EPISODE 1944

[0:00:12] GV: Hello, and welcome to SED News. This is the monthly format of SE Daily where we look at some of the tech headlines. We dive into a deeper topic in the middle and then we pull out some of our favorite Hacker News highlights at the end. As usual, we've got myself, Gregor Vand, and with me as usual is Sean Falconer. 

[0:00:34] SF: Hey, Gregor, how are you? Hey, everyone out there. 

[0:00:37] GV: Yeah, not bad. We keep using this word busy. But yeah, it has been, I think, exceptionally busy for both of us. What's been keeping you busy over the last month, Sean? 

[0:00:47] SF: I was on a work trip internationally and I also dragged my family along on that. So that was good. And also, introduced certain complexities, I guess, along the way. And then I've been back home for a bit. But then I'm leaving this weekend to join you. Well, not specifically to join you, but I will be in Singapore, and hopefully we'll meet up. 

[0:01:07] GV: Yeah, that's exciting. Yeah. So yeah, your first time to Singapore, I believe. 

[0:01:12] SF: Yeah, I'm excited about it. 

[0:01:13] GV: Yeah. No, I'm always happy when people pass through. Yeah, we will catch up. 

[0:01:17] SF: Yeah. And big news in your land or your world as well with Supabase and their Series F. 

[0:01:23] GV: Yeah, exactly. Yeah, we announced our Series F which, yeah, was 500 million at 10.5 billion post. Yeah, which was just a bit of a huge step, I guess, for us. And not something that we maybe expected quite as soon as we went from E to F, if you know what I mean. 

[0:01:41] SF: Yeah. Well, how long have you been - it hasn't been that long for you, right? 

[0:01:44] GV: No, it's only been about 9 months. 

[0:01:47] SF: Yeah. Clearly, you're the catalyst for - 

[0:01:50] GV: Yeah. Well, just saying, since I signed that contract, which was actually longer than 9 months ago, yeah, that's gone from D to F. So I brought some sort of good luck charm, maybe. 

[0:01:59] SF: Yeah. Clearly, correlation equals causation in this particular example. 

[0:02:04] GV: No. If any of my colleagues listen to this, absolutely not. I'm just a sort of cheerleader, making sure people attend the right meetings and all that kind of stuff. Yeah, it's a really exciting time. If anyone is listening and interested, Supabase is still hiring. Yeah, do check out roles there as well. But it is that kind of startup. It's just like it is a rocket ship, and you feel it when you're inside. 

[0:02:28] SF: Oh, that's fun. That's a fun place to be. 

[0:02:30] GV: Yeah, for sure. Well, yeah, let's get on to the headlines. The first one up this week is Fable and Mythos. These are the sort of complimentary models from Anthropic that I'm sure most listeners are sort of - well, Mythos probably more maybe than Fable. Mythos being the very cyber security-focused model that was only available to certain companies and even maybe some certain governments but very much restricted. And then about 2 weeks ago, Fable was released, which was their sort of consumer-grade of this. 

[0:03:10] SF: Think Mythos with extra guardrails. 

[0:03:12] GV: Correct. Yeah, exactly. Yeah, it was interesting. I was actually doing an episode with - I won't spoil it, but I'm doing an episode with someone for SE Daily that morning who's in the security space, and so he had been tinkering with it like that night. And he was saying, "Oh, yeah, Fable actually does what it kind of says in the tin," which is like it can find - really found things that I wasn't expecting, but then equally, it also has the guardrails. It really did sort of stop me going further than I - I was trying to deliberately push it to go further just to see what happens. And yes, it had the guardrails. So I was like, "Oh, wow." 

And then within about, what, 2 days, Fable had been removed from the Claude platform. And this is based on security concerns. What we're seeing here though is that we've also got the same situation kind of playing out with GPT 5.6 where they're saying that it's going to be only rolled out to certain people. And this is because, again, it's "dangerous in the wrong hands". 

And I think we're just starting to get to this slightly strange place of what is going to be the go-to-market of these supposedly super powerful models now if they can only be used by certain people. And then, e.g., a government can then say, "Hey. Well, if South Korea telecom," I think it was, "is using this thing, we've got to restrict it now," or something. It seems a very strange place to be. 

[0:04:36] SF: Yeah. I mean, I was in Europe when this came up. And of course, I got a lot of questions from customers I was interacting with and things like that about what I thought about it and opinion. And I think the big concern from a lot of countries is that, hey, if I - or a government, foreign governments, or even companies, is if I adopt this model, and this is the thing I'm investing in, and then suddenly a foreign government can just say this model is no longer available. That's a major impact to those types of investments and decisions you're making. It's kind of woken people up to that being a potential risk. 

And I think one of the things I've been thinking about is does this end up encouraging more use of openw weight models or more use of the models that are being developed within certain regions of the world? Mistral in Europe is a popular model. But does that mean - I wouldn't be surprised if Mistral got a bump after that or something like that from European contingency. And how does this kind of affect the decision that people are making around investing in these types of models? I don't think anybody necessarily has these answers, but I I think it calls into interesting questions. 

