EPISODE 1978

[INTRODUCTION]

[0:00:12] GV: Hello, and welcome to SED News. This is the monthly installment of Software Engineering Daily, where myself, Gregor Vand, and Sean Falconer with me.

[0:00:22] SF: Hey, everyone. Hey, Gregor.

[0:00:24] GV: We take a spin through the tech headlines. We then do a bigger topic in the middle, and then we take a look at our favorites from Hacker News that have popped up through the month as well. Both of us have been pretty busy, I think, since we last were doing SED News, but we did actually see each other in person as well in SF, which is very, very nice always to see you in the real world, not just on a video screen.

[0:00:49] SF: Yeah. We went, I don't know how long it was, well over a year, having never met each other, and then now three times in a couple quarters, so that's pretty good.

[0:00:58] GV: Yeah. So, yeah, I was over in SF for the Supabase Conference, although we call it a festival. I don't know if Sean, you popped your head in for a couple hours. Could you attest to it being more of a festival than a conference?

[0:01:10] SF: Yeah, I guess it had a festival-type feel to it. But you could probably also label it a conference, but definitely in good energy and vibe there.

[0:01:18] GV: Nice. Yeah, managed to briefly meet Steve Wozniak, which was quite -

[0:01:22] SF: That's amazing.

[0:01:23] GV: - quite a bit of a pinch me moment. But actually just seeing him speak on stage was actually great as well. He was very funny.

[0:01:29] SF: What was he talking about?

[0:01:30] GV: It's funny. We had our CEO and CTO interviewing him, and they kept trying to ask about AI and things, and he kept changing the topics back to really just his personal life and stories from before. Bringing up things to do with Steve Jobs, and took a dig at Oracle, which was quite - Obviously, we quite enjoyed that, because he was saying how much he loved PostgreSQL. Yeah, it was really just whatever he felt like talking about, but that was nice in a way. He was doing this on his wedding anniversary, and he has, apparently, wine that he has his name on or something, so they were drinking the Woz wine on stage. It was a nice way to round out the conference, I guess. Yeah.

[0:02:13] SF: Yeah. That sounds awesome.

[0:02:14] GV: Yeah. I'm sure a lot of people listening have done this already, but riding a Waymo on the freeway, like an early access thing, apparently. I forgotten I'd signed up to early access on my last trip, and then I needed to take a car to the Stripe office, which is down in a place called Oyster Point, which is South San Francisco, and Waymo claimed it would do it for me, and I was like, "That sounds great." Yeah, took a Waymo on the freeway there and back. Absolutely loved it, got to say. I'm such a convert to Waymo, or just self-driving for these purposes. But I know some people -

[0:02:49] SF: I haven't done that yet, so I'll have to check it out.

[0:02:52] GV: Yeah. It was still a strange moment, even though I've ridden Waymo a lot and still do, or when I go to SF, but the moment it was pulling onto the freeway, I was like, "This is quite weird." But then you get quite used to it, so yeah.

[0:03:05] SF: A little faster reaction time, slightly worse consequences of things going on.

[0:03:09] GV: Yeah. Yeah. Well, basically, it just sits in the middle lane and doesn't go over the speed limit. That's what every driver should do, quite frankly, or rather weaving around. But yeah, what else have you been up to in the last few weeks?

[0:03:21] SF: Well, I've had a nice little break from work travel until today. I just got to Atlanta a couple of hours ago for - I'm here for IBM Tech Exchange this week. So, if I sound a little bit different, I'm using my travel mic, so that might be what's going on. But I have six talks this week, four on Tuesday, so hopefully, I have a voice. It's probably good that we're doing this early in the week, rather than later in the week, because I might not be able to speak by the end of the week. I did do a fun solo trip with just my daughter, my four-year-old daughter, to visit my parents in Canada recently a couple of weeks ago for Canadian Thanksgiving, so that was really fun.

[0:03:55] GV: Nice.

[0:03:55] SF: Than back and then we have our big conference in - Confluence version of what you did with Supabase is coming up here next week.

[0:04:02] GV: Yeah, of course. It's crazy, because even though Confluent is part of IBM, you actually still have a Confluent-specific conference, so.

[0:04:08] SF: Yeah. Then they're back-to-back, so that's fun. TL1 got to exist, but -

[0:04:13] GV: Yeah, amazing. Yeah. I mean, we've got lots of happening, even though it's probably going to get released maybe three weeks since the last time we did this, but we've got a lot - yeah, a lot to get through. Start with it.

[0:04:24] SF: I do. It's great.

[0:04:25] GV: Yeah. Main headlines, and we're going to roll in on, I think it was the last episode we did a A&R acquisitions and raises special, so to speak, but actually, we're going to roll in some more of that into our main headlines this week as well. There's a lot still going on, and it's good to keep up to date with the movements as well. Top two main headlines, really, the PHP vulnerability that seems to have affected 70% of the Internet. I found this very interesting, just the fact that this had been two bugs that had actually been discovered in 2024 already.

[0:05:00] SF: Yeah, I didn't realize that until I was digging in more, because I knew we were going to talk about this. I didn't realize that it actually just been discovered several years ago. But they just didn't consider it an exploitable thing at the time, so they didn't -

[0:05:14] GV: Exactly.

[0:05:14] SF: - prioritize patch. It turned out, they were grossly mistaken, I guess.

