Comment by colechristensen | Hacker News Reader

Comment by colechristensen | Hacker News Reader

Parent Full thread

colechristensen

>I think bugs and flaws in systems like this are normal at this stage

I don't think you should be able to license a vehicle to drive on public roads with such "bugs and flaws".

I don't have a choice, I have to interact with public roads. My local, state, and national government owes it to me to keep experimentation off of public roads that modern life is impossible without.

The problem with robot cars is they fail in very alien ways. They will be made illegal after a small number of horrific incidents where a car will kill people in ways a licensed human never would.

I think the usage of the phrase "intelligence" in "Artificial Intelligence" needs to stop. It was cute in the 90s but today people are starting to actually believe it when it's really a vulnerable shallow approximation that is vulnerable to profoundly unintelligent failures and it will continue to be until there starts to be a debate on if a synthetic system is actually sentient.

It's a good point about whether they should be on public roads. At the same time, it's crucial to their success that they are and experience outlier situations.

The hope is that by experiencing an outlier and making a mistake, the algorithms will be able to adapt to ensure that mistake never happens again. Humans today don't have this kind of networked, massively parallel mind to avoid the mistakes of others. Again, that's the hope.

I think it'll be a long time before learning algorithms will take over the wheel completely - but I think we are at or near a time when these systems can and should let us know of low confidence, insufficient data to make a decision, or equal confidence of multiple outcomes. A networked AI should be able to tell us far enough in advance (e.g. while driving) for humans to step in and make the final call, based on their own rationalization, empathy, etc.

> it's crucial to their success ...

I don't care about a company's success if it requires endangering members of the public who have no choice in the matter.

If we were talking about, say, astronauts and test pilots who know the risks they take and make the choice, I'd have no problem with them risking their lives. We're not.

>I think it'll be a long time before learning algorithms will take over the wheel completely

They did. A system called "auto pilot" drove a man at full speed into a barrier and killed him.

I have a limited trust of other humans on the road. The deal is that I understand them. They behave in a way that makes sense to me, even (or especially) the mistakes they make. As another driver or a pedestrian, I can communicate a lot with a car just by looking at the driver and interpreting body language and attention.

> The hope is that by experiencing an outlier and making a mistake, the algorithms will be able to adapt to ensure that mistake never happens again.

I think people are making a huge underestimation about the number and types of edge cases. If you have to teach a car not to drive into a static object at full speed, the issue sin't going to be about fixing this one problem, the issue is that if this problem exists 1000 others like it do too.

I come from cold climates and country roads, I think the fact that these vehicles are being designed and tested in the bay area (or Arizona) makes them incredibly dangerous because of the fair weather bubble that the people design them in are living. Make your engineers live and work in rural northern Minnesota and maybe you'll get a bit more trust from me.

Reply on news.ycombinator.com