Autonomous Driving – Brains not brawn

End-to-end machine learning is not the answer.

• SoftBank, Nvidia and Microsoft are pouring $1bn into a system that relies entirely on machine learning to drive vehicles in the hope that if the model is big enough and trained with enough data magically, the answer will pop out at the end.
• I suspect that when it comes to the problem of driving cars, they will be disappointed.
• SoftBank is leading a $1.05bn round into UK-based autonomous driving start-up Wayve and I suspect that it is providing almost all of the funding.
• Microsoft is an existing investor and together with Nvidia, provides credibility to the investment even if they are risking very little in the transaction.
• This is because most of the money that Microsoft and Nvidia are investing will come back to them in revenue for Azure or sales of chips to run the models in the vehicles.
• While I think that there is nothing wrong with Wayve as a company, I think that trying to crack an infinite problem with a finite solution is not going to work.
• This is because all systems based on deep learning and neural networks have no causal understanding of what it is that they are doing and as such, are unable to deal with any situation that they have not been explicitly taught.
• Hence, for tasks where the dataset is finite and stable, deep learning can perform tasks to a superhuman level of ability.
• However, the road is neither of these things which is why I continue to think that deep learning on its own is insufficient to provide the causal understanding that is required to drive the road as safely as humans.
• I have long argued that the best solution for the road will be one that uses a combination of rules-based software and machine learning working together.
• This is because software can reason but it can’t learn while deep learning can learn but it can’t reason meaning that if they can be properly integrated, then they should be able to complement each other.
• This is how I have long thought that the autonomous driving problem will eventually be solved, but both Wayve and Tesla, with the latest version of its Full Service Driving (FSD 12.1), are going in the opposite direction.
• Consequently, without a system that has some understanding of causality and an ability to reason, I think that neither Wayve nor Tesla will arrive at a solution that can drive vehicles as well as humans.
• This issue is not limited to autonomous driving systems but is also rampant in large language models (LLMs) which consistently make things up, get things wrong and are generally unreliable as a result of their lack of understanding of causality.
• Wayve’s business model is to license its software to vehicle manufacturers rather than go direct to market with a fleet of robotaxis which is the right choice.
• However, I think that this will be problematic as I am not convinced that its solution will be better than Waymo, Mobileye, Cruise or any of the other very good Chinese offerings.
• Consequently, competition will be tough and prices low meaning that only the biggest and best solutions survive.
• This is why I think that the market for both robotaxis and autonomous driving software will not be nearly as big as SoftBank’s valuation of Wayze seems to indicate, leaving me very cautious about the sector from an investment perspective.
• The only sub-sector I like in this space is lidar where Ouster is my top pick as it is not dependent on the automotive sector to thrive and is in better financial condition than its peers.
• I have a position in Ouster which is discussed in more detail here.