Hacker News Reader

Comment by aphextron | Hacker News Reader

>Now, I have nothing against using Linux for everything else in a car (infotainment, navigation, etc.) but I would not trust one that will have anything to do with basic controls.

Comma AI has been shipping self driving software based on Android for 4 years now. It's rock solid and completely safe, with millions of miles driven [0]. The idea of everything in automotive requiring formally verified aerospace grade software is just nuts. It's better to have flexible, adaptable software whose failure modes are fully covered than to rely on an illusion of perfection that will never actually exist.

[0] https://comma.ai/

View on HN

When you're talking about AI then your software suite looks very different. The actual running of the vehicles engine and subsystems is something that is very well known, and can be handled by embedded software/hardware quite easily. It's not that I don't think more complicated stacks can handle it, it's that I don't think they're needed. That said, no matter which way you skin it, if you increase complexity then you increase risk.

>That said, no matter which way you skin it, if you increase complexity then you increase risk.

This isn't really true, for eg. if you add redundancy technically you increase the complexity but also reduce risk (point being complexity will also be introduced when making the system safer, simplification just makes analysing and screening for bugs more effective, but there are other risk factors)

25 million miles is completely inadequate testing to make any such statement. The prevailing traffic fatality rate in the US is 1 in 125 million miles. 25 million miles means that the information is extremely insufficient to make any statement about comparative safety. In fact, making any such statement would be mathematical malpractice and criminally irresponsible. This also ignores the fact that unless those miles were all driven without any human correction you are actually testing the safety of a combined human+software solution which had better be better than just a human alone otherwise the system actually reduces safety.

>you are actually testing the safety of a combined human+software solution which had better be better than just a human alone otherwise the system actually reduces safety.

That's the entire point. Comma AI is a level 2 system. It is not intended to, nor will it ever be fully automated. It is a tool to augment and enhance human driving capabilities. Full autonomy at Level 3 and above is a 20+ year pipe dream at this point. Anyone advertising such a system (Tesla...) is just fraudulent.

> It's rock solid and completely safe

What is your source on this? I don't think even GeoHot says that.

> It's rock solid and completely safe

It sounds... Unbreakable!

Reply on news.ycombinator.com