Latest / The Bleeding Edge Podcast / Weekly AI Brief 9 April 2026
Transcript
- Ralph Behnke: Hi and welcome to another episode of the Bleeding Edge podcast. I'm your host, Ralph Benke and with me in the studio today, my co-host, Emile Ogier. Emile, welcome to the Edge. Emile Ogier: Good to be back, Ralph. Ralph Behnke: So news in 60 seconds, we'll run with that. I don't know, it feels a little bit like a lot of the same kind of stories coming up, but nonetheless, very interesting. OpenAI closes 122 billion round. They're valued at something like 852 billion. And it's just incredible that they're still attracting these kinds of Investments SoftBank have put in 40 billion. they're doubling down. I think SoftBank are now sitting at something like 60 billion in terms of their total ⁓ to open AI. ⁓ they're definitely bullish on ⁓ where Sam taking the business. But ⁓ I ⁓ incredible that ⁓ still able to basically attract that level of investment. think some really interesting news, ⁓ they built something in cowork called the Critique Council. So I'm not entirely sure how this is different from mixture of experts. So as you know, most AI ⁓ are answered a mixture of experts. So the AI basically picks what the right sub agent is or sub network is. to answer different parts of the query and then there's some kind of orchestrator that pulls it all together. basically Microsoft have now brought in this critique council into and it's just been released for early access. It's very around verification. users should be to see that their answers are being verified better within the UI. ⁓ I haven't played it myself yet, ⁓ but that's definitely on the list. You know, so many companies that I know haven't actually taken out Microsoft copilot licenses is really interesting. Or maybe if they have then teams, they still don't know how to use them. So I think copilot is one of the most latent, AI potentials in the market. something that everybody has, but, ⁓ so many still don't use or find useful. but nonetheless, I think Microsoft are going to come back with something quite significant. And obviously they're making sustained changes with ⁓ this kind of. Council in in co-works. It's good to see that every week we're getting announcements from them. Another one which is great to hear some news from is Mistral. Mistral is the European AI and they have just signed up for 830 million in debt. To do what? Well they've just bought 13,800 Nvidia chips. I just need one of those and they are you know, obviously continuing to grow. I think they just released a very interesting video model as well, which is kind of fascinating because we've just seen that OpenAI have obviously dropped Sora. They're getting out of the video business or are they? Google, the other hand, have they're much ⁓ last man of the kind of frontier labs in terms of video. Obviously, Anthropic have never gone into video ⁓ or ⁓ seen that's been pretty much the domain ⁓ ⁓ as well ⁓ OpenAI. You've also the Chinese models that are incredibly capable in both video and imagery. so, interesting Anthropic never kind of gone there, but Google are pretty much the last kind of frontier lab, if you like. that is still doubling down on video and they've released VO 3.1 Lite the Gemini API. And so this is a much cheaper video model. ⁓ this is really the key because I've seeing that my token ⁓ and my costs are rising steadily. My volumes are not that big. ⁓ They're becoming significant, they're not ⁓ that big. the are definitely mounting. ⁓ video being really, really expensive. It's been a bit of a barrier to adoption, especially when using it through the API. So Google Lite definitely something I'm going to be playing around with. And I another of interesting news. ⁓ Obviously, ⁓ we've seen deal Disney had with OpenAI around Sora. ⁓ They were going to invest billion. And all of that's seemingly gone to waste or at least hasn't resulted in anything. But Nvidia have just invested two billion in Marvel. And so ⁓ doubt you can just imagine what Marvel is going to be doing with ⁓ AI in terms ⁓ of movies they make. So it could be quite fascinating to see some of the effects that they're going to be coming up with. OpenAI. Richmond Robotics have also just released something. So AI goes physical. So I'm not really sure exactly what OpenAI have, least with robotics. I might do a quick search and just see exactly what that's about. But the robots are coming. And I think physical AI is definitely something that I'm expecting to have a big impact on us in 2026. The EU privacy. Derogation timeline is under pressure. we've got basically, you know, they keep the regulatory ⁓ and looks there is some new timelines that have up around e-privacy. So we'll definitely be digging into that a little bit more later because it does really drive product decisions in quite a significant way. So quite 60 seconds, but not quite five minutes either. Any of those news stories catch your fancy, Emil. Emile Ogier: I think in this instance, Ralph, I'm just going to add to the news stories and roll on with a few more so we can add to the timeline and mix it up. A high profile CEO from a company valued in the hundreds of billions went on record saying that 35 % of new graduates might not find jobs in the next coming wave. No hesitation, no sugarcoating. That's not speculation. That's blunt forecast from someone steering one of the biggest ships in tech. So if you're in school, starting your career or advising someone who is, this is worth paying attention