back to index

Elon Musk: Treat All Input as Error | AI Podcast Clips


Whisper Transcript | Transcript Only Page

00:00:00.000 | (gentle music)
00:00:02.580 | - So the magic of deep learning
00:00:10.160 | is that it gets better with data.
00:00:12.680 | You said there's a huge inflow of data,
00:00:15.040 | but the thing about driving,
00:00:17.560 | the really valuable data to learn from is the edge cases.
00:00:22.160 | So how do you, I mean, I've heard you talk somewhere
00:00:26.960 | about autopilot disengagement being an important moment
00:00:30.640 | of time to use.
00:00:32.400 | Is there other edge cases,
00:00:33.680 | or perhaps can you speak to those edge cases,
00:00:38.040 | what aspects of them might be valuable,
00:00:40.100 | or if you have other ideas,
00:00:41.560 | how to discover more and more and more edge cases in driving?
00:00:44.640 | - Well, there's a lot of things that are learned.
00:00:47.800 | There are certainly edge cases where,
00:00:50.280 | I say somebody's on autopilot and they take over,
00:00:53.520 | and then, okay, that's a trigger that goes to a system
00:00:57.840 | that says, okay, did they take over for convenience,
00:01:00.600 | or did they take over because the autopilot
00:01:03.560 | wasn't working properly?
00:01:04.820 | There's also, like let's say we're trying to figure out
00:01:07.280 | what is the optimal spline for traversing an intersection.
00:01:11.720 | Then the ones where there are no interventions
00:01:16.800 | are the right ones.
00:01:19.080 | So you then say, okay, when it looks like this,
00:01:21.800 | do the following.
00:01:23.720 | And then you get the optimal spline
00:01:26.080 | for navigating a complex intersection.
00:01:30.220 | - So that's for, so there's kind of the common case.
00:01:34.760 | You're trying to capture a huge amount of samples
00:01:37.720 | of a particular intersection, when things went right,
00:01:40.480 | and then there's the edge case where, as you said,
00:01:44.680 | not for convenience, but something didn't go exactly right.
00:01:47.480 | - Somebody took over, somebody asserted manual control
00:01:49.520 | from autopilot.
00:01:51.440 | And really, like the way to look at this is view
00:01:53.440 | all input as error.
00:01:55.320 | If the user had to do input, if there's something,
00:01:58.080 | all input is error.
00:01:59.320 | - That's a powerful line to think of it that way.
00:02:01.760 | (upbeat music)
00:02:04.340 | (upbeat music)
00:02:06.920 | (upbeat music)
00:02:09.500 | (upbeat music)
00:02:12.080 | (upbeat music)
00:02:14.660 | (upbeat music)
00:02:17.240 | [BLANK_AUDIO]