[0:05:19] GV: Yeah. Well, I guess this is then the whole AI security story where OpenAI and Google, in some AI form, figured out how to exploit both of these. The biggest problem here was that it's just a complete remote code execution attack. It doesn't really need a user to do anything. It literally could just be you go and find a PHP-powered website and basically, could take over it, was the worst case of what could happen. Yeah, it basically meant that a lot of government sites across the world will have to be taken offline while they patched, because there really wasn't any other option other than to just take things offline until fixed.

[0:05:55] SF: Yeah, and like Wikipedia is PHP. The big thing is the sheer number of WordPress websites out there.

[0:06:03] GV: Yeah. Still dumb founds me, like how much the Internet runs on WordPress, basically runs on PHP. Fun fact, didn't know that PHP does stand for a Personal Home Page. Didn't know that.

[0:06:12] SF: No, I didn't either. I programmed in PHP for years, but I had no idea that it's just that. The LAMP stack was so popular in that probably late aughts to mid aughts period, where there was just an explosion of companies and websites and of course, popular CMSs, like WordPress, and it's just went everywhere. Nothing wrong with it, I guess. We've moved on from the technology perspective in terms of what the modern web languages are now, but it's still everywhere all over the Internet. Exploit there suddenly becomes a huge, huge problem. Wiz described it, I think is the most severe vulnerabilities in software history.

[0:06:49] GV: Wow. Yeah. Yeah, it's probably - anyone listening, probably you've been affected in some way in the sense that something you might have accessed over the last week or so is taken down for a day or so while this thing got patched. Apparently, though, it was the PHP community, and in the back-room channels, they had actually been - I don't think responsible disclosure is really a thing so much these days with AI, but at least people on the good side were aware of this and able to coordinate a bit more, especially with this super important, I think government websites and so forth before it was really, really known in the wild. That's a good thing for in terms of the community.

I still find this a bit shattered for it, just because I hate PHP in alignment. There's a reason I tried it and then just put it down very, very fast, but I know that doesn't make it insecure, or just - I just really, really - It's the only language I really just hate.

[0:07:46] SF: It's an odd language. I mean, I guess it has some - the fact that you're using dollar signs for variables is reminiscent of earlier web languages and things like that. But I think part of the reason why so many people ended up using it, one, you had the LAMP stack, so you had this open-source version of MySQL, Linux, and PHP, where you could spin up websites. You just didn't have that any option, as many options back then. It was either proprietary stuff that was fairly expensive, or you could go with the LAMP stack. Then there was built up significant community around it.

Even though the language has certain weirdness, where there's a dozen different ways to do searches of arrays and things like that, it basically has something built in to almost do everything. There's some version of it. You might not be able to guess it, because it doesn't follow the standard naming conventions and they use different naming conventions all over the place, but it's there. It's somewhere in the language.

[0:08:42] GV: We'd call a language at this point.

[0:08:44] SF: Yeah. I think that's not exactly the case, but similarly with Python was never intended to be the language it is today. It was really supposed to be something, like a wrapper for stuff that people were doing in C++ and we're going to build full applications in it. JavaScript certainly was never intended to be.

[0:09:01] GV: Java, to say JavaScript. Yeah.

[0:09:03] SF: I mean, it blows my mind that JavaScript turned into the dominant beast that's turned into.

[0:09:08] GV: There'll be a lot of people probably listening going like, "What JavaScript? What are you talking about?" I do unfortunately like JavaScript, so yeah, I can agree to disagree on languages as every good programmer should do at some point. Great. Moving on to the other main headline, way back, we talked about some espionage cases in tech, but this is a pretty wild one. OpenAI and an employee was discovered to have placed a sophisticated recording device behind a television in a conference room. I still love the fact it was concealed in toilet paper. Of course, a lot of people have taken to Reddit saying like, "Oh, of course, I would never think to look at. Most TVs just have something in toilet paper behind them, don't they?" Great concealing.

[0:09:54] SF: Yeah. A strange choice.

[0:09:55] GV: Very strange choice. Yeah, maybe there's some device that needed some, I don't know, stop two connections in each other. Who knows? Apparently, it was picked up just during a routine sweep of conference rooms. That's interesting that OpenAI even do this at all. I guess, they should be on guard. I happen to go to, I think this exact office a couple weeks ago, and there's a lot of security. I think this is before it was known. It's a lot of interesting, a lot of security to get into that office in the first place. It doesn't have a name on the door. You have three different things you have to do outside before you get inside. It's already quite a lockdown place. This Saudi national had placed this device in the back of the television. It probably was an executive conference room.

Yeah, this raises a lot of interesting questions of like, the why here. I mean, it does feel state sponsored in a sense, because this person apparently was a go between Saudi Arabia and Silicon Valley. That was how he had positioned himself, money in Saudi Arabia, opportunities in Silicon Valley and now working for OpenAI. It seems a very risky thing to throw it all away on this. Yeah, I guess this is the stakes that are there now with the frontier labs.

[0:11:10] SF: Yeah. I just wonder what the plan was. Also, how did they track the device that employee - Did you do and cover that?

[0:11:17] GV: They did a sweep. Then, there's some screenshot of showing the guy behind the television.

[0:11:23] SF: Yeah. I mean, it's just probably cameras in there.

[0:11:25] GV: I think it's just a camera. Yeah. I think they managed to hyper zoom in and show this guy. I mean, they must have seen, once they had discovered the device, they must have then tracked who was going in and out of this room. Yeah, it definitely took a -

[0:11:38] SF: Who had a roll of toilet paper with them.

