Latest / The Bleeding Edge Podcast / Weekly AI Brief 17 April 2026
Transcript
- Ralph Behnke: and to another episode of the Bleeding Edge podcast. I'm Ralph Benker and with me my co-host Emile Ogier. Emile welcome to the edge. Emile: Hey, Rolf. Thanks for being good to be back off to two technical hiccups. I'm glad we finally here. Ralph Behnke: Hey Rob, thanks for being back after two technical setups. We just need AI to do the show for us because I'm sure it will have less issues or maybe it won't. at the news, there's been some exciting developments. So let's dive right in. Actually ran a bit of a different protocol this week. ⁓ made a lot of mistakes with AI and resulted in some pretty interesting Emile: Hit me. Ralph Behnke: Well, outputs really. Once I built an app by mistake that actually allows you to basically clothes on models. So I built an app that had a model ⁓ and you put items, clothing on it. That's not at all what I intended to do. I was just running query in a particular. and anyway went off and built the app but this week I changed research to Claude I was using chat GPT for all of my news research ⁓ and I into Claude and ran a couple of prompts forgetting that it was actually connected to my email account and it the most incredible scan of ⁓ all of the news in my inbox ⁓ so we're going to be ⁓ about some of those stories today. And I might have just basically finished off my last use case that I was using chat GPT for. So I now well and truly committed to ⁓ Claude. So you go, Anthropic. You may as well send me the T-shirt. So let's jump It sounds very cool. I thought you were going to say when you made the clip. ⁓ Emile: It sounds very cool. thought you were going to say when they accidentally made clothes, you accidentally put clothes on it. Ralph Behnke: ⁓ Yes, I was looking at naked models. I was messing around with a looking at retail fashion trends ⁓ and to figure out how you actually use AI in retail fashion. so anyway, that's how that came about. so news items top of the one today ⁓ is Anthropics Project Glasswing now probably most people have heard this kind of Claude Mithos preview and this is ⁓ Anthropics ⁓ tool for cyber security. It's a cyber security model that can outperform most human experts at finding software vulnerabilities. I know you've probably got some stories around this one. So we'll leave it at that. This is going to be insane. Why? Because it probably marks the moment that we've got a capability that is definitely outpacing what literally % of all humans can do. Maybe actually % given the vulnerabilities that it's already found. So that's an interesting news story. ⁓ is back in the race. they've got news spark. And this a story that's of been the making for a while. Meta has got their super labs. It's headed up by Alexander Wang ⁓ and he was the founder Scale Meta board ⁓ ⁓ they've come in with a model and it's ⁓ fourth place just behind Gemini, Claude and GPT 5.4. So really interesting to see them launching that OpenAI, we actually about this last week. They've raised $122 billion at about $852 billion valuation. But one of the things that's been quite interesting is they have shares on the secondary market that have gone unsold. And is that that's basically cash that's now there ⁓ is available ⁓ to and buy those shares, but people are waiting to invest in Anthropic. And that me to the fourth news item. Anthropic have passed, imagine this, ⁓ 30 billion annual recurring revenue. And that obviously largely at the of open AI. So that's kind of fascinating because they are also beating open AI, which their recent figure, I think, is something like 24 billion in recurring annual recurring revenue. And obviously, Anthropic have also announced some pretty big expansion plans Google and Broadcom for multi gigawatt next gen TPU compute. TPUs those tensor processing units that Google make. You can't buy them, but you can definitely rent them. And so that's going to help because Anthropic have been struggling a little bit over the past month or so to serve all of the demand with Claude Opus 4.6. And so they definitely need some expansion. another item in the news about Anthropic Claude code source code was accidentally leaked. there's something like 512,000 lines that are available on NPM and so yeah, you can go Check their source code and build your own code apparently I might do that this afternoon So those are the top five news stories a couple of other ones that are probably worth an honorable mention SpaceX files for what could be the largest IPO in history. So This is the 1.75 valuation of SpaceX so Basically, yeah get cash ⁓ That's going to be an interesting one to invest in. I'm sure you will want a piece of that action email. GLM 5.1. This model is just going from strength to strength. A lot of people have never heard of Z.AI or GLM 5.1, which is their LLM model. I've trying to play around with it. It's basically come in a lot of engineering circles as that is absolutely comparable with Claude code. But as I say, I've been so in Claude recently. It's actually been difficult to even run other models alongside it. And else we talk about? ⁓ this is an interesting one in healthcare. AI psychiatry startup gets Utah approval renew prescriptions via a chat bot. Wow. So you can go ⁓ and basically get AI to give you your drugs. I think That sounds like really good idea. But who knows? Obviously have to watch that one to how it plays out. And then I guess, final two Gen Z is using AI just as much, but liking it a lot less. Very interesting. Z, obviously young people, AI rates are unchanged last year. I just know from my own teenage kids, are in incredibly suspect of anything to do with AI whenever I talk about AI they told me to or not so politely ⁓ up ⁓ Asked me whether I'm more in love with AI than I am with my family Of course I'm not But you maybe a little bit and so I think some interesting You know just generational stories out of that much negativity and anger kind of Parallels with social medias earlier arc I think then ⁓ one Google AI overviews gets answers wrong 10 % of the time How can that even still be but there you go bit of a watch out for us AI still go very wrong and ⁓ I've some interesting experiences this week, which I'm sure I will share on This episode so Emil lots of to pick from there. There must be something that caught your attention Emile: Always too many things, Rob. I'll just focus on a few. So with the cloud code leak, it's quite well publicized. People are focusing on mostly what they can copy and security vulnerabilities. But I think there's something underlying that's even more interests that people aren't really speaking about. it's, if you dig deep into the in between the lines, you can actually see. Ralph Behnke: always doing things. I'll just focus on a few. So with the cloud code leak, it's quite well publicized. People are focusing on mostly what they can copy and security vulnerabilities. It's something underlying that's of public interest that people aren't really speaking about. it's if you dig deep into the code, between the lines, you can actually see what Emile: what ⁓ strategy is ⁓ they've got a new agentic model that's effectively monitoring you in the background, predetermining what your behaviors are, your preferences, your likes and your priorities. So I suspect what the future is going to look like, it's going to be even crazier than the world we're living in now, which is getting all these wonderful agents to do our bidding. But instead of ⁓ out Ralph Behnke: Anthropics future strategies and they've got a new agentic model that's