EPISODE 1856

[0:00:12] GV: Hello, and welcome to SED News. I'm Gregor Vand.

[0:00:16] SF: And I'm Sean Falconer. 

[0:00:17] GV: And this is a certainly different format of SE Daily podcast where we basically take a spin across the last few weeks in terms of big headlines. We have a main topic in the middle. We look at Hacker News highlights, and we just kind of give our thoughts on what's been going on in the tech and predominantly software world. How's your few weeks been, Sean? How's it been going on? 

[0:00:40] SF: It's been good. I mean, I think last time I chatted, I was in the thick of coming off of Snowflake's big conference and Databrick's big conference. And then the challenge with going back-to-back conferences is that then I come back and I have a mountain of work to actually catch off on because it's hard to do the work while you're at the conferences. I think the last chunk of time has been essentially just trying to recover from a work perspective on all those things that were piling up. And then getting ready for a personal trip to Hawaii soon. 

[0:01:10] GV: Oh, awesome. Nice. Yeah. I like Hawaii. 

[0:01:12] SF: What have you been up to? 

[0:01:13] GV: Yeah, it sounds like we've had sort opposite schedules then. I was sort of heads down for a while. And then, yeah, the last couple weeks have been quite busy from a sort of events perspective. Yeah, we had Super AI, which is like this big conference that got put on over here in Singapore. It was a bit weird, I'm not going to lie. It was a sort of going into a nightclub for like two days straight. I had to sort of take breaks. But it was interesting. Just a lot of agentic stuff. I would say it was more sort of corporate leaning, perhaps. But good keynotes. There was Dwarkesh Patel, another AI podcast that's very popular these days. And so had a quick chat with him. And Edward Snowden was also a keynote speaker, but by video link for obvious reasons. 

[0:01:55] SF: Yeah, I haven't heard Edward Snowden in - it's been a while. 

[0:01:59] GV: Yeah, quite. Yeah. Yeah, it was a very interesting event. And lots of people had flown in from all corners of the earth for this thing. It's always nice just to meet a bunch of new people from around the world as well. 

[0:02:10] SF: Yeah. And Flink Forward Asia is happening in Singapore, I think, next week. 

[0:02:15] GV: Oh, yeah. 

[0:02:14] SF: Which is coming out very soon. I guess Singapore is a hotbed for tech conferences. 

[0:02:18] GV: Yeah. In this part of the world, for sure, it's kind of the place. Yeah. Facilities are kind of - you can just go down to this place called Marina Bay Sands. I believe it's actually owned by Sands of Las Vegas. And yeah, you just go down there any day of the week and there's something going on down there. It was kind of fun. But yeah, thanks to the Singapore government for the ticket. I did not pay a thousand bucks for that ticket. I am a real startup founder. I don't just blow a thousand bucks on these things. 

Anyway, let's get on to the headlines. The big one we're going to talk about to begin with is around Meta. There was a lot to talk about Mesa this week. But Meta and copyright basically. And they've been in a lawsuit. And at least the first sort of round of that is that they effectively won their argument that - it's around books. It's the idea that publishers are not very happy that basically chunks of text come out in responses. Obviously, this means the model has been trained on these books. And yes, the publishers are not very happy about this. But it seems like Meta has one on the basis that - I think the TLDR here was simply that their argument wasn't strong enough yet. And Meta, I imagine, had gone to town on their lawyer side. But yeah, what was your take on this one, Sean? 

[0:03:33] SF: Yeah. I mean, my understanding was they were kind of able to argue it was under sort of fair use, which it was the same kind of argument that Google made over a decade ago about Google Books, which was also determined fair use. And I guess like it's a very complex issue. I'm certainly not a lawyer. Some of the fair use arguments kind of make sense to me. If I can't prompt the model, essentially to regurgitate the entire novel word for word, then are you really getting anything more than perhaps you would be able to get from Wikipedia or CliffNotes version that also falls under presumably fair use? 

[0:04:09] GV: Yeah, I think that's a good example, yeah. 

[0:04:10] SF: But then at the same time, if I am an online content creator, if I own Reddit, for example, I can block third parties from crawling and ingesting that content and training models on it. It's kind of strange in some ways that if you're in the digital world and your thing exists digitally, I can prevent essentially model providers from essentially getting that data and using it for training. But then if I am an author of a book, I don't have the same easy control over it. Even if they didn't make the right argument, they have to essentially take this thing to court in order to be able to make that argument. 

It's very complex. I don't pretend to be the person who has expertise on it. I do find it kind of - there's some irony in it, though, that there's protections basically put in place from a technical control perspective. But if you're not in that world, suddenly you don't have the same control to protect your IP. 

[0:05:05] GV: Yeah, and I think this is the point. A lot of us in tech - technically I do have a lot background, maybe unusual, but I don't ever sort of pretend I'm more knowledgeable on law these days than anybody else in tech who's not like an in-house counsel. I think the point is that we're not lawyers, but we still have to understand on what side is sort of right or wrong in some ways. I believe what was said here is that there was no meaningful evidence of market dilution. That's a fancy way of saying they don't believe that this is so the judge saying, like, "I don't believe an LLM is going to stop people from buying the book." It's kind of like a translation of that, I think. 

