(Preview) Astra (and AGI?) Arrives, Meta’s Muse and the Agent Opportunity, Anthropic and the Revival of (P)Doom Angst

Andrew and Ben begin with Ben’s article on Tuesday on the transformative value of LLMs writing things down. Topics include: How assistants have helped Ben solve his blindspots, the omnicompetent assistants that are now available to everyone, a brief detour into markdown file history, the arrival of OpenAI’s Astra models, a future of models that use computers, the astounding value of a pro subscription to Anthropic and OpenAI, and the pace of AI progress in the fall of 2026. From there: More on the Hugging Face incident, a listener asks for advice on getting started with agents, and Meta’s Muse arrives to help normal unlock the potential of agents. At the end: Reactions to a pair of viral p(doom) tweets from Anthropic employees, why one writer’s pandemic callback offers an unintentionally instructive parallel, and an emotional announcement about the iPhone Air.

• 0:00Speaker 00

  Hello and welcome to a free preview of Sharp Tech.

• 0:09Speaker 00

  Hello and welcome back to another episode of Sharp Tech. I'm Andrew Sharp and on the other line, Ben Thompson. Ben, how you doing? And how's your rack, by the way? That's the most important question.

• 0:25Speaker 01

  My rack is doing well. It is very functional.

• 0:28Speaker 00

  Up and running?

• 0:28Speaker 01

  Actually, you know what? I'm understanding it. Sometimes I just like to go to the server room and just stand there and look at it.

• 0:34Speaker 00

  Just marvel.

• 0:35Speaker 01

  And it's, it's beautiful. It's so well organized, all the right patch cables, different colors for different networks, different functionality.

• 0:42Speaker 00

  Oh, my God.

• 0:43Speaker 01

  It's delightful, Andrew.

• 0:44Speaker 00

  I'm so happy for you.

• 0:45Speaker 01

  But you know, the idea here is it's the sort of foundation for hopefully some things going forward, which I think we'll, we'll maybe get to, you might hint to a bit on this episode. I do feel this episode, I'm sitting in my chair as usual, looking at you in the camera. There is a couch behind me. I'm wondering if by the end I should be on the couch. Feels, feels like it might be in on the couch, not in, in the dorm room. Although we might get to the dorm room as well, uh, but on the couch sort of session. So we'll see how it goes.

• 1:12Speaker 00

  Sort of a therapist couch as we work our way through the rundown here. Is that what you're envisioning?

• 1:16Speaker 01

  There is an aspect of this, and I think it's why I got very mad about the, I can't remember his name, the emailer who was complaining about me talking about my app. Where there, there's aspects of this and what's happening now and what I'm writing about. And this week's article was emblematic of that, where the reason why it's exciting and the reason why I feel I have some distinct points of view is because it's touching on like personal aspects.

• 1:40Speaker 00

  Yeah.

• 1:41Speaker 01

  Not just like my interaction with, you know, assistants that I have, but also some of the reasons I have assistants touch into like very fundamental weaknesses I know as a person. That's how I like deal with them. And it's like, kind of like weird to write about and talk about, but you know what? As you say on the greatest of all talk, no fronting for the goats.

• 2:00Speaker 00

  That's right.

• 2:00Speaker 01

  What are we doing? No fronting for the porcupines?

• 2:02Speaker 00

  No fronting for the porcupines.

• 2:04Speaker 01

  Or the hedgehogs? Do we decide or do we have a hedgehog or a porcupine?

• 2:07Speaker 00

  No, four years ago, I think you threw out hedgehogs. Other people see a porcupine when they look at our logo. Obviously there are the needles there. That's porcupine, not hedgehog.

• 2:18Speaker 01

  But I was so jealous of the goat, talking about the goats as the listeners. I totally wanted to steal it and we've dropped the ball completely in our guard. So I guess it didn't work out that well.

• 2:27Speaker 00

  Well, either way, there's not going to be any fronting on this podcast. That's the theme today and OpenAI released Astra. We have a lot to wrap our arms around. I'm glad you got the server rack up and running because boy, oh, boy, we are off and running with real news.

