[philwelch]

> Driving on the open road requires real intelligence. Not the pretend intelligence that modern AI gives, but real understanding of situations and terrain.

People used to say this about every single thing that computers can do better than people.

In my college town, some pedestrians got ran over by a driver who later pled insanity due to "caffeine-induced psychosis". I think you're seriously overrating the predictability of human failure modes.

[emodendroket]

> People used to say this about every single thing that computers can do better than people.

But computers can't do it better than people! That's what drives me nuts about this debate -- it's just accepted as a premise that either the self-driving cars are much safer than human drivers, or the path to getting them there is very close and no serious obstacles remain. Neither is true and it's not clear they will be. https://blog.piekniewski.info/2017/05/11/a-car-safety-myths-...

[baddox]

I think what is clear is that virtually no one is okay with large-scale public deployment of self-driving cars that aren’t clearly statistically safer than human drivers, so when we’re talking about public deployment of self-driving cars, it’s implied that we’re talking about when (if) they reach that level of safety.

[emodendroket]

Frankly, that's not clear to me. A lot of people seem quite eager to put the first vaguely plausible thing all over the road because of an exaggerated idea of the incompetence of human drivers (which is understandable, but harmful in this context)

[philwelch]

The question is, which human drivers?

Let's say we have a self-driving car that is as safe as the 20th percentile human driver. Do we allow that self-driving car on the roads? Do we selectively revoke licenses from 1 out of every 5 drivers and replace them with a car that's at least as safe as they are if not probably safer? Do we replace breathalyzer interlocks for drivers with DUI convictions with an AI driver and just revoke their licenses permanently?

There isn't a trivial solution to this problem. At some point, some AI driver is going to cause an accident that would not have been caused by a 95th percentile human driver. At the same time, human drivers do shit like this all the time: https://www.youtube.com/watch?v=oidHSzukSss

[emodendroket]

OK, nice anecdote, but if you follow my link you can see that if we just blindly take the average of all drivers and accidents per mile driven and then compare it to AI performance AI is nowhere near average, let alone the 95th percentile. Here is another example of what I'm complaining about: an argument that simply takes as its premise something not yet proven to be true.

[philwelch]

Like what, “self-driving AI will never improve past the state of the art in mid-2018 so it’s useless to speculate about it ever doing so”?

[emodendroket]

It's not entirely clear it will in the foreseeable future improve to the point where it is a serious prospect to be safer than human drivers.

[darpa_escapee]

To me, this seems like an a priori assumption that accepted as the truth. Then, when it's questioned, the person questioning is made out to be some kind of luddite.

Which is an odd phenomenon to me. I can't even get Google Assistant to understand me 3/4 of the time, yet I'm supposed to take it on faith that autonomous cars are inhumanly safe?

[emodendroket]

Google Assistant is a funny example, because my wife has a foreign name and I can do absolutely nothing to get it to understand when I ask to call her, even attempting to imitate its weird pronunciation of the name. The only thing that works is hand-typing "Xxxx is my wife" into the Assistant and then referring to her exclusively as "my wife," and that gets reset with updates periodically

[Piskvorrr]

Questioning the capacity of Friend Computer is treason here on HN, Citizen! Report yourself for immediate termination.

In other words, much of the autonomous robot debate here is based on handwaving, wishful thinking and No True Autonomous Scotsman (...would run over a human).

[mavhc]

No, require data to be public, require a billion km simulated driving test for every software version that's released to a car without a safety driver.

[lucideer]

> People used to say this about every single thing that computers can do better than people.

Did they though? Think about things that computers can do better than people: they're mainly things that we completely predicted computers would be better at (arithmetic, precision manufacturing, drafting, telecommunications routing). Beyond that, you're left with things computers are only better at dependent on priority, the canonical example being service jobs where economics trump's QoS; computers are much worse than a cashier, but comparatively cheaper by a margin that makes the quality compromise worth it.