[0:05:41] SF: Right. Yeah. If you want to read Harry Potter, are you going to ChatGPT and saying, like, "I'm already paying for this. I want to save some money, so I'm just going to have it tell me the story in ChatGPT." I certainly am not, but I can't speak for everyone. I mean, I think that's a fair argument. 

[0:05:54] GV: Right. Because I think, to what you mentioned just before, that you can't sort of prompt to get the whole book out. And exactly, you can't just say, "Oh, I want to read chapter one right now," and then chapter one pops out. It doesn't work like that, to my understanding. I think, yeah, CliffNotes. For those that remember CliffNotes - 

[0:06:10] SF: Yeah, it's a dated reference. 

[0:06:14] GV: CliffNotes were kind of like cheat sheets for when you were studying literature and things. 

[0:06:18] SF: Yeah, if you couldn't understand, I don't know, the Shakespeare play that you're supposed to read in high school English class, you could get the CliffNotes version where it explained it to you in plain English. 

[0:06:28] GV: sExactly. I definitely use those. Here we are. We're going to see how this plays out. And obviously, we're touching on it here because this does affect all of us in software. If suddenly, "Oh, the model has to stop referencing any books tomorrow," well, a lot of products just don't have that kind of concept that that might not come out. So then what are the cascading effects of people's platforms or products where suddenly this sort of content that was just assumed that would be available is gone, for example? 

[0:06:58] SF: Yeah. And I think this kind of talk is one thing that points to a larger sort of theme, and some of these things we're going to be touching on through the course of our conversation, is there's such a land grab right now around data. The big competitive advantage for people building models or even those building applications on the model is really what data do I have access to? What data can I either use for training purposes? Or if I'm building applications on these models, how do I get the right contextual data into the prompt in order to get something relevant for my business application? 

And the people who are the most successful of that are the ones that are going to kind of win the market. That's really where the competitive advantage is. It's less, at least currently, around new truly innovative techniques in terms of how these models are constructed. It's a lot to do with essentially the data and how organized it is for training purposes or prop assembly purposes. 

[0:07:58] GV: Yeah, exactly. And we've been seeing a few things like this pop up, especially in the last few weeks. We're going to get onto that, as you mentioned, Sean, and the main topic around basically walls that are going up. We'll get to that shortly. The other kind of main, I guess, announcement headline that made mainstream news as much as tech news was Meta, again, making a 14.3 billion investment in Scale AI, which gives them a 49% stake in that business. And that's always a great number. If you see 49 or 51, you know it's effectively saying this is effectively equal ownership. It's just that someone's decided there's a reason to take one off one side and put it on the other. 

[0:08:40] SF: Is that the same for Microsoft and OpenAI is the 49%? 

[0:08:43] GV: That's a good question. I'm not super sure. 

[0:08:46] SF: I don't know. It sounds familiar, but don't quote me on it. 

[0:08:49] GV: Yeah, exactly. It could be that kind of similar. In this case, Scale AI, they have high quality training data, and they are a vendor to all the big players, to my understanding. And this is just a blatant land grab by Meta. But crucially, because they haven't gone over that 50%, it's not an acquisition. Again, let's just put the legal hat on for five seconds. It's the idea that this is not going to be scrutinized sort of from antitrust. And I'm sure it's just a whole bunch of time and effort that would be needed to fully acquire or majority acquire a business. And this is "No, no. We're just 49%. Lots of money in your bank account. But you know, you can still supply other people if you want." Yeah, I mean, you're kind of on the ground over on that side, Sean. What do you make of this? 

[0:09:36] SF: It's a similar structure as I think Microsoft's investment in OpenAI, whether that's 49% or not. I can't remember the exact percentage breakdown. Or Amazon has a stake in Anthropic as well. And a lot of these giant tech companies are investing in these companies to access the AI capabilities without necessarily triggering any sort of antitrust reviews. It's a bit of a hack around that system. 

But as a consequence to the move that Meta made, Google, I believe, paused their Scale AI projects within hours of the announcement. OpenAI is also winding down their relationship. Elon Musk's xAI project also halted some of their projects as well. A lot of these companies are pulling out of this. And this kind of goes back to even the earlier conversation that we were talking about, it's all about data. That's what Scale AI does is it gets data ready for training these massive models. They have all kinds of people deployed around the world that are involved in the cleanup process and the labeling process. There's a lot of human labor that goes into preparing the data, and that Scale AI has been able to address that. And they were working with all the biggest companies in the world to help prepare the data for training. And now, of course, Meta gets to strategically kind of own that data funnel. 

[0:10:55] GV: This is though maybe still the bit that maybe some of the audience as well are scratching their heads on a little bit. Like, I'm still just trying to fully understand, okay, Meta own, let's just say effectively own, Scale AI, but at the same time - and Scale AI is providing the data. What changes when - is it around like how they're going to structure the data? Is going to be very skewed towards, say, LLaMa models versus something else? Why is there such immediate pullout from these companies? I'm still trying to get that concern. 

