Hacker News Reader

Comment by candiddevmike | Hacker News Reader

What will it take to stop them? I believe it's a settled matter that their self driving is fundamentally flawed due to relying on ML and cameras instead of lidar. Will it take people dying for the government to step in? Lawsuits? I don't think our politicians are well versed enough in how this works and the level of risk letting this kind of thing happen.

View on HN

Is that a permanently settled matter? When humans drive, they are relying almost exclusively on visual inputs.

I don’t see a reason to believe that humans are so far superior at visual processing that no self-driving based on video could best them and many reasons to think that we’re inferior in certain aspects of the task.

1) Human eyes are superior to any camera system in a Tesla. both in resolution and dynamic range. 2) The human visual system is backed by the equivalent of strong, general AI that intuits and understands the entire world around the vehicle and all objects interacting with it. It is in theory possible for a video-only system to work, but ML is nowhere near the level of development needed for that to be a reality.

Sure they're superior, but you don't decide whether something is flawed by whether it's superior or not.

Our road systems are not designed for the full dynamic range of the human eye, which is why we drive with headlights and we have street lights on highways rather than relying on night vision for all driving.

> The human visual system is backed by the equivalent of strong, general AI that intuits and understands the entire world around the vehicle and all objects interacting with it.

I think the Human general AI vastly overestimates it's ability to drive and respond to conditions that suddenly appear when driving.

> It is in theory possible for a video-only system to work, but ML is nowhere near the level of development needed for that to be a reality.

[citation needed]

Look at the car's display in this video: https://youtu.be/RN5Qoei7v1k?t=2281

The "lane" in front of it is wildly moving around, so it's definitely misinterpreting the visual inputs.

As an owner of a tesla, the rendering on the screen and what the car actually does with the inputs don't always correlate. The rendering is purely for your own viewing pleasure. Hell, there's a disconnect where if you have the new autopilot computer but the old entertainment screen computer, it didn't even have the capability to render anything worthwhile but you still had self driving

But in this video, as soon as the FSD got engaged the car started driving towards the train... It doesn't look like the car had much understanding of what was going on.

I've seen this several times in vieos with the new FSD, when you enable it suddenly for example when you're already stopped at a stop sign, it's not sure what state it's currently in and why it's stopped and assumes it should drive forward when you enable it from a stop as long as there's no physical objects in front of you.

If you want to see longer FSD videos check out https://www.youtube.com/user/pilotjc78

The driver enabled AP at that very moment, and disabled it half a second later, you can hardly take any conclusions from that.

Given that the rendering on the screen is supposed to be showing the inputs the car thinks it is receiving, it is even more concerning to learn that what the car actually does isn't correlated with the inputs it is receiving.

That seems intentionally tortured logic. Why would it be a given for the rendering to show the inputs as received rather than being an entertaining caricature loosely related? Surely you don’t think the animations of the car’s battery being charged are realistic in any engineering or physical sense.

Because Tesla says that it is a rendering of what the car thinks it is seeing, meaning that the rendering is supposed to roughly correlate with the information upon which the FSD systems are acting.

In contrast, the animations of a car battery being charged don't claim to be realistic in any sense, and they deliberately attempt not to be realistic: oversized images of batteries or other standardized iconography, stylized arrows or other indicators of charging "flow", etc.

I'm distracted by how unsafe the driver's use of the touchscreen is in that video.

I'm fact I'd argue the opposite is true. It's settled that the endgame of self driving will rely only on visual input, since that is what humans need.

ML shows no signs of achieving intuition, which allows us to predict various scenarios in real time and adjust accordingly.

Does “instead of LIDAR” mean that LIDAR solves that in some significant way?

With camera-based sensing, you need AI to interpret the inputs in addition to AI for deciding how to react to those inputs. With LIDAR, the job of detecting objects seems to be much simpler.

Taking a relatively simple to work around example, you need AI to distinguish lens flare from an object approaching/departing.

With LIDAR the job of detecting some objects becomes much simpler but for other objects becomes impossible.

Nope. I don’t think we’ll have Level 5 autonomy with safety on par with human drivers until we have general AI. At that point, whether or not that AGI had LIDAR supplementing its visual inputs becomes irrelevant.

Doesn’t Wayno already have level 5 autonomous vehicles operating in the Phoenix area? So far they’ve had about 65,000 driverless miles[1] (and 6 million miles with safety drivers). That number should increase pretty quickly now that they’re scaling up their service. So far it seems like the main problem is that they get rear-ended by inattentive human drivers.

1. https://www.engadget.com/waymo-indepth-details-selfdriving-a...

ML is intuition. Consider the difference between an expert system solution to chess and an ML solution to go. Go is all intuition.

>> Is that a permanently settled matter? When humans drive, they are relying almost exclusively on visual inputs.

I really don't understand how so many smart people are repeating this argument over and over. it's one the most flawed arguments that was used by Elon Musk to sell his lack of ability to retrofit lidars in every single car.

We also got a single brain, so why can a computer not do what we do?

You could short Tesla stock. It is well proven that the free market will correct mistakes like this. /s

It will take a few serious accidents that most humans would have avoided.

Reply on news.ycombinator.com