GTC Recap Automotive (1 of 3)

Tuesday (3/19) brought together how far AI has come, how far it still has to go, and the fact that as humans we have something that AI will never replace...ever.

The focus for me on Tuesday at GTC was the automotive industry from the next gen autonomous driving startups ("AV2.0") (Part 1), to the established auto OEMs (Ford, Mercedes) embracing AI/GPUs for their aerodynamic analysis (CFD--Computational Fluid Dynamics)(Part 2) and Ford doing LLM/RAG on their factory repair manuals (Part 3).

AV2.0

First and foremost Raquel Urtasun (Waabi) was one of the most inspiring founder/CEO I saw at the show she is thoughtful and poignant in her presentation and answers.? Raquel as well as Wayve’s co-founder and CEO Alex Kendall emphasized the AV2.0 (Autonomous Vehicles 2.0) era which is an AI focused approach to autonomous vehicles to achieve “generalization” the concept that the AI system performs well in a situation it was never trained for and has the ability to handle statistically rare scenarios.? The reason this is so critical is because previous self driving approaches (“AV1.0” as referenced by the AV2.0 crowd) struggled to deal with the “rare” cases that only happen one time in on hundred million miles of driving.? The issue is the test fleet size to find these “rare” is very large and the fleet has to drive a lot of miles, and then when the autonomous vehicle performs in an expected manner then some engineer has to program in a fix for the exception (this is a significant simplification of the process, but still valid at a high level).? This becomes an endless game of whack a mole with even a sizable test fleet.

As Raquel brought up in both her talk and panel if you look at the statistics of significant driving events at first glance they are very rare.? For the US in 2021 there was 1.37 deaths per 100 million miles traveled which makes a fatal accident fairly rare per mile and yet there were 39,508 fatal accidents in that same year in the US (source: iihs driven by US DOT data, links below).? The point is these exception events are so rare they are hard to find in testing, but the sheer volume of miles driven means they happen with tragic regularity over 100 times per day.? The solution emphasized by Alex and Raquel was two main parts, using AI for the autonomous vehicles software and two high fidelity simulation that pulls in both real world data/scenarios and scenarios created by generative AI to create test scenarios no human could ever dream up.? The simulation has the potential to quickly perform millions and millions of miles of testing at a dramatically lower cost than having the equivalent? real world fleet.? To be clear real world testing never goes away, but the self driving AI is much much more mature and closer to true generalization when inevitably the rubber hits the road.? I hope for the whole self driving industry which is currently deep in the trough of disillusionment that AV2.0 can build these safer vehicles that people can trust and provide dramatically safer transportation, we will see.

On a lighter note Raquel mentioned “Waabi World” which is their high fidelity simulation environment, but I couldn’t help immediately thinking of the Griswolds heading to Walley World and AI trying to process either silicon cooking spray coated sled sparking through traffic or an ironically wood paneled station wagon stuck under a logging truck on the highway…

Next up: AI and GPUs collide with massive data in Computational Fluid Dynamics (CFD)

https://www.iihs.org/topics/fatality-statistics/detail/state-by-state

https://www.nhtsa.gov/crash-data-systems/fatality-analysis-reporting-system