[0:11:41] GV: Yes. A roll of toilet paper peering behind the television, all looks pretty - Yeah. I mean, it doesn't feel very sophisticated in total for such high stakes. Yeah.

[0:11:52] SF: Yeah. It's very strange. I think, given, we've covered this bunch, but given all the concerns over the Chinese labs and national security, the arms race around AI, there's this point of a potentially larger issue of where AI security needs to be some sort of national security matter now. The US wants to keep its lead, and that's important for many reasons. Does it need to be some national concern at a government level?

[0:12:18] GV: Yeah, absolutely. Well, clearly, I mean, given, yeah, Saudi Arabia, probably state sponsored there. Yeah, it feels like, perhaps, it should become more of a national concern. This was, I think, something that we debated a bit either in the last SED News, or the one before that. Yeah, always something I'm sure we'll come back to.

Moving on to the acquisitions and raises, A&R, so to speak. I know A&R is something in the music industry as well, but I thought I just re-appropriate it for our purposes. Acquisitions and raises, the first one that has appeared in the last few weeks is Chroma is joining Replit. I mean, again, I think it's - we should always home in on the words being used, because joining it usually suggests acquihire. Yeah, this does look an acquihire that Chroma, which we had episodes on with just not that long ago, actually with the CTO, Hammad. Hammad has actually left Chroma now and has gone to OpenAI. I think it's the CEO, Co-Founder, he's, I think, some of the other team, they're joining Replit.

One of the founders of Replit was apparently a key investor in Chroma. There's a link there clearly as well. On the tech side, what's your take on Chroma as a product and a technology? Then, what do you think Replit could do with that?

[0:13:38] SF: Yeah. I mean, Chroma was, I think they were one of the top vector databases, especially when rag and vector databases were really taken the world by storm. A lot of those companies raised pretty big rounds and grew really quickly a couple of years ago. I think the challenge has been that vector search basically turned into a feature within many, many different databases. As a result, they probably, I think people churn from these different platforms, growth slowed down significantly. Vector search is still something that people often want to do when they're building AI systems, but there's many, many ways to do it.

Even then, there's lots of ways where you might not even do vector search if you're doing search, because you can rely on other ways of searching and doing document lookups and just exploit the fact that there's a larger context when there's other ways of dynamically getting information than just semantic search. Overall, the market shrunk down on them. I think this has been the consequence of it for many of them. It was easier for the existing databases to pick up the feature than it was for some of the dedicated vector databases to really carve out market and continue to grow.

[0:14:48] GV: Yeah. Chroma as clouds offering will shut down with this. Yeah, so it's technology, the open source, but rolled into Replit and thus with the remaining team moving over there as well?

[0:15:01] SF: I mentioned this on a prior SED News, too, where I think that - just because things are moving so quickly, there are a lot of these moment in time, like by products that will, in companies that will grow. They seem like great businesses for some period of time, but then they get essentially eaten up by other product areas. I think it's very hard to have a company that is a feature at this point, because it's either needs to grow into a platform, which some of them might, or it ends up getting eaten up by the existing platforms.

[0:15:35] GV: Yeah. Hyper example of growing into a platform, at least at the moment is Vercel, who just seem to be releasing product after products, clearly pivoting far away from the roots of where the frontend to your backend somewhere else. Basically, they're just cobbling together all the bits that make a backend. Yeah, seem to be doing a great job at that at the moment.

[0:15:56] SF: I think they probably saw where the market was going in. There was less of a moat around their original psychic capabilities.

[0:16:04] GV: Definitely. The next one is Deno, the runtime. Open-source runtime being effectively acquired by CloudFlare. This was definitely not something I think I saw coming exactly. Deno had been struggling for a while to keep up the traction, I think, that it got initially. I don't think I would have predicted that CloudFlare would have vacuumed it up. This was super interesting.

[0:16:30] SF: I thought originally that it was more of a talent-based acquisition. I'm sure that's a piece of it. But when you read up on the articles around the acquisition, it seems like they're going to continue the project and continue to invest in it. It seems like maybe CloudFlare was using a wildly, I don't really know. It's pretty hard to keep a language runtime as a successful solo company. That eventually needs a landing spot. Either it has to go some sort of open-source foundation in company backers, or it needs to go and be part of some strategic reason for a company.

[0:17:04] GV: Yeah. It's going to have a working group currently set up. I'm aware of people in industry who've now been asked to be part of that. Because obviously, Deno has used just - it's been a huge part of the Supabase story, in fact, but also was actually a huge part of one of the areas of a Supabase product, where a lot of feedback was, can be a developer saying, "I don't want to use Deno." We were pivoting away from it already. It's interesting. Then this news dropped as well. Then next one, slightly different area, but a company, Sundial has been acquired by Meta. This is in the data and analytics space.

[0:17:42] SF: Yeah. I didn't know a lot about Sundial, but it seems like they're planning to add that into their existing AI portfolio. I think it's another situation, where through acquihire, the company's no longer going to be serving outside customers, essentially. Chroma is doing that with their cloud service. Sundial is doing it with their customer, which actually includes OpenAI as well. They'll lose access once the deal closes. This isn't new just because of AI. There's always been risk with investing in a startup product, because you don't know the longevity of the company.

I wonder, though, if that becomes even more of a risk now, just given the accelerated pace of technology. Seems like companies raise more money, but they also seem to accelerate the pace of the product not being successful, too.

