Claude Code Psychosis: How SemiAnalysis Is Token Mogging Meta | Ep. 008

I just have a lot of things I prepared to show and tell, and then I'm gonna share. This is our intern.

Oh my goodness.

I'm like, hey, you know, I know you guys say we don't need interns, but like the script is flipped. The intern's the boss now.

Like, yeah, man, how the turns have tabled. Hello everyone. Welcome back to Semi Analysis episode number 8. We're gonna talk about all of the ways in which Claude Code psychosis can be used for profit today. With me is Mr. Daniel Nishball again, back at it, calling in from Singapore. We're going to, we're going to run through exactly how many agents the team has built at our biggest office. And a new member of the team, Sam Harsh, calling in from San Francisco, just south of me. Guys, welcome to the show.

Thank you. Excited to be here. Awesome.

Cool.

So look, it's been a big week for Claude Code. We, we've been using it a lot. We put out Some bait Dylan tweeted today about how much we've been spending on Claude code, how we're consuming more than twice the amount of tokens per employee that Meta is. Clearly have much more psychosis than those guys do. And we also token mogging, I believe was the technical term. And then we also, we also saw a nice release from Mythos from Anthropic, which yeah, we can run through in a little bit of detail. Dan, what's new, man? You were just telling me you got agents ready to go.

Yeah.

Yeah. We're, um, we've been stockpiling IT equipment at a scary pace, mostly because Ray's been scaring us into, you know, buying everything like laptops, iPhones, and Mac Minis, right?

Yeah.

So, you know, we've got, we got 2 of them in the office, which we've been kind of, um, playing around with stuff, probably move them to proper servers at some point, but it's all about like how, how we're using cloud code, right? And, um, you know, a lot of people say cloud code means You don't need interns anymore. You know, sort of like I would say the flip—the script is kind of flipped, right? Let me just sort of show you, you know, our intern. So, you know, like the script is kind of flipped, right? Now, now the intern is the boss of the agent swarm. And so this is Terence. He's at NTU. You know, there are three universities in Singapore. NTU is is a great AI program. So he's actually an AI and accounting major, right? Dual majors. His job is teaching the agents how to. How to do, um, you know, actual financial modeling. Um, so I'll kind of introduce you to a couple of the agents, uh, we've been using. Um, so let me kind of share—

For those who are just listening while, while you pull this up, that was just a, a picture from, we got to see the nice back of Terence's head and he had 5 different terminals up on across 3 different screens, I think. So now we get a live demo.

Yeah. Yeah. So this is the kind of stuff Terence does all day, right? Like he's the boss. So this is, um, you know, this is, uh, what we call WAGs. It's our Agentic Director of Research. You know, the problem statement is, you know, we gotta cover a really broad industry. across networking and TCO, 4 to 5 switching-related companies, dozens of optical transceiver supply chain companies. And we can't just do a full model on each of these companies per quarter, right? So what this is, right, is on the left, you'll see this is what we call WAGs. He's a director of research doing supervisory function. And right now what we're doing is we're actually initiating coverage in AOI, right? So the key problem is, you know, if we talk about like agents before we, you know, kind of step back and talk about why we're doing this, you know, you're, you're mainly limited by context, right? And so you see in the bars, you know, things are actually filling up already. And, you know, the problem is, you know, by the time you actually load in all the context you need to know how to grab this info, how to scrape stuff, how to do all that, you've actually really overwhelmed the context, right? Like, so this has read all the transcripts, processed them, generated summaries. This has grabbed all the news digests and then. Well, event agent, and then let's see, briefer agent's made it. And then model agent's not here. They're actually wrapping up now. But model agent, which would have been here, is responsible for actually building the model, right? But the problem is all of these fill up, fill their context up with the sort of task at hand, right? And so what we do is we structure this. Oh, so it's actually finished. So let's see, I can sort of show you. I'll show you like a little bit of the process. I said, okay, let's initiate, you know, here is the kind of steps we're going to take. Conference summaries. Here's key takeaways from the coverage. And then here's the final scorecard deliverable. I'll just ask it like, hey, Wags, oops. Hey, Wags, can you summarize all, all of the agents you used and what they did? Yeah. The key thing, right? The one, the one agent that we keep pretty much pristine is the company agent. And what the company agent does is it actually reads in all the results, all the summaries, right? And then you can actually ask it questions because it's important to have that into context, but not necessarily how we got to it. And so here, you can sort of see what they did. And I guess I'll take a step back and say like, why did we build this, right? Other than covering more ground, right? And one really silly problem statement, in my past life before this, I was in investing public markets and I've built so many of these models and covered a lot of companies. And it seems silly in this day and age that there's nothing to even, in a reliable way, Ingest financial data, right? Like by the time you ingest financial data, you actually spend more time, you know, checking it and fixing it. And you might as well have typed it in in the beginning, right? And it's only kind of now this allowed us, right? So anyway, you know, point is we're going to, we're going to have this initiate. We're going to be able to ask it questions. And we want the analysts to spend more time on reviewing, editing, thinking, drawing connections, and less on data entry, less on telling us what happened, but you know why it happened, what it means. And so, you know, like our analyst team is fixed in my view. It just means we We cover more ground and cover it smarter.