The only possible exception I could think of that's come up in recent discourse is diagnosing patients, but even that, while encroaching on a role that has traditionally been revered as a career, is still something that seems at least on the surface to be quite predictable given the nature of what's required to make diagnoses (simultaneous access to a trove of data and knowledge).

Beyond the above, I think it's pretty reasonable that there's a broad range of things computers will not be better than humans at for a very very long time, if ever.

[baddox]

But this is a cliche in the AI field: that AI is defined as the things that computers can’t do as well as humans right now. As soon as computers match human ability, well, clearly that doesn’t count as “intelligence.”

And I don’t think the problems computers have proven themselves useful in solving have been what most people expected. Chess, Go, facial recognition, Jeopardy, image classification (hot dog or not), captchas (clearly, since they’re designed specifically to resist computer solutions), etc. seem to me to be things that, before computers proved to be decent at, would have been widely considered to require intelligence on the level of humans.

[foldr]

People only thought that chess required intelligence because they had no idea how much raw computational power computers would obtain. To anyone who understands the rules of chess, it's obvious that a machine which can perform a near-exhaustive search of the state space for a few moves ahead is going to be able to play chess better than a typical human.

[mavhc]

People only thought that driving cars required intelligence because they had no idea how much raw computational power computers would obtain. To anyone who understands the rules of driving cars, it's obvious that a machine which can perform a near-exhaustive search of the state space for a few seconds ahead is going to be able to drive better than a typical human.

[foldr]

>To anyone who understands the rules of driving cars, it's obvious that a machine which can perform a near-exhaustive search of the state space for a few seconds ahead is going to be able to drive better than a typical human.

But there aren't any formally-specified rules for driving cars, and this isn't obvious.

[Piskvorrr]

Driving has a universal, formal, self-contained, non-contradictory, simple set of rules that all the road users unconditionally follow? Can I see it?

Nope, despite a myriad of road codes, the actual traffic doesn't follow a set of formalized rules: a chess rook can't just decide that it will start disintegrating all of a sudden, as opposed to a vehicle. You could probably approximate the ruleset if you made it self-modifying...which will then demolish your second point about near-exhaustively searching the state space - good luck doing that before the heat death of the universe, as you're essentially simulating the whole environment. Oh look, there's also weather. How's that exhaustively searchable? Asking for the Nobel Prize committee.

For the sake of discussion, let's say that a miracle happens and you managed to do all that - but sorry, it's useless again, the few seconds have already elapsed and you need to do it again. And again. And again, ad infinitum.

Now, I could envision "by our current technology, we can't yet, but we're hoping for a miracle in this specific spot" - but "assuming a massive miracle happens every few seconds, for each vehicle" is completely removed from reality: why not have teleports, if we're in magical wish-granting land already?

[mavhc]

it's ok, you only need a near exhaustive search, and to be better than humans.

For highway driving Waymo had 6 disengagements in 2017, street: 57.

Total driven: 352000 miles

1 disengagement for "a recklessly behaving road user" 5 for "incorrect behavior prediction of other traffic participants"

Seems like predicting other people is almost perfect, the others were more internal problems.

Driving a car when a weird thing happens isn't that complicated: You stop, braking at the minimum amount required to do so safely, to avoid cars behind you hitting you

[wyattpeak]

That just isn't true. It was widely believed through the seventies and eighties that chess inherently required creativity, and a machine could never beat a grandmaster.

Rather, I suspect, tasks which computers start outperforming humans in we reanalyse as "completely procedural". Nobody called chess procedural in the middle of last century.

[gaius]

Well, a computer playing chess is simply enumerating all possible boards. Which is the same way that a GAN produced “art”. Humans do it creatively because that’s how you do it if you lack exhaustive computing power and memory. In neither case is a computer mimicking that process of a human, they simply arrive at the same outcome by a brute force means.

[tialaramex]

No. Computers don't play chess by "simply enumerating all possible boards". That would require ludicrously more compute power than we have, and, of course, it would also _solve_ chess, rather than just allowing the computers to play once it would (if it could ever be done) show that the game itself has a solution, a best way to play, like Tic-Tac-Toe.

