,
11 tweets,
3 min read
Read on Twitter
1\ Here is a thread of incredible quotes from the AutoPilot talk given by @karpathy. In $TSLA's own words, they don't even know how to approach the problem. The product they've been selling is vapourware.
2\ "adventures of trying to get this to work"
"[Advanced Summon] is the best...when it works."
"Even for one task, this can be a very difficult problem, if you want it to work all the time"
[Long list of vehicle types that challenge or break their classification system]
"[Advanced Summon] is the best...when it works."
"Even for one task, this can be a very difficult problem, if you want it to work all the time"
[Long list of vehicle types that challenge or break their classification system]
3\ [Long list of contextual challenges: crosswalks, neighborhood type, static objects, toll booths, tunnels, sirens, turn signals]
"The scope of this problem very quickly becomes quite daunting"
"The scope of this problem very quickly becomes quite daunting"
4\ "I'll talk about the struggles and some of the ways in which we're *approaching* these problems"
[Long discourse on different ways to divide the architecture across tasks. Sounds like they don't know which is best]
"There's a massive space of in-betweens"
[Long discourse on different ways to divide the architecture across tasks. Sounds like they don't know which is best]
"There's a massive space of in-betweens"
5\ "Exactly how you lay out the architecture, there's a large un-known there"
"Exactly how do you lay out this architecture?"
"It's not exactly obvious"
"There are papers that have recently started to come out..."
"Exactly how do you lay out this architecture?"
"It's not exactly obvious"
"There are papers that have recently started to come out..."
6\ "How should I allocate the heads to the backbones?"
"We also try...but it's very complicated"
"It's not obvious how you study them"
"Where it gets even more complicated for us..."
"This is what the network could look like."
"We also try...but it's very complicated"
"It's not obvious how you study them"
"Where it gets even more complicated for us..."
"This is what the network could look like."
7\ "It's a bit of a beast"
"This paper I found a few weeks ago..."
"A lot of it relies on intuition. I'd love to see other approaches.."
"The difficulty of these tasks"
"Some objects take a long time to train"
"There's no real language in the community"
"This paper I found a few weeks ago..."
"A lot of it relies on intuition. I'd love to see other approaches.."
"The difficulty of these tasks"
"Some objects take a long time to train"
"There's no real language in the community"
8\ "How you come up with these task weights is unknown"
"Some of the problems you run into"
"There's a lot of open questions here"
"Where do you take your model? I don't know."
"This heterogeneity is tricky"
"It's not obvious"
"Some of the problems you run into"
"There's a lot of open questions here"
"Where do you take your model? I don't know."
"This heterogeneity is tricky"
"It's not obvious"
9\ "It's a beast of a network, and we have a finite team"
"Two interns are working on the architecture on the bottom" [I think this was a joke?]
"It's really not obvious how you iterate on networks when multiple people are changing the tasks, the weights.."
"Two interns are working on the architecture on the bottom" [I think this was a joke?]
"It's really not obvious how you iterate on networks when multiple people are changing the tasks, the weights.."
10\10 [Discourse on how when they optimize one part of the network, it degrades other part of the network unpredictably]
"Political drama on the team"
"There's no easy way to do that"
"Some of these models become not-reproducible"
"Political drama on the team"
"There's no easy way to do that"
"Some of these models become not-reproducible"
11\11 Just as incredible as what was said in this talk, is what was *not* said.
Almost no bragging about obstacles overcome, or goals within sight.
Just @karpathy humorously lamenting how inchoate the field is, and how ill-defined most of the problems are.
Almost no bragging about obstacles overcome, or goals within sight.
Just @karpathy humorously lamenting how inchoate the field is, and how ill-defined most of the problems are.
