Hacker News new | ask | show | jobs
by goalonetwo 561 days ago
This i what is so tricky with Self Driving. People "feel" it is almost there because most of their rides are mostly ok. However to make a system truly driverless you need to master the long tails of difficult events and FSD is nowhere even close to do that.

Going from 99.9% to 99.99999% is what makes a system truly driverless and where most of the work is. Waymo is way way ahead of FSD for this.
1 comments
floxy 561 days ago
I know essentially nothing about machine learning, nor about what approaches to ML that Tesla or Waymo are using. Is there a Metcalfe-type effect possible here? Where the better FSD becomes, the more people use it, finding more edge-cases, so the driving get better, so more people start to use it. With the end result that the learning starts to get better/faster with time, in a positive-feedback like mechanism?
link
friendzis 560 days ago
The answer is very strictly no.

Your comment/question seems to assume that somehow accurate classification of performance is possible, which is demonstrably false assumption.

First, if somehow system-internal watchdog was able to detect erroneous outputs for training, it would also be able to stop those outputs propagating to control system live, leading to zero errors. Such a watchdog requires self driving to be quantified analytically, leaving implementation entirely testable and therefore control system not passing the testsuite (i.e. exhibiting erroneous behavior, i.e. still having "edge cases") not deployable in public. Many words to say that that in practice with driving being somewhat loosely defined you need humans in the training loop.

Second, vehicle operators are not only unqualified to accurately monitor behavior of these autonomous systems, in part due to not being formally trained on safety systems in general and the system they are monitoring in particular, but on top of that incentivized to underreport erroneous output.

My prediction has been that we are going to see increased number of accidents with ADAS deployments, not less, before the number can start dwindling.

link