Historically AI chess (e.g. "Deep Blue" or Stockfish) is played by machines using one heuristic to estimate how "good" positions are without truly knowing, not so dissimilar from how humans evaluate a chess position. and then another heuristic to try out moves to get to further positions. The machine considers possible plays and how they affect the heuristic "value" of the board, preferring those with more value. Human Chess AI authors design the two heuristics used, though they often aren't very good at actually playing chess because it's a different skill.

Google's AlphaZero AI plays chess differently again, it had no preconceptions of how to play Chess, instead it learned through self-play - it knows the rules of the game but began with no idea what's a good or bad move, it adapted its own heuristics based on how well they'd won or lost. It actually recapitulated most of human chess theory history over its incubation period of thousands of games, discovering ideas like the Sicilian Defence for itself, new attacks would at first see overwhelming success, and then, playing versions of itself that had seen these attacks, they'd be defended more effectively.

Alpha Zero plays a radically "more human" style of chess than most modern human Chess grandmasters, huge multi-move strategies in which pieces are sacrificed to take positional advantage. It looks like something humans were doing last century - except Alpha Zero does it much better than they ever did.

[foldr]

A typical chess position has fewer than 100 possible moves, so a modern computer can do quite a deep exhaustive search of the state space. You won't beat Kasparov just by doing that, but I'd bet it's enough to beat me.

[tialaramex]

"Just by doing that" you can't even begin.

The problem is that you lack an evaluation function. Let's consider two of those 100 possible moves. Your rook could take this opposing pawn, or, your own pawn could move forward one space. Which is better? Why? Neither of them immediately wins the game, but we must pick something. In a smaller, tighter game, like Tic-Tac-Toe we could crank our exhaustive search until we discover that this opening move leads to a possible win... but the search space in Chess is categorically too enormous for that.

Both Google's Alpha Zero and simple human play encourages the belief that a good evaluation heuristic is essential. The evaluation heuristic looks at a board position and it doesn't recommend a move it says something like "I rate this position 0.418" where 1.0 is "I'll definitely win on my turn" and -1.0 is "My opponent wins on their turn". Google's engine contemplates relatively few possible moves (for a computer) but the results are striking because it's looking at _good_ moves more of the time rather than wasting a lot of time thinking about moves that are a bad idea.

This seems obvious, but, well, learn chess and see for yourself.

[rayiner]

> People used to say this about every single thing that computers can do better than people

Even if that’s true (and I doubt it is), there is ample precedent (AI winter) for the industry dramatically overestimating what computers can do.

I bet if you time traveled and showed Siri/Cortana to an AI researcher from 1960 they’d be incredibly disappointed.

[pcwalton]

> I bet if you time traveled and showed Siri/Cortana to an AI researcher from 1960 they’d be incredibly disappointed.

It's a common misconception that the 1960s and 1970s were a time of unbridled enthusiasm in AI. In fact, there was a ton of pessimism back then too: for example, ALPAC [1] was so pessimistic about the future of natural language processing that it got the US government to pull most of its funding.

I think if you were to show Siri, Alexa, etc. to some of those folks they'd be pleased that we've gotten as far as we have, while acknowledging the obvious fact that there's plenty more to do.

[1]: https://en.wikipedia.org/wiki/ALPAC

[ghaff]

Not sure why people are disagreeing. That seems a blindingly obvious comment. There's so much today that would seem almost like magic to pretty much everyone living in 1960. But coice assistants? (And even just voice recognition.) almost certainly seemed like relatively "easy" problems. Perhaps less so to AI researchers than the general public but still.

[rayiner]

I don't know about "magic." A modern smart phone might be "magic" to someone in 1860--it operates based on technologies that didn't exist back then. But by 1960, the building blocks of modern computing were already in place: digital von Neumann computers built out of transistors, radio communications, signal processing, etc. AT&T used frequency-division multiplexing of multiple voice channels in phone transmissions in 1918. The mathematical framework for modern technologies like LTE was in place by the 1950s and 1960s. Would it really have surprised anyone that transistors would continue to get smaller and faster, allowing higher complexity, higher-throughput signal processing, which would allow Facebook?