[0:11:26] SF: Yeah. I'm not a hundred percent sure they're either why the immediate reaction was to punch Scale AI as a fashion. 

[0:11:34] GV: I mean, a few of the alternatives at this point. 

[0:11:36] SF: Yeah. I don't know who else is in that. I mean, there's a couple of other companies, like LabelBox and stuff like that, that are in sort of the - there's a bunch of data labeling companies. But as far as I know, there's no one sort of operating at the scale of Scale AI, but there's a bunch of companies kind of focusing on that problem set. I don't know if the plan for the Googles of the world is to go and leverage those competitors in the space, or maybe they're going to build out, knowing some of this themselves, because they realize that they don't want to be dependent on a third-party vendor to provide this. 

I don't fully understand what the impact would be if all those companies were using - they're already using Scale AI. But presumably, when I give data to, say, AWS and Amazon, it's not like just because it's running in an Amazon server, anybody at Amazon can go and just look at that data. 

[0:12:22] GV: Exactly. Yeah, yeah. 

[0:12:24] SF: I'm not sure why you feel that compulsion to pull out. 

[0:12:27] GV: Yeah. It could also be - well, it's a pure economic thing. You know that that thing effectively is your competitors. And then you say, "Well, these big multi-million-dollar contracts, yep, there are potentially even billions at this point." Like, "Oh, they're gone." It could be a kind of power play where they're trying to maybe have Meta double-check their decision on that one. But who knows? 

[0:12:46] SF: Yeah, I mean, it could be also just because the space is so competitive. And presumably, all those companies using Scale AI have some proprietary data that's prior to that process. They do have concerns of something nefarious going on. We talked last time about the corporate sort of spying and espionage. So just because you sign a contract doesn't mean people aren't going to necessarily break it if the right incentives are put in place, especially if they feel like they can get away with it. Maybe there's just enough potential risk there that they want to go and seek somebody else to do that job for them. 

[0:13:20] GV: Yeah. And I think this is maybe a good time to move kind of onto our main topic, which is just the idea that big tech, the walls are going up. And why is this sort of significant? Well, I think it's fair to say maybe pre-AI or pre-GPT, we just didn't maybe see so much of this where the big tech was like really going at each other. They didn't really have a specific piece of lands, so to speak, to fight over. They were kind of like, "Well, you know, Google has its like ad thing and so does Meta. But Meta also has this other thing. And we're kind of all doing our dances around different things." 

[0:13:57] SF: Yeah, I think cloud is maybe the closest between major cloud providers. But all the biggest companies are kind of multi-cloud anyway. So there's lots to go around there. Certainly competitive. But I agree, I don't think we've seen this kind of level of competition since maybe the early days of social when Facebook or /Meta became really bid. There was a real existential threat to Google's busines and they really put a lot of resources and time and effort behind Google+ and the other failed Google social projects and stuff like that. I forgot what the circle - maybe that was Google+. But Google Buzz was another one. 

They had all the different social experiments, and none of them really took off. And then eventually, they kind of gave up on that as a business and went after other things. But since then I don't think I've seen - I think AI, at least in my lifetime, since I've been working in the industry, is probably the thing that I feel has the highest competition, and companies are just throwing crazy amounts of money at it. Either they're trying to steal each other's talent with offering massive amounts of money, incentives for people to come over, for talented people to come over. And it's like this land grab where I think that they see this as the future. There's going to be winners and losers, and they want to make sure that they're on the winning side. 

[0:15:17] GV: Yeah, for sure. We're going to dive into a few of these specific battles going on. This is sort of in the realms of the walls are going up. Who have we got against each other and on what grounds? The one that's also touched the big headlines, main headlines very recently, OpenAI versus Microsoft. And this is around legacy agreements that they had around Microsoft, as you've touched on, Sean, owning a significant chunk of OpenAI for a long time. 

However, there was this interesting clause in there called like the AGI clause. And this is around like at what stage does technology get to that stage. And the thing is, it's a sort of ironic problem because achieving AGI would automatically terminate the partnership. But surely you want to reach that stage if the idea is just to advance technology. And there's been something mentioned where many executives at Microsoft back in 2019, they thought this clause was nonsense. 

But again, reportedly, Satya Nadella was like, "No, we're too far behind on this. I imagine transformer, et cetera. We just need to do this deal. We'll figure out later." And well, here we are later, six years later, almost. Yeah. Yeah, how's this look? 

[0:16:33] SF: I mean, I think the challenge just on the AGI clause thing is there's not a clear definition of even what AGI is. 

[0:16:42] GV: Let's just make sure that AGI is - 

[0:16:44] SF: Artificial general intelligence. Essentially, we've reached the place where we have human-level intelligence. And some people argue like, "Hey, we can have a chat bot past the Turing test," which was sort of the original idea of a test created by Alan Turing many, many years ago of where essentially the idea is you have somebody that's interacting behind like closed door asking questions to either a human or some sort of computer. And if it's a computer and the human can't tell the difference between the answers coming from another human or from a computer, then essentially the computer's passed the Turing test. 