[0:18:33] GV: Yeah, absolutely. Definitely, acquihire. As I say, anytime we see the word join being used, often indicates that the final one, yeah, it is to do with Supabase, actually. Supabase acquired Turso. This was a proper acquisition, acquiring the technology, but the team has also moved across the entire - think there's about 10 of them. Also, remotely distributed teams, so quite easy culture fit there. Yeah, Turso is basically databases for agents from a technology standpoint. It's effectively an S3 file, so you don't need any compute to keep these things in storage and then it needs very little to then run them.

Yeah, Supabase is going to be continuing the Turso cloud offering, but it's also going to be rolled into a new product called Spark. Obviously, I can't talk too much about what that is, but anything you can find online, which is already there, so this new product, Spark, is going to be a incarnation of Turso. It's an interesting technology, because they basically re-wrote SQLite in Rust to achieve. It's an SQLite interface, but it is technically not SQLite. It's a rewrite. And they're doing the same with PostgreSQL, so it's a rewrite of PostgreSQL that they were working on. That's also what Supabase is quite interested in, because it is a PostgreSQL-first company. Interesting. I'll say, I was not be able to say too much from the inside, but it is definitely interesting to see this acquisition happen.

[0:20:03] SF: Yeah. Between PostgreSQL and SQLite, you've got the good chunk of the market.

[0:20:10] GV: Yeah, it is interesting. Yeah, certainly which parts of the market want PostgreSQL, or do they want SQLite, or do they almost want both? Yeah, that remains to be seen. Yeah, very interesting times in that space. Then yeah, we've got a pretty chunky raise from a company that I had not heard of, but definitely deserves some attention. They're called flapping airplanes. Of course, nice name that means absolutely nothing. They just raised a series A, a 1 billion of funding at a 6 billion valuation for series A, which is pretty chunky. They're only one year old.

[0:20:45] SF: I think their seed round was over a 100.

[0:20:48] GV: Yeah. Seed round was over a 100. 180, or something. Yeah, they're one-year-old company. They basically believe that GPUs are really not in any way being utilized to their full potential. And so, they believe that they can train - especially training, they can do in a much more efficient way and utilize the GPU in a much more efficient way. Yeah, pretty interesting. It started by, one of the founders also founded this underground student-led accelerator called Prod. There's only been a little bit about this written recently. This is all via TechCrunch. So, thank you to TechCrunch on this one. Yeah, basically, Prod is this underground tech accelerator run by students at MIT, Harvard, Stanford. Cursor is one of their graduation companies, but it's sort of you can't really apply to Prod. Prod finds you is how it's built.

Yeah, flapping airplanes is, I don't think it came exactly out of Prod, but let's say that one of the founders is part of that. Clearly, yeah. I think investors are very interested in the people as much as the technology.

[0:21:56] SF: Yeah. I mean, the idea of better GPU utilization makes a lot of sense. There's a lot of ways right now. I think, essentially, what they've developed is this idea of a mega kernel that can run memory bound inference alongside compute bound training, so that the GPU is rarely waiting. It's always being utilized in some way. It's more of a less sexy than AGI and some of the things that other AI companies tout, but it's solving a real problem. It's more of a classic, I say, like computer optimization problem than we are typically hearing in the news and things like that now.

I didn't know a lot about the company either, but it sounds like what they're doing makes a lot of sense. Probably hard problems to solve there. If you can solve them, then it justifies the raise and the valuation.

[0:22:46] GV: Yeah. I know a billion dollars can go pretty fast these days as in B -

[0:22:51] SF: Especially where GB is.

[0:22:52] GV: Exactly. But still, you'd expect something to come off this. We'll be following along. If you check out their X announcement, I think on the, I don't know if it was about the razor, just so about the product in general, but they have a little video. Of course, DOOM is part of it, which is great. It's basically a video showing that they are training data and serving, and it is playing DOOM all in one kernel on a B200. That's fun to watch that for anyone interested. I love that they still reach for DOOM, despite the crazy advances of technology there. Nice to see.

Onto our main topic. We're looking at ways that AI products have died. Or you could call this the AI graveyard. It was inspired a bit. There was an article in TechCrunch that about this and looked at some of those examples, but very much came across a spectrum. We're going to look through four ways that we've seen AI products effectively die over the last year or two, given the speed at which products have come about, and then they have fallen away. Both products that were just - that was the product of the singular company that did it, or products that have been within other products that you might know of. Notion is a good example. They have various product offerings and made big splashes with some of them, which we'll talk about, but actually, they get folded pretty fast, which is what people don't talk about so much. We're going to go through some of these and see how these happened and what that maybe means for just the AI product landscape in general over the next few months.

We have this theme of the platform ate it. Effectively a standalone product, whose feature became a checkbox in somebody else's app. The first one in there is Relay. That was actually launched in 2021. It was a AI Zapier alternative.

[0:24:51] SF: I think they shut down because Google and OpenAI and others basically built the same feature into their products.

[0:24:57] GV: Yeah. Yeah.

[0:24:58] SF: They didn't have much of a competitive mode, essentially. It's not that hard to recreate what they had.

[0:25:03] GV: Yeah, exactly. Yeah, when we talk about a platform eating it, exactly the large platform is basically just replicated it and really just shut down.

