Finally saw the replay last night. I hung on every word of the chip presentation. I’m the kind of nerd who read Microprocessor Report in the ‘90s for the pure joy of learning whose translation lookaside buffer had the lowest miss rate. And my interest in neural nets dates back to Rumelhart’s “Parallel Distributed Processing.”
With that said, the single comment I found most noteworthy was Elon’s saying that Tesla’s cars now shipping (by which I assume he meant the Model 3) are designed for a million-mile useful life, the same spec as for large trucks. Doing this obviously makes achieving a gross cost that supports a $35k price more difficult. The same is true about their currently shipping all the systems redundancy needed to achieve reliable autonomy. And in terms of profitability, his comment that the cost of developing autonomy was essentially the company’s whole expense structure was also telling.
This is not meant as a criticism: I want them to succeed and I am not a short. My high level conclusion from all this – which has been pretty obvious all along, but the comments in this presentation somehow drove it home for me at a more emotional level – is that Elon really has a completely different view of what business he is in, and of which success factors are important and which are not, than any other car company. A view that is differentiated by much more than electrification. He is pursuing a certain vision of the future of transportation, whether in space, in tunnels, or on surface roads, that is not evolutionary. And he is pursuing his vision even if it takes him off the path of maximizing business success and survival in the nearer term, that path being one where he would refresh his luxury car to have more luxury, cost-cut his affordable car to have more affordability, fix the bugs and deficiencies in his driver-facing software, and use his expense structure to generate more sales either through advertising or more frequent styling changes to stimulate repeat buying.
No, he is pursuing a vision of what customers ultimately will want, even though they don’t realize it yet (a la Steve Jobs, only more so). I will say that I, for one, shudder at the thought of buying a car only to let a bunch of probably vandalous strangers ride around in it while I am having dinner. I similarly shudder at the thought that the most fun-to-drive car I’ve ever owned (and I’ve had some good ones) won’t be shipped with a steering wheel in a few years. But then I’m 60 years old, so I’m probably not the best indicator of trends in the future.
It will be interesting to see how it all unfolds! FWIW I thought both the hardware and software presentations were very good. It seems like their on-chip memory architecture allows the multiply-accumulate array to have much better utilization than would GPUs that fetch their data off chip. I would also love to ask them about the number of distinct neural nets they now use and plan to use, and their specific roles, particularly in path planning/driving policy, which they implied is now mostly heuristics implemented on conventional CPUs, and relatedly I would ask about the different levels of abstraction that the different nets operate at: e.g. pixel processing vs. processing various representations of identified objects.