Hacker News Reader

Comment by seibelj | Hacker News Reader

Driving on roads with human drivers, bikers, pedestrians, animals, freak weather events, maintenance and construction workers, and all manner of situations is - in reality - a problem of unbounded complexity. Unless we create infrastructure that radically reduces the complexity, like invisible “tracks” and banning as many variables as possible (other humans) then nothing short of AGI will get us close to the current below-average human driver.

Self driving cars is Silicon Valley hubris and hype at its finest. I genuinely feel proud about the computer scientists who have wrangled millions of dollars from investors, truly excellent work, I wish all engineers could get that kind of money. Hopefully the billions keep flowing.

View on HN

Humans also fail at those things and we get accidents. We do not need perfect cars that deal with every conceivable situation in the best way possible. We only need them to deal about as good as your average human driver.

Yes, self-driving software will be much dumber than a human, but it has the advantage of constant attention, lower reaction times and the ability to access more data. Perhaps their failure modes will be different due to that, but as long as the accident rate and severity is comparable that's good enough. You also need to factor in the value and lifetime-hours gained by freeing up human drivers from the task.

> nothing short of AGI will get us close to the current below-average human driver.

If people try to climb icy 20% inclines with summer tires every winter is that the result of general intelligence or just zombie mode driving? The correct choice for an autonomous car in that situation would be to refuse or look for alternatives instead of doing the dumb thing.

> Unless we create infrastructure that radically reduces the complexity, like invisible “tracks” and banning as many variables as possible

This is the road Chinese companies chose. Biggest interest are from public transport companies. Most work with very simple directional radio beacon and obstacle detection (not obstacle avoidance.) Works as an experiment in a few cities.

The interest from the state is a very distant hope to automate low speed traffic in cities, after the state lost all hope to fix local driving habits (https://www.youtube.com/results?search_query=chinese+drivers). Second to it is also a distant hope to do something with road congestion if traffic will be forcefully directed by some capacity optimising computer program.

My own take on that would've been trying to automate high speed highways instead. Some cars are quite capable of 150kph+, but most of drivers aren't. China has the biggest network of elevated/isolated expressways in the world, and adding computer guidance should be the easiest option out of all others.

There are lots of places where the variables are constrained enough that the current state could be useful. An AI shunt truck would be super useful and a much easier problem than a self driving car. We don't even have those yet and nobody seems to be trying to make them or similiar. I think most of the reason is there's no chance of hockey stick growth for niche commercial applications so VCs don't care.

It's not of unbounded complexity, though.

That doesn't imply it's not complex, it is very complex, but it's way, way off being a problem whose solution maps to any other problem, i.e. only solvable by general AI.

Remember also, the goal is not dealing perfectly with every possible situation. The goal (at first) is dealing slightly better with the average predictable situation than the average human driver.

> The goal (at first) is dealing slightly better with the average predictable situation than the average human driver.

The average human driver is terrible at driving. The goal should be higher; i.e., to be slightly better than the top 10% of human drivers. One of those skills will be refusing to drive when the weather is too bad.

I totally agree, just the absolute minimum and the initial goal is matching or slightly surpassing that level. Once that happens I guess you'd expect it to get much better pretty quickly as well, which is good.

On the other hand we do let the worst human drivers (that don't immediately kill someone in the first hour of driving) onto roads too. The same low bar should be applied to autonomous cars as they are still scaling up. The bar should be raised as their percentage of the overall fleet increases. This treats them as a collective of junior drivers that we expect to improve with time.

> a problem of unbounded complexity

A loosely used term here. Especially in terms of it's practical meaning, not theoretical. In this sense, both Dota and Starcraft II were also problems of unbounded complexity, and Alphastar is doing pretty good there.

> nothing short of AGI will get us close to the current below-average human driver

What are you basing this assumption on? Most prominent scientists in this field don't seem to be agreeing with it. It is by no means self-evident or backed by any substantial direct evidence.

More generally, I just can't understand this attitude you have. How does your thinking process work here, what is that attitude good for and where is it coming from?

Here we have world's brightest people trying to tackle a very important and useful problem, it's hard to know when it will be tackled, but it seems it will be eventually. Some people are disagreeing about when it will happen (perhaps you think it will not be for a 100 years), but at least they are doing something and are making progress. Investors similarly know that it is a hard problem and no one knows when it will be done, but they want to give their money for this problem. And here you are just complaining about the whole situation? What would you have all these actors do, just do nothing? Stop making these cars? Stop trying to solve the problem, stop bringing the progress forward? Or what, working on it in secret and never telling anyone so that god forbid any one person would ever happen to see their presentation and build a faulty (a too optimistic) view on the current state of the technology? Where is your negativity and sarcasm coming from?

Starcraft bot and a self driving car are not even remotely close to the same problem

Starcraft involves an actively hostile, adapting opponent. Daily traffic generally does not and when it does they are liable for any resulting damages, not you.

> Here we have world's brightest people trying to tackle a very important and useful problem,

No you don't, but world's best paid people are working in it for sure =D

The "AI industry" has a problem. Every few weeks there comes a "big name AI researcher" and call solving another "world's biggest problem," but a 5 minute research into his claim would, yet again, reveal just a hello world on opencv.

This is from where the negativity towards the ai industry is coming from.

Saying that you have "world's brightest people" sounds like a big stretch to an industry outsider, and plainly funny to people with some solid level of computer literacy.

I haven't heard of any anecdotes in the form of a big new breakthrough being a HelloWorld in opencv, and I think that any such cases can be more readily blamed on the media and bloggers inflating the actual research papers or results than actual scientists and engineers. But let's say you have experienced such episodes.

So now what, what is your point, what is your goal, what is your vision for this field? What would you have happened better? To just complain about it? Should they stop working on self-driving cars? Should they stop sending press-releases, should they stop selling cars with this functionality, what is it that you want? Do you want them to be paid less? I actually don't understand. What would be your best scenario for this field that would be better than what is going on right now (that big players and many of the brightest (and yes well-paid) data scientists are working on making a safe self-driving car a reality, while many investors are believing in it and are voting with their dollars)?

I have no suggestions for you

Reply on news.ycombinator.com