Pod Insights Sharp Tech with Ben Thompson (Preview) The Futures for Cursor and Perplexity, Multi Modality Dreams and Apple Possibilities, Normies Contemplate an AI Future About this episode As the dust settles after keynote season, Ben and Andrew answer mailbag questions and talk through the future prospects for Cursor and Perplexity, multi-modality possibilities for AI devices, Apple's advantages if they deepen their partnership with OpenAI, more on Meta's investment in Scale AI, the business logic of chasing superintelligence, and takes on an AI conversation between Bill Simmons and Chuck Klosterman.
Transcript Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.
0:00 Andrew Sharp? Hello and welcome to a free preview of Sharp Tech.
0:08 Andrew Sharp? And we'll begin with this note from Sirab who writes, "Is Perplexity today's Snapchat? It's a bit of a contrarian take, but they don't own any infrastructure. Everything runs on AWS and Azure. No custom inference stack, no vertical integration. Likely skating by on credits from cloud providers looking to chip away at Google. Fine strategy for a B2B SaaS startup, but not for what perplexity wants to be. On the model side, there's little real depth. It's mostly a wrapper around third-party APIs. GPT-4, Claude, Mistral. Their so-called proprietary models are just modified Llama.
0:51 Andrew Sharp? No foundational research, no benchmarks, no publications, no standout research team. Even parts of the search rely on Bing. The business model is equally unconvincing. High compute and API costs, no platform lock-in, and no meaningful monetization path yet. With all due respect to their headline-chasing jabs at competitors, how far can an easily copied, polished UI really take them?" So Ben, what do you think of the take from Sirab there?
1:25 Ben Thompson? I don't know how contrarian this take is because it's a take that I don't really share. I mean, I think that the two sort of startups that are talked about are Cursor and Perplexity. And I'm pretty intrigued by Cursor, have been for a long time. I'm even more so after talking to Mike Truel for a Scheckery interview. And I think a big part of that is how, number one, just the space that they're in. I think it's a place that naturally lends itself to sort of aggregating the AIs. And they're doing so much on their own, their own proprietary models. And they have their own unique data source, just programmers that are using their IDE that helps them understand how to better tune models. And they have some degree of leverage over.
2:10 Ben Thompson? Over the other model providers, like particularly, I think Anthropic because you know Anthropic's coming out with Quad Code. And I think they feel a bit threatened by Cursor, but Cursor drives a massive amount of Anthropic API usage. So it's not like they're gonna, they're gonna cut them off.
2:23 Andrew Sharp? Well, for anybody who's not familiar with Cursor, can you give people the elevator pitch for Cursor? And what they want to do for the next five years here?
2:31 Ben Thompson? Yeah, so Cursor is an IDE, an integrated development environment. It's basically where programmers write programs. And, and so Cursor started out as being a way to leverage AI models while coding. And the initial here was sort of Microsoft, GitHub Copilot. And at first it's sort of like just autocomplete. Like a lot of coding is doing a lot of boilerplate. Like, like you, you know what you want to do. You just gotta type a lot of text to sort of get it out. And there's long been like debates about different coding languages. Some are more verbose than others, some are less verbose. If they're more verbose, they're sort of easier to understand and they have higher readability. If they have to do more typing and like there, there, there's what?
3:16 Ben Thompson? I mean, I'm very just scratching the surface, but there's all these sorts of trade offs. And so they started out as like, look, it's just easier to use AI to do some of this stuff in, in this development environment. And they started out doing their own, eventually they just use Microsoft VS Code. Microsoft has their own edit text editor or IDE, which is called VS Code. It's open source, so you can take it and branch it. And they, why not just use that? It's already a standard, it has millions of users. And so they, they built on it, but what they've really done is they, they've really elevated and leveraged up from where they started. So before it's just a way to easily sort of bring in these AI models into your environment. But they started realizing, well, that's kind of slow and expensive to send out calls all the time. And it's go out to the cloud, it has to come back.
4:01 Ben Thompson? And so they start doing stuff, their own models and, and like for their own autocomplete. And they start seeing how users use it and what changes they accept and which ones they don't and where it goes wrong. And so they have their own data feedback loop that they can use to train their models to make them better. And then they still call out to the big models, like if you want to write a big segment of code beyond just sort of what they can do in autocomplete. But they don't like, you don't like send out your own prompt. They'll like rewrite your prompt to do it better, to get the results that you want. And they know how to do that because all the people that are using Cursor, they have all that data and the feedback. And they sort of know how it looks.
4:43 Ben Thompson? And now they have their own models and they can plug into different ones on the backend. And you could also like plug in your own API key to use the ones in the backend. But it's just, it's gonna work better if you use through, through their service. And of course they take a margin on top of those calls and, and they, they, you know, they have a subscription. It's a SaaS offering, so you pay and you get like their autocomplete. And you also get the cloud ones up to a certain level, then you could pay for more.
5:09 Andrew Sharp? But their advantages over competitors should compound over time because more people are using it.
5:15 Ben Thompson? It should. It's sort of like search back in the day where everyone using search, Google could see how they use search, see what results they picked and circle that back in and gave them sort of a, an advantage over time. And Mike Trill specifically analogizes what he sees their advantages being to search back in the day. And what's really interesting about this is Cursor's thesis aligns with my thesis, which is I think there is this middle area between people that just dismiss AI and people that think like the world is changing tomorrow and everyone's gonna be out of a job.