And for certain types of question and answer, certainly you could argue that something like ChatGPT could pass the Turing test now. And I think people have kind of tried to prove that. But at the same time, there's things that these models are like incredibly stupid. There's ways of tricking the model that would never ever trick a human. 

[0:17:37] GV: And even just like basic arithmetic and this kind of thing. You would expect to ask a human, most humans, "Hey, what's one plus one?" and get the right answer. And there's enough evidence to show you might not always get that answer from a model at the moment. 

[0:17:50] SF: Yeah. So there's these types of challenges. How do you even prove that you've reached AGI would probably become some sort of legal thing again? How do you enforce that? There's no clear set mathematical definition of what that is. I think that's a challenge. But the relationship between OpenAI and Microsoft has continued to get, I think, more and more contentious over the last couple of years or last 18 months, certainly. OpenAI has accused Microsoft of anti-competitive behavior multiple times. In a lot of ways, it kind of reminds me of the old browser war days where there was a lot of accusations against Microsoft in terms of forcing OEMs to have Internet Explorer installed. 

And Microsoft certainly has a history of anti-competitive behavior. Famously, they got brought up in the early 2000s on antitrust charges where they tried to essentially break apart Microsoft as a company. And eventually, those charges went away. But it really damaged Microsoft, their sort of - 

[0:18:47] GV: Reputation. 

[0:18:48] SF: Yeah, their reputation, sorry, during that time. 

[0:18:50] GV: And we've seen the Windsurf acquisition, of course, which sort of plays into this whole thing as well. 

[0:18:55] SF: Yeah. OpenAI acquired Windsurf, and then Microsoft has GitHub, and Copilot and all their IDEs. And suddenly you have this competition that's happening between these two companies, where there's a significant amount of stake from Microsoft's perspective and sort of investing in OpenAI. And OpenAI has been running on Azure Cloud. I think they've been trying to pull back some of their dependencies there to be less vendor-dependent. There are all these things that are happening sort of behind the scene. 

And then I think also, OpenAI over the last year or so has started to really pay attention to the enterprise and be less just about a consumer-facing application. I think enterprise is where Microsoft historically has really thrived as well. And that's where a lot of the dollars are. That, of course, creates more tension between the two companies. 

[0:19:44] GV: Yeah. We're going to move on to the next battle. This is Salesforce versus Glean. Now, what does this even look like? Salesforce owns Slack. That's kind of where we're going to go with this one. Glean is, let's say, more of a startup. This is interesting. We're actually seeing a kind of incumbent go against one of the more early guys. We did have an episode on Glean back in April with yourself, Sean. And this is funny because I remember listening to that episode, I happened to be in London listening to this episode. I went straight into a meeting with a friend who works at a very large PE firm. And we got straight onto the topic of AI and they said they just rolled out this white labeled search all our internal information. And I said, "Oh, who is it?" "Oh, it's Glean." It suddenly all kind of made sense. That's the context here. We don't need to name the name here, but the largest P from the world effectively is running on Glean. 

And here comes Salesforce or Slack saying, "Oh, you're not going to actually be able to, with any sort of practical means, now catalog Slack messages via the API." There's been like a pretty onerous rate limit on that. It's not like full block, but it is incredibly hampering. And it seems like Glean was the main target for this. Yeah, what do you make of that? 

[0:21:00] SF: Yeah, I think it's unfortunate because Glean, I think, was born out of Google. Google had created internally a product called Moma, which was to try to solve this kind of heterogeneous search problem of internal documentation across Google. Google has been a company over 20 years, 100,000 plus employees, there's stuff everywhere, how do you make it findable, essentially? 

And then a bunch of smart engineers from Google left, started Glean to take that idea and build a product around it. And it solves a real pain point for enterprises. Anybody who's worked in a large company can identify with this challenge of like, "Oh, someone said something to me. Where the heck is it? Was it in a doc, an email, a Slack message?" The enterprise search problem is really challenging for most businesses. And Glean, they really did a good job of addressing it, where it can index essentially all these different disparate data sources, give you one interface, and then search into it. 

And then they've done a lot of stuff since then. Now they have a bunch of AI tools. You can chat to it, ask questions and it'll go use sort of a RAG-based application behind the scenes to pull in the context and provide a proper response. I don't have to necessarily just search it and click on links. And then they also have the agent support now where I can build my own workflow and use my own internal docs and stuff like that. Really, really cool stuff, cool company. 

And I think it's kind of unfortunate from a competition standpoint that Salesforce is penalizing their ability to do that. But it kind of all goes back to the data problem. Like all these companies want to own the data because they own the data, then that's where the AI serving is going to have to be dependent on. And Salesforce is making a huge play around becoming a data cloud company, and they're going after lots of Snowflakes of the world. And they just bought Informatica. They're trying to get more and more data sort of in their gravitational pull as a company. 

