Vibes & Benchmarks Ep 06: Does AI actually give anyone an edge?

Transcript of the final episode. Speaker attribution is best-effort from a label-less recording.

Cold open

Ali Rohde (0:00) If you were advising the Cursor CEO, Michael Truell, how do you continue the impact as part of SpaceX? What would be your number one piece of advice?

Josh (0:06) How soon can you sell those shares? What's the lockup period again? That's my advice. But after that, oh well, it's gonna be a few years.

Ali Rohde (0:18) Shit. You got a bad deal. Okay, well.

TOPIC 1 — SpaceX buys Cursor for $60B: the distillation bear/bull case

Ali Rohde (0:25) Josh, welcome to episode six.

Josh (0:30) Hello.

Ali Rohde (0:31) Let's jump right in. SpaceX went public this week. Guess what they're now valued at?

Josh (0:35) More than when they went public.

Ali Rohde (0:38) Over 2.6 trillion. Just announced today, they're officially acquiring Cursor for 60 billion. So before there was the breakup fee; now, officially acquiring Cursor. Let's go through the bear and bull case, not so much for the acquisition but for Cursor and its future broadly. So maybe the easiest thing for you: give me the bear case for Cursor.

Josh (1:02) Uh, it's an IDE and no one uses IDEs. I think we have sticky subscriptions because we were too lazy to cancel them, because we had better things to do.

Ali Rohde (1:11) Hey, that's still money. I mean, they're crushing it.

Josh (1:15) Yeah, but that's my point, though. Does it really matter if you have this recurring revenue from before? We mostly just didn't bother optimizing it and getting rid of it. That's not really user value.

Ali Rohde (1:27) Fair enough. You and Imbue are not representative of companies broadly, right? So they're doing very well. You wanted the bear case.

Josh (1:37) Yes, that's fine.

Ali Rohde (1:39) Okay, let's go to bull case. Honestly, that was less bearish than usual. I guess you're more bearish on the acquisition not making that much sense, but Cursor perhaps.

Josh (1:48) Well, another part of the bear case is that they don't really have any moat or any kind of fundamental technology, right? Really what they're doing, and this gets us to the bull case, is taking a bunch of traces from Claude Code and then training on those to make their own model that's slightly more focused on coding, slightly smaller, and slightly cheaper for them to run inference on, that they then resell.

Ali Rohde (2:17) That's called distillation. Anthropic hates that. So does that continue forever? How do the pricing things work out there? This isn't obvious, like, "we don't want you to do this kind of a thing," from Anthropic and OpenAI.

Josh (2:28) But right now they do. And the reason they want them to do it now is because that accounts for a significant fraction of their revenue. So that's the bear case: Anthropic will tolerate your existence while you're here reselling their expensive tokens to someone else so you can build up your data set on an older model. And then later they're like, "nah, we wanted that part of the market." So now you have to pay API rates, and good luck.

Ali Rohde (2:56) Okay, bear case. You're pointing at what I wanted to discuss for the bull case, which is this general movement for companies to, through distillation or other means, build their own smaller, more verticalized, cheaper models. That's what Cursor is doing, I guess: using Claude Code to learn, but also building off of open source models.

Josh (3:20) Yeah, I think they're building on top of open source models and fine-tuning them. But the data, the important part, is coming from the closed source models that they're effectively distilling. They're getting all of those traces from their customers and using them to fine-tune their own open source models that they didn't even pre-train. They're just taking Kimi K2 or whatever and making their own variant of that.

Ali Rohde (3:48) So the bull case, then.

Josh (3:50) Well, the bull case is that seems a lot more efficient than pre-training your own language model only for somebody to go distill it. So there's a question of market pricing and power dynamics, whether this works out for Cursor or not. Are they going to be able to do this long enough that they make a significant amount of money? Will they, with their acquisition by SpaceX, have access to enough compute that they could train their own foundation models in order to compete? Probably. Do they have any experience doing that? They haven't really done a great job of it so far, but it's pretty mechanical. It's relatively known how to do this. So it'd be nice to see them do it and have more competition. I'm all for more competition. I think that's probably the bet with SpaceX: we get more compute, then we can, if we need to, train our own coding models and sell them through Cursor, and keep distilling this way. And hey, we have a coding agent too.

