5 ms·

FSD is currently using neural-network style AI, and it's frankly amazing to use and watch and is massively useful.

by 93po 3y ago

stetrain 3y ago
In what way is it useful? What value is being provided? In my experience it requires constant supervision and readiness to intervene at any moment. There are plenty of reports and photos of it running wheels into curbs with little time for the driver to react. Given that, while using it you do not regain any time or attention that you would have otherwise spent driving. That doesn't mean it isn't impressive. A car that can drive itself like a 15-year-old on their first outing with a fresh learner's permit that needs constant coaching from a parent or instructor is very impressive, just not useful. I will say that in clear conditions on long highway trips, basic Autopilot does have utility. It does allow you to divert some attention from keeping the car between the lines and matching the speed of the car in front, and use that attention to keep an eye on the large traffic picture, and arrive to your destination slightly less fatigued. Using FSD on city streets seems like the opposite of that to me, an increase in stress and workload that currently provides no practical utility.
mplewis 3y ago
Is that the thing that keeps crashing people into highway barriers?
- CamperBob2 3y ago
  That, and inadequate sensor diversity and coverage.
cbeach 3y ago
Any recent version of FSD (i.e. 12.3.x) is technology close to magic. There was a time when HN would recognise technological acheievement on its own merit, without allowing personal politics to cloud our judgement. But sadly, we're in a perverse era of political tribalism where FSD is bad because https://elonbad.com/ https://elonbad.com/
- nojvek 3y ago
  FSD is cool as a great demo. But the optics and facts are that it’s got people killed. Multiple people over the years. Silly mistakes causing crashes. It will get better but definitely does not live up to the “full self driving” marketing hype. That kills the magic. Meanwhile look at Waymo. They don’t make a lot of noise. They take safety really seriously and keep on improving actual “self driving cars” city by city. Zero people dead. I’ve sat in both. FSD was a great demo, but Waymo truly felt like magic. No driver at all!
  - mbrumlow 3y ago
    I don’t think we will ever have any self driving tech that inter mingles with non self driving cars that will result in zero deaths. While sad, mile for mile FSD is better than humans.
    - Mawr 3y ago
      FSD only drives in a subset of conditions humans drive in so the comparison is invalid.
    - acdha 3y ago
      > While sad, mile for mile FSD is better than humans. For this statement to be correct, we’d need to have full disclosure of all travel using FSD at any point, accidents which happened anytime FSD was active or had been recently deactivated (for example, that guy who fell asleep counts even if FSD deactivated a minute before the vehicle crashed), and be able to compare that to the same trips driven by human drivers. You especially need to avoid including incidents in the human stats which are in conditions where FSD would do even worse.
    - alemanek 3y ago
      How the heck did you reach the conclusion that the current state of FSD is safer than a human driver. The Tesla Community FSD tracker has it at 157 city miles per disengagement at the moment. The people collecting this data are the Tesla enthusiasts as well. It has a very long way to go before it is better than a human. We won’t know the true stats until it is allowed to operate without supervision. Source: https://www.teslafsdtracker.com/Main https://www.teslafsdtracker.com/Main
everforward 3y ago
I don't think they lead the pack on that, though. Everybody in the self-driving space is using AI to some degree. E.g. Waymo was at 17,311 miles per disengagement (human takeover) in their 2023 report, and they're not even the top. Zoox was the top at 177,602 miles per disengagement, which is shockingly good if they're not gaming those numbers with tiny service areas or something. I don't think Tesla publishes their disengagement data, but what I can find crowdsourced from their users is pretty bad relative to the above. The most optimistic number I could find was from 2022 at ~400 miles per disengagement. That's not even very good for 2022; Mercedes-Benz was at 1,400 miles per disengagement, and I didn't even know they had a self-driving division. Nissan was at 149 miles/disengagement, which makes Nissan their closest competitor by capabilities (the next highest after Tesla was QCraft.ai at 863 miles/disengagement, no idea who they are).