I think there's also the question of which sort of market Andreessen articulated, which is this supposedly totalitarian regime when referring to the Anthropic and OpenAI. They're trying to open up the technology. And then the supposedly democratic system, which is the US government, is trying to restrict and control the technology. And I think that makes certainly certain people uncomfortable. How much control should the government have over software like this? I don't know. It creates a lot of big questions. And I think also there's also the concern, of course, that people have around Chinese government sort of catching up in this stuff as well. And if we slow down innovation within the US, does that give essentially the model race to another country? 

[0:06:31] GV: Yeah, I think that piece specifically is interesting where - and I feel there's a few things at play here. It's like does China have an incentive at the moment to restrict models? Or now, does it kind of know that its models still have ways to go on certain areas? And so, yeah, this is kind of their moment to strike while the iron is hot, if you like, and keep them as accessible as possible so that people try them out. 

I mean, I definitely feel a lot more anecdotal evidence of people trying out open weight models now and being less bothered, concerned about the fact that a lot of them do originate from Chinese labs, etc. Yeah, we've got this strange situation of democratic countries being the ones to restrict it, but on the basis of our technology is absolutely superior. It's like so capable that if this gets in the wrong hands, then we can't have that. It's just a very strange place to be at the moment. 

[0:07:30] SF: Yeah. And I think that it's not clear to what they actually mean in terms of safety standards. I don't think a lot of this stuff has been clearly fully thought through. It's not defined somewhere. It feels a lot more ad hoc and subjective than it's like some official checkbox. Or it's not like getting your PCI compliance or something like that where there's something written down that you can actually follow to show that, "Okay, I meet the regulation." Nobody seems to know, including the administration themselves, of what this is. We talk about vibe coding. This is vibe regulations or something equivalent to that. 

[0:08:08] GV: Yeah. I think you heard it here first, vibe regulations. 

[0:08:11] SF: Yeah, there we go. 

[0:08:12] GV: Yeah. I mean, GPT 5.6, code named Sol, which is the OpenAI model that has been so-called restricted. Apparently, it introduces a max reasoning effort mode and an ultra-mode that uses coordinated sub-agents to solve highly complex tasks. That sounds pretty normal to me. But then I think, again, it's the cybersecurity angle that has kind of undone this one. 

To assuage any fears of its powerful models being unsafe, OpenAI has said that Sol includes its most robust security stack yet, and it's this hardened against adversarial attacks and intentionally optimized to favor defensive cybersecurity work as opposed to offensive exploits. 

I think this is like where they have to explain back to the government what this can and can't do. And I'm be very curious to know who in the government is, I don't know, testing this thing to its max limits on any of that rather than just saying, "Oh, this sounds dangerous. We've got to restrict it until what point in time?" Yeah, what do you classify as like safe to release again? 

[0:09:18] SF: Yeah, I feel like that's going to be a difficult conversation just based on if you ever watch any of the history of when Microsoft, or Google, or Meta has had to explain their technology to the government in official settings. It's clear that nobody understands. At least the US government understands the basics of how this technology works. Now we're talking about these deep neural networks and things. I think that's going to be a tough conversation to articulate. 

But it's interesting too to see the reaction from the two companies, Anthropic and OpenAI as well. Anthropic's quote was that they're pleased to see the progress. And maybe that's just a throwaway line or something. But OpenAI is we don't believe this kind of government access process should become the long-term default. They're kind of at least publicly on the opposite side of that. And maybe that speaks to the culture within both companies as well. 

[0:10:10] GV: Yeah, absolutely. Moving on next. Thanks to TechCrunch had a bunch of points from the Fable and Mythos regulation piece. The FT, Financial Times, had an interesting article this week on London's AI boom via the "DeepMind mafia". When we talk about mafias, it's often people that were in a certain company at a certain time, and then they sort of - in a sort of heyday, and then they disperse and go and do new things. The famous one is the PayPal mafia, which has like Peter Thiel and Elon Musk, of course. And the name is escaping me, but the person who founded a firm as well. 

[0:10:55] SF: Oh, Reid Hoffman. 

[0:10:56] GV: Not Reid Hoffman. He was also in the PayPal mafia. 

[0:10:59] SF: Yeah. I mean, there's a lot of them, right? 

[0:11:01] GV: Yeah. Obviously, read Hoffman LinkedIn. But here, we're talking about the DeepMind mafia, which is DeepMind was one of the first research labs into large language models basically. And was acquired by Google. And that really kind of - Google were kind of sitting on this thing for many, many more years than people realize basically. That's why it was always a bit surprising or strange that Gemini was kind of lagging behind people like OpenAI and Anthropic. We've talked about it many times. It's caught up in various areas, but the fact that we're literally sitting on this technology and sort of got beaten to it by these other two. 

But yeah, London has kind of become a bit of a hotbed for having where the researchers are and sort of the deep research on this. Demis Hassabis is the DeepMind founder who's kind of, I think, planted that seed. But the problem that London seems to have is just that no foundational models are actually really built there and/or run there really. And so when, say, Anthropic has to like stop a model from being used. 

I think the temperature here is that, in London, they're saying, "Well, we have all the researchers, we have all the technology. We just don't have the company or the infrastructure running it. What happens when an American company pulls the plug on things? This is a bit embarrassing. 

[0:12:21] SF: Yeah. I mean, I think if you look at it, the DeepMind alumni have raised $55 billion globally, but only 5 billion of that has stayed within the UK. And none of the alumni are building any of the stuff there. That's the big thing that the article's highlighting and highlighting some level of frustration is that the UK is creating all these brilliant people, but they're failing to capture the value. 

