instead, their AI is a general-purpose AI, which is expected to understand the rules of the road. This concerning lack of safety nets or redundancy in the system means that Tesla’s AI has to be nearly 100% accurate and reliable to even come close to the safety levels of competitors like Waymo.
So, in reality, Dojo wasn’t going to help Tesla leapfrog the competition — instead, it was designed to keep Tesla’s bafflingly stupid self-driving approach in the race.
But why did Musk disband Dojo?
According to him, it’s a simple case of brain drain. Peter Bannon, who had headed the Dojo project since 2023, left Tesla for Density AI and took the majority of Dojo’s engineers with him. Density AI was actually created by the previous Dojo head, who left in 2023. And Tesla’s original head of chip design, Jim Keller, left all the way back in 2018.
It seems Musk just can’t keep his AI talent, which isn’t surprising. He has a tight, insecure, authoritarian grip over Tesla’s AI division. He has overruled AI experts and engineers multiple times, to utterly devastating results (read more here). Musk has threatened to leave Tesla and develop AI elsewhere unless he is handed $56 billion in Tesla shares, taking his current 13% stake to 25%, giving him even more control over the company. If you are a leading talent in the AI industry, Tesla is possibly the least appealing place to be. Your boss is an egotistical dick who has pushed you past your natural limit but is still happy to overrule you at a moment’s notice because his understanding of AI doesn’t extend very far beyond the Dunning-Kruger effect. All while he takes all the recognition and kudos for your work.
Yeah, I’d be leaving too. Naturally, with the entire Dojo team moving onto finer pastures and other engineers having very little interest in working under Musk, the project met a dead end, and so it was wrapped up before a single supercomputer was delivered.
This is a broader issue at Tesla, given that critical talent across the entire company, from the head of engineering for Optimus to the VP of software engineering and countless others, are intentionally fleeing Tesla. But that is a story for another day.
So, now that Dojo is no more, how will this impact Tesla?
Well, it’s a death sentence. And that isn’t hyperbole.
Tesla is currently building AI supercomputers using off-the-shelf chips. As such, they don’t have any cost advantage in developing AI. But they needed that cost advantage to make their vision-only self-driving system even a viable competitor in the self-driving race. Therefore, Tesla’s FSD has a fatal disadvantage, as it will be catastrophically expensive, if not completely impossible, to make the AI accurate enough to meet safety standards for fully autonomous cars.
Now, that alone shouldn’t be a death sentence. After all, Tesla essentially pioneered the EV market and has maintained a near-monopoly over it for over a decade. As long as Tesla expands its lineup with sensible and popular cars, continues to improve its EV technology, and maintains its incredible brand image, it could lose the self-driving race and still be okay.
What’s that? Musk has halted vehicle development, and the only new model launched in over a decade was the largest egotistical flop the automotive industry has ever seen?
Unfortunately, Tesla technology hasn’t improved in half a decade, while Asia and Europe have developed significantly better battery, motor, and manufacturing techniques, with many of their EVs now undercutting Tesla in price with the same or better specs. And, by heavily funding, supporting and associating with the most unhinged authoritarian leader the US has ever seen, as well as performing a double Sieg Heil on national television, Tesla’s impeccable brand image is now in tatters. Everything has led to the nosedive of Tesla sales and revenue worldwide.
Yeah, that would do it.
By abandoning Tesla’s fundamental business principles and chasing fictional futuristic unicorns, Musk has bet Tesla’s entire future on AI, while Tesla holds a dead man’s hand.
Dojo was Tesla’s only slight advantage in the AI industry. That is, assuming the team could actually deliver on the batshit pie-in-the-sky numbers Musk previously claimed. In fact, I highly suspect the team wasn’t disbanded due to brain drain but rather because they couldn’t deliver any significant optimisation, making off-the-shelf chips a better option and rendering Dojo obsolete. By terminating the project now, they can cover up their deeply embarrassing mistake and blame the less embarrassing issue of the talent exodus.
Either way, Tesla’s only perceived AI advantage is now gone. All their routes forward have closed. They can’t win this race. The AI speculation that holds Tesla’s valuation at such an astronomical level will fail, and as I have covered before, that will kill Tesla. It’s not a question of if — it’s a question of when. How long can Musk convince people that the emperor’s clothes are still fresh and exciting?
Thanks for reading! Don’t forget to check out my YouTube channel for more from me, or Subscribe. Oh, and don’t forget to hit the share button below to get the word out!
Sources: The Verge, Tech Crunch, Electrek, Will Lockett
Good to know that engineers, at least, still have a grasp on reality.