And most of these companies are not interested in that data flowing out. They're only really interested in the data coming in. And for their products to work, they kind of need to own the whole world to make it work. I think that's unfortunate. But hopefully, there's some resolve to that where Glean figures that workaround or other products in the same space. I think Notion is also looking to go after smart - 

[0:23:11] GV: I was just about to mention Notion. Yeah, I went to Notion - or trying to like make big inroads over here. I believe their APAC office is in Australia. But they put on this very on-brand event over here. Rented out this very nice space in the National Gallery, and it was called Cafe Notion. And it had like jazz music and fancy breakfast canapés, which I thought was just a very nice spin on this kind of event where you're not pushing it to the evening with drinks and things, you actually push it to the morning and just make it very nice. 

All on-brands. But I think the big standout for me with them during the whole presentation was A, we're going for enterprise. And they make a big case about OpenAI in theory all running on Notion now. And then B, there's data integration. This is their big push. It was, "Oh, you can integrate all these platforms. And these ones are coming soon. But you can already integrate Slack and so on." 

And I'm just wondering where did they go when their whole offering is saying, "Look, we know you've got disparate data. You might not be all in Notion yet. But what can we do about that?" "Well, we can integrate." And I'm like, "Well, now we're one of the biggest providers of the integration that would probably help you," because I think most companies, if they're using Notion, it's not like they don't use Notion instead of Slack. That's just not a thing. It's actually Slack was trying to kind of recreate Notiony bits inside Slack. 

[0:24:26] SF: Yeah, yeah. 

[0:24:26] GV: I don't think it really worked very well. What does a Notion do in this case as well? 

[0:24:31] SF: I'm assuming they have the same throttling challenges, right? All the companies that are integrating there, besides Salesforce, which probably has a workaround through some internal API. I'm not sure what will happen. Or you have to get into a place where you have to end up with more of a strategic partnership with Salesforce where they unblock you. You're not just using sort of a public-facing API endpoint. There's an API that has a higher limit, or it's a different API, and stuff like that. 

[0:24:58] GV: It's interesting. I'm just on the basis that Notion roll in - they're pushing very much, "We can clean up these 10 SaaS platforms that you use. And even if you need to keep paying for them, you won't actually interface with them. All the data will just be pulled into us." And as we're going through right now, data seems to just be the thing that actually companies use to be kind of more open to exchanging because it seems like a sort of fair trade, like, "Well, this data for this purpose, and you need it for that person. Sure, let's have APIs." That's the name of the game. But as companies like Salesforce, clearly, I guess, looking at this and saying, "But why can't we be the people to give you the best context from our own data?" And context is the value here, seems to be, is what we're being told. So much data. Context is the value. 

The next one we're going to move to, it's more of a rumor, but it has been strongly, I guess, reported. And you touched on it earlier, Sean, Meta versus OpenAI. And this is in the sense of people. In theory, Sam Altman came out in public saying that Meta had been offering $100 million signing bonuses, which is obviously, I don't think we've ever seen a number like that in terms of at least publicly stated for signing bonuses. Meta haven't commented on this, and I don't know if that sometimes means it might be true. But I don't think we have seen sort of this kind of level for a while, at least or in terms of competition. 

[0:26:22] SF: Yeah. I mean, even if it's not a hundred million, I'm sure it's a lot to poach talent from some of these other companies. And certainly, I think there's been three fairly prominent researchers that moved from OpenAI to Meta recently. And I'm sure there's a movement all over the place. It's almost like you're in the space of professional sports or something like that where people are getting these huge contract owners, there's like a multi-year contingency to that. It's like $100 million, but over 10 years to join or something like, or based on your performance and things like that. You get into a place where you're not just acquiring companies, you're acquiring the talent to build the company that you want. 

[0:26:58] GV: Yeah, exactly. 

[0:26:59] SF: And it really goes back to this arms race that's happening in AI right now, whether it's for data or essentially the talent to make things happen. 

[0:27:07] GV: Yeah. And as you called out, it's 100 million. Okay, sure. That's like a headline. These things are never structured that way. It's not just like, "Are you joining?" And then on a Monday and on Tuesday, 100 million lands in your account. There's many ways they sort of structure this in terms of options, and as you say, kind of over a certain amount of years, and performance basis, and so on and so forth. But obviously, it's not unusual to see someone like Sam Altman just take the number just to kind of make a splash. And he's obviously trying to say, "Look, we have the best engineers. This is what people are trying to pay for them." It's marketing effectively. 

[0:27:41] SF: Yeah, that's another angle, right? In terms of the competition, the competitions, it's all the way up the stack, essentially, all the way at the hardware level. if you look at the hardware level. NVIDIA, AMD to some extent, are essentially providing all the chips. But a lot of the cloud providers and also big model companies are also looking now to figure out - they're investing essentially in their own chip designs because they don't want to have this vendor dependency. And a lot of AI startups are going multi-cloud because they don't want to have too much dependency on a single vendor. It's protective moves. They're trying to diversify their stock portfolio to some extent so that they're not beholden to any one company. 