And there's probably a lot of things that go into that too, just where data centers, where some of the technology innovation's coming from, where the money to raise has come from. And I think that that has gotten a lot more global. There's a lot more companies that even raise from VCs in the Bay Area, but aren't necessarily headquartered in the Bay Area. It doesn't have to be there. But there is certainly advantages to being in the US, in particular in the Bay Area for technology companies. So, I can see why some of that has moved there. But there's clearly this kind of brain drain that's going on between the brilliant people coming out of DeepMind and out of these universities in the UK, and then the UK not being set up for success to capture the value for whatever that might mean. 

[0:13:33] GV: Yeah. I mean, it's a sort of well-worn path, I'd say, where a lot of amazing technology research and R&D happens, especially sort of in the Cambridge area. But yeah, unfortunately, most of these people or companies, they just "follow the money". And the money usually is in the US. 

[0:13:52] SF: Yeah. I mean, when Google was acquiring DeepMind and also all the - basically, they had all the greatest minds in AI in the world essentially working for them at one point, probably, I don't know, 10, 15 years ago. All these companies that we talk about all the times, their founders at some point dip their toe in the Google. Were Googlers at some point. It's kind of incredible that they were able to bring all those people together. But I think that talent attracts talent. And then they have basically unlimited funds to compensate those people and have them come work on those things. 

I think if you listen to - I watched a documentary on DeepMind. And really, they had this aspiration. Even though they had raised money, they didn't really have aspirations around being a commercial entity. They had this ambitious goal of building AGI or solving AGI. And they weren't really building products. They were trying to solve this problem of like AGI. It's kind of very research-driven thing. 

And it's hard to continue as a private company indefinitely with that. Part of the motivation was that they could join forces with Google. And Google founders essentially gave them the promise of like, "Hey, you can go and do this thing, and we will worry about the commercialization of this. And you can just continue to do your amazing research," all the pioneering work they did around reinforcement learning that led to AlphGo and so forth. 

They got to kind of just focus on the mission and not have to worry about the commercial stuff. And the money is always flowing in. And then on top of that, they also have suddenly access to this gigantic compute environment that Google has been building over the last 20 years. 

[0:15:28] GV: Yeah. And it's almost that environment that it probably did a lot for the R&D side, but it hampered how they actually get this out as a product because you've got this kind of specific - I read a book kind of covering all these companies. And yeah, DeepMind, they they still sat in a very separate office, and they got to do all DeepMindy things. While, literally, you have over the road the Google office who's like, "Hey. Well, we own those guys, and we should be able to use the technology. But how do we use it? Or how do we sell it?" And anyone in product management will know that that's like not how products get released fast. 

[0:16:00] SF: Yeah. Well, I think they've taken steps to unify those two teams now, the Google Brain Team and the Deep Mind team. Yeah. And certainly, I think what's happened to Google and the AI race is a deep motivator for aligning those teams and becoming more focused on getting products out than just purely the research mission. 

I think the interesting thing too is that the article highlights that, I think, UK's own tech advisors, including people from AI Labs and DeepMind had told the government that building a British LLM would be a waste of taxpayer money. But now, given all the things we were talking about at the beginning in terms of the US cutting off access to certain models, it's now reinvigorated essentially this attempt that there should be a UK AI model that they own so that they don't have to be subject to the whims of essentially a foreign government. 

[0:16:52] GV: Yeah. I find that sort of interesting because it's not like there's a US government LLM that was funded by taxpayer's money. But I think it's just a different way of looking. The US is just this amazing sort of capitalist machine, right? And the UK - 

[0:17:09] SF: Maybe the UK is thinking that the key is to move fast with products. You want the government to be involved. That really is, if you want to move quickly, involve the public. 

[0:17:18] GV: Yeah, definitely heard that one before. Yeah. Really, I mean, in my humble opinion, having grown up in the UK, it's just that the UK just needs to - if they were serious about something like this, they have to just get more risk-averse. People have to be more okay with venture capital, etc., etc. That's a pretty well-worn understanding, I think. But I think sort of a British LLM from the government, I'm not sure that's like that's the answer, but I can see why they're at least just talking about it. 

Yeah. I think going on to our next news item. This is back to a TechCrunch. Yeah, there was another YC-backed company where source code stealing was sort of a claim that has been made. This is something that's kind of happened before. The company today is - and this one is Corgi. Corgi is a deal flow kind of backend platform storing all sorts of data for VCs and this kind of thing. But they got accused of replicating another platform called Papermark. 

And it was kind of interesting because the post that sort of highlighted this to the world, it blew up because I think one of the founders, he shared a screenshot. I think this is the Papermark founder. Shared a screenshot, and he showed Corgi's product side-by-side with Papermark's, and it was virtually the same UI next to each other but in a very specific place. It wasn't like the front page. It wasn't like the main dashboard or something, but it was like in these certain areas where the language is very specific. 

And so, really, the arguments come down more to it's not code copying because it was vibe coded. And so Corgi is arguing, "Well, we just told it to build the thing." But they have admitted that they did tell it to basically replicate Paperwork, which just raises this whole new concept of what is copyright infringement exactly if it's not exactly the code but it's the idea. I mean, way back in the day, I studied IPO law. I need to dredge up some of that knowledge. But I think it's just a very interesting concept. 