effectively monitoring me in the background, predetermining what your behaviors are, your preferences, your likes and your priorities. So I suspect that the future is going to look like it's going to be even crazier than the world we're living in now, which is getting all these wonderful agents to do our bidding, but instead of figuring out, thinking what we need. Emile: thinking what we need, they figuring out what we need automatically. So it's going to be a whole new world on top of this layer that we're trying to familiarize ourselves with. Pretty, pretty crazy stuff, really. Hopefully that makes sense. Got any questions that we know of? And along with cloud code, ⁓ do love it myself. I am trying to get more involved in it and I don't mean to burst your nirvana bubble, but there are other things that are. Ralph Behnke: day figuring out what we need automatically. it's going to be a whole new world on Pop! The Slayer that we're trying to familiarize ourselves with. Pretty crazy stuff, Hopefully that makes sense, got an equal amount of know-of. And moving along with the through, I did note myself, I am trying to get more involved with it. I don't mean to get in a blown-up bubble, but there are other things that are... Emile: And Propeco doing that I think is opportunity for us to maybe temper the excitement and look at things on balance and I'll cover that new story in a bit. But moving on to my first story, Rolf, unless you've got any other questions on those points or viewpoints, I'll crack ahead. Yeah, so I think we'll keep off this slide just with a... Ralph Behnke: I think it's us to make a good team for excitement and to look at things on balance and I'll tell you a in a bit. maybe to my first story, I'll let you know the questions on the point or viewpoint. No, let's go. Yeah, so maybe we can decide just a bit. Emile: little bit of an exciting story, a little bit of a concerning story. It's about behavioral response issues to AI, much like what you mentioned with your children and your family and younger people. There's, you know, there's a bit of kickback responses and there's also some negative behavior in the workplace. Just imagine walking into your office and making a cup of tea with the person sitting next to you, completely unaware that they've secretly building a digital clone to get you sacked. Ralph Behnke: The of the concerning story is about behavioral response to AI, much like what you mentioned with your children, your family and younger people. There's a kickback in responses and there's also some negative behavior in workplace. Just imagine walking into your office and they're sitting next to you, thinking unaware that they've secretly built a digital phone to get you sat. Emile: That's exactly what's happening in China right now. As desperate workers have started quietly training artificial intelligence agents to learn their colleagues daily routines and automate them out of a job. The tactics are absolutely ruthless because they secretly document every single task their deskmate performs and feed it straight into the to prove to their management exactly how fireable that person is. Ralph Behnke: That's exactly what's happening in China right now. Desperate workers have started quietly training artificial intelligence agents to turn their colleagues' daily routines and automate them out of the job. The tactics are absolutely ruthless because they secretly document every single task their guest made, performs, and feed it straight into the model, and prove to their management exactly how horrible that person is. To the terrifying degree that corporates are hard to beat. Emile: It's terrifying new breed of corporate survival where employees are weaponizing automation against each other just to avoid being the ones being put on the chopping block themselves. This invisible technological arms race got so intense that developers actually had to fight back by creating a viable counter network known as the anti-distillation network. It operates as digital camouflage designed to deliberately confuse these rogue models and stop Ralph Behnke: These are weaponizing automation against each other, just to avoid being the ones being put on the chopping block themselves. This invisible technological arm of grace called Springtents, the developers at the head of Python, are creating a viable counter network, known as the anti-distillation network. It operates as a digital chemical art design, deliberately confused the growth models of predatory co-workers. Emile: predatory code workers from successfully copying human workflows into a machine learning script. We are watching the modern workplace devolve into a silent battlefield where the biggest threat is no longer the algorithm itself, but the colleague who knows how to unleash it. If the daily ground now requires defensive software just to your peers from turning your career into code, how long might it be until the entire office is just bots trying to outsmart other bots? Ralph Behnke: successfully copying human workflows into machine learning scripts. We are watching modern workflows evolve into a silent battlefield where the biggest threat is no longer the algorithm itself, the quality of those hard-to-unleashers. If the playground now requires the best software to stop the pillars from turning your career into code, how long might it be until the entire office is just bot-struck outsmart other bots? Well, I've been quietly recording every word, Emile, and I have to admit I have now built the Emile bot. ⁓ ⁓ no longer need you on the show. Emile: You for coming. Ralph Behnke: Thanks for coming. with that, we'll switch to the email bot. Yeah, it's a zero sum game, unfortunately. Ridiculous that that kind of culture in corporate. Clearly, if can automate somebody else's job, they can automate your job. So I just think that sounds ridiculous. I can't even imagine the culture that they exist in that is causing them to behave in that way. just very odd behavior. And think going to kind of go, I wrote an probably about eight to nine months feels like years back in AI time. But it was really about the fact that AI works best when we really start to pair it with some really capable people. And hopefully ⁓ the point that can to prosper and thrive with the people that they've got. and actually more without having lose their entire workforce, which I think is incredibly dangerous. What happens when have moved everything over to various bots and AI tools, and we then completely hostage by Anthropic or OpenAI or ZAI or whichever platform that we've with our operations to, which essentially outsourced our entire workforce to them. And they start doubling their prices. You we know that this is a strategy that companies use regularly to actually increase profits. obviously won't be able to exit. You know, we have choices these days to move to another service provider, but it's going to be that much harder when your entire operational execution is managed by a handful of companies. I just think it's a very dangerous place to be and the combination of human and AI still feels to me the most robust. But again, you know, maybe there is actually an extinction in terms of work that will be transpiring over the next five years. It certainly feels like a huge potential for to be the case. But somehow, you know, working AI in the way that I have been, installing it in teams, at the kind of interface between humans and the kind of the physical part of I'm still whether it's actually possible to, you know, wholesale outsource that we're doing. I think there is a lot probably, you