Claude Code Psychosis Semianalysis Token Mogging Meta

I just have a lot of things I prepared to show and tell and that I’m going to share. Um, this is our our intern.

[laughter] Oh [laughter] 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 Semianalysis episode number eight. Uh, we’re going to talk about all of the ways in which Claude code psychosis can be used for profit today. Uh, 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 uh, new member of the team, Sam Harsh, calling in from San Francisco, just south of me. Uh, 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 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 uh, token mogging, I believe was the technical term. And then we also We also saw a nice release from Mythos from Anthropic, which uh, yeah, we can run run through in a little bit of detail. Dan, what’s new, man? You were You were just telling me you got the agents ready to go.

Yeah, yeah. We’re um, we’ve been stockpiling I think 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 two of them in the office, which we’ve been kind of um, playing around with stuff. Probably move them to proper service at some point. But it’s all about like how how we’re using Claude code, right? And um, you know, a lot of people say Claude code means you don’t need interns anymore. You know, it’s sort of like I would say the flip the script is kind of flipped, right? Let me sort of show you uh, you know, our intern. So, um, you know, like the script is kind of flipped, right? Now now the intern is the boss uh, of the agent swarm. And so this is this is Terrence. He’s a NTU. You know, there are three universities in Singapore. NTU is is a great AI program. So, he’s actually um, an AI and uh, 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 so we’ve been using. Um, so let me

While you’re doing this, while you’re doing this, while while you pull this up, that was just a a picture from We got to see the nice back of Terrence’s head and he had five different terminals up on across three different screens, I think. So, now we get a live demo. Yeah, yeah. So, this is the kind of stuff Terrence 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 got to cover a really broad industry across networking and TCO, you know, four to five switching relay companies, dozens of optical transceiver supply chain companies. And we can’t just do a full model on um, you know, each of each of these companies per quarter, right? So, you know, what this is, right? Is uh, you know, on the left you’ll see uh, this is what we call you know, Wags is a director of research doing supervisory function. And what what now we’re doing is we’re actually uh, initial 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 is read all the transcripts, process them, generate summaries. This is grab all the news digest. And then well, event engine. And then let’s see, brief agent’s metadata. And then model agent’s not here. Actually wrapping up now. But a 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, it’s so it’s actually finished. So, let’s see I can show the show you I’ll show you like a little bit of process. I said, “Okay, let’s initiate, you know, here is the the kind of steps we’re going to take, conference summaries. Here’s key takeaways from the coverage. And then here’s a 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 And 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 actually reads in all the results, all the summaries, right? And then you can actually ask it questions. It’s important to have that in the context, but not necessarily how we got to it, right? And so here, you know, you can sort of see see what they did. And I guess you know, I’ll I’ll take a step back and say like, you know, why why did we build this, right? Other than, you know, covering more ground, right? And you know, one really silly problem statement, you know, in my past like life before this, I was in investing, public markets. And you know, I built like so many of these models and covered like a lot of companies. And you know, it it seems silly in this day and age that there’s nothing to even um, you know, 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. You might as well type it in in the beginning, right? And it’s only kind of now this has 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. We want the analysts to spend more time on reviewing, editing, thinking, drawing connections. And the less of data entry, less of telling us what happened, but you know, why it happened and what it means.

And so, you know, like our our analyst team is is fixing my view. It just means we we cover more ground and cover it smarter. Make sense? Sam, what are your impressions watching this happen? Have you been using agents at this level? Um, not at this level. Um, it’s a bit bewildering to me, certainly. I was burning some tokens yesterday. It felt good to uh, break that in under the Semianalysis banner for the first time. Um, but nothing like this. Uh, 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 um, to make it all go as quickly and reliably as possible. Yes. Um, yeah, so it’s a good question, right? You know, one one one problem we’re here is So, one problem we had um, is um, like moving moving data sort of between the agents, right? And um, so what we have to do is the way the way it sort of works is cuz they can only talk to sort of the the lead agent, right? 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 ingest all this context. So, I mean, I’ll just sort of see if this What was the large uh, 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 in its context. I’m not even sure this is going to work cuz 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, uh, 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 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, you know, like just imagine how long it would take my analyst 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 analyst to, you know, go through initiate on this and come up with this answer. Is it okay? AOI landed 324 million, you know, million 1.2, likely Meta, Amazon, mostly forward-looking. You know, obviously like, you know, I’ll have to check this 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, you can see why like um, you know, what we’ve written about is we’ve written about um, you know, probably I think there was some initiation work I did in other sites. So, maybe the total cost is like 10 or 15 dollars. But it’s either this or I have an analyst spend like a day on it. Um, and so you can you can imagine if it was like 10 or 12 dollars 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. And it’s why you’re seeing that that inflection in in GPU rental price, which we talked about last time.