[0:25:12] SF: Yeah. I mean, this goes back to what I was saying earlier about even the vector databases. I think a lot of other things, like these moment in time, project offerings, where they can't really defend the space that they have from the platforms, because they're a singular feature and it's either not a complex enough problem that's not difficult for another company is motivated to recreate, or it makes a lot more sense for it to be part of a larger platform, because more and more, especially in the enterprise, companies just don't want to have to deal with a million different vendors.

[0:25:46] GV: Yeah. Another one is Huxe. H-U-X-E. This was built by a former Notebook LM developer, which was - and it was intended for podcast style audio from your content, which is how Notebook LM got its big break out of Google. Anyway, I still remember this like, yeah, two, two and a half years ago, everyone was suddenly creating these AI podcasts of their content and throwing them around, because they thought that was an easier way to digest information. Interesting that then, I guess, one of the people left Google to then just found a version of this on its own. But then, platforms like Spotify actually just shipped a similar feature within the platform. And so, having a dedicated platform to this just didn't have a lot of likes. I was surprised that a company thought it could just do this, or just didn't spend enough time then building itself out into a wider platform.

[0:26:42] SF: Yeah. I mean, even the audio feature has been built into one of the chat-based systems, like ChatGPT and things like that, too. It just ends up being like a small feature in the UI of some existing product.

[0:26:55] GV: Yeah. Then finally in the platform ate it as such, ChatGPT Atlas. I had to look this one back up again. It was actually a browser, basically.

[0:27:05] SF: I think we talked about it. If memory recalls, I think we were not bullish on it.

[0:27:12] GV: Yeah. I like though that they keep the blog posts around. They got to say, this product has been deprecated, but here is the blog post that we put out about it. It's fun that you can - I mean, you could find it on my back machine, but just find it right there so you can see what they were trying to achieve with it. This is only back in October last year. Yeah, I guess a year. But they did shut it down in August. Some of the features got folded into a desktop app and Chrome extension. Yeah, the whole browser play perplexity went big on that as well. I really feel like, yeah, the whole browser play has just disappeared almost.

[0:27:49] SF: Yeah. I think that disappeared. I also think going back to what I was saying, the accelerated pace where these companies can learn that something's not successful, you think of what is the bar for success that a company like OpenAI is pretty high. If something isn't hyper scaling in the next two quarters, then it's probably potentially going to see us with depth, because they'd rather spend their resources on things that are going to be those hyperscale opportunities.

[0:28:18] GV: Yeah. We work with them quite a bit. Again, can't say too much, but I can definitely attest to the fact that products do appear and then they disappear quite fast, even nothing's particularly built. It was like, "Hey, we've got this idea," and then don't hear about it for a while.

[0:28:35] SF: They've had nine different agent frameworks, or takes on agents as well. They keep trying, they keep iterating and things like that to try to figure out.

[0:28:43] GV: They have the scale at the moment to do that. They have the scale and the money, it seems to just really throw stuff at a wall and see what sticks.

[0:28:53] SF: Yeah. It's like the Google labs era of Google on steroids.

[0:28:57] GV: Yeah. Yeah, yeah. Yeah, very interesting there. Then there's a another, a next category is basically that, so you can subtitle it like, you can't afford your own success, i.e. you literally cannot afford to run the thing that is a runaway success in theory, but just too expensive to run, or there's an example here where both expensive and ends up with a bunch of legal liabilities. The first one is Sora. People will go, "Oh, yeah. Sora, I remember that." That was the video model from OpenAI and they even got to the point where they had a deal with Disney, who were going to license their, I guess, content into it. It wasn't very formed. It just said a 1 billion investment from Disney that would include licensing.

Yeah, they basically just couldn't sustain how much money it was costing them to run this for users. Apparently, one user managed to rack up a million in one day, like a million dollars of usage in a day. I think there was even a 15-million-dollar in a day reported as well. Yeah, they just couldn't sustain what this was actually going to look like commercially. Similarly, character.ai shutting down basically, partly to do with how much it was burning. I think it was burning 8 to 9 million dollars a month, but it then had a ton of legal liabilities from, unfortunately, people saying that it did people talking to it came back with answers that were quite unhealthy, unfortunately, for a lot of people. Yeah, just ended up in Chapter 11, I believe, on October 12th. That just shows if you do have a - you want to create the next AI chat experience, or whichever, it is very, very clear now that safety is a big concern that companies should have.

[0:30:49] SF: I think it was a lot of minors that used character.ai.

[0:30:53] GV: Yes, exactly. Yeah.

[0:30:55] SF: Then the controversy around it was pretty - if you read into it, it's pretty awful and pretty scary.

[0:30:59] GV: Exactly. Yeah. I do see that as a "good riddance." It's interesting to call out just how a company that just doesn't take its responsibility seriously enough, yeah, it just ends up getting legally blitzed.

[0:31:14] SF: I still think the outside of engineering, of course, it's still trying to find, like what is the killer app. There's the chat experience, which is that - ChatGPT is obviously a killer experience. Just general purpose, question and answer exploration. Then there's agentic engineering, but one of these other attempts for things like create a - especially in consumer, create some social thing that I can interact with and things like that, they just haven't quite found what that killer experience is. For mobile, there was Instagram and Uber that came along that were these mobile first companies and it took several years of mobile before they were able to think from a mobile first lens and do something that wouldn't have been possible if people didn't have computers running around in their pockets. I think we're still struggling to find that with AI.

