The Information's TITV
00:55:06 · Akash Pasricha, Dan Shipper
[00:55:06] S0: Codex It's is having that still personal personal. [00:55:08] S2: It is... Yeah. Like, that is a personal first person narrative. Okay. But it's not creative writing in the sense that I don't necessarily look at it as something that I would wanna write a short story with. And I also don't necessarily think it is the best at nailing a particular person's style. I think Claude is actually much better at at just doing the style or voice. But in terms of how it structures and lays out a series of thoughts, it is the best. [00:55:32] S0: Okay. So... And so let's go rapid fire through the the rest of the labs here. Just give me, you know, two words here to describe the type of work that you think each lab is prioritizing for. So let's let's go with meta, actually. Consumer [00:55:50] S2: consumer shopping. Consumer shopping. Okay. Or, like, consumer interactions, you know, what... Whether that's, like, keeping up with your sports team or or finding stuff to buy or whatever, like, all the kinda, like, stuff that happens in your day to day life as a as a regular consumer. [00:56:06] S0: What about Anthropic? [00:56:08] S2: Power creative tools. Power creative tools? [00:56:12] S0: Yeah. Not power enterprise tools. [00:56:14] S2: I mean, I think that they care about that, but the the, like, the feeling that I get from Anthropic is, like, it's for it's for pros. It's for creative pros. [00:56:21] S0: Okay. Yeah. What about OpenAI? [00:56:23] S2: Knowledge work. Like like, more more specifically for the general knowledge worker that just needs to get shit done.
00:56:23 · Dan Shipper, Akash Pasricha
[00:56:23] S2: Knowledge work. Like like, more more specifically for the general knowledge worker that just needs to get shit done. [00:56:31] S0: Non coding knowledge work? [00:56:32] S2: Both. But I think a lot a lot of it is... I think the the CHECHVT harness is extremely good for noncoders. [00:56:40] S0: Grudbeck Rock. [00:56:43] S2: People who want to get conspiracy theories from their AI. Or or or... You know, actually, think there is there is something, like, really interesting, which is there are some things that none of these models will tell you except for Grok, and that is kind of important. You know? Like, for example, it's very common right now in tech. I'm not naming names, but it's very common right now to take peptides. Claude's not gonna tell you how to dose your peptides. Grok will. So do with that what you will, but I think that's actually an important thing to be able to to do if if if that's if that's your back. [00:57:19] S0: Okay. I wanna ask you a question about Grok. I don't... I didn't miss any of the main labs. Did I? [00:57:23] S2: I think you got it.
00:57:23 · Dan Shipper, Akash Pasricha
[00:57:23] S2: I think you got it. [00:57:24] S0: I I hit got most of So so let's stick with Grok here for a second, though, because Elon had an interesting tweet. He was talking about the economics of GrokBot and how they're basically losing way too much money on Grokbat as of now. And the sense I was getting is that's not gonna get a whole lot better in the near term here. I mean, it kinda leads to a question around to what extent the labs are subsidizing these models and whether or not prices will come up, whether or not you think these personal agents will be good enough for people to pay any, if not a lot of money for. What's your what's your read on this?
00:58:03 · Dan Shipper, Akash Pasricha
[00:58:03] S2: I mean, the the history of the way that model pricing has worked is that it has just come down dramatically over time. And even if the model tiers... Like, they they with the recent OPUS five five release and and the recent recent Ash releases, they did reduce prices by, I think, like, 50%. Before that, they were holding the tiers steady, but the but the amount of tokens that that were used in each tier went up and up and up and up. So what... So model so model prices have have continued to go down and usage has has continued to go up. Obviously, the labs are highly subsidizing their their their model costs. And, like, training and... And the way that we build models today is extremely inefficient, and they're constantly finding ways to make that more and more and more efficient. And I I I think that there's going to be a category of personal agent that you're not gonna wanna pay for, just like, you know, in a in a similar category to you don't wanna pay for Instagram generally. And there's a whole host of pro agents for professionals or prosumers that people are absolutely gonna pay for and already are. [00:59:14] S0: You launched your own agent, the Every agent. What does it do and what are the economics of it? Are you making money on it?
00:59:14 · Akash Pasricha, Dan Shipper
[00:59:14] S0: You launched your own agent, the Every agent. What does it do and what are the economics of it? Are you making money on it? [00:59:23] S2: Yes. So every... The Every agent is an agentic coworker that lives in Slack. It can do any task you give it. So you can delegate writing tasks, you can delegate coding tasks. It it works in the way that we work. So, you know, at Every, we have a a whole bunch of people who are constantly testing new models, trying to figure out workflows that work at the edge to push the models to their limits. The Every agent gives you all of those things in in one package that you can just add it in Slack and it helps you do work. And it also helps you spread that work to the rest of your team. One of the things that we find at Every is we have so many people in our audience who are AI pros and love using it and think they do great work with it. And it's extremely hard to get other people in their company to to work in that way. And what we found is the best way to do that is just do the work in Slack. And so if you just at the every agent and and do your work with it, you you kind of bring your coworkers along automatically without having to explain it them because you're showing it to them. [01:00:17] S0: And is it profitable?
01:00:17 · Akash Pasricha, Dan Shipper
[01:00:17] S0: And is it profitable? [01:00:19] S2: Is it profitable? So... Yes. Because the way that we the way that we charge for the Every Agent is you pay for your tokens. It would it would be it would extremely inefficient and really not possible for us to subsidize tokens to the to the extent that the labs do. Even if you raise, like, $50,000,000, you'll still run out of tokens, like, really fast if you're subsidizing. So the way that the way that the Every Agent works is we charge you for the tokens that you use, but we don't make a markup on it. And this is actually quite different from a lot of the agents on the market, and it's been, like, a really important thing for me to be, like, we don't make a markup on tokens because what that means is we're aligned with people that use the the the agent. We're aligned. We wanna make it more efficient, and we wanna find the best model to do the work at the lowest possible price for you. And we never have to think twice to be like, but, you know, we're making 30% margin on it. It'd be really nice to to, like, keep these keep these tokens flowing. [01:01:15] S0: Right. [01:01:15] S2: And instead, what we do is we charge you a subscription. So, as long as you're in every customer, you pay you pay a subscription for you or your team, and, you pay the market rate on tokens, but, that goes down and down over time.