Hacker News Reader

Comment by kalleboo | Hacker News Reader

> It's enough to look at FSD 1.0 (Less than A year ago) vs current FSD 10 to see the crazy rate of improvement

We still haven't gotten basic technologies like speech recognition and OCR to work without proof-reading, and those have been around in consumer technology form since at least the 80's and have had an order of magnitude improvement every decade. These "fuzzy" technologies have an asymptotic improvement curve that never reaches 100% and self-driving is several orders of magnitude more complex.

View on HN

You're right, autonomous driving will never reach 100%, however I'm convinced it will be 10x human in a few years, mostly due to:

* Better sensors - ~10 cameras around the car vs 2 cameras on a limited gimble.

* Very high consistency - Speed, distance from other cars, optimal stopping distance from a switching stop light, better speed/distance estimation of other entities (eventually)

* No distractions - a computer will never drunk drive / look at their phone / argue with the passenger / have road rage.

IMO These advantages will out-balance the very rare ambiguous soft situations a computer will have a hard time handling, let's say a custom written no-entry sign or a cop waving their hand.

"These advantages will out-balance the very rare ambiguous soft situations"

Tesla has driven into a solid concrete wall, that does not seem ambiguous to me

Humans have done that too. Meanwhile Tesla’s autopilot has many instances of saving humans who would’ve other crashed.

There is always the argument of Tesla being safer than drunk, intoxicated drivers. First, I would not be too sure about that, but I would rather the governments implement a DUI check as well as some sort of alertness test before an engine can be started than trust the current fsd software. Put speed limiters as per the given location and the DUI check in all cars , then let's compare the numbers again.

And how can a Tesla save any driver when any driver is supposed to be in full control of the vehicle at any given time? If a driver needs Tesla as a saviour, then he has no business being behind the steering wheel.

We don't need more government in our lives. However instead of all that overbearing safety equipment, why not just have automatic driving?

> "when any driver is supposed to be in full control of the vehicle at any given time?"

Because this is not reality. Nobody is in full awareness and control of the vehicle at all times, that's just how humans work. Even if we were, we still don't have the same reaction speeds. If you actually watch the Tesla examples, you'll lots of aware drivers who just didn't notice or react in time to a collision.

Your entire premise is faulty. We can't work with what should be but rather what actually is.

You're making a claim that in a counterfactual world, FSD might be safer than human drivers. Even if this were true (which obviously there's no evidence of), why is that an argument against FSD in the real world?

Have you ever heard of an alert, sober driver just driving into a wall for no reason? surely 99% of the time thy were drunk, or texting, or falling asleep at the wheel.

I am not using FSD untill it's better than an average professional driver, like a bus driver, who have to be sober, etc. If something goes wrong there is not enough time to intervene.

Yes, people have driven into walls, buildings, cars, lakes, even other people, while being completely aware.

Many accidents (including the ones that Tesla's have prevented) include completely normal drivers who made mistakes. If AI can make fewer/no mistakes, and react better to mistakes by others, than it's a better overall outcome.

Do you currently drive a car? Is there any evidence that you're an above-average driver? If you answered "yes" to the first question and "no" to the second, then why wouldn't you trust a self-driving car as soon as it's known to be better than the average human driver?

You are doing bait-and-switch:

I am not better than average driver in skill, I am safer than average.

Being 'better than average' by comparing to national crash statistics means better than a half-drunk guy who is sleep deprived to boot.

I don't drink, I don't need to drive for 8 hours straight, my vehicle is in good condition, I dont have to deal with a gravel path through the mountains, etc. Those are the people that dies and they are in that statistics

So? There's no fundamental problem with identifying static 3D objects with vision. If anything it will be one of the things that become more consistent than humans sooner than later. I think the control/planning part is probably much harder than vision/perception (maybe except "soft" perception).

Nothing will be 100% perfect but computing power and machine learning has drastically improved in the past few years and changed everything.

Also the basic tech you mentioned like OCR and speech recognition is actually so good now that it runs on your mobile device, can tailor itself to the way you talk and write, and can even enable real-time translation to other languages across mediums like something out of a sci-fi movie.

Does anyone ever use speech recognition to dictate a legal document where a milkion dollars hangs in the balance? No, it would be dumb.

So why should that layer, after meticulasly checking this document, trust his life to another ML system, just like the one that just failed him?

Lawyers make mistakes too. Perhaps you’re surprised to hear that but there’s an entire section of law to deal with this. And yes lawyers use software dictation all the time because it’s faster and more accurate than a human assistant.

In fact ML does already do a better job in tasks like crawling through case research, while people then take over for strategy and deduction.

This split between low-level mechanical vs high-level creative thinking is a good distinction between ML vs human application, and most driving falls in the former.

"lawyers use software dictation all the time"

And they don't check the output? Are you sure about that, because the few I've spoken to say they won't touch it with a barge pole.

If they do check, then they don't trust the system. In driving you don't to hit pause and check the output.

You're conflating issues. Software is already better at transcribing than humans. [1] And lawyers will verify the contents whether it's done by software or a person, because neither is perfect and their job is the higher-level intention of the document, not the copying of words. But they still like the machine saving the human from tedious work.

1. https://www.nuance.com/dragon/business-solutions/dragon-lega...

We're never going to achieve perfect self-driving cars, and that's okay. Waiting for perfection wouldn't be the prudent option, anyway. As soon as self-driving cars are better than the average human, using them becomes a moral imperative, because not using them would result in a net loss of human life.

Not how it works. If there was a way to prevent all smoke related cancers but it randomly kills 1 million person a year would you say it is also a moral imperative? After all it is a net gain of lives.

Total casualties is not the only the to consider. You have to think about responsibilities, control, separate cases etc.

In addition, currently FSD is NOT safer than manual driving when comparing on equal conditions.

I think we’re on the same page. My primary point was that we don’t need a 100% perfect solution for autonomous driving to make sense.

From a regulatory perspective, I do think that a system that results in a ‘net savings of human life’ should be permitted, even though removing the human from the equation would likely change the distribution of driving deaths, such that some people (like exceptionally careful drivers) may be marginally more likely to die in such a system. Those people should of course have the choice to drive themselves.

From the perspective of an individual, a ‘net savings of human life’ may not be a good enough threshold. I personally would not want to put my child in a self-driving car unless I felt confident that it was safer than driving myself, and the knowledge that the AI was “statistically better than the average driver” would not instill sufficient confidence.

Reply on news.ycombinator.com