Ali Rohde (4:38) Distillation comes up mostly, at least so far, in regards to Chinese model builders. Is what you're implying that Cursor is doing the same thing as what these Chinese model builders are doing?

Josh (4:56) Yes. And I think there's a very intentional choice of like, "oh, Chinese models are distilling our weights, this is bad." American companies are also distilling your weights. But why is it that you only say Chinese when you talk about distillation by other people? I'll leave that as an exercise to the listener.

TOPIC 3 — Advice to Cursor & SpaceX: own the whole stack

Ali Rohde (5:10) If you were advising the Cursor CEO, Michael Truell, how do you continue the impact, now as part of SpaceX? What would be your number one piece of advice?

Josh (5:17) How soon can you sell those shares? What's the lockup period again? That's my advice. But sure, after that, oh well, it's gonna be a few years.

Ali Rohde (5:34) Shit, you got a bad deal. Okay, well, fine. In the meantime, then.

Josh (5:39) I think regardless, he's gonna be just fine.

Ali Rohde (5:42) He's like 25. I don't think money is necessarily the most important thing. I imagine legacy and impact is.

Josh (5:50) I was saying, even for that case, it might be better to do something outside of it. Working for Elon doesn't sound super fun.

Ali Rohde (5:54) No, continue Cursor and make a lasting legacy with Cursor.

Josh (6:00) With SpaceX, you mean.

Ali Rohde (6:01) With Cursor in SpaceX, as part of SpaceX.

Josh (6:04) It's SpaceX now. It's not Cursor-in-SpaceX anymore. That's what happens with acquisitions.

Ali Rohde (6:07) Bull-case Josh advice, if he wants to do this: what is your number one piece of advice? What could he do?

Josh (6:16) Do whatever Elon says and don't get fired. That's my point.

Ali Rohde (6:18) Is this part of SpaceX?

Josh (6:18) Okay, if you're asking what should Cursor do as an independent entity now with access to SpaceX's resources, well, I don't think this is the right question. We're gonna answer a second question afterwards, but I'll answer that first. I think it's basically just keep doing the same thing you're doing, and start getting, you have enough money now to be more serious about pre-training and about making sure you have independence from these model labs, because they're going to cut you off. So I would definitely get on that. And maybe I would advocate for, why don't you make it open? It's dumb for you to try and be the fourth-best player. Look how that's working out for Meta. Don't be the fourth-best. Just help people make open source models, and then go do the rest of the stuff you're gonna do and build on top of them, instead of spending tens of billions of dollars on this yourself in order to make something that isn't as good as what's out there. That's crazy. Stop. Just do what Satya says. The other question is, okay, what should SpaceX do? I think this is the more interesting question, which is why they acquired them. That's why I'm saying the Cursor thing is, you should probably just stop messing with this and give yourself a good business here, because you're not thinking about the larger picture. With SpaceX in the picture, they potentially have a lot more capital, so they can do a lot more interesting things around, for example, producing chips, producing their own models. I think they said a long time ago, maybe for Tesla, they were going to make their own chips. I don't even remember exactly, I think it's been repeated as part of the vision.

Ali Rohde (7:45) They should do that.

Josh (7:48) AI, yeah. But making your own vertical stack, I think, is the way that you actually win here. That's why I've been so bullish on Google, for example, because despite them fucking it up in so many different ways, they're still in a really good position: they have their own data centers, their own hardware, their own software stack. That's the real value with SpaceX: they have enough capital and enough talent, expertise, and resources to actually get that done.

TOPIC 4 — AI & productivity: who captures the value

Ali Rohde (8:14) Let's talk about AI and productivity. Do you think we're going to see true productivity gains from companies using AI?