[0:32:06] GV: Yeah. Another subheading topic is agents effectively killed the interface. I think there are going to be many, many more of these in the future as well. But two that stand out, Notion Mail, the first one. They launched that in April 2025, but they shut it down, I believe in June this year, or they announced they were going to shut it down in June this year, and then closed it completely on September 22nd. Notion said that more than half of its users just ended up managing email without even opening the inbox, because agents did it for them. That's interesting how they thought they could come up with this interface that was going to be so much better than Gmail.

We've seen that story before, a superhuman, which also sidebar, got acquired by Grammarly. I'm not really sure how much. I think the superhuman product still exists, but I don't think it's probably getting a ton of active development on it. Yeah, it was a bit of an odd one for Notion to go hard into this, because I guess, they were matching AI with email interface. I also ran a startup that was in this space, but quite quickly, we saw that having an agent work against the inbox was probably going to be where things were going. It was also just an incredibly competitive space, to be honest. It is interesting that Notion also found, again competitive to the point they thought that keeping that product up wasn't in their best interests.

[0:33:30] SF: Yeah. I think that agents that only exist in whatever their system of record is, is kind of something - it was popular a year ago, but it removes a lot of the value of the agent itself, because you're limiting its capabilities. I think one of the reasons why Claude Code and other agents that run from your terminal have been so successful, so popular is because you can point them to multiple disparate data sources than they're able to pull in data from those different places. Reason about basically connect ideas together that would have been really hard to do previously without wiring a bunch of APIs together into a dashboard, or pulling data and centralizing it somewhere and then trying to pull a dashboard together, or something like that. You were able to federate that out. But if the agent exists in the system of record, you can't do the federation piece, its knowledge is always going to be capped by whatever that data exists in that system.

[0:34:26] GV: Yeah. And interestingly, recent have actually spun up a new product in the last couple of months, or at least they've launched it in the last couple of months. Yeah, it's just literally called Inboxes. Literally is the inbox for agents. I mean, I look at this and kudos to that. A really, really nice landing page for it, I got to say. But I'm still a little bit confused on the use case. The use case that they seem to present is killing the - like a front, which is a customer support, or a Zen desk. The examples they show is you can basically build your own kanban from inbound email, like people asking you to do stuff. Yeah.

Or you can also, people emailing in for jobs, or something. Yeah, I don't feel the use cases are quite explained yet. I would still support that there is a case for email protocol being used by agents, because a lot of it is very clunky to use on an interface and like, yeah, you could - obviously, you bring up the federated piece, Sean. Yeah, I'm curious to see how this Resend inbox for agents, hands out. Maybe they'll be the feature of next year's AI products that died. But great to see still Resend trying other things than just being a sending protocol. Sorry, not a protocol, but a sending platform, that they do very well.

Then, we've also got in this bucket, we've got, which is agents taking over interfaces. There was a company, Yupp, Y-U-P-P, which was a model variant of this. It was 33 million, A16Z crypto-backed playground for comparing 800 models, but shut in March 2026. The founders said that models had improved so fast that the need for human comparison had faded. Yeah, that's an interesting one, where you think you're in the right space at the right time. But actually, things as we've seen over the last couple of years, do move at such a pace that it's very hard to actually keep up with what is relevant for a developer to need.

[0:36:36] SF: I wonder what the ultimate business plan was there. Surely it couldn't have been just, "Hey, our business is going to be model comparison." I don't really see how you would turn that into a business. There must have been something, they had planned to go beyond that. I'm curious about what that was and what the plan was and why it didn't work.

[0:36:54] GV: Yeah. I mean, if I go into their - they've got a site up saying what they did and obviously, that is all shut down now, but they said, provided free access to the world's most powerful AI models. Over a million users across 200 countries participated. Yeah, it doesn't really answer that question. Having users participate to evaluate and give that human evaluation. It feels like almost it was traced a bit of a weird, maybe trace data play, where people would - If I look at it actually, that's a little bit, it looks like it's a gateway that they might take the trace data and then maybe somehow monetize that, I don't know, from -

[0:37:32] SF: From the human evaluators.

[0:37:33] GV: Exactly. Yeah. What we touched on in a slight different forum, but what we touched on a couple of weeks ago, a couple of episodes back on the Chinese proxies. Yeah. I don't know, maybe they got into hot water there, because it was a proxy-ish thing. Yeah, wasn't clear what the end goal business plan, any of that. Yeah. Yeah.

Then finally, we've got a topic here on AI that failed. This is also AI that has never really materialized. yeah, we've got a couple of actually hardware ones. We had the humane AI pin, which raised 230 million.

[0:38:14] SF: Former SED episode as well.

[0:38:15] GV: Yeah. Did you do that one, Sean?

[0:38:17] SF: Yes, I did it.

[0:38:18] GV: All right. Okay. Yeah, but they unfortunately, their assets, apparently, were sold to Hewlett Packard. 416 million in February of last year. Yeah. I mean, well, given that you spoke to them, what's maybe your take on that one?

[0:38:32] SF: It was an ambitious project. A lot of smart ex-Apple people went there to try to do something big, change the world. Of course, doing anything big comes with risk. I think it's really hard to do something where you're trying to rethink what is the fundamental interface that someone's going to interact with technology. That's what they're doing. It didn't depend on your phone. It was a closed system, so you're not getting any benefit off of the existing phone ecosystem. It had no UI. It used a laser.

