Hacker News Reader

The difficulty of creating L5 autonomy and the provable impossibility of the halting problem are not comparable at all.

There is absolutely nothing fundamental that makes L5 autonomy impossible, while the halting problem is provably impossible, as normally formulated.

I don't know how close or far away L5 autonomy is, but it's definitely theoretically reachable, while the halting problem is always going to be impossible.

Halting problem comparison is silly, but L5 autonomy as defined ("all conditions") is probably not attainable.

A car that would complete all the trips I'm willing to complete, and some more, but refuses under some circumstances (e.g. whiteout blizzard conditions) would be a L4 vehicle under the ISO definition.

When we are actually have cars that sufficiently close to what humans can do in terms of range of condition, I suspect the level 5 definition will be updated to be more like:

Can drive in approximately the same set of conditions as a human (professional driver?), possibly being unwilling to drive in certain (seldom encountered) conditions where most humans would, but offset by being willing to drive in conditions few people would.

Importantly, the car must avoid completely giving up on driving mid trip (as opposed to deciding "too dangerous, turn around and go back at next opportunity"), unless conditions are comparable to those in which a human would give up mid trip (which are pretty limited, as humans seldom just stop and give up on the road unless the car is broken down, or fully stuck. At worst, in some really bad conditions, humans may pull other to wait for the storm/extreme-fog/etc to blow over.)

Or perhaps at that point we won't need the definition anymore. It becomes a bit arbitrary and market-y at that point. Assuming that we don't end up with a single vendor.

"My BMW still drove during the snows in February, but my neighbor's Tesla said it wouldn't drive 2 of the days." "Yah, but mine doesn't insist that the windows are perfectly clean and pristine before starting a trip!"

We probably should never have had level 5, but split L4 into a couple levels: heavily geofenced/restricted vs. relatively unlimited applicability with some restrictions.

L5 autonomy is a absolutely a long way off (if ever practially acheivable) and Elon is a bullshit artist, but I don't understand your comparison to the halting problem. Could you elaborate?

I interperate your comparison to the halting problem to mean that even if we ignore the feasability of a particular solution, it is literally an impossible problem to solve.

My understanding of L5 is driving without external human intervention. In one sense we already have the technology to do that - our brians. It would never be feasible (or ethical), but if we could put a human brain inside every tesla, wouldn't that achieve L5?

> In one sense we already have the technology to do that - our brians. It would never be feasible (or ethical), but if we could put a human brain inside every tesla, wouldn't that achieve L5?

It would achieve L5 of a sort, but people don't usually mean "L5 autonomy" in the sense of "capable of crashing the vehicle deliberately to protest the horror of their existence".

I think equating to the halting problem is silly, but the ISO levels are a bit of a mess.

* Level 5 is a vehicle that never needs a human to take over and can drive in "all conditions". Humans do not meet this driving standard.

* Level 4 is a vehicle, that within a set of vehicle-defined conditions can drive the vehicle without a human ever having to take over. It refuses to drive unless the conditions are met.

So a level 4 car could be a vehicle that can drive in a 1 block area of residential streets only... or something that can drive in way more conditions than I could safely attempt, but refuses to drive in say, whiteout blizzard conditions at night.

I mean, he has a pretty clear financial incentive to say stuff like this given that he runs an auto manufacturer that heavily invests in vehicular automation.

Not saying he's right, but find me a company that doesn't polish their own turds, even just a little bit. Everyone trying to sell something is painting the best picture of their product possible.

There's polishing a turd, then there's calling a pile of shit a chocolate cake.

I really hope this becomes an actual idiom.

Yeah, but even this doesn't seem very thoughtful. He recently acknowledged his predictions around full self-driving were wrong, and said it was because he'd failed to appreciate that it would essentially require artificial general intelligence.

Then he claimed we'd have that solved by 2023.

Not sure what the upside is of being known for repeatedly making non-rational predictions and being wrong.

> Not sure what the upside is of being known for repeatedly making non-rational predictions and being wrong.