know, more than people realize that AI can handle and will handle. But ⁓ yeah, would like us to see we're really building a very clear role for humans in of this ⁓ and really level of understanding of how AI is actually working because otherwise we will have lost the plot that we won't understand how anything in the world is actually being run. And I think that's a very dangerous place to be. You know, we get kind of X-class flares from the sun all the time and it takes out significant portions of our infrastructure. have a ⁓ delicate ecosystem powers the internet and it's to outages. We've seen it with Iran bombing data centers in the Middle East. There've been a of outages with AWS. I don't know why I'm smiling at this, but it's just one of these stories that just makes me think we are too fragile to go all in ⁓ with this level of automation. So anyway, that's my view on that story. Emile: Yeah, Rolf, I'll share your concerns as well. substitute our trading convenience for our own capability. It's one to watch for sure. Ralph Behnke: Yeah, I'll show you as soon as this will be... substitute or trade people for convenience or own capability? know. It's all up to watch, sure. Yeah, we need resilience and resilience needs redundancy and I think, you know, we are the redundancy. We still need to know how to do the job. But anyway, we're going to go deep down that rabbit hole. And, you know, as I say, they'll just be, you know, with one X-class flare or one well-placed bomb away from discovering that, you know, we just can't rely on it. Emile: Exactly. Ralph Behnke: know for myself, we some big outages last year and I was in a community that lots of users freaking out because they literally couldn't do their job because their friendly neighborhood AI was unavailable. And this is like next level. We're not talking about internet outages. We're talking about people can't do anything. Like literally you're completely redundant. I think I've probably hit that point. not a lot I could do if I wasn't able to access my various tools. I have to really think about retooling, which would take some weeks, maybe even months. And there's lot that we wouldn't be able to do. And so maybe already crossed the Rubicon or that one. But let's move on to another story. We'll move away from the... Corporate skulldaggery that is going on in China and hope that that doesn't infect the rest of the world I want to jump into this project glass wing at Anthropic because I think it does kind of feed into a little bit of that story that you've just told because You know Anthropic have built this tool which they're calling it. It's built around something called the Claude Methos preview. It's an unreleased cyber security model And the company actually says it can outperform most human experts. And I just think that's an absolute bullshit statement. what you already quoted in an earlier conversation that we had is that the model has found kind of day zero vulnerabilities software that's been around for multiple And this in the first scan has gone off and found vulnerabilities that can be exploited to basically take over these systems. the fact that no human has found that tells me that it's already better than what we're able to do. So move Anthropic that actually released access to this tool ⁓ to vetted So if you are AWS, Google, Microsoft, Cisco, CrowdStrike, JP Morgan Chase, then you actually have access to this tool and you can go away obviously run various scans. ⁓ I'm hopeful that this is good because between Google, Microsoft, AWS, Cisco, this is probably a significant portion of infrastructure that all using, especially from an enterprise perspective. And hopefully they patching and fixing and closing loopholes so that those vulnerabilities don't But of course, this tool is so capable. mean, mean, essentially it sounds like it can literally hack anything. I don't know whether we are actually going to be able to build a resilient and secure and safe infrastructure, you know, given we've now got these kinds of tools. I from our own endeavors in web scraping, there were few websites that were very, very resistant to me their information. And I... basically encouraged Claude on its journey to get that information and what I is absolutely nothing is safe. can put whatever blockers you want to stop bots scraping your site but ultimately Claude will find a way and this just feels to me like ⁓ the next level of you know, interrogation, if you like, or vulnerability extraction. there go, you know, Anthropic have now built a tool that absolutely is able exploit anything and everything. So we'll have to see, it sounds like a great opportunity though to sell this service to enterprise. So instead of of, you know, buying hacker services to test resilience and security your system. You can just pick up this Anthropic tool have run your various scans. so hopefully, know, they'll able to get some benefit out that. But just very interesting to see that this thing is so powerful that they weren't actually released it to market. So I'd be interested in your point of view on this one, Emil, like what have you heard? ⁓ Emile: Yeah, Ralph, well, to add even more color to the story is the original creator of Cloud Code is saying publicly that the response to the system is alarming his actual kids, which causes a lot of room for concern. And one of the key objectives they gave the model when they ran it was Ralph Behnke: great code is saying publicly that the response to the system is allowing his actual code, which causes a lot of concern. And one of the key objectives of getting them on with a rabbit was they need to know if it's broken out of its harness. And it know, we put you in a sandbox. Emile: They need to know if it's broken out of a tarnished and was, you we've put you in a sandbox, you shouldn't get out, hadn't ⁓ any impenetrable weaknesses for over two decades. you know, I think looking at that all holistically, ⁓ is room for concern and ⁓ Ralph Behnke: that haven't had any penetrable weaknesses for over two decades. you know, think all of you can have that all understand that it is room for concern and I don't know where we go too much. These moles are appearing to be way more powerful, becoming way more powerful, are quicker than we haven't seen. Emile: I don't know where we go to next. These models are appearing to be way more powerful, becoming way more powerful, far quicker than we had anticipated. Ralph Behnke: Yeah, definitely a pivotal moment and worth just explaining what a zero-day vulnerability is. It's a security flaw in software that the people responsible for fixing it don't know about yet. So the name actually comes from the idea that the software has had zero days to work on a fix because they didn't the problem exists. So that means no patches, no updates. No official defense available the moment that the vulnerability is discovered or there ⁓ you go. Joel. And last thing to add on that ⁓ is discoveries all made on the first run. Well, this is it. mean, the first scan basically identifies this. And I have say, we recently ran ⁓ a scan for client just on the website. It one of the things. Emile: Thanks, Rolf. And last thing to add on that is discoveries were all made on the first run. Ralph Behnke: we did it. I always run scans for my clients on whatever ⁓ and systems have and actually we found ton of stuff that was quite powerful really just you know 10 minutes worth work and it just kind of goes to show There are billions of websites out there, which means there are billions of vulnerabilities, or maybe even trillions at this point. even