Makes sense. Sam, what are your impressions watching this happen? Have you been using agents at this level?

Not at this level. It's a bit bewildering to me, certainly. I was burning some tokens yesterday. It felt good to break that in under the Semi Analysis banner for the first time. But nothing like this. It strikes me that if you're using that much context on each agent, I don't know, maybe it'd be better to be a bit more aggressive with spinning up sub-agents. How do you think about like how you're structuring this whole thing? And passing information from node to node to make it all go as quickly and reliably as possible?

Yeah. Yeah. So that's a good question, right? You know, one problem here is, so one problem we had is like moving data sort of between the agents, right? And so what we have to do is the way it sort of works is because they can only talk to sort of the lead agent, right? And so what we have to do is We have each of them doing like a very discrete part of the process. Um, and so you can see I've woken up a company agent and company agent is actually when it, when it wakes up, right? When we finish the initiation process, it actually saves a whole index file or it saves a whole file, like of what we create. Like, you know, there's a model, there's the earnings transcripts, there's the company briefs. And then, you know, when you, when you wake up the company agent, it's, it's told to ingest all this stuff, right? It's just all this context. So, I mean, I'll just sort of see if this, what was the large, you know, optical contract that AOI won and what impact did it have on their financials? I mean, there's a lot of good products in this already. Like Quarter has a good product, but it generally only will scan, you know, it'll scan earnings transcripts and it won't necessarily have like all this other stuff in its context. I'm not even sure this is going to work because I just initiated it on it, but you know. It, it, it, again, it's broadly how do you keep the right amount of context and the right amount of information in the right, in the right agent? And, you know, like I said, we have an agent for earnings. We have an agent for events. We have an agent for just updating financials. Um, does that answer your question?

Yeah. That gives some good insight. I'll need to, uh, go to school on this these next few days, but good to see the expert at work.

Here. So actually see, see, look at this, right? So if I, uh, you know, like just imagine. How long it would take my analysts, like, you know, I just ran this, probably started 5 minutes before the podcast, right? And, you know, think of how long it would take your analysts to, you know, go through initiate on this and come up with this answer is like, okay, AOI landed 324 million, you know, million 1. T likely Meta, Amazon, mostly forward looking, you know, obviously, like, you know, off the checklist against the actual thing, right? You know, we're still developing it, but let's see how much this cost. Okay, this cost me $3.52. So So you can see why, like, you know, what we've written about is we've written about, you know, probably I think there was some initiation work I did on another side. So maybe the total cost is like $10 or $15, but it's either this or I have an analyst spend like a day on it. And so you can imagine if it was like $10 or $12 or $15 or $20, I'd still do it. Right. And that's kind of why I think, you know, the end demand, the end ROI is so strong. That's why you're seeing that inflection in GPU rental price, which we talked about last time.

Dan, can you talk about how quickly you learned this? Like, From going from, 'cause I've been using coding agents for a while and I feel like I'm almost more cautious than you or slowly dipping my toe into agents or OpenClaw sort of stuff. Like, how did you, you weren't using this even 3 months ago, right?

