cliff harris
Member
I'm not worried for several reasons:The "dominance" part I legitimately don't understand. Self-driving is hard right now because no one has done it. Waymo sort of has done it, but they control all the vehicles and the vehicles only work in specific areas.
But if Tesla comes up with a general solution and then releases it out to the world (through software updates to the cars), I don't understand what's stopping a competitor from just training their self-driving program off of "behave the way this Model 3 would in this situation". It feels like tesla's unique access to data + capital won't be that much of an advantage once everyone has the distilled result of it in their car.
I think OpenAI even at one point had some clause about not allowing ChatGPT to be used to train other models. And they control the whole stack -- they can actually rate limit / ban people for bad behavior. But the FSD computer is just sitting in all of our cars.
1) Tesla AFAIK use ASICS, meaning the chip is literally designed with the code in mind. That means you need not only all the code, but also the hardware. Its not exactly easy to look at a 9-billion transistor chip and just get someone to copy it for you. The chip design will be intellectual property, and you can't simple copy and paste the hardware. How do you persuade TSMC to anger Tesla by even attempting such a thing?
2) Tesla's end-to-end NN is assuming certain information about not only about the position of every camera on the vehicle, but the shape of the vehicle too. Even if you perfectly copied code and hardware and sensors, if you vehicle is 1" longer than expected, its going to be...interesting. Also you better have exactly the same performance as a Tesla too, because the control of brakes and wheels will be in that network too.
3) Even if magically you copy all the vehicles, all the sensors, and all the code, and the hardware, you still are going to have to be able to update it all to keep up with any changes. If there are changes to any law, or any feedback from organisations like NHTSA, then you need to re-train the networks from scratch. I don't think regulators will approve a system where the company says "It works for now, but we dont know how, and we can never change anything".
Not only is it legally impossible for anyone to copy FSD outright, its also almost physically impossible, and way, way, way too impractical. No company would ever consider attempting it. It will be dramatically simpler to license FSD from Tesla.
LLMs are different. You can scrape the open web for training data relatively quickly. Text data is laughably small compared to video data, and errors are acceptable because LLMs like ChatGPT don't control life-threatening chunks of metal at 70mph. The compressed text of Tolstoy's war and peace is 1MB. Even a low-res movie 2 hours long can be 3-4,000 times bigger.
The only possible competition would be a theoretical scenario where, for example, Apple buys General Motors, retrofits a ton of cameras to their existing fleet, spends $10-15Billion on compute, and works for a few years to build up an equiv-tech system. That would be to get to where Tesla are already right now today. Not impossible, and there may be huge pressure for something like that, given the alternative is that Tesla owns self-driving globally as a monopoly.