Pod Insights Sharp Tech with Ben Thompson (Preview) All About AI Companions: Personalization Upsides, Dystopian Downsides, AI Therapy, Advertising Tension, and Lots More About this episode After last week's conversation about AI for companionship, Ben and Andrew answer emails about the future of personalized LLM answers, speculative costs and benefits, the tension between building trust and building a product with 3 billion customers, and much more.
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? Obviously, the other side of the spectrum is they all just sort of converge and
0:13 Ben Thompson? Well, they all do sort of converge because they're all trained on the internet. And there's a certain sense where we're stuck with the 2010s for freaking forever. Because that's like the corpus that built all this.
0:25 Andrew Sharp? Yeah, that could be. But the ability to say the unpopular thing or give you the fact that would be verboten in polite conversation is a capability that should distinguish LLMs from other experts today and should be extremely valuable. But if it's being gated today and is going to be gated into the future, then that will be really, really frustrating. So again, hopefully there's diversity in the market and people can choose whichever they prefer. Why does Grok get worse? I've seen people complaining about Grok.
1:01 Ben Thompson? All models tend to get worse because they get optimized. And a lot of it's just like a cost issue. So they get distilled and they get various versions that run more efficiently. And you do tend to lose something along the way. Like the first time you use a model, they're probably running it very expensively and losing a bunch of money on every query. And then they get more efficient over time and they run tests to say, oh yeah, it's still getting the same sort of answers. But the universe of answers is effectively infinite. You're never going to be, you know, LLMs as they are, are a compression algorithm. Like it's compressing the internet, basically. And you further compress it, you lose bits. It's like an MB3.
1:41 Andrew Sharp? So they're optimizing for efficiency a couple of weeks after the model releases.
1:46 Ben Thompson? Yeah, continuously, continuously. Like an MB3, like 512 kilobytes per second, sounds like the real one. 256, sounds like the real one. 128, ah, my high-end speaker, this starts here at 64. Okay, this sounds like garbage. It's a similar dynamic.
2:03 Andrew Sharp? Interesting, okay. Well, the more you know. David says, and the subject line here was AI companions in advertising. Ben and Andrew, I love the show and I had some thoughts about your discussion of AIs as personal companions. Specifically, I think that having a high trust tool like that is at odds with advertising as a business model. On one level, many people get upset with seeing ads that are targeted at their personal troubles, like seeing ads for a divorce lawyer after they ask an AI for marriage advice. But on an even deeper level, the
2:37 Ben Thompson? I mean, it also could be useful.
2:41 Andrew Sharp? I don't know if I'd be mad at that.
2:45 Andrew Sharp? But on an even deeper level, the advertising incentive means that now we have to wonder if the AI actually has our best interests at heart with its answers. Am I seeing divorce lawyer ads or getting divorce advice? Because that pays better than marriage counselor ads. And while these examples may be relatively few, once you're thinking about this concern for one question, you'll be thinking about it for every other question and that erodes the trust. You need to be a truly useful companion. I don't really care if Instagram is showing me posts that advance its economic interests, but I will worry a lot if I think my AI tool is tailoring its answers toward an economic incentive. Ben, what do you think?
3:29 Ben Thompson? No, I agree. This is a massive challenge. It's part of the reason why it's not super clear what an advertising model should look like. For these sorts of things and the tension here is the reality that you only become a large consumer player via an advertising model. Because you get this dynamic of you can continually raise prices without having being concerned about price elasticity, because the users are going to use it no matter what, because it's always the same price for them. That's just how the economics work here, but pushing against it is exactly this. And you know, I think to the extent that that Sam Altman and OpenAI are authentic about not wanting advertising. This is a big part of the reason. It's not just that it's it's grubby and nasty.
4:14 Ben Thompson? It's that do we lose something ineffable about people's willingness to use this?
4:20 Ben Thompson? And it's contrasted by the fact I have a friend that has plenty of money and he's like, can you generate the Studio Ghibli picture for me? I'm like, you do it. He's like, that's $20 a month. Right? Like there's so it it so no, it's a I it's a very valid concern and point. It will be hard to manage and it's going to run up against this reality of of of advertise. There's a reason advertising ends up winning everywhere and someone's going to do advertising and that's going to be who you're competing with. And are you? Are you going to have sufficient agency to say, no, it's worth it to me to pay because that's trustworthy and aligned? It kind of ties to the Apple discussion, right?
5:04 Andrew Sharp? Hmm.
