4 ms·

The idea that data will somehow magically translate into FSD is laughable. It relies in the delusion that we just need to train neural networks with the proper

• _coveredInBees 5y ago
  It isn't laughable at all. The real problem of FSD is the ridiculous long-tail of scenarios in the real-world that you simply cannot account for or manage well. At this point, Tesla has a huge upper hand because every vehicle in their fleet can constantly collect and provide new semi-labelled training data every time there is a user disengagement or an unforeseen action taken by the driver. Tesla has built out amazing infrastructure to capture extensive amounts of "hard" examples from their fleet, turn them around into labeled data for training very efficiently and then utilizing simulations to further broaden the distribution of such quirky long-tail events in their training-set. In the absence of AGI, this is a very effective "brute-force" approach and they have a huge upper hand over every other player in this space. I say all of this even though I am very skeptical that anyone will achieve L5 self-driving with where the state of things are today. But Karpathy and team are very pragmatic and making lots of good decisions coupled with excellent engineering and infrastructure development.