Mar 7, 2024 · 1h 3m · bg2-pod
Ep4. Tesla FSD 12, Imitation AI Models, Open vs Closed AI Models, Delaware vs Elon, & Market Update · Bg2 Pod
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of the BG² Podcast, hosts Brad Gerstner and Bill Gurley analyze the paradigm shift to end-to-end imitation AI in Tesla's FSD V12 and physical robotics, debate open-source versus proprietary AI models and regulatory capture, and evaluate Delaware corporate governance risks alongside Big Tech market valuations.
Statements from this episode (23)
Gerstner: Nvidia 2023 data center revenue beat initial consensus by over 4x
“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 twenty two billion for the year. Right. Guess what it ended up being? Ninety six billion.”
Gerstner: Nvidia EPS forecast expanded from $5.70 to $25 in one year
“The EPS at the beginning of last year, the earnings per share was expected to be five dollars and 70 cents. And now it looks like it's going to be 25 dollars.”
Gurley: Tesla's end-to-end AI approach has much better chance of achieving FSD
“It seems to me this approach has a much better chance of going all the way and of being successful and certainly of being maintainable and reasonable.”
Gerstner: Tesla FSD 12 model improves 5x to 10x faster monthly than prior systems
“In fact, I think what, you know, we learned was that the rate of improvement of this is order of magnitude five to 10 X better per month as a model versus the rate of improvement of those Prior systems.”
Gurley: Tesla uses generic open-source AI models for FSD V12
“And these AI models that Tesla's used for FSD-II are these generic open source AI models that you can find on hugging face, you know, and they obviously customize them.”
Gerstner: 99% of vehicle data never gets sent back to Tesla
“And in fact, I think 99% of the data that a car collects never makes it back to Tesla.”
Gurley: Waymo and Cruise cannot compete if neural networks solve autonomy
“If let's, let's make an assertion. If this type of neural network approach is the right answer. Yes. And I want to reason you once again, you know, Occam's razor seems that way to me, then who could compete in one of the companies or two, you know, several of …”
Gurley: BYD and Chinese OEMs are most likely to rival Tesla
“If I had to handicap anybody, it would probably be BYD or one of the Chinese manufacturers.”
Gerstner: Tesla earns $2,500 per car with 7% FSD penetration
“If you look at the unit economics of a Tesla, right? With no FSD, they're making about two and a half thousand bucks on a vehicle. If you look at it today, they have about seven percent penetration of FSD.”
Gurley: Operating a robotaxi fleet service requires $100B in CapEx
“I've always thought this model that we're going to build this incredible car and our business model is going to be to run a service like the CapEx. Like if you just build a 10 year model, the CapEx you need, like they would have to go raise a hundred billion.”
Gerstner: Tesla's autonomous driving engineering team is roughly one-tenth the size of Cruise's
“I think the engineering team working on this at Tesla is about one-tenth the size of the teams at Cruiser.”
Gerstner: Llama 3 and GPT-5 Will Launch Between May and July
“The first H 100 only trained models, I think, will be Llama three and chat GPT five. So we're hearing rumors that both of those models are going to come out in the May, July timeframe.”
Gerstner: Next-Gen AI Frontier Models Will Arrive in Q2 2025
“All the rest of the frontier models are kind of in this holding pattern because they're waiting for the B 100 to get launched to this Q three, Q four out of Nvidia. Which probably means the next iteration of the frontier models will come out in Q two of next y…”
Gurley: Startups Switch to Multi-Model Setups in Production for Cost
“When I look at what I see going on in the startup world, they might start with one of these, you know, really well-known service models that's proprietary, but the minute they start thinking about production, they become very cost focused and on the inference …”
Gerstner: OpenAI Should Rename to ChatGPT and Focus on Google
“If I was Sam Altman running OpenAI, I think I might rename the company ChatGPT and just focused on the multi-trillion dollar opportunity to replace Google, right? Because I think winning at both, beating Google at consumer and beating Microsoft at enterprise y…”
Gerstner: Zuckerberg will commoditize AI pricing by releasing cheap frontier models
“Zuckerberg is going to drive the price, right? He's going to give away frontier esque models on the cheap. Okay. And that's going to be highly disruptive to your ability to stack margin.”
Gurley: Open-source AI models will outperform proprietary models in the enterprise
“So there's a lot of different elements that, that make me think that the open source component models, Are gonna be way more successful in the enterprise, and it's a really tough thing to compete with.”
Gurley: Banning open-source AI would cause the US to lose to China
“The one thing that would cause us to get way behind China is if we played without open source and they had it like us.”
Gurley: Google Gemini acts like an autocratic model restricting information
“The Google Gemini model was much more similar to something autocratic that you might equate with a communist society, like that's intentionally limiting the information you can have and painting it in a very specific way. And so, yeah, I, I'm more afraid of th…”
Gurley: Tesla pay package plaintiff lawyers are seeking $5B–$6B in fees
“The lawyers that, that pursued this case are asking for five or six billion dollars in payment.”
Gurley: Musk's 2018 Tesla compensation package was among most shareholder-aligned ever
“It may be like one of the most, you know, shareholder aligned incentive packages ever, which is, is exactly what you would think Delaware courts would be looking after.”