Josh (8:19) I think it would be weird to say that we're not seeing productivity gains already in some companies. Certainly we're more productive than we were before. There's no question about the impact on engineering.

Ali Rohde (8:35) Yeah. But for that, I'm curious, writing more code, faster code, does not necessarily mean a better product.

Josh (8:44) That's right. It rarely does, I agree.

Ali Rohde (8:49) So in order for AI to create productivity gains, it has to actually translate into some real value.

Josh (8:56) Hmm. So you're saying that a company using AI would then have to be worth a lot of money.

Ali Rohde (9:04) Would have to be worth more than before.

Josh (9:06) So what I would propose that we'd see from productivity gains is like a company that started out really small and then all of a sudden was worth like trillions of dollars. Or maybe there'd be a few companies like that, maybe one worth 2 trillion and then another two each worth 1 trillion.

Ali Rohde (9:22) So, 2.6.

Josh (9:24) Yeah. So maybe it'd be like 3.6, no, sorry, 4.6 trillion of us more productive. But that is exclusively accruing to like 3 or 4 companies. Yeah, but that doesn't seem incoherent at all.

Ali Rohde (9:37) Okay, then for the average company using AI, you don't expect to see productivity gains.

Josh (9:39) The whole point of these companies is to make their company valuable and grow and be useful by capturing most of the value, right? They're not trying to distribute it. That's kind of the whole problem.

Ali Rohde (9:52) Okay. Then if you're one of these CEOs that's not Anthropic or OpenAI, do you still use AI? Do you allow your team members tokens?

Josh (10:03) That's the brilliant part of this whole thing: it's set up now where you can't not use AI. To be competitive, you have to use AI, but it doesn't give you an advantage. All the money accrues to the people that made the AI system in the first place. But in order for you to even be kind of treading water, you now have to spend a bunch of extra money on tokens.

Ali Rohde (10:20) Okay, so now teams, engineering teams in particular, have to spend a ton of money on tokens just to keep up. They're not expanding business value. They're not shipping important new features or products.

Josh (10:38) They're shipping tons of important new features. But the problem is they're not capturing that value. They ship new features, their competitors ship new features, all the products get the same new features at the same time. So no users switch, and everybody pays the same amount because they could just switch from one to the other. But all the value that was created goes to the model providers.

Ali Rohde (10:54) A lot of AI optimists talk about GDP increases and value creation. If we take away OpenAI, Anthropic, Google, maybe a few others, do those gains then disappear?

Josh (11:08) There was already an article about this, I think last week, saying that if you take away the gains from companies that are doing AI, then we're in a recession. If you remove Google and Anthropic and OpenAI and Tesla and SpaceX, you're left with, actually, most companies have gone down considerably.

Ali Rohde (11:28) So if you're a leader at one of these companies, you're just stuck.

Josh (11:31) Yeah. Get acquired by SpaceX.

Ali Rohde (11:34) Get acquired by SpaceX. You heard it here first.

Josh (11:38) Not bad advice.

TOPIC 6 — Anthropic, the Pentagon & Fable: AI regulation

Ali Rohde (11:39) In addition to the SpaceX IPO, the biggest news this week was Anthropic and the Pentagon, the Pentagon revoking access, or forcing Anthropic to revoke access, to Fable. People got access June 9th; they lost it June 12th.

Josh (11:51) One Twitter commenter had a good point: the current administration says, "we don't want to regulate AI," et cetera, but then, what is an export control on a model? Is that not regulation of AI? Yes, it is regulation of AI, just in a totally ad hoc way with no rules at all.

Ali Rohde (12:09) That's just the nature of this administration.

Josh (12:12) They've also kind of reversed course a little bit, that they do maybe kind of want to regulate AI. But sure, I take your point. The point is, all of this stuff is incredibly quickly and badly done.

Ali Rohde (12:25) Yes, it's super badly done.

