Face it, self-driving cars still haven’t earned their stripes | Hacker News Reader

Face it, self-driving cars still haven’t earned their stripes

garymarcus.substack.com

16 pointsnsoonhui16 comments

Open articleView on HN

Self driving trains, on the other hand, complete hundreds of thousands of passenger-miles every day without incident.

That derailment by Youngstown, Ohio was a huge ecological disaster

And what kind of train was that?

And it wasn't self driving

I'm guessing you have in mind subways operated by UTO?

I wonder if this is a good application for recently developed multimodal/vision-language models that demonstrate "emergent" properties like an ability to classify objects unseen in their training data.

That would definitely help. Honestly though I feel like level 5 self driving requires a theory of mind, and I'm not convinced that we're there yet.

I keep coming back to things like you have a pedestrian on a street corner. What are they going to do? Are they just standing there or are they going to cross? Are they paying attention to their surroundings or not?

Yeah basically. We need a decent world model to handle the edge cases, and current ML approaches are 100% not there yet.

I am unconvinced. Mammals suck at edge cases unless they have the training. Machines can machine learn edge cases by the boatload every day, and learn faster than any mammal.

Also, most driving is not edge cases. Most driving is 40 - 70 in 5th or 6th gear on highways.

Finally, humans will always need to be in the loop with self-driving cars because humans are the passengers. Question is, what can we put in a self-driving car that will permit responsible human intervention?

Because lots of idiots think they are better drivers than they really are. A subset of them will think they ate better than the machine, and most of the time, they will be wrong.

Well, I might be an idiot, but if the bar is currently set at "never hit a fire truck", I'm still several orders of magnitude better than the machine.

And with many fewer hours of training and experience, at that.

Most driving for some people*

Most of my driving is national speed limit country/2 lane roads. Because of where I live I'm rarely on big A roads or motorways. This kind of driving has many more edge cases, for example yesterday I had to go into the other side of the road to avoid a lot of ivy hanging off a stone wall that was torn down by the wind. I'm not sure I'd trust a self driving car with that (would radars see it?)

Well, it turns out the edge cases are where the self-driving cars crash into stuff.

> Machines can machine learn edge cases by the boatload every day, and learn faster than any mammal.

And yet here they are making the same mistakes for years, often ones which humans would not make (or would recover from, such as the robot cars blocking emergency vehicles). It seems likely that at some point in the future what you wrote will not be wrong but that hasn’t happened yet.

When missing an edge case can maim or kill multiple people, "good enough" is not nearly good enough.

Reply on news.ycombinator.com