Comment by ericd | Hacker News Reader

Comment by ericd | Hacker News Reader

Parent Full thread

How will it deal with an accident up ahead where some drunk bystander is trying to direct traffic? How will it know to ignore the drunk guy? What if it isn’t a drunk guy but a sober person directing traffic? Does the car obey in that case?

Most people wouldn't instantly know how to handle those cases either. Many people would obey the drunk guy. Maybe that's the right thing to do.

None of those are edge cases because every time it drives it will encounter some novel edge case that has never happened before and it will have to perform better than a human.

That's why these things could never be rule based, there are too many small exceptions. When they're not overfitting/able to memorize all your training examples, neural nets learn heuristics, just as people do. Different people learn different heuristics. Granted, they don't have much of the same context about the world that people do, it will take a long time to build enough examples for them to infer all of that. But Tesla's fleet is getting more driving experience every day than you will get in your entire life, and every time they train on one of those exceptions, the entire fleet will benefit.

Don’t even get started with liability. Once you take away the steering wheel the manufacturer is on the hook for every single mistake and every single accident. You’d be insane to be a manufacturer and sign up for that.

If drivers no longer carry their own insurance, this is probably going to be handled by insurance at the manufacturer level, and baked into the price. The insurance will demand certain processes to prevent large-scale bugs being rolled out.

I don't see any fantasy here, just a lot of work.

But Tesla's fleet is getting more driving experience every day than you will get in your entire life, and every time they train on one of those exceptions, the entire fleet will benefit.

Humans can reason about things that haven't to them happened before. Today's machine learning systems cannot. As you say, to react appropriately they must have been trained to do so using human annotated data.

The argument against FSD is that you would need an infinite number of annotated examples, and an infinite number of subroutines for behaving in any identified situations, because the space of driving is effectively infinite.

Until machines are able to do general reasoning about things they've not experienced before then FSD is not happening. By the way, Demis Hassabis thinks that this sort of transfer learning is the key to solving AI.

> But Tesla's fleet is getting more driving experience every day than you will get in your entire life, and every time they train on one of those exceptions, the entire fleet will benefit.

That’s not how machine learning works. You can “train on exceptions” as much as you want, and you have 0 guaranteed results. It can help, it can make no change or it can cause unexpected regressions.

It certainly can cause unexpected regressions, but of course that's being monitored during training. I'm not saying this is some real time learning system, nor that the system behaves perfectly on an exception case as soon as it's added to the training set. My point is that their engineers are getting an incredible amount of data on tricky driving situations to train their models on, far more than a human will see in their entire driving career, and when their model improves to deal with those, it improves for everyone. And at some point, these systems will be much better at driving than any human alive.

It’s one of great misunderstandings (or more likely - very effective way of getting money) to claim that machine learning is a data problem. We’re so far away from that point, that we have literally no idea what’s needed to make ML a data problem. Algorithms are extremely simple, and it’s all more or less curve fitting.

Media (and surprising amount of tech people as well) tend to claim that ML learning is like human learning - repeat something enough times and you’re done, you know how to do it. ML is no where close to that point.

You have any idea what the current state is of neural networks and their practical use? Most work only in very predictable controlled environments. Once you add one new ingredient things will fall apart. A 'complex' ai currently has less than 100 different things to deal with. And you need a hell of a fast machine to simulate millions of gens. In case of a self driving car you need approximately 1.000.000.000 different inputs. If you would the Moore law to predict the moment we deal with that amount you would have to wait about 500 years.

I do have some idea, I use them in my work. What are you calling inputs here? And gens?

> But Tesla's fleet is getting more driving experience every day than you will get in your entire life, and every time they train on one of those exceptions, the entire fleet will benefit.

Prove they are getting better. They run into the sides of trucks and off ramps on the freeway quite often (and don’t you dare blame the driver... it’s “full self driving” remember?)

Can you prove a machine learning algorithm does the right thing in novel cases it hasn’t encountered before? Nope.

And also, don’t reply with “well can humans”. That is a lame rebuttal. Computers will be held to an almost 100% non-failure standard before society accepts them. And that will never happen because of, well, reality.

Reply on news.ycombinator.com