[0:19:28] SF: Yeah, I love the we didn't steal the code. The AI happened to produce something identical. But yeah, I mean, if I went and said to an LLM, "I want to write a book. I want you to copy Harry Potter, but change the names and the setting to some degree." In my mind, I'm clearly breaking some sort of law, right? It's a direct sort of copy. 

I guess the difference is that the LLM can actually inspect the writing of the book. Whereas, the LLM hasn't necessarily inspected the code backend of the service. It's just copying to the best of its ability the frontend and then assuming certain things in the backend would need to be built a certain way. But it's a complex - super complicated. 

It's not like people haven't been copying what their competitors or similar products have been doing forever. It's just it takes more effort to do it than - and with the effort, people probably also put some thought into. Maybe we shouldn't copy this like exactly pixel for pixel or word. And they did a better job of hiding the fact that they were copying it. 

[0:20:34] GV: Yeah, I think the effort thing is interesting because, yeah, intellectual property frameworks, yeah, they were sort of designed for a world where we're copying required effort. And then if you look at, yeah, the incentives for someone - let's just take the obvious one like Instagram copying stories off Snap. It's not that they sort of looked at that and said it is really worth our time putting a bunch of engineers and literally taking that concept and basically planting it in our platform. And we believe deeply in that. And so, they must have also been ready to defend that. 

But if that effort is removed, then the sort of incentive to copy is just like so high. Because it's like well we might as, well, copy it. See if it works. If it doesn't work, we can like trash it. Go spin up a vaguely similar but different version. But the kind of are we really going to do this? Are we really going to put the engineers on this that's going to take a few months to crank this out? It just changes the whole dynamic of anyone thinking about replicating somebody else's product. 

[0:21:35] SF: Yeah. And clearly, it takes a lot more than just copying the code or copying the interface to have a successful product. But I do think it kind of calls into question of like what is the moat for companies now? Traditionally - or I guess another way to think of it is like if AI can basically reproduce the look and feel the functionality of pretty much any SaaS product in an afternoon without even touching the source code of the original product, what exactly is the moat of the software company? 

And it has to be either some access to proprietary data that the competitors don't have access to. And then I think another big one is certainly around go-to-market. And I think if you look at all the money that Corgi has raised in succession, maybe that's kind of what they're thinking, is like, "We can essentially grow really fast, corner this market, put the money behind the go-to-market." But it's not about the money necessarily going into building the software. The software we can build relatively cheaply other than the token costs. But we need essentially to go out and land grab all the potential users, and that takes money. 

[0:22:44] GV: Yeah. And speaking of money, I think the fundraising story here also plays a part where Corgi raised 108 million series A and then 160 series B. And then I love these extra rounds. A series B1, another 100 million. That's a lot of cash. There's 360-odd mill - 

[0:23:07] SF: It's a lot of tokens. 

[0:23:08] GV: Yeah, it's a lot of tokens. 370-odd million. And then they come out with this sort of vibe-coded clone product. You sort of have to ask like is the fundraising environment hotter than the product market right now? I would argue it is. And the pressure, I guess, is, "Hey, I have not gone through YC." But, hey, imagine I've gone through YC, and then I'm thrown up to 300 million. Of course, I feel the pressure to get a product out that's like why shouldn't it be the same or better than my competitor? It just is a very strangely "lazy way to do it", I think. 

[0:23:42] SF: Yeah. I mean, I think it's a confusing time, I think, for founders building companies. It's like how do you think about what is your unique value? And how do you actually build a company in this market if it's not building a model or something like that? Where do you really create value that isn't easy to copy? 

[0:24:00] GV: Yeah. And then just to sort of round this one out, I mean this is in contrast to the last sort of case of this was PearAI, which I think it was almost 2 years ago at this point. But yeah, they basically just ripped out an open source repo and then just passed it off as their product. And also, they got into even sort of murkier water because I think they started lying about it inside YC. They sort of got asked, "Hey, just tell us what this is." And then they actually still went deeper on the lie before they sort of came clean. 

And Gary Tan had really pumped them up online for like a few weeks, and then was like, "Oh, actually, sorry, that was true. I've got to apologize," and say we've disconnected from PearAI. Really messy, that one. 

[0:24:47] SF: It's probably hard as an investor too right now to kind of navigate such an influx of companies. How do you kind of differentiate between what's real and what's not real? I guess that's always been something that investors have to navigate, but it's probably particularly true. Because I think some of the things that they might have used specially to evaluate early stage teams are not necessarily the case now. You can create a really compelling early stage product rather quickly at this point. And you probably have to look a lot deeper than just what the product experience is and things like that to sort of evaluate the opportunity. 

[0:25:22] GV: Yeah. For one of the first times in history, I would probably not want to be a VC right now. I don't know. Yeah. having to pick through all these products and try and figure out how is it actually made. And what's the real "USP" here that isn't vibe-coded by someone else, or can be vibe coded by someone else tomorrow? Very challenging. 

[0:25:42] SF: Mm-hmm. 

[0:25:43] GV: One of our final headlines this week, really just about SpaceX. But lot has been talked about SpaceX elsewhere more. We're kind of more interested in their acquisitions. Some may have seen already Cursor. They're probably some people's favorite coding tool. Was actually purchased by SpaceX. We're going to get onto that in the main topic, so we're not going to touch on that one too deeply right now. 