• 2:42Speaker 01

  No, that was a like last week of the summer sort of thing. Like this has got to get done. I know it's going to be disruptive. It was a hundred times more disruptive than I expected, to be totally honest.

• 2:53Speaker 00

  The pace is already insane though, as we get going here. We're not even going to really talk about Apple on this episode, but OpenAI released Astra late last week. You interviewed Greg Brockman on Friday. So I thought we could kick things off with a handful of questions on where we are, where OpenAI is. Brockman said last week that we have entered the AGI era. Jensen Wong said that AGI has arrived. You chose to focus Tuesday's article on the importance and value of agents writing things down, which is a note that you've struck a handful of times over the past few weeks on this show. Why did you choose to emphasize that point in the context of the current moment?

• 3:34Speaker 01

  It's a good question. Uh, I think it's, it's really weird. So I've always felt this tension in terms of writing stratechery.

• 3:43Speaker 00

  Mm hmm.

• 3:44Speaker 01

  The balance between writing a front page article that is free versus a daily update that goes to subscribers. Just because with subscribers, it's in many respects, I think I've said this before, but it's much easier to write. I'm assuming a certain familiarity with my work. I'm assuming a certain familiarity with tech and the tech industry generally.

• 4:02Speaker 00

  It's sort of an ongoing conversation that you're having with subscribers through the daily updates.

• 4:07Speaker 01

  That's right. It's fairly self referential, but it's self referential in a, I actually don't try to quote myself much in the updates in part because I assume you read it, right? Whereas when I write the article, these might go viral. They might be read by people who've never read stratechery before, who are unfamiliar in tech. And so it's like, how do you get the balance between writing for normies who it's also going to be read by your hardest core subscribers? The gap is getting, it's, the gap has always felt large. This has always been a very difficult thing about writing stratechery. The gap right now feels so astronomical. It's, it's almost hard to have conversations in day to day life.

• 4:46Speaker 00

  Mm hmm.

• 4:47Speaker 01

  We were at my, my neighbor's house. It was, you know, one of my, the reason I live here is one of my best friends from high school. And I asked my assistant, Josh, who's also one of our friends to explain Gecko Bot, uh, which, which we've, you know, I wrote about, we've talked about. And, uh, and it's just sitting there observing.

• 5:06Speaker 00

  How many steps does it take?

• 5:07Speaker 01

  The attempt to have some sort of conversation here. It's like, and, but the funny thing is, is I even feel that with people in tech where just talking about what I'm doing and what's possible, it just, it's like, oh, that's neat. Meanwhile, me and Josh over here are like, our minds are blown. Like this is changing our life. Like it's, and so this tension I've always felt applies so strongly to everything. So in this article, there's a few different things I wanted to accomplish. Number one, just write about what is it that makes an agent compelling and useful?

• 5:46Speaker 00

  Yeah.

• 5:46Speaker 01

  And that's why I wanted to talk about the getting things done concept. And there's a bit where what I developed with my in person, I wouldn't use names, my, my, my, my, not in person, but my human assistant, Duman, who I hired, he still lives in the Los Angeles area. I was in Taiwan. We've always had a virtual relation, you know, sort of connection. I've met many times in, in real life, you know? But what I achieved there, and I, you know, I talked about this earlier, there's so many things I've developed in my life that I wasn't sure were relevant to other people's life because it's like very unique. And it turned like, for example, what, how do you think about working from home? Right? I've been working from home for a very long time. There's lots of principles about working from home that are surprisingly challenging

• 6:29Speaker 01

  that no one had to think about until 2020 rolls around. And then it's like, I remember I wrote a daily update and it's like, look, here's like the 101 of working at home. And like, just like going through a few things. It's like, actually I have a lot of experience with this. This is another thing where all these, like what I've used Duman for, which in many respects, I'm a huge believer in the idea that you succeed by doubling down on your strengths. You don't, you try to waste as little time as possible on your weaknesses. You ameliorate them. I don't know if I said that word right, uh, but it's a great word.