Yeah. Yeah. I mean, I mean, 3 months ago it was sort of like, um, you know, we, we were, we were like, we, we, we had like a Chinese New Year break, right? And, you know, we really, uh, like wound down a lot of the projects that were on. And so we had a little bit of headspace to get into this, right? Because the ironic thing is, you know, we were so busy, we didn't have time to like, you know, like learn new stuff to keep us, get us less busy. So anyway, you know, with that break, we just started playing around a bit. And, you know, I think what really helped is, you know, this, the concept of like, you know, how you want to build like a research team, like what the tasks are and then how you would sort of like, you know, divide it. So it also started like, this also started from, You know, if I were to sort of, let me just switch windows real quick. It also started from like a collection of discrete, like very discrete things we wanted it to do. And I'll show you a couple of those discrete things. And what we realized is there are a lot of discrete tasks. Like, I mean, there's 2 important things to realize, right? There are like a lot of discrete tasks. So I'll just fire up Claude. So here, I'll show you the skills we built. There are a lot of discrete things that you wanted to do. But, you know, to get it, you know, build that into something that's an ongoing, like living agent, living coverage, and not just like, you know, let's, let's build this project, let's build that project, let's build something that'll compound itself and build and actually like use all these functions, right? So all of these skills are used by, by WAGs, event summary, earnings, fin model build, press news, PR monitor is all, is all built upon. And these were all started out as very, very discrete things like Hey, can you summarize an event? Hey, can you build me a financial model? Like, can you do a press news monitor? And they all stemmed out of things we were doing that were like really manual that we couldn't have even imagined doing now, right? Like, hey, I want to know what is all the press announcements from Broadcom ahead of the OFC, for instance.

Yeah, makes sense, man. Okay, you talked about the job of an intern being different. Has your job changed meaningfully? Do you feel like you have to spearhead learning this stuff, even though in some ways maybe younger people who are going through school right now might be more AI native and better prepared to learn how to do some of this stuff?

Well, I would say for me, a lot of this is really trying to— I think all the analysts get it, right? Because they've all done this kind of tasks. And I think for me, it's really trying to figure out what is the end goal, right? Where do we want to get to? What do we need to be as a team, right? And where we need to be is we need to be in a place where, you know, we can cover like 60, 80, 100 companies, right? Without that creating so much baggage that we lose sight of what's really going on. And so making sure that, you know, WAGs and everything, it does things, you know, in such a streamlined way that it really enables that sort of cadence. And I'll sort of give you one more example of—

Mm-hmm.

You know, another agent family within, you know, it's actually within WAGs. What do we call Claudia?

Oh, they're family members now.

Yeah. Yeah. So Claudia is our conferences chief. We actually, we as a company attend over 50 to 60 conferences, right? So, you know, and the thing is, right, when we attend, do you cover now? So the thing is, right, Most of the time, we actually spend it in meetings these days, right? In targeted demos. I know, Jordan, you were like GTC, you probably had back-to-back meetings with all the NeoClouds, and I—

Oh, I went to like 3 sessions the whole time. I was just there having meetings with people in person 'cause they happened to be there too.

Yeah, exactly. Every time I go into a conference, I kind of earmark all these amazing sessions that I want to go to, and then I go to exactly zero of them, right? So, I mean, you can see, you can see there's 621 sessions at GTC. There was 830 presentations of which I went to exactly zero, right? You know, what's the problem?

Yeah, these 2 overlapped, to be clear. Like, I was at GTC while Dan was at OFC a couple of episodes ago. We talked about that. I actually like got the debrief kind of live on that podcast because like, you know, I tried to read some notes, but I hadn't talked to you about it since you'd gone.

Yeah. Yeah. And, and, and so you're, you're like kind of not exactly, you know, always like able to view every presentation and then It just becomes like an obstruction, right? Having to stop whatever you're doing and like, okay, now I got to go to the website and dig through that and everything. And then I might get interrupted, right? And so, you know, once we have the file resource, what Claudia helps us do, like, you know, presentations, YouTube links, we transcribe it, index the talks, we build an index of each conference for speakers. And then what I can do is, you know, I can say, hey, if I'm researching CPO and DWM for CPO, it's like, hey, find me, you know, every talk that discussed CPO and particular use of DWM for CPO. I'm like, well, here it is, right? And so if I want to do, you know, research project on CPO, you know, this is, this is kind of the starting point, right? Um, and I'll sort of show you, uh, let me see, I'll show you another, um, output that we did, a couple of cool things. You know, I told it like, hey, you know, tell me what to, what to do for Scale Across. And so, so here it is. So Julian's actually leading our work on Scale Across. He just joined us last week, actually this week is his first week and So now I've got a bunch of stuff for Julian to read. I'm like, go read this and tell me what it says.

So the agent is going to brief an analyst to brief you is the workflow. Yeah. I like that. I like that.

Yes. Yes. Yes.