They're trying to rethink everything from bottoms up to have a non-invasive piece of technology that serves you when you needed. There was other problems, too. Their launch was not very successful at the initial launch, because it didn't work as advertised. The demo looked great, but the actual reality of the device was not very functional. It just took way too long to do compute. Models were much slower. They can only put so much stuff on the device and then you have to round trip across the network. Your network connection is not very good. You're like, five minutes of just falling stuff. There's a lot of problems to try to navigate with a lot of these hardware devices that you're supposed to be using live in the wild. A lot of variables in this.

[0:39:47] GV: Yeah. Like I said, similar, it's still going, but Rabbit. The Rabbit r1, which is, I say, was, is this product, this orange box thing with a bunch of buttons on it where you can do all sorts of AI things and meant to be a AI companion. Yeah, really got pretty flat reviews. I remember watching a Marcus Brownlee review of it, who just didn't seem to like it. He did review it as quite unfinished as a product. Rabbit then said, they were building a thing called a Cyber Deck, which some people then said, well, it's basically a net book, but let's see it. But actually, we've not seen it. That hasn't materialized.

Actually, if you go to the Rabbit's website now, they're really playing on their own agent OS. Actually, maybe more of a software play, which probably would be more economical, it sounds like, if they're not really hitting out the park on anything hardware related. Or just trying to monetize, basically, what they maybe have on device on the r1, I don't know. Yeah, definitely a bit of a muddled, muddled company, even if they seem to be trying very hard on certain things. Like, hardware is hard, as they say.

[0:40:59] SF: I think hardware and voice is tough. Even if you look at Alexa and the Google Home devices, they've been largely given up on. They tried to turn them in the app platforms. It never really took off. For whatever reason, that form factor. I think one is it's hard, so the performance a lot of times isn't very good. Then people get frustrated. Then I think we made this comment before they end up turning on to just fancy devices to turn the lights on and off and set timers.

[0:41:28] GV: Yeah, yeah, exactly. Then just around out this one, I mean, we've touched on certainly one of these before, but Apple's Siri AI just delayed repeatedly. Actually, we touched on this before, but there was a legal settlement on iPhone 16s, because they claimed they were going to ship with this amazing Siri AI, which wasn't really the case. It was all a bit fake on stage. There was also a product called Microsoft Recall. I don't know if Recall has even - has it actually launched in the end?

[0:42:00] SF: There was a touch of controversy around it, because the idea was it takes periodic screenshots, so that you can search your history. I don't know what the timing on it is, but it's like, as I'm working, or as I'm using my computer, it's taking screenshots of what I'm doing. Then I can search that history. It's almost like your browser history, but it's your entire desktop and then visual. People are up in arms with the idea of that. I mean, they come down to sometimes with some of these products, and I think we brought this up before when we've talked about some of the stuff that has come up with, and it's just like, who's this for? Who was asking for this? What problem was that person trying to solve, where they thought this was like, "Oh, you know what would really help me is that if I had something automatically take a screenshot of my desktop every 30 seconds, or something."

[0:42:46] GV: It's funny. What's interesting there is there, I think it's Tobias Lütke. He's built his own version of this a while ago, because he says he has this crazy library of every 30 seconds, I think, for him, or every minute. He's got a screenshot from the year 2000 since, right? Like, for him, I can imagine that's insanely helpful, where being on the forefront of having access to AI models when they came about and so on, he's been able to turn that data into something quite interesting.

Yeah, the average person, and especially when it's a company like Microsoft pushing this on, consumers are saying, "Hey, guess what? We're going to take screenshots of your screen every -" Yeah, I can see why they didn't get a great reception. When someone is like, "Who's it for?" Well, it clearly might be for a esoteric CEO of Shopify, but it's not for the -

[0:43:42] SF: That's a small -

[0:43:44] GV: It's a very small subset.

[0:43:45] SF: - total addressable market. 

[0:43:46] GV: Yeah. I'm curious if they were inspired by that, and then just didn't do their market research. Yeah, very interesting.

[0:43:54] SF: Yeah, it's supposed to be in Windows 11. Periodically take screenshots of your screen, you create a searchable visual timeline, or past activity.

[0:44:01] GV: Yeah. But then, there's also now certain Brave browser and ad guards saying that they've added functionality to basically counteract this. Great. Well, yeah, that's a very interesting spin through, and for our main topic of, yeah, AI that failed effectively, but failure is part of the path of making technology at all.

[0:44:23] SF: Yeah, the other thing too, is every cycle has a graveyard. There's lots of famous dotcom, pets.com is the classic one that people always refer to, that lots of companies didn't make it there. The majority of companies that are started never go anywhere. Most of them you've never heard of, because they never went anywhere. Most companies fail. It's a fun spin, but it's not like it's an AI specific thing that people should give you the line.

[0:44:47] GV: Yeah, that's a good point. We've focused on AI today, I guess, given that most companies in the last two years are pretty much AI focused in some respect, even if it's just in name to make sure that investors are interested. It's definitely not an AI problem per se. It is part of the landscape that comes with technology and venture investing and startups in general. 

Yeah, moving to the final piece of the show, Hacker News highlights. What did you turn up in the last week or so?

