All right, you guys.
I just finished watching a 45-minute live stream of Musk driving around with WholeMarsCatalog in a V12.x-equipped Tesla. This popped up on the Investors's Forum, here's the
link to one of the posts.
The whole thing was streamed with what looks like WholeMarsCatalog's cell phone, so it's a little shaky. In the 45 minute drive, there was one intervention.
Now comes the crazy stuff. In no particular order:
- Musk said, repetitively, that they didn't tell the NNET what things were. That stop sign? Nope. The traffic circle? Nope. Other cars? Nope. How to drive smoothly? Nope. Tesla just showed it videos of how good drivers drive and told it to do that.
- Musk stated that there's 300k+ lines of C++ code in V11.x that do tell the car what to do in traffic circles, when cyclists come by, how to handle intersections, how to handle left and right turns, and all that jazz. That code is gone.
- The Hardware load on HW3.0 that he was driving allowed for 50 frames per second of driving NN work. They had to limit it to 36 frames per second because that's the limit of how fast the cameras can run. He stated that HW3 is more than capable of Doing The Job. And, in fact, wasn't working as hard as it does in V11.x.
- His words: It's NNET, all the time.
Guys... From time to time I've made mumbly noises around the FSD-b threads that there's been gradual improvement, but thought that, with software, breakthroughs can happen that take the idea that each version is 1.2X better than the previous (leading to infinite numbers of 1.2X better before we get to 9 9's better than humans) and toss those ideas out the window. If what I just saw tonight is real, and I have no reason to believe that it's not, at least one major breakthrough has happened.
I have no idea if this is leading to robotaxis. But it sure looks like a step in that direction.
Couple of minor points.
Musk says that 12.x isn't ready yet, it's still making mistakes from time to time. OK, I get it: It's a massive re-write, so what else is new.
Musk discussed that infamous, "Must come to a complete stop at a stop sign" fight with the Regulators. He said that real, live data, from real, live drivers showed that humans only came to a complete stop 0.2% of the time. Even when a human thought that they had stopped.. they hadn't. They pointed this out to the regulators: They demanded the complete stop. So, now you know who to blame. (Some of us had thought that around here, but here it is from somebody who was actually there.)
I'm in shock. That was one
heck of a drive.