[0:28:25] GV: And you have also come across the fact that Google has donated A2A to the Linux Foundation. How this sort of play into the walls going up? Or are these actually what was coming down? Or what is this? 

[0:28:37] SF: Yeah, it's kind of the opposite in some extent, at least from a surface level where we talked about Google Agent2Agent a couple of times on here. It came out a handful months ago. it's kind of a similar idea to Anthropic's MCP, but focused on inter-agent communication. And just recently at the Open Source Summit North America, the Linux Foundation announced the formation of the Agent2Agent project, which involves like AWS, Cisco, Google, Microsoft, Salesforce, SAP, and ServiceNow, perhaps others. 

I think it's good from we were kind of moving towards consolidation, like even Cisco, which had a somewhat competitive product is integrating A2A support into their agency's core. I think that's good from a standards perspective. Overall, it's good that companies that are looking to invest in using something like Agent2Agent. Now you have neutral governance, remain vendor agnostic, it can be a community-driven project. 

But strategically for Google, there's probably multiple reasons, I'm sure, to do this. It's not necessarily something they're going to directly make money from. And more adoption of Agent2Agent and other such standards overall is better for the companies that are already succeeding in the agent market because it essentially becomes another thing that people can build against, less barriers to entry. You want to reduce as much as that as possible, make it as easy as possible for people to build stuff. And if you own the GPUs, you own the cloud serving, or you own the models and the tokens being generated, then you want more people essentially building. 

[0:30:11] GV: Yeah, absolutely. And yeah, sort of on the same vein, we've got an episode coming up in the future around MCP security, actually. And actually, during the episode, we talked a lot about there is no sort of one place at the moment to go for validated MCP servers, for example. And maybe this is sort of the same thing. We're not the same thing. But if Google are kind of handing this off to the Linux foundation, they maybe hope they're going to be a good steward of that protocol, for example. More so than, say, Google doesn't look neutral. Unfortunately, Linux kind of does. Yeah. 

[0:30:47] SF: Yeah, exactly. I think one of the challenges with one of the MCP servers right now, too, is that the majority of them aren't managed experiences from a vendor that you trust. It's source code that's available from the vendor that you have to run yourself essentially, so you're taking on the burden of doing it. Or you go to one of these MCP aggregators that are running that on your behalf, and then you have to know like, "Do I trust this aggregator or not? How are they running this?" All that type of stuff. I think those are just signs of it being such a new thing. It's like a year old. It's going to take a little while. There's surprisingly few companies that have an actual managed MCP server that you can just use. 

[0:31:31] GV: Yeah, that's precisely what we touch on in that episode, where a lot of it's around, "Do you know who provided this server effectively?" Yeah, look out for that in the future. Yeah. 

I think we've hit the high notes on walls in tech. Obviously, we could have covered at least double that. But I think we've kind of covered what the last three to four weeks have shown us in terms of what's going on. I'm no doubt, in next month's news episode, we'll see some developments here as well as - yeah, who knows. 

Let's move on to - I think it's almost my favorite part of SED News, where we just get to look at Hacker News. Throughout the weeks, we just sort of have a scratch pad of things that we've maybe seen and kind of interest us. I might kick off with one that I just love these little projects that people do for no other reason than it's cool and fun, but actually they have Interesting uses at the end of the day. And this one was called my iPhone 8 Refuses to die. Now it's a solar-powered vision OCR server. And it was like I have to click this and see what this is. 

Thank you to the user that submitted that. This is basically someone who has rigged up, I think it's an iPhone 8. And their argument was this thing was sitting doing nothing, but it's actually incredibly powerful, especially with vision OCR from Apple, which is an on-device vision model. They've rigged this thing up like a power unit, which is powered by solar. This whole thing runs effectively from its own power. And this person says they've got like a lot of image processing needs. They don't explain what that is, but they say thousands of images, I think, at least per week that need to be analyzed and labeled. 

And this person claims it is doing it fantastically well. It's all self-powered. The iPhone 8 is remarkably stable and powerful for this purpose. And also, he says it's also a great conversation starter. This thing just sits on his windowsill, and people are like, "What is this thing?" Yeah, very fun. Very fun. 

[0:33:31] SF: Yeah, I also love to get a kick out of these little side projects and stuff like that. And a lot of this like side project stuff is kind of like how I started my interest in computer science, software engineering, and stuff like that many, many years ago. I have less time, unfortunately, to do those things now. But I loved how he talks about the amount of over-rengineering that he took this little side project and stuff like that. It's fun. 

[0:33:55] GV: Yeah. Super fun. I almost got it confused that he was using it for like birdwatching because of the fact that the phone was like on the windowsill. And I thought, "Oh, is this like you're using the camera as well?" But it's not that. I don't think it's like - he just happens to place it on the windowsill. He's clearly feeding images to it from - I think he's got like a microserver also sitting there, which is also powered by the solar, yeah.

