[LMB] self-driving vehicles OT:

Damien Sullivan phoenix at mindstalk.net
Wed Sep 1 15:06:31 BST 2021


On Wed, Sep 01, 2021 at 04:28:25AM -0500, Raymond Collins wrote:

> I think after both the software and hardware glitches are solved then we
> will have a reliable smart cruise control system for cars. It's just a
> matter of market forces convincing manufacturers to build a better and
> safer autonomous vehicle.

No, it's a lot more than that.  Navigating from point A to point B is
easy.  Driving in an urban environment is *hard*.  It's not a matter of
'glitches' and just trying hard enough, it's a matter of hard AI
problems: predicting behavior, theory of mind, reading ad hoc detour
signs.

https://xkcd.com/1425/

Predicted deadlines keep receding into the future.  The *executives* in
charge of these programs are pessimistic.  It's not a matter of
"convincing" them, the market incentives are already substantial.  It's
a matter of what's possible when you're asking insect-level computes to
navigate human-complex environments.

> On Wed, Sep 1, 2021, 3:38 AM baur baur via Lois-Bujold <

> > > Humans are still a lot better than the "self-driving" car systems.
> >
> > there are already indication that the CURRENT asssist system (which are
> > not full level 4 or 5 autopilots) do reduce the numbers of accidents in
> > partly-protected motorways whne compared to human driving

That's easy mode.  Humans can drive through Boston streets, with
pedestrians and bicyclists, in the snow, at night, with road closures.
"Self-driving" cars get tested during the day in dry climates that have
no weather; last I heard, admittedly a few years ago now, they failed
outright in rain or snow.

Human drivers aren't perfect but they do a decent job in a huge variety
of appalling conditions.  The AI doesn't.  And the AI has to be not just
safe -- that's easy, just don't move if you might possibly hit something
-- but also effective,  actually getting somewhere.

-xx- Damien X-)


More information about the Lois-Bujold mailing list