[0:45:19] SF: Yeah. So, there was a fun one caught my attention, which was that washing machines took decades to catch on, which is a bit different than the typical ones I highlight, but I thought it was interesting. Essentially, apparently, and I had no idea of this history, but washing machines came out in the 1910s, but it took until the 1950s to win over American homes. Even though laundry was the most hated household chore, and I still, I personally can't stand. I don't actually mind putting the laundry in the laundry machine. I don't like the folding and organizing afterwards. There was a couple different reasons for this. One was that the early machines still needed someone to feed the clothes through wringers by hand. It saved you effort, but not - it was still very time consuming.

Then there was also this social status issue, where in the south, many white families kept black laundresses to do the heavy manual work, because they felt like it was beneath them at that time in the era, the US. There was that factor as well, where it wasn't a problem for these people who could afford to have people basically do it for them. Then it was really wasn't until after World War II that they became much more common. There was laundromats that were built up as well for people that lived in places that were small on space that needed somewhere to wash their clothes. That's what ended up accelerating the sales of the washing machine, but it took almost 50 years for them to take off.

[0:46:50] GV: Yeah, super interesting. Yeah, I always find any of that domestic history, complete sidebar, but the fact, because living in Asia, and a lot of people eat out as opposed to cook in their houses, and then in Europe, it's the opposite. There was a time in Europe when eating out was the norm, because people didn't have - but actually didn't have any kitchen in your home was seen as very upper class. So, then everybody wanted a kitchen in their home. Whereas, I think, we would often see eating out as a more expensive and trendy thing to do.

Laundromat version of this, and I guess in World War II, and just this boom in domestic appliances being bought by these families that the men had come back from the war and they were getting jobs and making money. I'm obviously just pulling on my Mad Men knowledge here, from my watching Mad Men so much. But yeah, that's super interesting.

Yeah, on my end, there was one that I just went by points to see. I was just scanning around and I saw that this is when I pulled it yesterday, it had over 1,500 upvotes points sent in by user. Sokam Young. Yeah, this is great. It's a local Wikipedia. This guy at edjefferson.com, if you go to edjefferson.com/localwiki, basically, you do need to enable it to find your location from the browser. I think maybe you can enter manually. But it just creates a very normal looking Wikipedia front page, but completely tailored to a certain radius that you are sitting at at that point in time. Yeah, it's pretty cool. Because I was just sitting in a cafe and went to it and then suddenly had a Wikipedia front page with all the landmarks and just interesting houses and buildings that people have written articles about.

It's a more interesting local travel guide, I guess, because you just never know what people have written about from especially from a buildings and history perspective. I think it's awesome. I'm going to use this a lot more just if I'm somewhere and check it out.

[0:48:55] SF: Yeah, that's cool. I'll have to check it out. The last one I had was this one that was called bring back the cool save file screen. The author, I think his name was Aaron, but it was posted by surprise talk on Hacker News, was looking back at the save file screens from the sixth console generation, so Dreamcast, GameCube, PS2, Xbox, because most of those consoles at the time needed a separate memory card. Each one had a menu that saved, or saves as these small icons -

[0:49:22] GV: I remember these.

[0:49:23] SF: - drawn from the game itself. Dreamcast used this little portrait of the squares over a cloudy background. Xbox had this alien looking green menu. Then modern consoles, they just show the games cover art, which is fine, but they're not that interesting. The author, this is like a plea to basically bring back the old screens, have some fun with it and do something unique.

[0:49:46] GV: That's interesting. I'm looking, because PS2, I remember very vividly the PS1, which was very 2D. Then PS2 had these 3D icons. I'm just looking at the one for PS2 at the moment. Yeah, Tony Hawk's Pro Skater is this little bubble thing that just says, Tony Hawk's on it. Yeah, memory cards. I mean, if anyone, after a certain point in time, nobody will know what these are. Yeah, just literally to save on your console, you had to have a memory card. If you ran out of space, you had to buy another memory card. There's only two slots as well. Then I think, maybe PS3 enabled you to plug an external hard drive in, or something as well. But then by that point, they had memory. Clearly, I get excited about this, so I'm going off on a tangent on that one. 

Finally, my one, just a funny, quirky one, this is posted by someone called J-L-L-Y Hill. It was a very, well, yeah, very detailed article about a bug in Figma that wasn't really a bug. It's basically, if you pressed shift plus the number two on a Windows keyboard, if you use the number two at the top of the keyboard, you would zoom. It would do what you expected. Whereas, if you did it on the numpad, it would actually act as a down arrow. People were like, "This is a bug. Why is this?" Turns out, it was actually very window specific that the way that Windows keyboards evolved from the IBM days was to do with Numlock, and Numlock was this thing that meant that if you wanted to use the numpad as arrows, you could do that. This was in the era of spreadsheet software coming into play. But then, you didn't really know if Numlock was on or not, so then they developed this new thing.

Mac never had this. Mac just basically had arrow keys and number keys. It was like a bug that goes all the way back to those days. Another fun fact was the guy, David Bradley, didn't write the article, but was mentioned, he invented Numlock, also invented control-al-delete. The person who wrote this said - That's the story for another day, so I hope that we maybe get that story sometime, where control-alt-delete come from.

[0:51:54] SF: Yeah, that's interesting. I wonder where what the origin. I'm sure we can look it up, but what's the origin of control-alt-delete?

[0:52:01] GV: Yeah. Well, yeah. That's it for another SED News. Yeah. I guess, we'll be back in another month, taking a spin around the headlines and on a topic in Hacker News again. But thanks for tuning in.

[0:52:13] SF: Yeah, thanks, everyone. Cheers.

[END]