[0:34:15] SF: I think it's like a laptop or something like that. I think it's a laptop. It's like not connected to anything. It's all like one internal network. There's actually bird feeders now that have cameras in them that will tell you which bird. 

[0:34:27] GV: That's why I was thinking about it, because I'm unfortunately a bit of a closet bird watcher. I was sort of looking at the thing going, "Oh, could I do this? Could I set up my -" I mean, I imagine you could. But I'd say, to be clear, I don't think he was using the camera piece of this. It was actually just using this as a pure piece of hardware that's surprisingly good at this purpose. And as he points out, he's feeding thousands of images for assessment, and he doesn't actually want these going to a cloud provider. This is all on-device. Yeah, pretty cool use case of that. 

He points out the fact that the vision OCR side of things was pushed by Apple pretty quietly. Maybe we'll see more with that. And obviously, Apple's not having a great time of it from the AI standpoint right now. People don't really associate them with any sort of strong AI offerings at the moment. It's kind of interesting that there's maybe some little hidden gems right now in terms of what's actually possible on-device for Apple. 

Sean, what did you find in your travails of Hacker News? 

[0:35:27] SF: I want to talk about this article, or highlight this article. It's actually written by a colleague of mine, Gunnar Morling, who is pretty well known from the Billion Row Challenge from a couple of years ago. And he's been involved in a number of open source projects. Now he's a principal technologist at Confluent. And he wrote an article called this AI agent should have been a SQL query, which was the number one article on Hacker News for some period of time. 

It was inspired by this talk by Seth Wiseman's, which was titled 'That microservice should have been a SQL query,' where he made the case for implementing microservices as SQL queries on top of stream processors. And then in Gunnar's article, he kind of explores that similar idea, but for AI agents. And he goes through this use case of processing research documents and being able to do agentic workflows all using Apache Flink and Flink SQL running on top of the stream processor. 

And he highlights a bunch of the various open source flips that have been contributed to Apache Flink over the last year or so, or a year plus, that bring in AI functionality, including one called FLIP-531, which is about Flink agents, which I actually co-authored. It's really interesting. I think he did a really, really good job of just kind of explaining where maybe you don't want to do all agents this way. But there's a lot of agents that are really just like, "Hey, give me input. I'm going to go do my AI magic box and spit out an output." Why do I need to stand up like a whole bunch of infrastructure to do that if I can just run that essentially as a stream job?

[0:36:57] GV: Yeah. And then I'm sure there's many examples of that where people are running things now through LLMs. And unfortunately, still RegEx or just an SQL query could have done it at least as fast, if not, certainly more efficiently. 

[0:37:09] SF: Or even traditional models. You see a lot of people are doing things like basic classification or sentiment analysis, and there's other models that can perform those tasks quite effectively. And I work in AI. I love building stuff on foundation models. But you don't have to do everything on them. You don't always need this massive power of this model to do things that have been - kind of solve problems using lighter weight techniques. 

[0:37:33] GV: Yeah, that's a great call out. Yeah, we've got an episode coming up in the future with Jigsaw Stack, which is a small model specialty company. Yeah, that's exactly what you've just mentioned, Sean. These are small models trained for very specific purposes, faster, more accurate for those purposes. Look out for that. That's a fun one. Very, very young companies. Great to sort of have a chat with someone thick in the trenches with that one. 

Okay, this is obviously security leaning. I always like to try and bring something security in if I can. This was ultimately on the Tailscale block. And it was submitted by user INGVE. And this is an article by - I believe it's the CEO, Avery Pennarun, and it's called 'Frequent re-auth doesn't make you more secure.' And I had to go into this one because I've worked a lot in this space and there's a product that we have under MailPass, which is pretty much exactly what he's talking about, which is we sort of took a look at credential management and it was like why are we constantly asking people to log in, bouncing them out of a service, changing passwords more frequently than is needed? 

And it also highlights the fact around device. Devices now can do a lot of, effectively, indirect checking. This is super interesting because, yeah, he's just talking about frequent logins are the wrong answer. You shouldn't need to - he says it's kind of from a bygone era of like internet cafes, which I think it's like pretty astute way of looking at it. We used to often, back then, use like shared computers, like in a school or in an internet cafe, if you can remember about that far. I think maybe schools and universities are maybe like the more sort of likely case there where you're like logging into your Gmail whilst on the school network or something. And of course, everyone just has a laptop now. 

[0:39:17] SF: Yeah, I think the last time I was in an internet cafe was I traveled to Europe for a conference in graduate school. And the computer I brought, the Wi-Fi wouldn't work for some reason. And so I had to go to an internet cafe in southern France and paid whatever it was to be able to use it for 20 minutes to check my email. 

[0:39:39] GV: Yeah. I think for me, it's also Europe. It's Croatia. It was just out of high school, interrailing, and arriving in a town and not having anywhere to stay and using it to go on one of these room share websites and trying to find someone at 6am in Croatia who will accept me in their home, which worked. That's a whole other story. 