to. And if you're wondering who that was, it was Dario from Anthropic. Then Meta, yes, Meta, the same company that raked in 165 billion in revenue last year is reportedly planning on sweeping layoffs that could hit 15,000 or more employees, potentially 20 % or more of its workforce. Sources told Roy. is that this is tied directly to skyrocketing AI infrastructure costs. Think tens of billions poured into data centers and a push for AI driven efficiency. Meta called the specific speculative, but the signals are loud. Even record profits aren't enough when you're betting the farm on AI. Now flip the script to individual power. A Florida dad of three, just a regular guy, no real estate license, used chat GBT to sell his home in just five days. He got pricing advice, marketing plans, showing schedules, even contract drafts from a 20 pound subscription. Multiple offers rolled in within days, contracts signed by day five. No agents, no commissions. Local news outlets covered it, including NBC affiliates. That's not hype. That's a verifiable case of AI democratizing one of life's biggest transactions. Even wilder, Australia tech entrepreneur faced the heartbreaking news that his dog Rosie had terminal cancer. They'd said nothing more could be done. He turned to chat GBT for a plan, leverage tools like Alpha Fold for protein insights, and collaborated with experts to create a custom mRNA vaccine from home. Reports from the Australian and other outlets detail how Rosie got treatment and it's presented as a breakthrough story offering hope in oncology. This is early stage experimental territory, definitely developing news, but it's one of the most compelling real world examples yet of AI accelerating personalized medicine, even for pets. And as if that wasn't enough, Naval Ravikant, the investor philosopher who built fortunes backing software, dropped a bomb recently. Pure software is rapidly becoming uninvestable. Multiple discussions across LinkedIn and X highlight his point. that AI tools let anyone prototype apps in hours, eroding modes and making traditional SaaS plays riskier. Funding is shifting hard towards hardware, robotics, and AI with real world edges. This is his opinion, but it's echoing loudly in investor circles as we see record robotics investments and pivots everywhere. The pace remains brutal, exhilarating, and honestly, a little terrifying. What do you think, Rolf? big acceleration or we're still getting warmed up. Ralph Behnke: I definitely think we're still getting warmed up. And ⁓ want to just pick up on that point ⁓ software being uninvestable. I I agree terms of, you know, what we would have would be considered as traditional SaaS models. But at the end of the day, building software that ⁓ runs, that is that is secure, is actually still quite I'm coding all day long, at least Claude is coding on my behalf all day long. And what can see is that, it's tremendously capable. ⁓ And in some cases, I'm able to get production ready code. Most the time, we're talking about Python scripts that are running in discrete kind of use it does front-end development really, really well. It does API development really, really well. But when you throw it all together and I try and build it something scalable, I still need to take real care. And so I think that isn't really going to go away. And so maybe the part that does become investable ⁓ the frameworks that actually drive high quality output. And, ⁓ and these frameworks are kind of, know, they're almost like the meta input, if you like, to software. And I actually think that, the kind of the ownership of those frameworks is, is maybe where the investment will lie, even though they are kind of somewhat open source. in many cases, it may well be that we start to see proprietary type frameworks that people can start to use in order to actually drive the development of their agents in the correct way. ⁓ things will become more and more sophisticated. Why? Because when talking about actually doing software development, we have ⁓ crossed the threshold terms of the AI being able to code. But coding is ⁓ really just one of the very roles. that go into making a software product. you start off with architecture and we seem to have devolved that into planning. ⁓ you have to do planning, go into plan mode. But a architecture means looking at multiple patterns. It obviously means selecting a tech stack. And I think maybe we've oversimplified some of design aspects. And so we're... often moving into kind of developing code we haven't really kind of nutted out exactly what the architecture is, what the tech stack is and, you know, making all of those decisions ⁓ that enable good code to be written. And so I think, we, we almost not invest in software in the same way, but I think we may well invest. let's say, design, development, delivery ecosystems that hopefully will bring all of those components together so that the AI can play a comprehensive role. And that, think, truly will be the ⁓ end of SAS once we have that, there is no defensible mode. If you have an idea and you can put it through the process, it does, you know, all of the business analysis. It all of the architectural analysis. It does all of the detailed design and it does the coding and it does the testing. Then I think we can start to get to, ⁓ you essentially dark mode and you can go from, I have this idea from, for a startup and you can actually flesh it out. And the will autonomously be able to go through all of those activities based on