Ep4. Tesla FSD 12, Imitation Learning Models, The Open vs. Closed AI Model Battle, Delaware’s anti Elon ruling, & a Market Update

**Bill Gurley** (0:00)
I would make the argument that every company in Delaware has to move to a different domicile because they could be sued in a future derivative lawsuit for the risk they've taken by staying in Delaware.

**Brad Gerstner** (0:15)
Oh my god, you're so right. You are so right. Oh, mic drop on that.
Hey, Bill, great to see you. I mean, people loved when you were here last week in person, so we gotta make that happen again. But now, where are you? Looks like you're in Texas somewhere.

**Bill Gurley** (0:41)
I'm back in Texas, yes.

**Brad Gerstner** (0:43)
All right, all right.
So what's on your mind? It's been a lot of action the last couple of weeks. What's going on?

**Bill Gurley** (0:51)
One thing that I've reflected on quite a bit is just kinda how lucky we are to be a part of the venture capital industry and the startup world, simply because things change so fast. And if you're a curious person, if you're someone that likes constant learning, it's really amazing. The stuff we're talking about, the stuff I'm listening to podcasts on every day, two years ago didn't exist. And now it's 80 or 90% of the dialogue.
And that's just pretty well.

**Brad Gerstner** (1:23)
Yeah, no, our brains really aren't programmed to work in kinda these exponentials, right? I mean, you and I both know every sell side model on Wall Street has linear deceleration and growth rates. Like we think really, we're really good at thinking in kind of these linear ways. I had that thought this morning that the biggest investment opportunities really do occur around these phase shift moments. I mean, Satya talks about all the value capture occurs in the two to three year period around phase shifts, but it's hard to forecast in those moments, right? I mean, that's when you see these massive deltas in these forecasts. And I just went back and looked at, for example, at the start of last year, the consensus estimate of the smartest people covering Nvidia day to day was that the data center revenue was going to be 22 billion for the year, right? Guess what it ended up being?
96 billion.
Okay, they were off almost by a factor of three or a four, right? The EPS at the beginning of last year, the earnings per share was expected to be $5.70. And now it looks like it's going to be $25, right? Like, over the course of your career, have you ever seen sell side estimates off by that much on a large cap stock?

**Bill Gurley** (2:43)
I mean, just like, you know, very, very rare. Like, you know, once a decade maybe, you know, that something like this happens.

**Brad Gerstner** (2:52)
So, you know, and I've had investors say to me when the stock was at 200, hell, you and I talked about this.
You know, should we sell it all at 200, sell it all at 300, sell it at 400? And now, you know, those investors are calling me every day saying, have you, you know, have you sold it yet? Our general view is that if the numbers are going up, so if our numbers are higher than the street's number for whatever variant perception that we have, right, then the stock is going to continue to go higher. At some point, the street will get ahead of itself and its numbers will now be higher or at the same level as ours. And at that point, I think it becomes more of a market performer.
But of course, some things will be wildly overestimated and some things will be wildly underestimated, but that sort of discontinuity really occurs around these moments of big phase shifts. So speaking of a big phase shift, right, we teased on the pod, I think at the start last time that I had taken a test ride in Tesla's new FSD12. And I said, kind of felt like a little bit of a ChatGPT moment, but I think we left the audience hanging. We got a lot of feedback. Hey, dig in more to that. So you and I spent some time on this both together and with some folks on the Tesla team. So roughly the setup, your background, I wanna get your reaction to it, is about 12 months ago, the team pretty dramatically forked their self-driving model, right? Moving it from this really C++ deterministic model to what they refer to as an end-to-end model that's really driven by imitation learning, right? So we think of this new model, it's really video in and control out.
It's faster, it's more accurate, you know, but after 11 different versions of FSD, I think there's a lot of skepticism in the world. Like, is this going to be, you know, something different?