[LMB] self-driving vehicles OT:

Marc Wilson marc.wilson at gmx.co.uk
Fri Sep 3 22:32:54 BST 2021


On Thu, 2 Sep 2021 19:27:29 +0100, Matija Grabnar via Lois-Bujold
<lois-bujold at lists.herald.co.uk> wrote:

>
>On 02/09/2021 19:00, Louann Miller wrote:
>> Not just Asimov, either. It turns out that before we had computers to smoke
>> test our ideas about computers, we were really bad at guessing what would
>> be easy and what would be hard for them. Probably based on over-using the
>> assumption that a human brain = really smart purpose-built computer. Boy
>> were we wrong.
>
>Very true. A nice representation of it is this comic: https://xkcd.com/1425/
>
>The joke is a bit out of date, we know how to solve that now, just use a
>neural net.

Hmm.  You have to be very careful with the training data.  There's the
classic example of a system designed to distinguish NATO tanks from
Soviet ones.

It worked almost ridiculously well with test images, but not in the
field - because all the pictures of Soviet tanks were in shadow.
--
The trouble with these "I'm a bigger asshole than you are"
contests is that they are winnable."  - Christopher J. Henrich



More information about the Lois-Bujold mailing list