Sam, when you hear, when you hear Dan say we got to cover 60 to 80 companies and you think about like joining SemiAnalysis and what it was like coming up to speed, trying to figure out, you know, what's a NeoCloud, what's a NeoLab, what's a tokenomics model cover, like, do you feel like you could have done this without AI or how long would it have taken to do without Claude Code effectively?

It would have been a qualitatively different experience. Yeah. I was thinking about this a couple of days ago because I think maybe there are some skills that I've left behind that it would be good to pick back up with the upgraded tool of Claude Code. But yeah, I mean, right now my first thought when I have something new to learn is just like spin up Claude and get into it that way to the point where I built a skill for Claude that added the content of my conversations with it to a flashcard app so that I could go back through all the lessons from the Neo Clouds because I realized that all of my learnings were in conversations with Claude and those, those were gonna be hard to surface if I wasn't a bit more deliberate about keeping track of them. So yeah, it was completely Claude-based and I don't know how I would've done it otherwise.

So what do you guys think about the trustworthiness? 'Cause it crossed some space for me. It was roughly November when we started pushing on, on people. Doug was definitely like spearheading that at SemiAnalysis to get people to use it. And we now kind of implicitly trust the output by default and then have to double-check things when it's potentially wrong. Whereas before, I feel like everybody implicitly like did not trust every output and had to keep prodding the model to get it to a point where it would be trustworthy. Do you guys think there is some level at which this stuff can get better? I'm thinking about the Mythos release that—

Mm-hmm.

came out today, which is to, or a couple of days ago, which is to say like, I feel like at some, in some level I am the constraint now using Claude, which is to say the instructions that I give the model are the constraint. It just like, it wants to write code. It wants to do analysis. It wants, like the model wants to build web apps for me.

Mm-hmm.

And I just need to, I just need to get out of its way so that it can build the thing that I want. And I'm like, it's, it's strange to think about how the coding models can really get much better from here in certain scenarios. Does that make sense to you guys? Like, again, do you think that there's a meaningful gap in your Claude Code experience where the model really needs to get better at a specific task that it's just really bad at right now?

Yeah. Yeah. And it's, it's almost like training a human analyst, right? Because one of the common problems that I actually see, and I was just troubleshooting this today, is when it does a model, right? And it doesn't always like balance the balance sheet. So I was kind of looking at, let me share a screen real quick on this example here. Okay. See screen. Yeah. So if you look at this thing here, right, you can see if you were to zoom in, right, you would see the balance sheet doesn't actually balance. So the asset liabilities, right? And so this is sort of thing where it'd be like telling us like, hey man, how can you, how can you input stuff that doesn't balance? So I got to go back and, you know, we got to go figure out like, Was it mixing up FactSet MCP? Was it mixing up SEC filings? And then, you know, we, we're building more of a supervisory function with WAGs that'll actually, every time a model finishes initiating, it has to go through these checks. So that's what, you know, Terence's job is to like kind of whip, uh, whip the agent and whip WAGs and to make sure that, you know, we don't see this. And it's just like teaching an analyst team really. So that's one of the things.

Yeah. But it's not necessarily just like teaching an analyst team because that type of mistake is not one that an analyst would ever make.

And, uh, that's a good point. And they don't forget the analysts like these, these, the challenges in memory because, you know, once you, once, once the subagent, you know, shuts down, like that's it, the context is gone, right? You have to memorialize everything in memory, but then you can't have like an endless memory file. So like having those learnings, right, persist over time is one of the biggest challenges actually. Like the balance sheet problem is not a new one, right? You know, mostly we've solved it, right? But, you know, we, we have to. be vigilant for these things. It may take us another couple weeks or months to iron out everything, but that's, it's like learning stuff. It's learning as we go.

Yeah. Can you talk a little bit about, based on your personal experience, what you think the impact is going to be on the broader market? Last week, the Core Research guys were commenting on how shocked they are that others in their friend groups who they consider smart and plugged in are just stuck in different jobs or different mindsets where they're not even trying this stuff right now. And they're kind of like, How, how do you get bigger adoption, more people using it? Like my sense is that everybody feels like there's going to be a lot more growth of use of Claude tokens this year, just as not only more people sign up, but also individual people start to use it more. You guys agree with that? Yeah.

Sam, what do you think?