Claude Code with a very well-defined skill already significant capability in exposing And so definitely, know, with ⁓ using Claude for but this kind of project mythos or glasswing is next level and I I don't where it's gonna go, but yeah, it's pretty scary and pretty exciting ⁓ all at same time. If I had time, would open a cyber security business right now and try and license that software because I think everybody's gonna wanna run it. So let's move on to related story really. So I actually wanted dive into a little bit more detail on the timeline of Claude Anthropic versus OpenAI I just think this a race that is getting so interesting. a couple of kind key comparisons the two companies, and Anthropic. at matched stages. So OpenAI founded in December 2015, founded in April 2021. The first product from OpenAI GPT-3, ⁓ which came June 2020. And the first major product from Anthropic was the Claude API March 2023. consumer breakout, ChatGPT, was really November 2022. That's when everything suddenly, you we learned about LLMs. And I remember myself, know, jumping on there and going, wow, this is pretty impressive. The Anthropic Enterprise breakout came Claude and which was around kind of, yeah, I guess March 2024. And the other comparison, OpenAI hit 1 billion annual recurring revenue in June 2023. That was six months after the initial version of ChatGPT that was launched. And Anthropic hit 1 billion in Q2 2024. That was 15 months after their launch. They were definitely a little bit slower. And this is where things start getting really interesting. Um, Q1, 2025, we had open AI at 10 billion annual recurring revenue and Q3, 2025, and Tropic were at 19 billion. Um, so, uh, uh, basically they took them 14 months, uh, to get to 19 billion from an annual recurring revenue of 1 billion. Insane growth. mean imagine adding 18 billion in 14 months. That's just staggering. Valuation of open AI. So as we said earlier in the news story kind of 852 billion and the valuation of an anthropic is six. What let me see if I can see that number. It's 61 billion. So that's based on their kind of series F. September, 2024. So there must be worth way more than that by now. I need to get an updated number there. So obviously OpenAI have got 900 million weekly chat GPT users. I haven't actually got the number for Anthropic, but their route to market is obviously API first enterprise and dev. So definitely a very different focus, a very different kind of tool. But it's just fascinating to benchmark the two seeing that, open AI, you ⁓ kind of kicking around 24 billion annual recurring revenue and anthropic surpassing them. I've seen recently that open, I have also now launched, I think like a hundred euro or a hundred dollars subscription. I know, you know, a lot of users are on the Claude max plan, which is about 200 euros a month. you know, staggering to think that people are forking out that kind of cash when you can do so much on just a 20 euro or 20 pound or 20 dollar subscription on gpt but there you go i guess if you're a developer and you the tool then you know anthropix 200 euro a month subscription is the way go so Not sure what you make of those numbers, but I just think the market seems to be kind of moving towards betting on Anthropic more so than OpenAI. doubt people still see a future for OpenAI, but Anthropic just seem to have captured the enterprise market and embedding themselves there almost in way that... going to be very difficult to see them being unseated. Even ⁓ with mean, Google have got a of enterprise clients, but I know Google have found it hard to break in over the years and unseek Microsoft. know, how many companies do you go in where you're actually using Google workspace? Not that many. Over the years, 10 % of the corporate clients I've worked with. have actually gone down the Google infrastructure route for their office applications. so it's just incredibly hard. think chat, GPT, open are going to find it very well, they are finding it very hard to sell into that space, but somehow Anthropic have unlocked the enterprise market. And I that's really driving this incredible valuation of 30 billion in annual recurring revenue. Emile: Yeah, it's interesting stuff. I've got so much to speak on this. think I'm actually going to create a different set of news piece for a future episode. And more specifically, if they are, if they these IPOs to release within the same year, not only are they significantly large, but they're larger than any IPOs that have. Ralph Behnke: Yeah, it's interesting stuff. I've got so much to pick on that I'm actually going to create a different music piece for a future episode. More specifically, if they aren't putting these RPOs all to release within the same year, not only are they significantly larger than any RPOs that have come before it in the markets, we're not even talking about slightly more. Emile: come before it in the markets. It's We're not even talking about slightly more. We're talking about, to give an example for SpaceX and XAI, it's going to be more than double the biggest IPO in the whole of history, which only happened once in 2019. So just to wrap your head around that size is bigger than all IPO cumulatively put together. Ralph Behnke: talking about an example for basic XAI, it's going to be than double the biggest IPO in whole of history, which only happened once in 2019. So just to wrap your head around that size is bigger than all IPO you can put together. And not only that, not only we have one IPO, we might be having three. Emile: And not only that, not only are we having one significant IPO, we might be having three. And it begs the question, where's all that liquidity going to come from to pay for this in markets? And traditionally, when IPOs get released, companies typically release about 15 % of their share structure into the markets. And in the case of SpaceX, it's only going to be 3%. So it's artificially going to create. Ralph Behnke: takes the question, where's all that liquidity going to come from to pay for this in markets? And traditionally, when IPOs get released, companies typically release about 15 % of their share structure into the market. And in the case of space, it's only going to be 15%. So it's arguably going to create scarcity in the market. And those liquidity funds are going to pass down through investment funds, which Emile: scarcity in the market and those liquidity funds are going to be passed down to investment funds which by way of example because SpaceX is moving into NASDAQ one of the pre-requirements they made was that that they can move into markets within 15 working days now typically when companies come into the IPO it takes a year for the market to decide Ralph Behnke: I wear a good sample because SpaceX is moving into that. One of the previous ones they made was that they can move into markets within working days. Now typically when companies come in with an IPO, it a year for the market to recycle its worth before we can actually move into the index one. And it's good to remove into index one. It has to be purchased within the basket. Emile: what it's worth before we can actually move into the index fund. And as soon as it moves into index funds, has to be purchased within the barcels. So anybody that has a generalist fund will be investing by default, law, because these investment funds have to get involved and purchase the share price. Ralph Behnke: So anybody that has a generalist fund will be investing by default by the law because these investment funds have to get involved and purchase the share price. So hopefully that