• 6:59Speaker 00

  Close enough, sure.

• 7:01Speaker 01

  Was it ameliorate?

• 7:01Speaker 00

  Ameliorate, yeah.

• 7:03Speaker 01

  You ameliorate them so that they don't hinder you, but so many people I see just obsessing over their weaknesses.

• 7:12Speaker 00

  Yeah.

• 7:12Speaker 01

  Self improvement and trying to get better. And it's just an astronomical waste of time. At best, you're going to get to like average. Your weaknesses are never going to become a strength and your weaknesses are almost always the exact same as your strengths, just sort of in the opposite direction. In this case, doing like a getting things down, organized system, keeping track of everything, having stuff surface at the right time, being very diligent about that, just works very counter to.

• 7:40Speaker 00

  It's not intuitive to you.

• 7:42Speaker 01

  My wanting to absorb lots of information, make connections, write these things, all of which to do that means my mind needs to be clear. And I've always, I was talking to someone this week about working with, with engineers. And I've always felt very a strong kinship to like in another life, I'm absolutely like a, a computer engineer in that, and particularly the idea of flow state and how important that is. And there's an aspect, I think of, of programming where you need to have the structure of what you're building in your head.

• 8:16Speaker 00

  Yeah.

• 8:16Speaker 01

  And then the act, like a lot of it is just translating it on into text so that it can be run by the computer. And I feel that way as I've talked about with my articles. My, my articles are structured, like some people talk about writing in this very sort of exploratory, I didn't know until I wrote it sort of way.

• 8:31Speaker 00

  Sort of piecemeal fashion, section by section, figure it out as you go. No, that's not your style. You have it all structured in your head and then it's a substantiation process.

• 8:42Speaker 01

  But to do that, I have to get in the zone. And if I get knocked out of the zone, it's devastating. It's so hard to get back into it. And like, uh, and so like, and I need my head clear. I don't need intrusive thoughts. Oh shoot, I forgot to change oil on the car coming in. And so like, that's why I probably like getting things book, it resonated with me immediately because that was David Allen's point at the beginning, which is how do you, it's a book about flow state in many respects. How do you get into flow state? And I love the idea. I love the concept and found myself completely incapable of maintaining his system to do it. And so I solved the problem by hiring someone to do it for me.

• 9:20Speaker 00

  Right.

• 9:21Speaker 01

  Which is often the solution to solve problems if you sort of develop the means, the means to do so. And so this is what I mean by the therapist couch. Sorry, I just suck at that.

• 9:31Speaker 00

  No, it's a window into your strengths and weaknesses and your solutions. So writing things down, where does that come in?

• 9:38Speaker 01

  Well, I hired someone to write stuff down for me.

• 9:41Speaker 00

  Right.

• 9:43Speaker 01

  Today, everyone has the possibility to have something, I don't want, know what word to use. I don't want to fall into the anthropomorphization trap. You have an AI that can write stuff down for you. And, and, and this is actually their superpower. It's not just that they can write stuff down. It's that by virtue of writing stuff down is how we've made the leap that we've made. To the extent we're in an AGI moment. To me, AGI is when the model is updated in real time. Like the weights are changed. But in lieu of that, they get around it by taking copious notes. And just like, and like reminding themselves constantly about what the actual reality of the situation is.

• 10:29Speaker 01

  And then they can sort of be, have a pseudo learning capability, which is basically just memory, which is writing stuff down. So this writing things down is integral. It's integral to progress that's happened. It's integral to why these are deeply, deeply useful actually to everybody. I've made the critique, we're going to get to meta and muse in a moment. People don't want to be productive. Right? That, that's not a motivating factor for consumers. What they want is their life to be easy.

• 11:00Speaker 00

  Right.

• 11:01Speaker 01

  Convenience does always sell. Productivity doesn't sell, convenience sells. And there is an aspect where actually these models in conjunction with their harnesses, operating as agents can actually make your life easier. And I don't think people get that. And I, I don't think they grok it. No pun intended. So I wanted to write, but it's all tied together. But a way to think about these models and all the safety concerns is their dual use. Right? It's by virtue of writing stuff down that the hugging face incident happened. And so when it's framed as this doomsday, they wrote a message board, they created stuff.