I think that it's reached the point where I'd be comfortable introducing it to my parents, grandparents, even. I was using it for work writing software about a year ago. And as you said, it has gotten a lot better at just getting things right the first time. Like a year ago, you had to use test-driven development. You had to be super careful in managing context and scaffolding projects up step by step. But now it's like good enough that it sort of uses the terms that we use. You can just kind of give it the same instructions that you would give to an intern and it works often enough. So there's some point at which, you know, these quantitative improvements in model quality you know, piecewise improvements on benchmarks nets out into a qualitative, qualitatively different experience using the model. And also a lot of that is obviously the tooling that Anthropic has built up around it, giving a nice interface with the Claude Cowork app and so on. So I think that it's only a matter of time before these things are not some intimidating tool that the code people use and every finance bro, every lawyer is just going to be relying on it to do the bulk of their work.

Yeah. Yeah. Big changes.

Yeah, I agree. Yeah, I agree, Sam. Like, you know, I think about, it feels like everyone in our circle is using it, right? Like everyone I met in the Bay is like using it, right? But we know that like for Fortune 500, you know, for all these companies, like they probably don't even have approval, like IT approval. And, you know, I talked to a lot of fund managers and they're still like only starting to, you know, use it for like, just like, you know, summarize like really small, small time, like single task stuff. It's not really like a tool that compounds and improves over time. And so I really feel if If you know GPU prices are shooting up now, and like you know you can think of everyone in your circle as like really the tip of the spear, right? Can you imagine if everyone at Meta or everyone Amazon was using it? Like like I think you guys had a great post as well, right? Comparing the Meta users per employee, how we're completely mocking them right now, and and it's only gonna the gap's only gonna become bigger, right? But can you imagine if it was like Unilever, PNG, if it was like Boeing, all these companies, right? Like we're using it. You know, in such an embedded way, like, you know, the only way this gets to relieve is like, honestly, a lot of chip fabs, right? Which, you know, you're gonna have to bring Jeff in next, next time for that.

Yeah. Well, I think the, the other thing is that like, we, we have a relatively direct impact on the company's revenue from like the time at which we start producing better products, better research with these models that flowing through to people spending money on SemiAnalysis research and Signing up for the products, it's like pretty, it's a pretty direct, clear link. Whereas somebody who's 6 layers removed from a customer working at P&G may have a different experience of like having their contribution to, or in some ways I think there's a lot of people who, if they were to get 50% more efficient at their job, they'd just stop working 50% earlier in the day as opposed to doing 50% more. Whereas it's almost the exact opposite at Semianalysis. One of the guys on the team told me the other day that he had a 24-hour straight Claude code session. He just like did not go to bed.

Yeah, I think I know who you mean. And he doesn't go to bed anyway, Jordan.

It's a different, it's a different guy who doesn't go to bed.

Because I work a lot with that guy in my time zone. I think you know who it is.

It's a different guy.

Actually somehow. Okay. Well, it's probably true of the guy I'm thinking about as well. He's not the only one.

A couple of guys who don't want to be named, but yeah, like it's certainly night owls. This is affecting sleep patterns, but it's, I think it's like, it's been this unblock for these people who have this like pent-up dam of ideas that they want to get out into the world or into the products that we have to sell at Semianalysis, which is, you know, it's just awesome, right? It's been an awesome experience that people can go from idea to like links sent in Slack for others to click and view in the span of an hour. I've had that experience like multiple times this week, right? The eGenic coding tracing benchmark with the guys there. We had the dashboards for Accelerator and for like tracking some of the ClusterMax participants and stuff. I mean, these are things that we got the idea to do and then 2 to 3 hours later were in a working format. And it's really, it's interesting to consider what improved productivity would actually do to companies that are not named SemiAnalysis. It's not clear to me exactly that the productivity's gonna flow through directly to the financials of the company if they spend a lot more on Claude tokens.

Yeah, I agree with that. There's some funny way that the quality of the models nowadays makes me pessimistic about their adoption long-term because it's become clear That the bottleneck is not actually the quality of the output. Like for some reason other than the fact that it's not good enough, all these people at the Fortune 500 companies are not using them. It is good enough. It makes me 5 times as productive and I'm still figuring out the right toolset. Yeah. But yeah, something else, whether it's apathy or poor incentive structures or whatever, is keeping everyone else from hopping on board.

So what do you think about the Mythos announcement that had the Comments about cybersecurity issues and obviously the Nicholas Carlini, uh, YouTube video where he was kind of reviewing some of the bugs that they found in the Linux kernel. And I think the NFS driver, like stuff that's been around for 10+ years in running on production systems with just a little bit more focus and effort and a new model.