will make sense. I'll break it down more in a future episode, but ultimately what's going to happen is we might all be directly involved in investing in very expensive IPO's. Emile: So hopefully that all makes sense. I'll break it down more mechanically in a future episode. But ultimately what's going to happen is we might all be directly getting involved in investing in very expensive IPOs. I'm necessarily saying those who are investing in it won't make money in the future. But it's creating a scarcity problem. It's going to create an inflation price problem. Ralph Behnke: I'm necessarily saying because we are testing and we're to make money in the future. It's creating a scarcity problem and creating inflation problem. Yeah, I can see that happening. mean, just to tie these two stories together with kind of open AI and anthropic and SpaceX. I've got some interesting numbers here that I think frames what you're saying very well. So if you look at it two years ago, OpenAI held 50 % of the enterprise API market share today. And at the time, Anthropic held 12%. Today, OpenAI has 25 % and Anthropic has 32%. And that's driven mainly by Claude 3 and Claude code, actually. course, ⁓ also have the Chinese models that ⁓ way cheaper. You know, in some cases, 90 % cheaper than the API of of code. so going, you but the point is that those numbers are massive and will drive significant valuations for both of those companies because they have got such a share of the enterprise market. SpaceX, on the other hand, It has obviously got Grok AI embedded in it. SpaceX merged little while ago with ⁓ XAI, XAI the X platform plus Grok. And obviously, you know, Grok's is kind of sitting alongside ⁓ Starlink's cash flow because they're one the same company. So SpaceX, OpenAire and Tropic could raise something like 150 billion in 2026 alone. mean, that's going to suck a lot of investment capital out of the market. there is, I'm hearing kind of warnings that it's just going to suck the air out of, know, investment in smaller listings. So if you had planned your IPO for this year, I don't know, maybe you're going to have to wait or get on it very quickly because otherwise there might not be any cash for you. So I this is all going to play out. But as you said, I've got some news here that's quite interesting. This IPO will be the first large scale public pricing of an AI model embedded with a profitable infrastructure So this is obviously talking about Grok ⁓ embedded in ⁓ SpaceX. So I can't wait to see what's going to happen with all of that. It's pretty interesting, know, if you've already got an evaluation of 1.75 trillion, which is what SpaceX is valued at. And they're also looking at a 75 billion raise. So fascinating. I think Saudi Aramco was the previous IPO with that was the record. 29.4 billion. I mean, Emile: Correct. Ralph Behnke: peanuts in comparison to what we're talking about. So we're next to you're looking at 50 to 60 billion razors. Yeah. And, you know, let's just think about SpaceX. What we're talking about is the entity that goes public will basically include, if I just get my list over here, it's a staggering list. So Emile: And each of the AI labs were typically looking at 50 to 60 billion razors, just for context. Ralph Behnke: It spans orbital launches, Starlink, is 10.6 billion revenue in 2025, 54 % EBITDA margins. That's very exciting. 10 million plus subscribers in 150 countries. So essentially you've got XAI, Grok, they did an all stock merger, obviously in February. And so... you know, plus all the SpaceX revenue. think that's going to be very, very to see what the market ends up pricing it at. Interestingly enough, all the original XAI co-founders have departed now. I knew a lot of them had left. I didn't know that all of them have left. ⁓ But still is ⁓ burning a a month. Can you imagine that? That's that's what are spending that money on? Emile: Yeah, well, they've got the biggest data centers in the world, Colossus 1 and 2. Ralph Behnke: Yeah well they've got the biggest data sectors in the world right? Must be all that water and electricity bill that's costing them that much. Anyway so there you go fascinating story on SpaceX I love that have you got another one for us? Emile: I do indeed. Thanks, Rolf. So picture this, star of the fifth element and resident evil decides Hollywood is not enough and quietly drops what might be the highest scoring artificial intelligence memory system that we've ever seen benchmarked. out Milo Djokovic and a friend Ben have created a free tool. Ralph Behnke: Indeed. Rob. So picture this facade elements of Resident Evil. Besides, Hollywood is not enough and likely drops what might be the highest soaring artificial intelligence memory system that we've ever seen in Benchmark. Turns out, Milo Prokofievic and a friend Ben have created a prequel called Mempalace and it's currently rocketing towards 5,000 stars because of the attempts of one of the most stretching bottlenecks in modern Emile: called MemPalace and it's currently rocketing towards 5,000 stars because it attempts to solve one of the most frustrating bottlenecks in modern tech. The core problem it tackles is that the language models have a notoriously terrible time remembering your past interactions, forcing you to start from scratch constantly. MemPalace fixes this by mining all your previous conversations and projects. Ralph Behnke: The core problem it tackles is that the language models have been at a policy to equal time remembering your past interactions, forcing you to start from scratch constantly. Limpalance fixes this by mining all your previous conversations and projects, taking inspiration from the ancient human memorization technique of building a mind balance to spatially map out information. The creators discovered that proposing a strict inspiration Emile: taking inspiration from the ancient human memorization technique of building a mind balance to spatially map out information. The creators discovered that imposing the strict spatial structure alone improves the model states of retrieval by a staggering 34%. Imagine that. Ralph Behnke: alone improves the model state of retrieval by staggering 34 % of the imagineer. The data is particularly organized into wings for people and projects, for floor types of memory and individual rooms for ideas, making absolutely everything stretchable and fun at one time. The theory is where it really gets genuinely a bit eerie. Emile: Your data is meticulously organized into wings for people and projects, holes for broad types of memory and individual rooms for specific ideas, making absolutely everything searchable and findable on demand. But here is where it really gets genuinely a bit eerie and fascinating. to be fair, because they actually invented a brand new dialogue strictly for the machines to communicate. It's not designed for humans. This bespoke machine language is entirely illegible to humans, yet manages to compress months of sprawling conversation into mere 120 tokens. When tech reviewers tested it against personal notes and massive databases like Obsidian Vault, Ralph Behnke: that way, to be fair, because they actually invented a brand new dialogue, strictly for machines to communicate. It's not designed for humans. This bespoke machine language is entirely applicable to humans, yet manages to compress months of scrolling conversation into an year 120 times. When tech reviews tested it against virtual nodes and mass databases like Obsidian Vault, it should prove the data effortless and interconnected smoothly Emile: It chewed through the data effortless and