It can significantly move markets in the short term, which he seems to have become adept at doing over the last few years. And unfortunately, the stock market isn't really interested in long-term thinking, it's largely about breaking news and twitter rumors nowadays.

Well, he's definitely mastered market manipulation. I take your point there.

But, the degree of absurdity across predictions undermines even that strategy over time. I think he does himself a disservice here. He should get out of his own way and allow his actual achievements to speak louder than irrational predictions (and other distractions).

I totally agree, and that's the very definition of short-term thinking: doing what's best for today at tomorrow's expense.

I think he knows intellectually that a lot of his claims and predictions are bogus, but he also knows that his fanboys are all over it. Look at the amount of preorders for cars that - as it turns out - were years away still, if at all (thinking of the new roadster, pick-up truck and big trailer truck at the moment). Look at Tesla's stock which is based entirely on hype and less so on actual product, market share or financial results. Look at how many companies and universities around the world started developing a Hyperloop just because he mentioned it - I don't even know if there was like a grant for it or some other financial incentive.

Intellectually everyone can deduct that a long distance hyperloop is science fiction, ridiculously expensive, complicated, and will likely face long outages at any incident (see the channel tunnel, but like if it was 10-100x as long and a vacuum). But because Musk says it with Confidence, an army of fans jumps onto it.

Unfortunately you are doing the very thing that you accuse Musk of doing. L5 autonomy is not formalized (nor do I think it is able to be formalized) to the extent that would permit a rigorous proof showing it is isomorphic to the halting problem.

Your claim conflates a nebulous, squishy, human goal with a formally and rigorously proven mathematical problem. The only support offered is links to wikipedia and news articles, none of which help connect the two in an equally formal and rigorous fashion.

I know what the halting problem is and I had to study why it cannot work.

However, for L5, you just have a quote saying it doesn't work. We know it is mathematically possible for L5 to work because, well, humans perform at that level. We know that our vision, our ears, our hands and senses are enough input to solve the problem.

Do you have a direct connection between them or are you just using it as a metaphor for an unsolvable problem?

I needed a reminder, so I opened a new tab in Chromium and typed in 'the halting problem' and hit return and Chromium immediately crashed :)

Which is not only funny, it's exactly why the halting problem is so hard…

Humans actually don't perform at L5, we perform at a very high level of L4. (Still, the halting problem comparison is silly).

Elon has changed his tune (somewhat) about automated driving. He said in an interview that he now believes full auto-drive requires AGI.

Personally I think full auto-drive requires us to change to the roads to make them work for machines... but that's a different story.

I heard him say FSD will be fully functional by next year in the ted interview he did a few weeks ago.

I mean, they already ship it, don't they? How do their customers square these statements with what they were sold and can enable with the flick of a button?

Or is this FSD in the sense of "my car can drive itself home after dropping me off" type of thing?

Hmm... While I don't see any evidence that L5 autonomous driving is near, I don't follow your argument that L5 is equivalent to the halting problem. Can you explain?

I am not convinced that L5 is fundamentally impossible (unless we posit that humans are also not L5 autonomous, which I suppose one could argue, as they are prone to driving errors). Granted I subscribe to Universality, and assume that humans are not capable of hyper-computation.

That Waymo CEO quote is about feasibility of fully autonomous driving in snow / rain. Seems unreasonable to even expect that. But I don't think it's as intractable as the halting problem. There's no formal proof against feasibility of L5.

But as far as Musk, yes he lies and isn't shy about it. It's shameless and overt. Perhaps he justifies it as being part of his job.

he says outright stupid things all the time and only if he gets into a topic you know a thing or two about, you realize how wrong he often is.

Recently he said "complexity and cost of a car is greater than that of a humanoid robot".

Will never be achieved? Only if humanity dies ous quickly. If, literally, even the dumbest people can learn how to drive, I'm sure with enough time we'll be able to replicate that autonomously.

You make that he's a businessman who also tries to sell his product. We try to judge him as an engineer, but he's much more than that.