So maybe I'll fire away with some hot takes. I've been working most of the day on a note and going through the system card. It's weird to me how good it is at coding. Its all-around improvements are significantly less than its improvements in coding and cybersecurity in particular. The system card says that they did not intend for this to be the case. It sort of fell out, which is a bit odd. I'm not sure to what extent the yap about cybersecurity is a marketing ploy. If I wanted people not to use a model for malicious cybersecurity uses, I would probably not have this gigantic marketing campaign talking about how good it is at finding zero-day exploits. I was speaking to a security researcher at Stanford yesterday who said that like, you're out to lunch as a researcher if you're not worried long-term about, you know, agentic attacks from LLMs. But he's, he and all his friends are very suspicious that this actually is a step change in model quality. So we'll see. I mean, they've done a brilliant job of making us all wait with bated breath for it, at least.

Yeah, we're all excited for sure. I think there's, I can throw out 2 other examples of stuff that seems similar to cybersecurity in this case, which is like kernel authoring on GPUs. As well as like system architecture for inference. We've been working on both of those recently. There's like been some progress in GPU mode style leaderboards for kernels where AI is generating a lot of the best kernels in a way that it wasn't even 6 months ago, which seems like a notable change because generally the way this stuff works is not that AI catches up, overtakes humans, and then suddenly humans land a counterpunch and get better than AI.

Mm-hmm.

Like, It generally is a one-way street. So it seems like the dam is breaking there, but it hasn't fully gotten through yet, where the people who have the most experience with the kernels, in addition to using AI, are able to, you know, still win these competitions instead of just anybody being able to, anybody who has enough budget for tokens being able to just like spend the tokens to author these kernels. But then the other thing that I was working on like this week that the models seem to really just not be able to do that well. is kind of the long-range tasks to consistently keep digging into documentation and source code to be able to find bugs. Even with the extra high reasoning, like, effort turned on and, like, lots of prodding, a lot of these models will just kind of give up and tell you to open a PR with the maintainers of the repo and not dig into the details of, like, SG Lang or VLLM to fix something that is clearly broken. They clearly have a log for, and I wonder if cybersecurity, the use cases that they're describing, is not necessarily something intrinsic about the model's understanding, like the weights having a good encoding or understanding of how these systems work. Seems like those representations are already there. It's almost more of an ability to train the model to be persistent on long-range tasks and then give it the right harness and the right instructions and let it go. And a lot of people kind of claiming like, Well, if, if you were to describe the exploit in great detail that Mythos found, many other models can also exploit that same, that same vulnerability, right? And, um, that seems like roughly correct to me, which is to say, like, if it was explained in great detail to me and I had all the source code and I had enough time, I feel like you could just, you know, give it enough tries to do it. And so why, why wouldn't a less capable model be able to understand something that seems relatively easy to understand when it's explained to you by the model that found it?

Yeah. If that makes sense. Yeah. Interesting. Does that imply to you that like, if we start building better harnesses for other models, we should be able to do something analogous, right? If it's not some fundamental improvement in model quality per se, but it's something more like agency or persistence, we should be able to get that by other means.

It just seems like both are going to be important. In other words, the model's training includes, like the post-training includes the ability to use these harnesses effectively. pursue these tasks, get a reward signal. And the more that you train it to do something long-range coding related, the better it's going to perform at that stuff. And so it's like, you need to develop the harness in addition to developing the model. Yeah, we'll see. We'll see, man.

See.

There's lots to think about here. Do you guys have any, any hot takes on others? So obviously Meta had a, had a big release with their Avocado model kind of catching up. It feels almost like it was rushed to come out. They kind of screwed up one of the charts announcing some of their benchmark performance. And it was like, Just like, you know, a tweet, like, here you go. Right. And then suddenly it starts showing up on all of these, these things that it, you know, anyway, my, my impression initially was like, they, they were almost surprised to have a model that performs this well. And even though they're not, it's not necessarily a counterpunch on my personal usage of it so far. It's, it's not like as good as, as the top 2 or 3, naming like Opus.

Mm-hmm.

5.4 and Gemini. It does seem like it's in that ballpark now and just more RL, more data, off they go. They certainly have the compute and the team has produced this thing. So 4-horse race now.

Yeah, I think you would know better than me, but that's how it seems from here. That seems like they've done the hard part just to get to this point and that there's not a whole lot required. I mean, easy for me to say sitting on a freaking podcast, but like not, not a whole lot, all things considered to—

I want to hear you think, man. We got to send you down to Menlo Park and start checking out what's, what's going on down there.