integrated smoothly with platforms like Lord to pull up specific past projects within seconds. It really makes you wonder what happens when our digital assistants develop a flawless, permanent memory of everything they have ever told them. I don't know, Rolf, do you think we're actually ready for a future where our tools never forget a single detail? It does make the question, what information would you really be comfortable letting AI store in its purview? Ralph Behnke: platforms like code to pull up specific past projects within seconds. It really makes you wonder what happened when your art digital assistants developed a flawless, permanent memory of everything they've ever told them. I don't know, Ralph, do you think we actually created for a huge, aware art to code and refer to a single detail? It make the question what information would you really be comfortable letting AI solve and it's very, very, very, very, Emile: memory of Thomas. Ralph Behnke: Yeah, it's going to be a game changer for sure when everything is stored. You know, we're so fallible. We like hallucinate more than the AIs do. know this because my wife and I often sit and have a conversation and I say, all right, so this is going to happen. She goes, well, you haven't told me about that. I went, yeah, I told you. This is when I told you this is what was happening. We were here. We were in this place. She goes, no, no, we weren't. And course, vice versa, the same happened. We both just kind of, you know, probably don't listen to each other anymore. Who knows? Maybe you can go for 20 minutes and listen. Yeah. Yeah. So I can't wait for that because we'll be able to rewind. This is what I said, you know, told you so. Of course, it'll be a game, a kind of a zero sum game because we are both so fallible when it comes to remembering what we've told each other. But yeah, I mean, you know, this is absolutely the key and Emile: Maybe you can both switch on your mentalities. Ralph Behnke: To be honest, think this is the significant gap in AI because so much ⁓ of work that we do as humans happens in between those moments that we actually enter some information to a system or even an email. anything that we commit text, AI has a good chance of basically ingesting, understanding and responding to. a lot of what we do is also in the conversations. did we make the decision to raise that order or cancel that order? AI just doesn't have any of that operational knowledge. I think, I've said when ambient AI comes around and you've got always on recording of all of your audio, then that immediately fills the gap. It actually understands the conversation of how you got to that decision forget about your colleague that is sitting there trying to automate you out of a job. The will be able to do that for you because it'll literally understand your thought process. Now that's partly happening. One of the that I see that's very interesting is transcription tools. I'm amazed at how many people are still not fully utilizing transcription tools. But wherever I have a client meeting, for example, I will always try to use a transcription tool and record that conversation in whatever AI I'm using so that it's got a continuous memory of what we discussed. that just helps ⁓ to obviously make sure that actions and next steps, but also decision points and the context of what we're talking about in terms of goals and what we're trying to achieve. is basically held in a vault and I think that helps us just stay on track and it helps the AI stay on track. so MenPalace, these kind of tools, I think are really fundamental and actually they will become the true enabler of a kind of a... a generalized super intelligence because nothing without contextual memory. so I this is part of the really the next of evolution. As you know, in own development, I've using a very similar kind of system. We call it Omnicortex by Tony. ⁓ it remembers absolutely everything we hand off to it. And so that is now working not just in terms of, you know, individual memory, but actually memory of small of people that are, you know, working on developments. And so I think this is a very exciting evolution in itself. you have a team of 20 engineers and absolutely everything that they are doing. basically entered into a memory and off sessions. But only that, the evolution of it that Tony is now currently working on ⁓ is basically being able run ⁓ chat sessions and for chat sessions to interact with each other and inform each other about what they're doing. not only are we talking about memory, we're talking about kind of, you know, co-creation of memory and infusing of activity so that it's blended memory. He calls it breathing brain and so ⁓ don't quite know where we're going to go with that one but really really powerful in terms of keeping multiple agents in sync, multiple engineers in sync and kind of this you know omnipotent awareness of the body of work that's being engaged in. so I think that there's even a next level of this memory, which is kind of, you know, continuous co-memory creation that we're starting to see. And I think that that's going to be, that's to have a big impact on how enterprise actually, you know, leverages AI in the future as well. Emile: Absolutely, Rolf. Memory is key. Ralph Behnke: Absolutely involved with Mimly and Aiskei. No, if only if I could remember what I was going to say next. Why don't you try and remember and I'll cover another story. Go on then. I don't mean to be in manga, but I do think it's thing that done with you and I was saying earlier regarding them, I do absolutely love them. They are my favourite choice. I am going to keep them. Emile: Well, while you try to remember, I've got another story. And I don't mean to do Manga of these, but I do think it's good to have a balanced view. what I was saying earlier regarding Anthropic, I do absolutely love them. They are my primary choice. I am going to live with them. I think before we start putting all these. Ralph Behnke: I think before we stop putting all these AI labs on pedestals, we are aware of the thorns that we move along ahead and through the treacherous path of getting better at what we do. Without the redo, developers have been playing it down for weeks. Artificial intelligence coding systems are suddenly getting incredibly lazy, but everyone has dismissed it as bad vibes for Reddit learning until... Emile: AI labs on pedestals, we're aware of the thorns as we move along ahead through the treacherous path of getting better at what we do. With our total ado, developers have been complaining now for weeks that their artificial intelligence coding systems are suddenly getting incredibly lazy, but everyone just dismissed it as bad vibes or Reddit moaning until an AMD directly finally brought out receipts. Ralph Behnke: AMD directly finalized full-part proceeds. Steeler, Lorenzo analyzed a 7,000 core code sessions over three months. And what do you think they uncovered? Staggering 67 % degradation performance that proves that developers would have to do the rights all along. Anthropic recently implemented a thinking reduction feature and the session log showed that the moment they did the artificial intelligence started. Emile: Stella Lorenzo analyzed 7,000 fraud code sessions over three months. And what do you think they uncovered? Staggering 67 cent degradation in performance. It proves that developers were absolutely right all along. Anthropic recently implemented a thinking reduction feature and the session log showed that the moment they did the artificial intelligence started defaulting to the cheapest