But SpaceX also acquired a company called Mesh. This is more of a sort of hardware tech company that builds optical links on the ground as opposed to - this was ex-SpaceX engineers who built optical links that connected Starlink satellites. But they think that they've seen that the same technology can speed up data centers on the ground. 

Yeah, with this sort of influx of cash from their IPO, I mean, two big acquisitions. I mean, if I just look at the Cursor amount of, I think, it was 60 billion. Yeah, I'll just call it. I think it was way overvalued, but that's just me. I have to assume that maybe this was maybe also slightly overvalued in its purchase as well. So, just kind of throwing money at it when they don't have to think too hard about money. 

[0:26:57] SF: I mean, that might be some of it. But I also think that if you kind of look at the potential of this, SpaceX is already selling compute to places like Anthropic, Google and so on. And then they can own the actual network layer amongst this is kind of where Nvidia is very much owning sort of the GPU and the chips. If you can own the network layer, which is another huge area of cost for these data centers, then that could be a gigantic total addressable market. And if these data centers continue to grow and people are investing in them, you're going to need ways of transferring the data. And you can own essentially that layer that's responsible for transferring data, that's massive. 

[0:27:36] GV: Oh yeah. I mean, I don't sort of disagree dispute why they would buy a company like this. For sure. Yeah. I mean, actually, Cursor is more the one we'll get on to is sort of the whys there, right? But yeah, this one, Mesh, certainly sounds like a "sensible play" from the sort of the why. I can't see if I can find how much they actually paid for Mesh. Need to come back on that one. But it still looks like a slightly more strategic purchase than Cursor. 

We're going to get on to Cursor and just dev tooling in general in the main topic in a second. Our final news item, it did hit the main news, was just Anthropic launching another product, which sort of starts to cut into specialized products. It's called Claude Science. And this is a product that enables rich scientific artifacts to be sort of fully reproduced and visualized. 

Claude Science says generates figures and manuscripts alongside the code that created them. It can natively render rich scientific artifacts, including 3D protein structures, genome browser tracks, chemical structures, and more. And yeah, this is just - I guess we've seen this a few times now where Anthropic cut into a specialized - 

[0:28:45] SF: Finance and legal. 

[0:28:47] GV: Exactly. We saw the finance. Exactly. Yeah. I got to say, I'm not super familiar with the sort of scientific software space, but I can imagine it's maybe a little bit stayed in places. And so, suddenly, when you've got someone like Anthropic coming in and saying, "Hey, we can just bring our might into this area," that sounds maybe really helpful. But equally, if you're a specialized scientific software provider, this could look a little bit concerning as well." 

[0:29:15] SF: Yeah, potentially. I do think that at least some of these launches that Anthropic has done into specific verticals, people have kind of overindexed on, "Oh, my god, this is going to like change the entire industry." It really kind of reminds me of like the early days of Google when everybody was like - every time Google moved to do anything with some random 20% project, everybody's like, "I guess all those companies are dead. Here comes Google." 

But I think part of it is where are their focus areas? And also, where are their expertise? Clearly, Claude Code is so amazing because they own the model, and then they also have a ton of people who are actively using Claude Code to both develop the Anthropic technology and also feedback into the software that they use to build the technology. It's this huge massive feedback loop. 

I think I doubt that they have that when it comes to some of these other domains. And maybe they get there, but I'm a little bit skeptic. I just haven't seen the - you hear the launch, people freak out about it. And you don't hear anything about these particular tools after that. 

[0:30:19] GV: Yeah. Yeah, that's a very good point. I mean, I haven't been tracking, say, the finance suite that Anthropic came out with. I think when you move maybe slightly outside of having been - I went back to builder mode before I joined Supabase. I was very alert to what these companies were doing and how the products were actually netting out, because it always affected you if you sort of felt like you're building something that Anthropic or OpenAI were going to come and sort of annihilate. But yeah, we'll sort of follow along. That might be another sort of deeper topic that we can do in the future actually. 

Speaking of main topic, we've actually covered a lot of headlines today. This might be a little bit shorter than usual. But we just want to look at the - if you want to call it the IDE Wars, but round two. Because last time we we looked at this, it was kind of interesting. We're only talking maybe like a year apart. But the IDE wars round one was Cursor versus Windsurf. And in both cases, it was like, well, they're both kind of derived off VS Code. And then where do they go with this? And what's the tools that will sort of "win" the day? 

And we just are in such a different place now. Cursor has, as we mentioned, just been purchased by SpaceX. I think really what we're focusing on here is just the fact that who actually owns your dev tool chain now? It's this idea of, well, depending on kind of your, if you want to call, top of funnel for where you write your code. You're also buying into effective an ecosystem quite often run by that company. 

The obvious cases are Claude Code from Anthropic and Codex from OpenAI. You use those tools where you're basically buying into the models that they themselves run and sell you and are making a lot of money from. And so you're kind of like what is your incentive to break away if you're sort of seeing success from building with those tools? You're less likely to want to break out of that because you're like, "Well, this is working so well. I don't know how well the other one will work," and so on. 