[Retric]

I remember reading a dreamed up device that would hopefully show up some day written about in 1960.

Weight ~1 ton, cost ~1 million dollars inflation adjusted, non toxic, delivery date ~1990. What did they want? A 1 GB random access HDD.

A 32 gigabyte micro SD card for 10$ would have blown their mind let alone a smartphone.

[rayiner]

They were off on size/weight, but they guessed the capacity just about right (IBM 0681). I don't think extrapolating out pre-existing trends, for sophisticated people, be "mind blowing." Would your mind be "blown" if you learned that by 2048 you'd have a 100 petabyte drive using, say, magneto-resistive memory (or something else based on anticipated, if not fully developed physics)? Seems like hyperbole (and setting a low bar for peoples' imagination).

Reading stuff written in the 1960s about what today would be like, what strikes me is that technology is so incredibly not mind blowing compared to what we had back then. Even in the area of computers. Hell, we haven't even come up with an input device that beats keyboards, which were invented in the 19th century (electro-mechanical keyboards, not typewriters).

[coolso]

I think the difference is that we today have a lot more reference points for technological advances than people in the 1960s did.

Just continuing with the storage example, for decades now we've all been witness to data storage sizes growing massively, while the housing of said data storage has shrunk in size tremendously - as has the cost.

So when a couple MB of incredibly slow storage weighs thousands of pounds and costs millions of dollars, I do think the concept of tens/hundreds of GB of super fast flash memory contained within an object the size of a thumbnail would be mindblowing, whereas your example of

>a 100 petabyte drive using, say, magneto-resistive memory (or something else based on anticipated, if not fully developed physics)

wouldn't, just because we already all know how far technology has come since the 60s.

[ghaff]

It's not so much the hardware as the combination of things. GPS+the Web+smartphones+... But, yeah, physical infrastructure has a lot of friction. So we have amazing pocket devices with access to much of the world's knowledge. But traffic jams.

[Retric]

They where vastly off in terms of size, weight, transfer speed, latency, and cost. In 1990 you could get a cheap RAID array so pick do you want 100x that capacity for far less than that price and weight.

[slavik81]

That must have changed rather quickly. In 1972 Alan Kay wrote "A personal computer for children of all ages". Here is the abstract:

> This note speculates about the emergence of personal, portable information manipulators and their effects when used by both children and adults. Although it should be read as science fiction, current trends in miniaturization and price reduction almost guarantee that many of the notions discussed will actually happen in the near future.

The paper is a great read. He basically imagined that in the future we'd develop the iPad and some high quality educational software for children. Forty years later, we can proudly say we've successfully developed half those things.

[blacksmythe]

  >> if you time traveled and showed Siri/Cortana to an AI researcher from 1960 they’d be incredibly disappointed.

If you time traveled and showed 2013 me what self driving cars would be doing in 2018, 2013 me would have been astounded.

[YeGoblynQueenne]

>> People used to say this about every single thing that computers can do better than people.

Who were those people, who said those things (i.e. where they AI researchers, or computer scientists?). And what exactly did they say?

There have always been strong criticisms of AI (e.g. [1]) and opinions dismissing computers voiced by people who did not have an adequate understanding of computers.

The interesting thing is to see what the people in the know actually thought over the years and what they think right now.

Edit: to clarify, what AI researchers usually do is overhype the capabilities of their systems and claim they can achieve things that they never manage to show they can- completely the opposite than saying that "computers can't do that".

__________________

[1] "What computers can't do" by Hubert Dreyfus

https://en.wikipedia.org/wiki/Hubert_Dreyfus%27s_views_on_ar...

[a3n]

> People used to say this about every single thing that computers can do better than people.

And they were right, until they were eventually wrong. There will be this phase for automated vehicles too.

[Piskvorrr]

True. However, "eventually" is a long time - the assumption "level 5 hardware exists today" is IMNSHO premature by many decades.