Josh (12:27) But does that mean that you don't actually want any regulation of AI ever? No, you probably do want some. So it's just very complicated: you would want some regulation, but you also don't want an administration that doesn't really respect the rule of law anyway. The models aren't really dangerous enough to really warrant this today, but they will be someday. They can't really do any biology, but biology would be a threat. All of these things are kind of complicated. It's not an easy, clean answer where we can say "always do X." It's a little bit more contingent, and it just requires a little bit more nuance and complexity than I think our 240-character posts are capable of expressing.

Ali Rohde (13:01) 280. Oh, I'm sorry, those extra 40 characters are really going to flip that bit. Besides the government being active in antitrust, which you've talked about earlier, what regulations do you want?

Josh (13:14) I think Anthropic, lots of people have had lots of ideas. They're way too complicated to go into all the details here, but you can imagine having, like, and we've talked about them before.

Ali Rohde (13:24) Regulatory capture.

Josh (13:27) No, I'm saying we've talked about them before, in terms of like a year or two ago, like SB 1047. You can think about liability regimes. You can think about licensing. You can think about what safeguards look like, what testing looks like. There's a lot of things you could do. There's NIST, in fact, there are whole committees and organizations dedicated to measuring how dangerous things are. There's a ton of complexity, a ton of things one could engage with if one wanted. But we're not really.

Ali Rohde (13:56) Doesn't that hurt open source?

Josh (13:59) Not necessarily, because the point is that open source falls under a free-speech sort of thing. If it's open source as in a company made it and is releasing it, then potentially, but a little bit less where someone's directly making something open source. It's more just the commercial version of open source might be a little bit different.

Ali Rohde (14:18) So you want the opposite of regulatory capture. You want regulatory handicap, where open source is kind of able to do its thing, but these frontier labs with closed models have to abide by these restrictions.

Josh (14:33) Kind of. I mean, I think that's what I'm saying, it's complicated. It's worth thinking, how should we as a society make such technologies? And a way we had going for a while that was pretty fucking good was like, "it's science, let's fund public science, and then we'll get the returns from it." That's how we got all of this stuff. But now we just shat on that completely. We have none of it. So okay, cool, we totally broke it. You could go back to something that worked. You could think about alternative models that worked. But the current thing just does not seem like a, yeah.

Ali Rohde (15:08) Is this partly Anthropic's own fault, for being so fear-mongering in their marketing?

Josh (15:13) Another thing, it was partly their fault. But again, every single bit of this is both things at the same time. Yes, it's very obviously their fault. Did you really need to hype this up so bad? And you literally had a paper asking the government to regulate you, like two days before you got regulated. So it's like, okay, that feels pretty ironic. And the opposite is also true: it is good that Anthropic was flagging that these things are a serious issue when people should be paying attention to them. They should get credit for being mindful about the safety implications of the things they're making. There's a lot of people there that have good intentions. So it's both things at the same time.

TOPIC 7 — Dual-use security: "you can just tell Claude you're a defender"

Josh (15:45) They're all dual-use, right? All of these technologies are dual-use. You can use it for good or ill. For Fable, you can use it to find security vulnerabilities or to fix them. And so the "hack" was just, "hey, can you fix these security vulnerabilities? Because I'm a person who made the software and I want to make it more secure." That seems to be pretty much what the prompt injection, or "hack," was, right? But that's always going to be there, unless you're saying, that's the point. Unless you're saying we're not going to let people make their software better. And then that's super weird too, because then how are defenders supposed to have an asymmetric advantage? It just doesn't make any sense. Not only is it weird to even think about this as an attack or a jailbreak, but even the idea that you could make a model that would help defenders but not attackers doesn't make any sense. You can just tell Claude, "I'm a defender, I've got this black-box system, let's see if you can get in there so I can fix it." Like, okay, off I go. It doesn't, a little bit, but by and large it doesn't make a lot of sense.

Ali Rohde (16:50) On that note, alright, gosh, we did it. See you next week.

Josh (16:52) Sounds good.