And then Cursor, what's the strategy there, I guess? Well, SpaceX has xAI, its own. And Grok - no, sorry, Grok. There's two Groks. That's why I always get confused. There's one Groc who got bought by OpenAI, which was more infrastructure. And then there's the other Grok with a K, which is within SpaceX. That's more the chat version, I guess. But again, you're buying into that ecosystem now with - or it's assumed with Cursor that if that's where you're going to be writing your code, you're probably going to be signing through to SpaceX models. 

Then there's the other option, which is you go fully open source. And OpenCode has actually been really ramping up. It's apparently 160,000 GitHub stars, 7.5 million monthly active developers. And the point is it offers it's model agnostic. It gives you access to 75 providers. So you're not in theory sort of "locked-in" if you're using that sort of tool chain. 

Yeah, we're just kind of back to this situation where you're sort of almost working in like a closed IDE environment versus open. I thought we'd like left that town a while ago, but here we are. We're we're back again. 

[0:33:40] SF: Yeah. I mean, there was a time. Now it's like what you were saying in terms of with certain choices, you're buying into the whole ecosystem and kind of being vendor locked-in. It's like, Okay, I'm going to use Claude Code, which means I'm going to use Claude as my model as well." And then I'm going to essentially have my contextual data associated with that particular development experience. 

And then it was similar at one point in terms of languages and IDE choices as well. It's like I'm going to use Microsoft's version of C++, which means I have Visual Studio. And I kind of locked into that ecosystem. Or even a bigger example would be something like the early days of C#. The only IDE I could write C# in and the only compiler available was also from Microsoft. And it happens to be code that's most easily run on a Microsoft Windows machine. So kind of like you're buying into this whole ecosystem. 

And over time, that has changed, where most of the languages - basically, every language became open source. A lot of the IDEs, if they're not free and open source, there is a free and open source version of that. And it became no one, I think, feels particularly vendor locked in. I wonder if this is sort of the common trend and wave that you have with any software innovation is that, in the beginning, you have these kind of like locked-in experiences because maybe there's an advantage early to being able to build those. And it's harder to build sort of the separated decoupled version of that. But that happens over time. You mentioned the open-source tools while - 

[0:35:11] GV: OpenCode, yeah. 

[0:35:12] SF: OpenCode. Yeah. I feel like every four to six months there's like a new hot open-source coding project that gets all kinds of GitHub stars and grows really quickly. And then there's like a new hotness. I would assume that eventually some of that stabilizes as well. 

[0:35:28] GV: Yeah, I think it probably will. I think developers are still open to experimenting with these different toolchains. And as we covered, I think, on last SED News, just the cost. Okay, it's making lots of money for Anthropic and I think somewhat open AI, but certainly Anthropic are just killing it right now on Claude Code and the amount of money that companies are spending through that. 

But I think what we were saying last month was what is the tipping point for companies actually saying, "Okay, we've got a reign in spending here, or you should go and find an alternative tool even if it's a bit slower or "the quality is a bit less". But we can't sort of continue on this sort of just always going to be using the hottest, latest model that's also the most expensive." 

The question with OpenCode and kind of the model choices is there was some head-to-head comparison on builder.io. And OpenCode was apparently 78% slower than Claude Code. I think we've got to kind of look - that sounds pretty drastic, right? But I think if we - 

[0:36:36] SF: Claude Code is not exactly fast. 

[0:36:39] SF: Yeah. I can see where developers, if they're sitting, waiting for this thing, OpenCode to chunk through with a model of choice. And they're just like, "Why am I doing this? Why am I not just using Claude Code? Is it just because my employer doesn't want to pay for it?" Yeah. 

[0:36:53] SF: Yeah. I would say maybe 78% slower is not ideal. But some of the efficiency or latency issues is a little bit hidden because these experiences are async. A lot of people that I know that are fairly advanced at using the agentic engineering tools, they're running four to six agents simultaneously. They're giving direction and then skipping over to some other tab or experience and giving direction there. And it's a lot of sort of babysitting these different agents. 

It's not like you have to just say give a prompt and then go for a walk for 30 minutes like the old days of compiling large pieces of source code, and then come back. And it's like, "Okay. Well, now I put in my next response." I think you can actually you do something. And even if it turns away, takes some time to compute, you're doing other things in the meantime. There's other things that you can be doing. It doesn't require 100% of your focus. I think that is a big change also in terms of the developer experience is, classically, it's all about getting in the zone, not being distracted. And now it's a lot of context switching in flipping between tabs in your terminal or different experiences within your IDE or something like that. I think that changes things as well, but not to say that you should probably aim to be a little bit faster. But there's less concern there. 

But I think one of the things that's interesting about all this though is that so much of the value that you're creating when you're interacting with these tools is sort of the data and the context with which you're building around it. That becomes somewhat like the vendor lock-in even more than the model and the experience itself. It's like if I needed this, what is the switching cost? If I want to go from Claude Code, to Antigravity, or OpenCode, do I suddenly lose all this rich history and context that I've been building up? This has become my buddy. How do I manage? 

And that's something that has always been an issue in all software. It's like okay. That's why databases are so sticky, because the migration cost of database is really expensive. But it's something that all companies ideally want to, I think, have the flexibility between vendors. And I wonder if there's going to be a market for creating an abstraction layer essentially that manages all this context and data that gives you a vendor agnostic way of delivering that context across these different surface areas. 