Dan, can you talk about how quickly you learned this? Like from going from cuz 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 toe into agents or open claw sort of stuff. Like how did 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 projects that we were on. And so we had a little bit of head space to to get into this, right? Cuz 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. Uh, and you know, I think what really helped is um, you know, this the the concept of like, you know, how you want to build like like a research team. Like what tasks are and then how you would sort of like, you know, divide So, it also started like this also started from uh, you know, if I if uh, if I were to sort of uh, let me let me just switch windows real quick. It also started from like a collection of discrete like very discrete things we wanted to do.

Uh, and I’ll show you a couple of discrete things. And And what we realized is there are a lot of discrete tasks. Like I mean there’s two two important things to realize, right? There are like a lot of discrete tasks. Um so I’ll just fire up Claud. So here I’ll show you the skills we built. There are a lot of discrete um 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, pure 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 out of things we were doing that were like really manual that that we couldn’t have even imagine, you know, 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. Um okay, you talked about the the job of an intern being different. Has your job changed meaningfully? Like you 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 more better prepared to learn how to do some of this stuff? Well, I would I would say for me a lot of this is um you know, really trying to uh I think I think all the analysts get it, right? Cuz they they’ve all done this kind of tasks and I think for me it’s it’s really trying to figure out like what is the what is the end goal, right? What do we want to get to? Like what 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? Um without that creating so much baggage that we lose sight of what’s really going on. And so making sure that that that, you know, Wags and everything it it it does things you know, just in such a streamlined way that that that it really enables that that sort of cadence. And I’ll sort of give you one more example of uh you know, another another agent family within, you know, it’s actually within Wags. Uh what do we call uh Oh, there are families now. Yeah, yeah, so so Claudia is our conferences chief. We we uh we actually we we as a company attend over 50 to 60 conferences, right? So um you know, and the thing is, right? When we attend uh 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. Uh I know Jordan you were like GTC, probably back-to-back meetings with all the Neo Clouds and I I went to like three

sections the whole time. I was just there having meetings with people in person cuz they happened to be there, too. Yeah, exactly. You know, when every time I go into a conference, kind of you remark 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 know, you can see there are 621 sessions at GTC. There is 830 presentations of which I went to exactly zero, right? You know, what what’s the problem here?

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 cuz 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 um 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 and then I might get interrupted, right? And so you know, once once we have the file users, what Claudia helps us do like, you know, presentations, YouTube links, we transcribe it, index the talkability index of each conference for uh 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, all 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 uh a couple cool things. You know, I told like, “Hey, you know, tell me what to what to do for ScalaCross.” And so so here it is. So Julian’s actually leading our work on ScalaCross. He just joined us uh last week or actually this week is his first week. And so now I’ve got a bunch of uh 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? Yes. 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 Neo Cloud, what’s a Neo Lab, 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 Claud code effectively? It would have been a qualitatively different experience. Yeah, I I was thinking about this a couple 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 Claud code. But yeah, I mean right now my first thought when I have something new to learn is just like spin up Claud and get into it that way. To the point where I built a skill for Claud 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 Claud and those those were going to be hard to surface if I wasn’t a bit more deliberate about keeping track of them. So yeah, it was completely Claud based and I don’t know how I would have done it otherwise.

So what do you guys think about the trustworthiness cuz it did cross 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. Um do you guys think there is some level at which this stuff can get better? I’m thinking about the Mythos release that Mhm. 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 Claud 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 like the model wants to build web apps for me. 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 do you think that there’s a meaningful gap in your Claud 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, you know, 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, let’s see. Screen. Yeah, so if you look at this thing here, right? You can see if 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 I’m like telling us, “Okay, man, how can you how can you input something that doesn’t balance?” So I got to go back and, you know, we got to go figure out like was it mixing up fact set, MCP? Was it mixing up SEC filings? And then, you know, we we’re building more of a supervisory function with Wags. So actually every time a model finishes initiating, it has to go through these checks. So that’s what, you know, Terrence’s job is to like kind of whip uh whip the agent, 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 that these these that the challenge is the memory because, you know, once you once once the sub agent, you know, shuts down, like that’s it. The context is gone, right? You have to memorialize everything in a 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 you know, that’s um you know, it’s it’s it’s like learning stuff. It’s learning as we go.

