> Google makes a big deal about how the system is to do some degree interpretable, but there is no obvious way to integrate large language models with the kind

> Google makes a big deal about how the system is to do some degree interpretable, but there is no obvious way to integrate large language models with the kind of formal formal verification methods that are now routine in the design of microprocessors, USB drivers and large aircraft. Pasting that into an online grammar checker finds that the word "do" is superfluous. I guess Marcus really does underestimate the value of AI. More substantially, let me say that the claim that in "2022 [self-driving cars] are still only in very limited experimental tests" is an unfair stretching of the truth. Waymo, for example, has been providing driverless rides to paying customers (not beta testers under NDAs) on public roads since 2020, which can hardly count as an "experimental test".[0] Similarly the criticism he quotes "There is an infinitude of different Halloween costumes." is deliberately obtuse. A self-driving car doesn't need to distinguish between a witch and a werewolf walking into a road, just as human drivers on Halloween don't think "If I see a ghost tonight I mustn't drive through it because it might actually be a kid dressed as a ghost". He's right, though, that Google probably should have tried to come up with a formal world model that the robot would be working with, to confirm that an instruction like "pick up cat" would be rejected no matter how it was phrased. Perhaps they could have released a simulated environment and allowed researchers to try tricking the robot into carrying out dangerous (simulated) actions in it, with a reward for finding such edge cases. [0] https://www.theverge.com/2020/10/8/21507814/waymo-driverless-cars-allow-more-customers-phoenix https://www.theverge.com/2020/10/8/21507814/waymo-driverless...