e.w. niedermeyer: The Tesla also uses a "model" rather than hard-coded "rules," but it's a statistical model that fundamentally lacks any of the social and psychological insights humans constantly use without even trying while driving. Driving is a highly social task, and statistics only get you so far in those.

"It marks the second incident in two months involving a Tesla’s near miss-with a train while utilising its driver assistance system." You can tell Tesla is close to "solving self-driving" because they are down to the tough edge cases like *checks notes* trains nz.news.yahoo.com/tesla-autopi...

Tesla autopilot appears to veer electric vehicle onto train track it mistook for roadnz.news.yahoo.com Local police urged drivers to remain ‘vigilant while using Tesla's autopilot feature,’ noting that it ‘can fail’

I think it's underappreciated that for humans the process is Visual data -> mental model of surroundings -> decision Whereas for a Tesla it's Visual data -> some simple rules -> decision so sometimes the simple rules condemn you or somebody nearby to death.

The Tesla also uses a "model" rather than hard-coded "rules," but it's a statistical model that fundamentally lacks any of the social and psychological insights humans constantly use without even trying while driving. Driving is a highly social task, and statistics only get you so far in those.

This is at the core of AI failures, it uses stats to “infer” the actual answer. Doesn’t use known answers.

TFW you train your model on the boss’ favourite 80s movies backtothefuture.fandom.com/wiki/Railroa...

Railroad crossingbacktothefuture.fandom.com "The car was coasting by the Hilldale housing development now, the place where the railroad tracks crossed the main highway. He [Marty] could hear the crossing gates clanging up ahead as the barriers ...

I should have specified: a set of rules that the programmers don't really know either :)

Also curious if they classify cars by their behavior. Like you'll see an aggressive driver in the rear view and think "that person is going to aggressively try to get in front rather than do a clean merge behind me" and be more prepared when it happens.

One of the fun things about driving in SF is watching the (real) self-driving cars slowly get better at this. They got better about showing intent, about reading other drivers intent, etc. I'd love to know what things they have discovered about real driver behavior and how local it is.

Post