FSD is currently using neural-network style AI, and it's frankly amazing to use and watch and is massively useful.

by 93po 3y ago

FSD is currently using neural-network style AI, and it's frankly amazing to use and watch and is massively useful.

In what way is it useful? What value is being provided? In my experience it requires constant supervision and readiness to intervene at any moment. There are plenty of reports and photos of it running wheels into curbs with little time for the driver to react. Given that, while using it you do not regain any time or attention that you would have otherwise spent driving. That doesn't mean it isn't impressive. A car that can drive itself like a 15-year-old on their first outing with a fresh learner's permit that needs constant coaching from a parent or instructor is very impressive, just not useful. I will say that in clear conditions on long highway trips, basic Autopilot does have utility. It does allow you to divert some attention from keeping the car between the lines and matching the speed of the car in front, and use that attention to keep an eye on the large traffic picture, and arrive to your destination slightly less fatigued. Using FSD on city streets seems like the opposite of that to me, an increase in stress and workload that currently provides no practical utility.

mplewis 3y ago

Is that the thing that keeps crashing people into highway barriers?

CamperBob2 3y ago
That, and inadequate sensor diversity and coverage.

cbeach 3y ago

Any recent version of FSD (i.e. 12.3.x) is technology close to magic. There was a time when HN would recognise technological acheievement on its own merit, without allowing personal politics to cloud our judgement. But sadly, we're in a perverse era of political tribalism where FSD is bad because https://elonbad.com/ https://elonbad.com/

nojvek 3y ago
FSD is cool as a great demo. But the optics and facts are that it’s got people killed. Multiple people over the years. Silly mistakes causing crashes. It will get better but definitely does not live up to the “full self driving” marketing hype. That kills the magic. Meanwhile look at Waymo. They don’t make a lot of noise. They take safety really seriously and keep on improving actual “self driving cars” city by city. Zero people dead. I’ve sat in both. FSD was a great demo, but Waymo truly felt like magic. No driver at all!
- mbrumlow 3y ago
  I don’t think we will ever have any self driving tech that inter mingles with non self driving cars that will result in zero deaths. While sad, mile for mile FSD is better than humans.
  - Mawr 3y ago
    FSD only drives in a subset of conditions humans drive in so the comparison is invalid.
  - acdha 3y ago
    > While sad, mile for mile FSD is better than humans. For this statement to be correct, we’d need to have full disclosure of all travel using FSD at any point, accidents which happened anytime FSD was active or had been recently deactivated (for example, that guy who fell asleep counts even if FSD deactivated a minute before the vehicle crashed), and be able to compare that to the same trips driven by human drivers. You especially need to avoid including incidents in the human stats which are in conditions where FSD would do even worse.
  - alemanek 3y ago
    How the heck did you reach the conclusion that the current state of FSD is safer than a human driver. The Tesla Community FSD tracker has it at 157 city miles per disengagement at the moment. The people collecting this data are the Tesla enthusiasts as well. It has a very long way to go before it is better than a human. We won’t know the true stats until it is allowed to operate without supervision. Source: https://www.teslafsdtracker.com/Main https://www.teslafsdtracker.com/Main

everforward 3y ago

I don't think they lead the pack on that, though. Everybody in the self-driving space is using AI to some degree. E.g. Waymo was at 17,311 miles per disengagement (human takeover) in their 2023 report, and they're not even the top. Zoox was the top at 177,602 miles per disengagement, which is shockingly good if they're not gaming those numbers with tiny service areas or something. I don't think Tesla publishes their disengagement data, but what I can find crowdsourced from their users is pretty bad relative to the above. The most optimistic number I could find was from 2022 at ~400 miles per disengagement. That's not even very good for 2022; Mercedes-Benz was at 1,400 miles per disengagement, and I didn't even know they had a self-driving division. Nissan was at 149 miles/disengagement, which makes Nissan their closest competitor by capabilities (the next highest after Tesla was QCraft.ai at 863 miles/disengagement, no idea who they are).