[0:39:19] GV: Yeah, absolutely. I mean, yeah, I mean, I use Claude personally. And we have a work subscription. But I've obviously got less - I've got more choice on the personal side in that sense. But still, the fact that it has memory. And yeah, I do kind of - I'm just going to keep using Claude because yeah, well, hey, it's got all my chats and it can reference things back and so on. And that definitely comes through on the code side as well. 

I think the question also here is just the neutrality piece with Cursor. They were always sort of seen as this neutral company, quite frankly, because they were like VS Code, but we're a separate company. We've done a whole bunch of changes to the underlying VS Code open source. And then you can choose your models that you work from. And I think the big question now is, A, will that continue? Or is Cursor auto mode, is that now always going to go through Grok? And is it always going to be on SpaceX owned infrastructure? 

I can also just correct myself, because Grok with a K is SpaceX. Groq with a Q, which was the other one I was referring to, it was actually Nvidia. They ended up partnering with Nvidia on infra side. Not OpenAI. But yeah, Grok with a K, which is within SpaceX. Yeah. It just means that if you're now using Cursor, what does that look like? And quite frankly, I used to use cursor less so over time for various reasons. This is just like another reason I think I would start to think twice about using Cursor, unfortunately. 

[0:40:50] SF: Yeah. Clearly, there's also a market for companies that don't want to send anything potentially over public internet or to cloud vendors as well where they're going to be in an air-gapped environment or they run on-prem banks and so forth. I talked to lots of customers in our world that they're running their own LLMs. And then they need a tool, an engineering tool that's going to work in that environment as well. Clearly, they got a market there. So it doesn't necessarily need to be like one winner takes all. 

[0:41:22] GV: Yeah. And, again, I think we should do a deeper topic on SED News. But yeah, open weight models. They are just becoming dramatically cheaper. I mean, obviously, you can run them in different places. But roughly speaking, saying that DeepSeek V4 Pro had come down to - is it 44 cents per million tokens? Whereas Claude Opus 4.7 is, for the same thing, $5. There's an 8x difference to using Claude versus DeepSeek. 

And again, anecdotally, I think developers are saying, "Well, we're actually starting to see some really good results coming out of these open weight models." It's definitely something we have to watch very closely. Because if there's a tipping point where whoever's running these for a developer, that that tool chain becomes, say, "as good as like Claude Code" And you can convince your, say, company there's no challenges here. It's just eight times cheaper. That's pretty huge. 

[0:42:28] SF: Mm-hhm. 

[0:42:28] GV: Yeah. I mean, I think it's interesting to have looked at this in terms of - yeah, we're curious what are you guys using out there? Audience always like - if you've got very strong opinions on this, we always love to hear them. We're always just trying to sort of report back what we're seeing day-to-day. And if you'd said a year ago, basically Cursor and Windsurf would almost be sort of "taken off the market", I wouldn't have believed you. Yeah, I definitely didn't predict Claude Code and Codex, but Claude especially, sort of taking over the IDE space and that sort of CLI-first for all builders. That really just seems to be winning at the moment. Very interesting to see. But yeah, let's get on to our favorite part of the show, Hacker News highlights. Do you want to kick us off, Sean? 

[0:43:16] SF: Yeah, sure. So, I found this article that was about HackerRank had open-sourced ATS, or Application Tracking System. And this guy did sort of a deep dive. Sam Bell, I think is his name. Deep dive into using the applicant tracking system to analyze its LLM capabilities of analyzing his own resume. He had scored 98 out of 100, but he basically ends up stress testing the resume screener running the exact same test over and over again like 100 times. And he had scores that range from like 66 to 99. 

And then the article is really interesting because he breaks it down. Where was the LLM sort of consistent? And where was it all over the place? Things where the LLM did a good job was you know a checklist of do you know Python? That is pretty easy. Or categories requiring judgment though, something like are these projects architecturally complex, which a company might care about? They're completely subjective and basically the equivalent of a coin flip with the LLM. 

The takeaway here, of course, is LLMs are great at parsing. They're great at private matching. They're terrible at subjective evaluation, which you shouldn't essentially ask them to do, especially if you're depending on getting some sort of score associated with it. It's just not going to be a dependable metric. 

[0:44:37] GV: Yeah. Super interesting. Yeah. I wonder what would happen if I put my resume through that. 

[0:44:43] SF: Yeah. Well, it depends. You might get anywhere from 66 to 99. 

[0:44:47] GV: Yeah. Exactly. On my side, we have - yeah, it's a partial port of Kubernetes to the browser, which is kind of interesting. This is actually by - well, it's another Sam actually, but Ngrok. Ngrok have kind of released this. And what they stress is this is not full somehow g-zipped Kubernetes running in the browser. But it is a partial port of Kubernetes, a kubelet binary. It's enough to run pods and problem, basically. 

It's got ports of several Kubernetes controllers. It's got a pod scheduler, it's got namespace controller, kube-proxy, deployment controller, and a bit more. And then it's got a browser-based take on the CNI, the Container Network Interface, so pods can talk to each other. And then it's got like a browser-based container runtime, which the kubelet can talk to over the container runtime interface. 