and easiest actions available. Ralph Behnke: default into the cheapest and easiest action available. The data reveals that the model is now handling complex engineering tasks midway through and docking ownership for its failures at rates that sitting at absolute zero before early March. The most alarming shift is how this could be handled as an actual code back because it used to read through a file on average over six times before making any changes. But now that that number of comments has just... Emile: The data reveals that the model is now abandoning complex engineering tasks midway through and dodging ownership for its failures at rates that were sitting at absolute zero before early March. The most alarming shift is how the system handles actual code base because it used to read through a file on average over six times before making any changes. But now that that number plummets it to just two roles. Ralph Behnke: to do all, meaning edits performed blindly without even reading the target file. If I rock it up by 500%, it means that the tool you trust to build your software can continually modify crucial architecture without looking at the bill. This is not just a glitch, but a fundamental warning about our global resilience of black box systems. I can secretly degrade their reason to get overnight while we still pay full price. Emile: meaning edits performed blindly without even reading the target file have skyrocketed by 500%, which means that the tool you trust to build your software is routinely modifying crucial architecture without looking at the first. This is not just a glitch, but a fundamental warning about our growing resilience of Blackbox systems that can secretly degrade their reasoning depth overnight while we still pay full price for them. If the very tools we use to write the future of software cannot even be bothered to read the files, we could be in store for spending more time debugging these oversights. We have no transparency over them. So just a word of warning. Ralph Behnke: If there are tools we use to run the future of software, and I'd even be bothered to read files, we could be in store for spending more time debugging these oversights. We have no transparency over them. So, is that a word of warning? Well, know, is that a word of warning? I... maybe I'll share is ⁓ just how go about it. You know... when we're kind of using these tools. So personally, I never trust that the AI is going to read any file. And so in Claude Kern in particular, you you can build very effective skills. And one of the things that I do is when I actually start building any new feature, I will invoke a skill that a brainstorm. And job of that skill is actually to build a spec. ⁓ in that brainstorm, one of the things that has to do is review all information in relation to that feature and it builds the spec. The next thing that happens is I review the spec and I go every single line, I read it ⁓ and then, you answer any clarifying and correct it. And I think that human in the loop is absolutely critical to evaluate at that point, whether it's skipped a beat and whether it's been a bit The next thing it then does is it essentially runs a series of skills where it evaluates output ⁓ the brainstorm and turns it into an implementation plan. It at downstream impacts. It looks at unintended consequences and ⁓ then all of that back into the plan. And again, that updates the spec and a human to at it line by line. And so I think this is the way that I personally to ensure that the AI ⁓ ⁓ attention and a really kind of layered approach and so before I any I've probably done around four to six validations checks that typically result in a specification or some kind of paper being that then needs to be reviewed and at each point you know there are various that are either output by the model or identified ⁓ a ⁓ you myself in this case. And that's the way that I personally work. And so I just, wouldn't trust it anyway to just go off and autonomously, you know, reference all material. We know it's lazy. We know it's prone to hallucinations. We know that when you expect too much from the model, it might do and it might not. you know, sicko fancy this word that we spoke about in the last episode as well, people are obviously kind of going AI psychosis because the AI is sycophantic. Again, I a lot of that because I typically don't chat with AI in, quite this, in the way you know, it seems of these stories about people that have suffered psychosis, the way that they're it. know, in this kind of ongoing chat and it blows smoke up their arse and they think they're amazing and they've invented new paradigms in maths. That doesn't happen because I throw 12 books at it and, you interrogate the information ⁓ and then it through various skills and structured activities. And so getting slightly different results. And so I would say maybe those developers themselves are kind of suffering from those types of issues are being a little bit lazy. if this is something that has changed at an architectural level within Claude, then no doubt it's probably affecting even code that I'm building. But I think people have to put guardrails in place themselves and try not to rely on the AI figuring out a problem and an answer end to end. just think that that's still very much prone to error. Personally, I am still inserting myself every step in the process and I've good results with that. So let's see that plays But I don't think Anthropic will suffer that for long, given the amount of developers engineering teams that are now using it. If that really is the case, then I think that degradation won't go unanswered for Emile: Thanks, Ralph. That's really good advice. And just to be clear, I'm not saying don't use the model and saying, ⁓ whoa, it's here and there. I'm saying don't blindly assume that the model is just going to do everything for you. And when things start working well, don't assume that degradation can't happen. Keep checking your work. Treat it like a partner, not as a... Ralph Behnke: That's really good advice. just to be clear, I'm not saying only as a model, and saying all the world is here and near. I'm saying don't blindly assume that the model is going to do everything for you and bring the start of the world. Don't assume that degradation can't happen. know, keep checking your work. Treat it like a partner, not as a full handle. Emile: a full handoff to something, collaboratively work with your agents. Great advice, Rob. Thanks. Ralph Behnke: something collaboratively with your agents. Yeah, think expect the degradation and build safeguards is the way that I would recommend people go about doing it. So did have a for the time I had the best joke series that AI has written ⁓ somewhere in my notes, you know, I can't even find it. wait, here we go. Okay, right. Segment one humor. So Emile: you Ralph Behnke: Probably about 10 episodes ago, I wrote a protocol for AI to come up with jokes. I used to share a few jokes at the beginning of the show. It just never worked. This one, see if you can kind of keep track of this joke. So, OpenAI raised 122 billion on Monday. By Tuesday, nobody wanted to buy their shares on the secondary market. Don't know about that one. Anthropic source code leaked to 34 million people on the same day NPM got hit by a North Korean malware attack. There's just nothing funny about that. Meta launched a new AI lab called Superintelligence. Honestly, still a better week than your last sprint review. And that made me absolutely guffaw with laughter because... I don't know. Anyway... Emile: Put them. You Ralph Behnke: but I still haven't cracked the humor code clearly, I managed to get a bit of a laugh out of that. I'm ⁓ thank that. my ⁓ as a, know, AI inspired comedian is well and truly in peril. But I think we're kind of nearing the end of the show. ⁓ I wanted ⁓ just very talk Emile: Keep them coming, Rolf. I'm sure they're gonna get better with each passing week. Ralph Behnke: ZAI this one does need to be on everybody's radar. GLM is their current LLM tool. It's an open source model and ⁓ it's for eight hour autonomous work sessions. just as we were talking about inserting a human the loop ⁓ and checking the I don't know how you're going to validate the work as a human, the output ⁓ of AI. that's been running autonomously for eight hours, it's going to produce an enormous amount of output. So I guess I'll have to rethink my strategy on that one. But, you claiming that it can work on a single task for up to eight hours and outperform top closed models like GPT 5.4, and Claude Opus 4.6 pretty impressive really. So, you know, we're talking about a direct challenge to the kind of agentic loop limitations that plague the frontier models. this thing's open source. I mean, it's insane. So I'm going to have to play with that one. It's just too good. I don't know what the costs are, but I was listening to an engineer talk said AI a couple of months back. And he, again, he was kind of reporting that it was 90%. Certainly of the API costs that you would experience with Opus 4.6. So if you are looking to play around with something, I would definitely recommend checking out ZAI's GLM 5.1. That's my top tip. Emile: Thanks, Rob. Sounds cool. actually haven't heard about it. I think I'll try that this weekend. And on the basis of tips, I have one more myself. So I got quite a lot of feedback. Let me just turn back. Ralph Behnke: Thanks Rob, sounds cool, I actually didn't hear that though. And on the basis that I am more myself. lot of feedback they just turned back. Emile: So one's not explicitly about AI, but rather a valuable lesson about comprehension and understanding amongst ourselves. And I think it's fitting ⁓ our sense of value moves more towards needing to be able to be really clear about our own thinking, specifically if we really want to excel in this new orchestrated world of agentic architecture. Ralph Behnke: So this one's not exclusively about AI, but rather a lesson about comprehension and understanding amongst ourselves. I think it's fitting as our sense of value moves more towards needing to be able to be really clear about our own thinking, particularly if we really want to excel in this catastrophic world of agentic architecture. Have you ever read something by a certified engineer that felt like your brain was melting because you couldn't understand a single word they wrote? Emile: Have you ever read something by a certified genius that felt like your brain was melting because you couldn't understand a single word they wrote? So Steven Pinker stood in front of the crowded room with bright minds and asked them exactly why most people have absolutely no idea why their own writing fails miserably. He calls it the curse of knowledge, which essentially means that once you learn something, it becomes almost impossible to imagine what it was like not to know it. You end up writing for yourself instead of your Ralph Behnke: So Stephen Finkus, sitting in front of a crowded room, wrote a book with five lines and asked him exactly why most people have absolutely no idea about their own lives and lives miserably. He calls it the curse of knowledge. Essentially, once you learn something, it becomes almost impossible to imagine what it's like to not know it. You end up writing for yourself instead of for the reader, because you simply can't realize that others lack foundational understanding to be able build. Emile: reader because you simply can't realize that others lack the foundational understanding it took you years to build. We see this everywhere from tech founders trying to start up programs using dense jargon to academics to bury their most groundbreaking ideas under layers of impenetrable texts. So the trick is to step outside your own head and remember what it felt like to be a beginner who needed things explained in a simple and concrete terms without any of the fluff. Ralph Behnke: The secrets everywhere from tech founders trying start a program using dense drop-in to academics to bury a most groundbreaking idea under layers of penetrable tech. So the trick is to set outside your own head and remember what it felt like to be a beginner. And you need things explained in a simple and concrete way without any of the fluff. So next time you sit down to buy something important for yourself or others or your AI agent, ask yourself. Emile: So next time you sit down to write something important for yourself or others or your AI agent, ask yourself, are you going to write this to stroke your own ego or you're actually going to want it to be understood? So that's my tip for today. Ralph Behnke: Are you going to write this to stroke your own ego or are actually going to write it to be understood? So that's my tip for today. Yeah, it's an awesome tip for the day. know, like a five-year-old is one of my favorite prompts for to make some difficult, you concepts clear, especially when it comes things like... quantum engineering, often turn to that particular prompt tip right like a five-year-old, like explaining it to a five-year-old at least. so I dumbing it down, mean maybe just getting to first principles and kind getting to some kind of, know, communication is all about, a lot of communicate because they love the sound of their own voice. Of course we are those people, but actually you the reason you communicate is to inspire understanding in others. many years ago, I had a wonderful mentor that kind of coined the term speak to their listening. And the principle was, much like you set out, it's not about what you've got in your head, it's what they've got in their head. Speak to what will help them understand. And I think this is really kind of key. In the age of AI, probably more importantly so. I'm going to finish on just one last story. I about ambient AI earlier on and There's kind of an interesting tool which I want maybe interesting bit of hardware that's coming out It's called button. It's AI talk button and it's built by X Apple vision Pro engineers and so this thing is basically been announced this week. So not quite sure if you can it but it sounds really cool because It's a tiny press to talk AI assistant only listens when physically pressed. And so it relies on the user's phone for connectivity. Actually it's available for pre-order I see. And so it basically solves the ambient listening problem. You know, this kind of always on creepiness that people are a little bit worried about when open AI, know, Johnny Ives is obviously working on a kind of always on device. this a quite I don't know. It's just an interesting tool when you've got good idea. You just want to record something. Sometimes I try and my phone for it, but I just I find the barriers to basically cause me not to be able to record stuff easily. so button, there you go. An talk button is hitting the market available for pre-order. If you guys are listening, send us one. We'd love to try it out and share our experience on the show. Send us too. Shameless are, shameless. Well, I without further ado, that's a wrap for this week. Emil, thank you very much. See you on the beach. Thanks, see you next week. Adios. Emile: Send us too. Thanks, Rolf. See you next week.