Watching this video of regular folk (not tech journalists or other "reviewers", or paid PR) riding in FSD Waymo has done a lot to reduce my "but it'll never happen" feeling.
It may very well be that future FSD modes will require highly detailed maps, manual curation, limited areas where things can work, but even that would be valuable. e.g. if self driving trucks only work on long haul highway-only, that's still a win. If FSD taxis are mostly rolled out in medium to big cities in the beginning, it's still be a win.
Trying to boil the ocean all at once, and launching an enduser consumer FSD program that can handle wherever the driver takes it, may be the wrong way to do it, top-down instead of bottom up. Start with small geographic areas, and expand as you go.
Huh, I'm surprised you find those videos convincing about the future of self driving cars. To me, they don't seem much better than the demos people were giving ten years ago, in academic competitions, etc.
The Tesla videos scare the hell out of me. My own auto pilot purchase scares me whenever I use it. I can't even use cruise control in my Tesla without it randomly slamming on the brakes.
It's pretty straightforward to make batteries and an electric car. To make a computer vision product work? Based on everything I've seen of the field over the past 20 years, I am not impressed by Tesla's efforts in the least. They seem crazy, not competent.
Waymo is quite a bit better than they were 10 years ago. It would be harder for a casual observer to note how much Waymo has improved in the past 4 or 5 years because the gains have been made largely at the statistical margins of performance.
It's easy to build a robot that can do one thing right, and that's all you need to get a cool demo. It's hard to build a robot that can deal with the many 1000s of things that can go wrong while attempting to do one thing right.
> To me, they don't seem much better than the demos people were giving ten years ago, in academic competitions, etc.
From what I remember from ~10 years ago (and a quick glance at Youtube videos from then confirms it) the state back then was roughly:
- DARPA challenges where cars were driving on dirt roads at about walking speed
- High speed driving on completely closed of racing tracks (with no other cars present), e.g. BMW i330
- Simulated real world conditions on closed of tracks with the system significantly slowing down at intersections and pretty reliably producing slow speed crashes
Compared to what that video is showing, that's night and day.
I like that Waymo is confident enough in their system to put it out there for people to use. But honestly, I think that their reasoning for releasing it is exactly to reduce worries in people, that 'it'll never happen feeling'. I think they needed to put something (before it was even 99% ready) out, otherwise people would lose hope and investors balk at the progress. I really do think that it'll never happen, at least in the ecosystems current form. There needs to be special lanes or routes that self driving cars can take, special pick-up points, etc. Humans take risks while driving every day, the problem is, that we have an incredibly good picture of the risks involved in everyday things. Machines don't have this sense perfectly attenuated yet, and I have yet to see evidence that they ever will. With us, there is no variability in our senses. Our eyes mostly work the same every day, our perception skills doesn't change. Machines cannot even trust their senses, and their code has to reflect that.
Obviously, this is all speculation on my part, but when I see videos of waymo sitting in roadways and parking lots, blocking traffic, pedestrians, and just the overall flow of life, the more that I get the feeling that machines should not be allowed to participate.
But people also shouldn't waste so much time of their life stuff in traffic driving vehicles either. I mean, the ideal city would probably be self-driving trains, people movers, and free scooters, segways, or wheelchairs everywhere. No AI vision systems needed.
If you live in urban Japan, do you ever need a car?
To replace all 273 million vehicles in the US with self-driving EVs would cost around $10 trillion @ $50k per car. You could built a national network of self-driving lanes for a fraction of that. You could build rail lines and subways in every major metro area.
The major upside of a car vs train though is privacy. You may not mind taking the train in Japan, but in SF, it can be disgusting, and so private vehicles and American culture may be inseparable.
But that still leaves many other solutions, like we have bike lanes, and HOV lanes, we could build AV-only lanes, and prohibit AVs from working (requiring manual takeover) when you depart them.
On long drives to work, you'd just stay in the AV lane and chill, and when you got close to an exit, you'd have to take over. It would still be a win if on a 1hr commute you only did 15 minutes of actual driving.
> You could built a national network of self-driving lanes for a fraction of that. You could build rail lines and subways in every major metro area.
Most of them will be built by government making huge losses, they will go from nowhere to nowhere and they will not function as intended. The California high speed train is a great example. Not to mention $50K will be put by people but the national rail network will have to built by government (not that they have to but very likely they will grab this opportunity).
Human senses may be fairly consistent day to day, but variability across humans is quite significant. Even a single person's faculties (vision, cognition, focus) change drastically over longer periods of time, or when they're tired, sick, or worse (drunk). Machines at least have standardized hardware with self-test protocols and onboard diagnostics. There are innumerable scenarios in which machines have no good solution yet, but I suspect over time the heuristics will reach a threshold of "good enough for public use."
> It may very well be that future FSD modes will require highly detailed maps, manual curation, limited areas where things can work
This is the way.
We need a common standard of machine-readable markers along roads to make long-distance FSD viable. We have the technology, what we need is to agree on a single standard and get it installed.
For cities it's a similar thing, but there are a bunch of additional variables (called humans) to take care of.
https://www.youtube.com/watch?v=qAZ6tJSj9T4
It may very well be that future FSD modes will require highly detailed maps, manual curation, limited areas where things can work, but even that would be valuable. e.g. if self driving trucks only work on long haul highway-only, that's still a win. If FSD taxis are mostly rolled out in medium to big cities in the beginning, it's still be a win.
Trying to boil the ocean all at once, and launching an enduser consumer FSD program that can handle wherever the driver takes it, may be the wrong way to do it, top-down instead of bottom up. Start with small geographic areas, and expand as you go.