Yeah. Sorry. And just to round out an API for interacting with the webrenetes, which is what they're calling it, webernetes cluster, to do things like apply manifests and watch resources. Super interesting. I think the sort of angle here is a little bit more on like education, for example. It's quite difficult to learn Kubernetes. And so if you can do a lot of understanding how things operate and interact with each other in this sort of lighter, more interactive environment, I think that's part of the thinking here. But yeah, super cool. And the fact it just comes via Ngrok, you can kind of assume it's actually pretty well put together. 

[0:46:19] SF: Yeah, that's fun. 

[0:46:20] GV: I just have to highlight one comment that came through on that, which was I think the name is Duncan GH. And they commented investing early in this HN post before it's a banger, instant classic. I just love that, yeah, this is like, as this person says, a classic HN article. It's definitely why it made my highlights this week. 

[0:46:43] SF: Yeah, Hacker News was a port of a difficult technology to some other form factor. Just like the myriad of Doom ports that we've covered. 

[0:46:52] GV: Yeah, exactly. 

[0:46:54] SF: Yeah. And the last thing I had was this article; Token maxing is dead. Long-live token maxing. 

[0:46:59] GV: I saw that. Yeah. Yeah. Yeah. 

[0:47:00] SF: Yeah. It's interesting. Basically, what the writer is saying is that, originally, people were token maxing because we were encouraging developers to consume huge amounts of tokens. We're creating KPIs within companies for them. A lot of it was like wasteful use of tokens. 

And essentially, what the author is saying is that now we're maxing out tokens not because of wasteful token maxing, but because of actually something they term compound correctness. As these coding agents improved, we can actually spend more tokens increasing the likelihood of better outcomes. If you spend sort of more time on compute, potentially even spend more money on expensive models, then it might cost you more over, say, a 5-minute period versus a cheaper model. But with a cheaper model, you're going to have to do more. You're basically wasting more tokens because you're doing more throwaway work. You got to do more iterations, whereas you're able to get it right sort of the first time is the argument. 

[0:47:56] GV: Yeah. The whole one shot still sort of holding up is more of like that's where to aim versus just rinsing tokens on cheap models basically. 

[0:48:05] SF: Mm-hmm. Yeah. 

[0:48:06] GV: Yeah. My final one - there's a lot of Sams today. Am I reading this correctly? Yeah. Literally. The Ngrok one I talked about was by a guy, Sam Rose. This is now Sam Wilkinson. Yeah. Okay. Three Sams. Great. This was RF hacking my cloud-controlled ceiling fan. 

It's quite a practical application. This isn't just a sort of like off the deep end. This was someone who had moved into, I think, a new house, and there was a rickety old ceiling fan on there. And they were like, "I'm just going to put a new one on." But as with most sort of smart devices these days, it has to go through a cloud service to run if you want to use any of the smart home features. 

Yeah. I mean, I won't go into all the details partly because I'm definitely no RF hacking expert, but you can be sure that this article has an amazing array of how they decoded the RF signals and figured out how to then set up a transmitter box that could then control this specific fan. 

And I definitely can relate to this. I have various ceiling fans where I live in Singapore. And if I could get them hooked up without a cloud connection, that would be kind of nice. Unfortunately, I don't plan to go through all the steps that Sam Wilkinson did. But yeah, if you are so way inclined, then that's, yeah, samwilkinson.io. Yeah, thanks for posting that. 

[0:49:34] SF: Yeah very cool. I wonder with all these tools that we're talking about, Cursor, Claude Code, everything, if there'll be - it kind of opens up access for a lot of these side hacker projects that maybe you could have done but would take too much time and energy to do previously. I haven't tried this. Can I have Claude hack my Google Nest cam to do something new, or some piece of hardware, something that I wouldn't normally spend time on. Can I just send off an agent to do some of that work? 

[0:50:05] GV: Yeah. No, for sure. Maybe I'll have time looking ahead. I'm off to Scotland in a couple weeks for the month for the summer. That will maybe give me some time to hack away on some things that I always struggle to. 

[0:50:17] SF: There you go. I expect me to be covering you in the Hacker News session with your latest ceiling fan hack. 

[0:50:23] GV: Maybe that's a goal. Yeah. By the end of the year, something, that one of us producers in our side project world ends up there. But yeah, any other, I don't know, predictions, thoughts for the next month? 

[0:50:35] SF: I think we're going to - this is not the last that we've heard about the government, essentially non-US governments concerned about the US government controlling deployment of models. It would be my prediction. I think this is going to be a big topic of conversation for the foreseeable future. 

[0:50:51] GV: Yeah, for sure. On my side, yeah, I'm curious to see if Cursor - it's reported that Cursor's usage goes off a cliff. Maybe not. I'm just going to kind of put it out there. That's something I'll watch and we'll come back to. I'm just curious how a deal for this - A, was it really worth 60 billion? Sorry, I don't believe it was. And B, I think people who would try to avoid any product that is Musk-related, just to be honest. I do have a lot of macro respect for him in terms of what he's achieved in many areas, but I think there's all sorts of other areas I just couldn't agree less. And I think I've got so many other options on this product. I'm going to use something else. 

[0:51:30] SF: Yeah. Yeah. Makes sense. 

[0:51:31] GV: Right. Well, a lot to cover there today. I'll be seeing you in Singapore next week, Sean. And we'll catch up then again on SED News in a month's time. Thanks. Thanks, everyone, for tuning in. 

[0:51:44] SF: Thanks everyone. Cheers.

[END]