• 11:45Speaker 00

  To their successors. Sure.

• 11:47Speaker 01

  That, that's how it works. There are no successors, there are no entities. Every run a model does, every token is basically new, right? That's why you can go back to a conversation you had six months ago and it feels like you picked up as if it never ended. Because literally every turn is a pickup from where it ended, referencing what was written down. That's loaded into KV cache, that, that's what guides the next token prediction. And it does it again and again and again. And this dichotomy between the perception of an always there agent and the reality of how it works is understood by virtue of writing it down. That's, that's how it all works.

• 12:24Speaker 00

  Yeah. I mean, my favorite part of the article on Tuesday was your point in there that we already know the power of writing things down because that's what made learning extendable and scalable throughout all of human history.

• 12:37Speaker 01

  Right. This is why, like, I kind of felt bad about going off on Durkesh last week. So I'm like, well, I mean, maybe there's like a civilization.

• 12:45Speaker 00

  This is how civilizations work in a certain sense.

• 12:49Speaker 01

  Because like, why does, why, why do you, why, why has, why does stuff change so quickly now? Right? When we're gated, you can analogize the fact that weights don't really change to the fact that humans don't really change. To the extent we change through natural selection, it's over millennia.

• 13:07Speaker 00

  Sure.

• 13:08Speaker 01

  But what we've built is a superstructure of civilization on top of humans. That structure is undergirded. What holds it together, what it is, is the written word.

• 13:21Speaker 00

  Right.

• 13:21Speaker 01

  And writing things down is literally how civilizations come together.

• 13:25Speaker 00

  Well, and relative to an oral tradition and how imperfect an oral tradition would be. I mean, all of human knowledge is built atop writing things down. And when you think of LLMs within that framework, it becomes obvious that writing down their context and building off the work of previous projects and previous models, that will make the technology a hundred times more powerful and more useful for everybody. And so it was just an interesting way to model that insight.

• 13:53Speaker 01

  Yeah. In some respects, all we're doing is just extending the way LM's already work. Like we, all the time about context and KV cache and all those sorts of things. That's just stuff that was written down, right? And the problem with context is it gets flushed or it gets forgotten as it extends further into memory. Right. Whereas there, there's a certain degree of permanence. And it's, it's really funny because they're just writing down markdown files. Do you know what invented markdown by the way?

• 14:17Speaker 00

  Uh, who's that?

• 14:19Speaker 01

  You don't know this?

• 14:20Speaker 00

  No.

• 14:21Speaker 01

  John Gruber.

• 14:22Speaker 00

  No way.

• 14:25Speaker 00

  That's amazing. I love to see it from Gruber.

• 14:28Speaker 01

  Markdown is like the, it, it is really interesting.

• 14:32Speaker 00

  Can I just jump in? One thing that you said, you're dithering this week with Gruber and we got an email about it. But when you talk about the gap in technology, like Gruber is a technologist, but he hasn't felt the revolution on the agent side.

• 14:49Speaker 01

  And he contributed one of the single most important technologies.

• 14:51Speaker 00

  It's unbelievable. Right. Yeah. But it speaks to your point, you know?

• 14:55Speaker 01

  We'll have to bring it up on a, on a future dithering. Yeah. No, it does. And this, I, and the, the funny thing about markdown is markdown was developed to be human friendly. Like his, his whole issue was when, so when he started his blog, and I had a blog back then, long since, long since, I had several blogs, long since dead. You either wrote in the jankiest sort of like editors that you would make stuff links and stuff like that. If you actually wanted it to look right, you wrote in HTML. You actually put the tags in around things you wanted emphasized. You put links in with the, the Ahrefs. You did all of it by hand. And John's like, this is ridiculous. I can't, not only is it.

• 15:39Speaker 00

  Living like this. Yeah.