5:05 Ben Thompson? The reason people trust Apple, I would contend, is not because of their pronouncements about privacy or some commercials they run. It's they sort of implicitly get that Apple's aligned with me and my interests. They sell an expensive device. I get it. It's going to be it's going to be what I want. And I actually think Apple's been moving against this by virtue of doing things I don't necessarily want as a user and not investing in having the best possible product. Right? Like like this, this would be in a suppose ChatGPT did do advertising and Apple started training on user data, but provided an AI that came with my phone that did not have advertising.
5:48 Ben Thompson? I think they're in a great position in that regard. And it feels more privacy preserving and more respectful of me than an alternative where actually their AI just sucks, so I don't use it. And I'm like, what am I paying for when I give you $1,000 for a phone?
6:06 Andrew Sharp? Yeah. I mean, as far as OpenAI, isn't this only a challenge if they're counting on creating like an emotional bond with the users?
6:16 Ben Thompson? But it kind of feels like they're trying to do that.
6:18 Andrew Sharp? Betrayal. Yeah, well, that's true, particularly when you look at how obsequious the new model is. But I really can only imagine myself balking at ads from an AI if I'm like hopelessly down the rabbit hole, emotionally bonded with that chatbot. But beyond that, like I'm an adult.
6:38 Ben Thompson? You go back to early Google and this was a huge concern there as well. Like, people aren't going to trust the search results. And so you start out and it's super clear what's an ad. And they're like a different color and they're they're made like, you know, this is definitely an ad. And yeah, we got to make some money. Our users will get it and users did get it. Then over time, you could barely tell it's an ad anymore. And I I think that we we think about advertising.
7:05 Andrew Sharp? Too far, for sure.
7:06 Ben Thompson? It can. Well, we but we think about advertising as it exists today. And layer that onto ChatGPT. What I'm calling for is like punch the monkey ads. Like, like, let's be super clear, like, yeah, you're a free user, so there's ads there that are next to it. And, uh, and but all these concerns are real and very valid. And it's gonna be a very difficult road to walk. I just think the walking the road's inevitable, given the opportunity in front of OpenAI. But we'll see. I maybe maybe I'll be wrong.
7:36 Andrew Sharp? Okay, well, as far as monetization, I just wanted to read this tweet from a woman named Rachel Cohen on Twitter. We got heavily discounted Ubers in the 2010s. Gen Z is getting massively subsidized AI chatbots. I read that last week and I wondered, all right, are we gonna have a situation where all of society becomes dependent on AI chatbots? And it's gonna become really, really expensive early in the next decade. Do you have any take there?
8:05 Ben Thompson? It it's a good take. Uh, what's interesting about that is you could, on one hand, say, yep, and they still use Uber today, even though they're paying like 5X what they were.
8:13 Andrew Sharp? That's what I mean. As a millennial who got addicted to Uber and DoorDash and any number of other like VC funded apps in the mid 2010s. I'm still using all those apps and so they could pull that playbook again.
8:27 Ben Thompson? Right, but the but the counterpoint is you also still own a car, right? The the the talk of Uber was it's going to actually replace car ownership. Because, you know, you you're just gonna use use an Uber all the time. And it's gonna change cities and all their organized and all these sorts of things. And now those those dreams have been shifted to self-driving cars. And that's what that's how it's gonna actually pull it off. And do these but but there is a cost to cost like it's not used as much as it would be otherwise. And and I think that's just that's the advertising reality. There is price elasticity in everything. If something is more expensive, it is used less.
9:08 Ben Thompson? And in doubt, it might be so good that your most people are willing to pay it, but not everyone is. And that's that's why advertising always wins in the end. Because it's always free for the user. And yet you can still make more, you can increase your average revenue per user over time. Because you get better at targeting, you be you you you're so effective that advertisers are are willing to pay.
9:32 Andrew Sharp? So if chat GPT wants a billion users or two billion users or three billion users in the next decade. Three billion people aren't going to pay to be using chat GPT every day. And so you're trying to widen the funnel that way by keeping it.
9:45 Ben Thompson? Yeah, just your your upside with advertising is just astronomically larger than it is with subscriptions. That that's just that's because of this dynamic. And that's exactly what I was talking about before. You have this real tension between does advertising really make sense with this model? Versus advertising is the only way to become a three billion user company. That that that is growing revenue continuously over time.
10:11 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 sharp tech dot 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 at sharp tech dot FM.
Previous Apple and the Ghosts of Companies Past, Privacy at the Expense of Performance, The Cook Question Next (Preview) Meta and Its Many Plans in AI, Apple Gets Hammered in Court, The Costs of Waiting to Do the Right Thing Page updated November 3, 2026 · transcript and summary are machine-generated