Yeah. Can you can you talk a little bit about like 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 kind of shocked they are that others like in their friend groups who they consider smart and, you know, plugged in are just like 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 Claud 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?

Yes. Yeah. Sam, what do you think? I think that it’s reached the point where um 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 a 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 co-work app and so on. So, I think that it’s only a matter of time before these things are not so 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 I agree. Yeah, I agree, Sam. Like, you know, I I think about it feels like everyone our circle is using it, right? Like, everyone I met in in the Bay is like using it, right? But but we we know that like for Fortune 500, you know, for all these companies like they they probably don’t even have approval like IT approval. Like, you know, I talked to a lot of 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 and improves over time. And so, I really feel like if if, you know, GPU prices are shooting up now and like, you know, you can think of everyone your your circle is 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 uses per employee how we’re completely mogging them right now. And and it’s only going to the gap’s only going to become bigger, right? But can you imagine if it was like, you know, Unilever, P&G, 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 relieved is like honestly a lot of chip fabs, right? Which, you know, you can have to bring Jeff in next next time for that. Yeah. 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 is like pretty is a pretty direct clear link. Whereas, somebody who’s six layers removed from a customer working at P&G may have a different experience of like having their contribution to um or 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. [laughter] It’s a different It’s a different guy who doesn’t go to bed. Because I work [clears throat] 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 lot of guys who don’t want to be want to be named. But yeah, like it’s certainly night owls. This is affecting sleep patterns, but it’s it’s I think it’s like it’s 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 link 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 agenetic 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 two to three hours later we’re in a working format. And uh it’s really it’s interesting to consider what improved productivity would actually do to companies that are not named SemiAnalysis. I it’s not clear to me exactly that the productivity is going to flow through directly to the financials of a 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 five times as productive and I’m still figuring out the right tool set. But yeah, something else, whether it’s apathy or foreign 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’s 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 plus 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. It’s 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 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 two 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 the GPU mode style leaderboards for kernels where AI is generating a lot of the best kernels in a way that it wasn’t even six 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. Like, it generally is a one-way street. Mhm. So, it’s it seems like the dam is breaking there, but hasn’t fully gotten through yet or 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.

The other thing that that I was working on like this week that the model seems to really just not be able to do that well is kind of the long-range tasks to consistently keep digging into documentation 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 and not dig into the details of like actually lying or the LLM to to fix something that is clearly broken, clearly have a log for. And I wonder if cybersecurity, the use cases that they’re describing is not necessarily something intrinsic about the models understand 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 a 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, 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.

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 they’re showing up on all of these these things that they did. You know, anyway, my my impression initially was like um they were 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 two or three naming like Opus, 5.4, and a 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, four 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 close.

you speak, man. We got to send you down to Menlo Park and start checking out what’s what’s going on down there. I 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, 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 that the moat it is less high and so I wonder, you know, does the moat become all 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 and I always thought like chat GPT 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 to Gemini or other things. So so I don’t know. Do you like I I I would think it’d be annoying to switch off but to switch out because you know, you’d have to get all your all your knowledge files to be like you know, retransferred to another uh a set of agents and other set of models that know how to work properly with it. Uh like I have the same thought like I like I would probably not want to go and retry everything and then have two models working, but like what what do you guys think? Do you think, you know, this becomes the way you differentiate is not just like raw models like it’s just the ecosystem around it, the incumbency and the you know, user experience and the you know, user history with it. What do you guys think?

Think like I’m I’m surprised I’m 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

the way you think? Would it work the way you expect though or would it like now you 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? Mhm. 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 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 co-work or whatever. That is motivating people to pay people to install open claw for them. Or something like line up for open claw installs and and teaching lessons, right? Cuz that that’s like the hurdle. Not

Do 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.

I know you don’t, but it’s like that’s what everyone’s doing. You know, you go to Apple store and say, “Oh, you’re doing this all this Claude code stuff.” You know, like yeah, we are. But unfortunately 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 computes in the cloud anyway when I’m using these models, so I might as well just like run it out of you know, a a notebook. Um And we got to wait 3 months waiting for the Mac Mini. I can all guarantee you that Paul does not listen to this. It’s okay. Paul’s awesome. He 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 but are called Paul. Yeah, better [ __ ] 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 Terrence’s head. Shout out Terrence. Shout out Paul. We appreciate you all guys.

next time. We’ll check out Terrence another time. We need live demo. Live demo of five screens, six or seven agents just all going. Yeah, that’d be fun. Sounds good. Go guys. Thanks for coming. Good job. Bye. All right. Thank you.