5:53 Andrew Sharp? 2027, here we go. The clock is ticking.
5:56 Ben Thompson? Yeah, AI 2027 or whatever. I'm.
5:58 Andrew Sharp? I prefer to live in the middle as well, at least for my own sanity.
6:01 Ben Thompson? Yeah, it's probably just a, generally a good place to be. Just think about the 2027 when we get to 2027, if there's not much you can do about otherwise.
6:09 Andrew Sharp? We'll cross that bridge when we get there. Okay, so Cursor's in the middle with you.
6:13 Ben Thompson? And I think like, I think programming is a great example. It's incredible what AI can do and it can definitely spin up these toy apps like, wow, I did this with a prompt and it has a working app. Isn't this incredible? That falls apart very, very quickly. And now there are people like, you know, I think Anthropic is really pushing on this the most, that thinks we're going to solve all these niggling problems that arise and. Moving straight towards this agentic world where you just ask for something and it goes off and does the whole thing. Cursor sort of thesis, and this is a self-interested thesis because it's a reason why they continue to exist, even though they're not a core model builder, is actually having practitioners who actually understand how this works.
6:54 Ben Thompson? There's a distinction between building the logic and superstructure of an app in your head and actually substantiating it in code. And yes, the AI can take care of a massive amount of substantiation, but the actual understanding of how it works is still going to rest in the programmer. And over time, and by the way, the way to get to the AI doing it all is to have people doing it and getting that data feedback loop that is giving them a better product today. That's actually the route to having useful agents in the future as opposed to just trying to train a model. If you don't have the usage that they do, you're going to have a hard time, have a hard time getting there.
7:34 Ben Thompson? And then maybe in the long run, once apps are built there, as it does become more agentic, it's very interesting to think about cursor as being, I think, maybe a platform in the long run, where, you know, people assume the presence of this layer with all this logic capability in it. But that's like years down the road. And the years down the road is the point. Like their thesis depends on there being years down the road. Whereas this sort of anthropic thesis is no, actually, we're going to get there by next year. And cursor, cute story.
8:06 Andrew Sharp? Right. And if that happens.
8:08 Ben Thompson? It's a cute story.
8:09 Andrew Sharp? Cursor is new by 2027. Yes, one of many casualties in 2027 would be cursor.
8:15 Ben Thompson? Well, so the perplexity is often lumped together with cursor as this other sort of big, successful AI startup. And I'm just a lot more skeptical.
8:24 Andrew Sharp? I mean, I found this email interesting because if you go back 12 months, like I was using perplexity a lot more often 12 months ago than I am today. I think part of that shift is related to how much better ChatGPT has gotten with real time information. I would expect that improvement to continue. But then when you look at all the deficiencies that Sir Rob laid out. It's like, how many advantages are there really for perplexity? Like, what's their moat going forward in this environment? And so I wonder whether they're just gonna get acquired and that's sort of the life raft.
9:01 Ben Thompson? Well, that's the funny thing is everyone's talking about like Apple acquiring them. That's always been sort of the thing. And my question is, what are they acquiring? Right? Because the, they're not a model maker. Like, they are sort of, like, to me, they feel much more like a rapper than a cursor does. Cursor started out there, but they've really levered up into something more compelling. It is funny, I hear the most about perplexity. I've mentioned friends that I don't understand them. They have the means to pay for chat GPT plus and they don't, and they just use perplexity. Like, perplexity is deep research. Like, no, it's not as, it's, no, just use the, use the real thing, please. That seems to be their market. I'm not sure how attractive that market is.
9:43 Ben Thompson? Again, I don't, I don't know. I don't see the usage numbers. I don't, I don't know who, I guess it's part of the, I don't really know anyone that uses it much either. And it does, just in general, structurally, they're in a consumer market. And I feel like in consumer markets, integrated products tend to win. Like, like, you just have the one thing that kind of, it has the whole stack and it can subsume sort of all this sort of usage into one thing. And I think it's more likely in, I mean, this is obviously what I've been saying for a long time. So I'm, I'm, it's no surprise, chat GPT is gonna win it all in the consumer market.
10:22 Ben Thompson? Consumers will just do stuff in chat GPT. Chat GPT will do perplexity functionality faster than perplexity can do something else. And there's no real advantage that they have. Like, relative to Google, sure. The, you could see some advantages. Google doesn't want to give you the full, or they can't even give you the full AI experience in search, even just for cost reasons, to, to give that to billions of people at a time is, is sort of very difficult. And I guess from an Apple perspective, if they, you know, if they need to get out of, uh, you know, the Google deal, because the Justice Department tells them to, or whatever it might be, I guess. But like, what are you, again, what are you actually getting?
11:03 Andrew Sharp? All right, and that is the end of the free preview. If you'd like to hear more from Ben and I, there are links to subscribe in the show notes, or you can also go to sharptech.fm. Either option will get you access to a personalized feed that has all the shows we do every week, plus lots more great content from Stratechery and the Stratechery Plus bundle. Check it out, and if you've got feedback, please email us at email@sharptech.fm.
Previous Apple’s WWDC Retreat, Liquid Glass and One Question, Meta Puts $14.8 Billion Toward an AI Reset Next (Preview) Meta Continues Its AI Spending Spree, More Fun with OpenAI and Microsoft, ‘Apple in China’ and Related Matters Page updated November 3, 2026 · transcript and summary are machine-generated