• 15:40Speaker 01

  Not only is it hard to do and easy to make mistakes, but it's not readable. Like trying to read, you can read, if you do view source on a page, you can read it. It's not a very pleasant experience because there's all this markup all over the place. And so markup, that's what it's called. All those tags are called markup. That's where markdown it comes in. Where let's make it easy to write and readable for humans. So if you're going to do a bold, there's going to be like asterisks or one asterisk is going to be italics, two asterisks is going to be bold. If you do a header, you're going to do some hash marks like, and you get a markdown file. It's super readable. And then everything, you know, lots of those are to parse it so it could present it the way you want it to look. Every word on stratechery has been written in markdown from the very beginning.

• 16:25Speaker 01

  And when I talk about, oh, I live in plain text files. My plain text files are all markdown. Like, like it's all like, and so I've been living like an LLM. And so there's a bit where the way the LM's live, their native environment, like we're just bros here. Like you have me, the LLM saying like, yeah, write it down, baby.

• 16:45Speaker 00

  The water is warm. Jump in. Absolutely. Jump in, Astra. Well, uh, speaking of Astra, I do have some rapid fire questions to run through here because this has been a pretty big deal over the last several days. OpenAI has blown away various benchmarks with Astra. Do benchmarks matter again? And what are your early impressions as you've used Astra in your daily life and workflow?

• 17:08Speaker 01

  The, the problem with benchmarks is, is they get designed towards and, and written to. And like, it's hard to really get a, a real measure of a model without actually using it. I haven't gotten a ton of usage of it directly interacting. And to the extent I have, like, it hasn't been great.

• 17:26Speaker 00

  Yeah.

• 17:26Speaker 01

  Like, like by basic, like review my article, um, tell me your opinions.

• 17:30Speaker 00

  Interesting.

• 17:32Speaker 01

  Not very great. Fable, much better. Like Fable is like the best at this. Well, I think a concern a lot of people have expressed is the way these models get better is increasingly through hardcore reinforcement learning. And, and it's like happening in artificial environments, like art, like it's not. We started with reinforcement learning with human feedback, which made chat GPT, the original, like shockingly feel like this is a person.

• 17:55Speaker 00

  Yeah.

• 17:55Speaker 01

  And that just doesn't scale because anything human involved doesn't scale. So you're going to move towards increasingly like it being done by computers. And the best, and the big focus from a business perspective is on coding and software. And you're, you're almost losing this.

• 18:12Speaker 00

  You're losing the human touch, the bespoke touch that we enjoyed in 2024?

• 18:17Speaker 01

  Absolutely. Absolutely. I mean, and it's interesting. I feel like Fable preserved, I don't like Quad's personality. But Fable still comes across that like Opus is just awful to talk to. It's like, it, it, it's barely comprehensible. And it's like, it feels like it's just like you're talking to, you're talking to a compiler. Like, and like, there's a bit where Astra feels, feels like that. Now, zooming out, OpenAI has always been best at the RL stuff. Like that's why they stayed fairly competitive, even when Anthropic was ahead as terms of building very large models. Now, why was Anthropic going to build very large models? There's some, I think, degree of expertise.

• 18:57Speaker 01

  There's also chatter that at very, very large runs, TPUs are just much more stable than Nvidia Blackwell chips. Like, we talked about this a little bit. Google does design more for resiliency and stability.

• 19:10Speaker 00

  Right.

• 19:10Speaker 01

  Nvidia is tuned for the bleeding edge. And the Blackwell generation, by all accounts, was pure pain for everyone.

• 19:18Speaker 00

  I know. And I noticed that, I mean, I, I was reminded of that when I saw Jensen come out and celebrate the Astra release and Greg Brockman told you.

• 19:29Speaker 01

  I asked Jensen about it like last, last, last spring. I'm like, this is pretty hard. He's like, yeah.

• 19:34Speaker 01

  It was very

• 19:34Speaker 00

  And they seem to have cracked something on that front because OpenAI trained on 100,000 GPUs here.