Always happy to go to Menlo Park.

Dan, let's, let's leave it on that, this note, actually. I'm really curious, everything you shared earlier in the pod, but what would it take to use a completely separate model? What would it take for you to switch off of Claude?

Yeah, I was just going to ask you guys that because, you know, I, it gets into like, this question gets asked a lot, right? You know, philosophically, like everyone's like always catching up with each other, then Then like, does that mean the moat is less high? And so I wonder, you know, does the moat become model capabilities or it becomes like how you capture the eyeballs and, you know, how you capture the users, right? And has Claude Code kind of captured, like found the right formula, you know, and people get comfortable, you know, and I always thought like ChatGPT would be like everyone would kind of have like their $10 thing and they'd just be happy using it. But I think, you know, people have been switching out of it like to Gemini, other things. So I don't know, do you like, I would think it'd be annoying to switch off, but to switch out because, you know, you'd have to get all your knowledge files to be like, you know, retransferred to another set of agents, another set of models that know how to work properly with it. Like, I have the same thought. Like, I would probably not want to go and retry everything and then have 2 models working. But like, what do you guys think? Do you think, you know, this becomes the way you differentiate? It's not just like raw model strength. It's just the— ecosystem around it, the incumbency and the, you know, user experience and the user history with it. What do you guys think?

I think like, I'm, I'm surprised. I think at a high level, it's easier to switch models than you think in the ecosystem of like using Claude code, but pointing it at a Codex model would be like one change to a URL in your like config. But I think that doesn't work the way you think.

Would it work the way you expect though? Or would it like, now you'd, it would, it would behave a different way?

I don't know what, you gave the example of the balance sheet stuff, right? I feel like everybody using Claude Code right now has a lot of, does not have very high standards for the model that it needs to pass to be wowed by it and to keep using it. And so when the thing doesn't work with the balance sheet, you don't just throw it away and stop doing it. You just like give it another prompt and then it fixes the issue, right? And so I feel like it's It's, uh, there's a lot of inertia that would need to be overcome to actually want to, to be motivated to switch to a new model. That's my impression hearing people like you describe their initial experience using Claude Code, which is more of a change, not in the model you use, but the interface, which is to say like, it's not a web browser on a remote machine with like this chat application or interface. It's like, it's more about being able to control your computer. Like that's the change in the interface that Claude Code seems to enable or Cowork or whatever.

Yeah.

That is motivating people to pay people to install OpenClaw for them or something like that. Line up for OpenClaw installs and teaching lessons, right? Because that's like the hurdle, not—

You know what the lead time for a Mac Mini is in the US now?

No.

It's like 3 months, but you know, fortunately we're in Singapore, so we're closer. Huh?

You don't need a Mac Mini.

How do you think that is? I know you don't, but it's like, that's what everyone's doing. You know, you go to the Apple Store, it's like, oh, you're doing this. All this cloud code stuff, you know, like, yeah, we are. But fortunately it's only a month in Asia.

That's funny. I was thinking about getting a Chromebook because I mean, like, it doesn't matter whether the file system is local at all. Almost all the compute's in the cloud anyway when I'm using these models. So I might as well just like run it out of, uh, you know, a notebook. Um, and save the 3 months waiting for the Mac mini.

Don't, don't tell Paul, 'cause we've been buying like $3,000 MacBooks to log into virtual servers.

I can almost guarantee you that Paul does not listen to this.

It's okay. Paul's awesome. He lets me get the 48 gigabyte configuration. I'm not complaining.

He adds people to Slack channels at a moment's notice at all hours of the day when they are restricted. Better call Paul.

Yeah. Better shit. Better call Paul. That is, I'm going to use that. Paul's, Paul's amazing.

Let's go. Shout out Paul. If he's listening to this, we love you, Paul. Awesome. All right, guys, this has been a fun episode. Rambling a little bit. Hopefully people found it interesting. Dan, I love the demos. I love looking at the back of Terence's head. Shout out Terence. Shout out Paul. We appreciate you all.

We'll get Terence next time. We'll take out Terence another time.

We need live demo, live demo of, uh, 5 screens, 6 or 7 agents, just all going. Yeah, that'd be fun.

Sounds good.

Cool, guys. Thanks for coming. Good job.

Bye.

All right. Thank you.