Anyway, let's go back to the re-auth, which is the fact that, yes, device possession, which is kind of what we're talking about here, that you actually tend to now just always have the device. It's your device. You've probably unlocked it yourself. Why are you now also re-unlocking a platform? Why are you logging back into it? These very short session timings, like a day or even summer, like 30 minutes. Avery's point is like, "Okay, for banking, sure." There's probably a bunch of reasons, like just that nice extra, extra safeguard to have. But for most platforms, this doesn't make sense. And the other point he makes is something around passwords. Being asked to change your passwords on a specific schedule, usually by enterprise. That's kind of at least the classic case of that. It's something I've been absolutely against from the early days. 

I remember when we were running our startup, and my cofounder was like, "Oh, we need to change our passwords on -" back then, it was last pass. And I was like, "This is a really, really, really good password, and it hasn't been breached. And I'm just being forced to change it, I'm going to forget the password or it's going to become less secure." And that's his point. As soon you ask someone to change their password every couple of weeks. Yeah, I don't know. What's your kind of like experience with all these re-auth nonsense? 

[0:41:14] SF: I mean, I think that when you put all this friction behind using a product in the goal of increasing security, what it does is actually create additional security holes because people figure out ways of working around it because they have to get work done or they need to access whatever it is. And maybe in certain circumstances where it's not mission-critical that they get access to it, they just give up with it and you use something else that doesn't have that friction. And you end up creating a situation where I think people are - they run out of passwords, their good passwords. They start using passwords that's just easy to remember. Or they just start writing them down and putting them in places where maybe they shouldn't. 

And I think the other point that he makes in the article, too, is about how attacks aren't due to physical access to devices for the most part. Attacks getting in your email happen remotely. Logging you out on your laptop isn't necessarily helping prevent an attack that's happening remotely. It's kind of like people being concerned over - going back to the cloud providers in AWS having access to my data, breaches don't happen because somebody in a data center walks up to the computer that has the hard disk with the actual physical data in it and grabs it. They happen because someone left an API key in their GitHub repo and someone found it and they use that to get access to the data. Or someone gets access to a server that has log files in it that haven't been encrypted, and they grab the log files, and the log files happen to have a bunch of passwords dumped in it, and stuff like that. It's more of these remote use cases and these kind of password protection stuff of forcing people to log out, re-log in, really doesn't solve sort of the root of the problem. 

[0:43:02] GV: Exactly. And any modern platform, A, they're most likely using some version of derivative of JWTs. And most likely now, they're also using JWTs with refresh tokens, which basically means that, behind the scenes, this token is being refreshed. If someone is lurking around on your system, every hour, that thing's being changed anyway. This bizarre idea that logging you out and logging you back in has any effects on that is kind of nonsense, quite frankly. 

Yeah, I think it was just a pretty short article. I think he makes his points in record time. I really liked that one from Avery. 

[0:43:39] SF: It's the illusion of security. 

[0:43:40] GV: Yeah, we didn't touch on this. But yeah, he also mentions like MFA fatigue. And that's like an attack in itself now. Or basically do something to make your MFA pop up. And eventually, you just accept it because you're like, "Oh, I don't know why this thing's popping up. But I guess I should just use my touch ID because that's what I'm being told to do." And then, bing. They've got the access into the platform that they're looking for. Yeah, it's a very good one to look at. 

Anything else from your side on Hacker News, Sean? 

[0:44:06] SF: No, that's all. 

[0:44:08] GV: Cool. All right. That was a nice little tour around Hacker News. And in terms of looking ahead, we're going to obviously be back next month. This gives us a purely hypothetical. Do we have any predictions for what we might see playing out through July, I guess? 

[0:44:25] SF: I actually think things are going to cool off in July. Because in the US, in Canada as well, first week of July is kind of gone due to holiday, national holidays. Then a lot of people are on vacation and stuff - there's not a lot of conferencing going on in July. You have a sort of slowdown and the announcement cycles and stuff like that. I guess my prediction is there's not going to be a major, major announcement in July. I'll probably going to be 100% wrong on that. But that's what I'm going to say. 

[0:44:54] GV: What's my prediction? Interesting, the company Manus AI has been really blowing up, I would say, like over here. And I do feel it's just something though, where users are finding that they're spending too much money on Manus and they don't know what they're spending the money on for. I'm going to make a prediction that something comes out about Manus that is in that realm, that user frustration with Manus. 

I think it's an interesting product. I'm not trying to knock Manus here, but it was just seemed like an underlying theme where people are not kind of clear on how they're spending their credits there. Yeah, I'm going to go out on a limb there and just say something about Manus. Or if it's about something we talked about today, again, I'm just going to go out on a limb and say maybe the Scale AI pseudo-acquisition hits a roadblock. Let's see about that. 

[0:45:40] SF: Awesome. 

[0:45:41] GV: As always, great to catch up, Sean. I hope we've also given the audience just a nice little tour around what's been going on over the last three weeks. There was a lot to cover. Again, I hope we've hit some main areas for everyone and some fun things as well. Hope to see everyone next time on SED News next month. 

[END]