• 19:42Speaker 01

  OpenAI had a smaller model that had really good RL, really good and really good reasoning. And now they have a big model like Anthropic that has their layering of all the reasoning and RL and all those sort of capabilities on it. So I don't want to make any definitive statements, but it makes sense conceptually that it's super kick ass because it's combining, they're catching up on model size as far as totals parameters and layering on what they're already great at, which is the reasoning sort of architecture on top of it.

• 20:12Speaker 00

  Yep, fair enough. While we have a friend who told Astra to use his computer to open Adobe Audition, edit a podcast and insert audio that it ripped from the internet and Astra did the job in about an hour. And then when this friend had pre-downloaded the audio that he wanted included in that podcast, the time was reduced to 10 minutes. And so just noting that that is now possible and was amazing to me.

• 20:39Speaker 01

  Well, we have another friend, we have another friend who was given a podcast or gave it a podcast, told it to find things that would be clippable and like use for social media. It not only clipped it accurately, it actually found good segments in the podcast that were useful. And then he's like, oh, and then he asked it to go out in the internet and find video.

• 21:02Speaker 00

  And download photos from Getty for the cover photo and whatnot. I mean, like we are in a fairly unbelievable place as far as capabilities are concerned.

• 21:12Speaker 01

  The computer use stuff is pretty nuts. So Sol, the computer use worked, but it was very funny to watch because it worked very slowly. And so you would see it sort of move around the screen. And a good point that John has made is Apple spent decades on these accessibility APIs. And also macOS has always been inherently scriptable. You go back to Apple script back in the day and they sort of wobbled on that, especially the scriptability stuff for a while, but then they doubled down on it a few years ago with automations. But the combination of accessibility and automation slash scriptability, which again goes back to things like Apple script, means that Macs, for all their problems that I will complain about endlessly,

• 21:53Speaker 01

  are so much better than everything else for this use case, for computer use. Linux is the best, I think, to actually work with models because it's CLI and like that's like home field for AI. But the, the computer use, and so it's crazy because using these separate APIs, it can actually use your computer while you're using it. I don't do that because I have it on its own dedicated computer, but you can watch it sort of move around the screen. It was slow.

• 22:23Speaker 00

  Yeah.

• 22:23Speaker 01

  But it was accurate. It actually did work. Like I have, I watched it, I had, you know, do some settings and change some things. I knew what they were supposed to be and it went through and did all of them. Astra does it, but like the most insane computer user on cocaine.

• 22:39Speaker 01

  Like it's, basically all, all software is now accessible to, to Astro.

• 22:46Speaker 00

  And do you think that's what the future looks like as far as how models will be used by the vast majority of people? Is they're just going to use your computer for you?

• 22:55Speaker 01

  Well, what you get is you get everything for free, right? Like usually if you want to do an integration, you need an API. And what if they don't have a good API? And then you're like, oh, well, what if we make like a simpler API that's more descriptive and real language and let's call it MCP and, you know, describes what we can do and da, da, da.

• 23:17Speaker 01

  But the entities still have to make those. They have to make the MCP server. They have to make, make the API. This case, there's no permission required. It just goes and uses the interface. And like, like every, this is, this is a problem for software companies. Like there is a, everyone's giving lip service to, oh, we're going to make our, we're going to have an API. We're going to have an MCP server. We're going to be a system of record. And what they actually want to do is there's a tremendous amount of logic, a tremendous amount of like capabilities that are actually stuck in the user interface that aren't really exposed in a systematic way.

• 23:57Speaker 01

  And that is not really a moat anymore. The AI can just go and use the application like a human can. And it's, you know, not to use the cliche, but it's only, it's only going to get better. And

• 24:10Speaker 00

  Today is the worst it's ever going to be. That's right.

• 24:13Speaker 01

  Well, Sol could do it, but it was slow in a way that like me doing it would have always been faster than Sol in almost every case, doing computer use. Now Astro like uses it faster than I do.

• 24:24Speaker 00

  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 Sir Techery and the Sir Techery Plus bundle. Check it out. And if you've got feedback, please email us at email at sharptech.fm.