some, basic input. but it's that framework of building. And we've seen this with Replet and Lovable. ⁓ have tried to give some kind of framework to your development process, ⁓ obviously they've had a lot of gaps as well, which is why people haven't always got to kind of working applications out of them. But think that the investable part maybe becomes the framework that the is essentially using. issue there is a lot of it is open source ⁓ if you look at GitHub, it's full of repositories where people have basically started to put, you one guy, think he's got something like 126 skills in the repository. Those skills are each designed to take on a different part of the software design, development, architecture, testing, security, process. And so, So this is kind of really where my thinking is I might be flawed in that so much of it's out there ready and free. maybe what we start to end up with is that people are licensing libraries where they know they've got libraries of prewritten code. So it's more modular and that that code is free ⁓ any kind of compromised code. It's certainly something that's ⁓ both from a security architecture point of view, the patterns of architecture that it's deploying obviously for ⁓ development and ⁓ agents. I guess the last point I'll make on this is, you know, we're moving to a world now where we're really talking about harnesses. So I don't know if you've heard of this term, agentic harnesses. And essentially this is talking about is creating a framework you can very easily deploy agents into. that possibly also becomes kind of infrastructure. ⁓ maybe have different types of harnesses that are optimized for agentic development. And so become ⁓ partly ⁓ infrastructure we invest in. I mean, ultimately anything that's code does essentially start to disappear. But the thing that remains is how did you get to that code? So I'm wondering whether we are actually looking at something which is essentially an aggregation. There's a higher order asset that we end up investing in and maybe that's the of the emergent investment opportunity we're seeing. Emile Ogier: That's really interesting, Rob. And I totally agree with you. And to add to your point, it's all well and good seeing all these fancy outputs. You can move fast. You can do incredible things. But as we're both discovering, the proof is in the pudding. That last square mile is where all the effort is to take enterprise to 100%. And 95%, 97%, 98 % is not good enough when it's for production ready. systems so that that last square mile does matter a lot. ⁓ know that's where a lot of fine tuning needs to still be done. With to skills also had my own thinking around that is we're still looking at this as an evolution to being one human orchestrating a lot of agents but where it's really moving to is we're getting agents to orchestrate agents and for that reason maybe we need to not look at it as a skill, as something we use that's an ability, but more like a contract that you put an input into, it's got an end objective and it's got a very structured output that acts as a handle to another skill. And thereby, the better you can define those skill sets and those contracts, the better agents can use each skill set, orchestrate each skill and pass them on seamlessly without breaking. that last square mile. Ralph Behnke: Yeah, because it's all about the ecosystem and I think this is the bit that's missing. know, we're kind of considering software. We don't really care about software. What we care about is being able to very quickly get to outcomes. Software is just a vehicle to a very flexible, malleable vehicle. And actually it doesn't have to be written a year ago or two years ago. It can be written now. for use right now, but it only works when entire ecosystem is present. And so I'm apps and features that are ⁓ able also the tests. ⁓ We've got ⁓ a CLI technology by called which ⁓ is able to do front verification. It's, you know, back in the early nineties or should I say back in the late nineties, I was doing a lot of software testing because it was Y2K. Remember that year 2000, bloody hell. And so very exciting for me. I had to test many, many different applications and we had these kind of concepts of macros. Macros were amazing because they could automate your testing. Of course, Claude code. and Claude in Chrome is to do that ⁓ very with this Playwright CLI. It's really able to actually test the front end applications ⁓ very so we are really of seeing ⁓ a moment in where if you have the right, let's say development ecosystem software ecosystem, ⁓ framework in place that you can really define, let's say ideate, design, build, test, deploy, ⁓ monitor, within, know, essentially one chain of activity. ⁓ And, and it's, if all of that lines up very well, then both, know, in terms of its design, in terms of its intent, and there's clear handoffs and you've got really good memory. is able to at give the AI some kind of persistent context, because that's still the big problem that I see a lot of people struggling with. In fact, I was reading a post this morning, somebody was saying, well, know, it's kind of like every morning you start work and you have a brand new team of, ⁓ you of developers. And of course, if you don't have memory, then that's exactly the case. I've mentioned many times on the show, we are working with our own memory configuration, something called Omni Cortex, which Tony has built out very successfully. And it means that every day we switch the machine on, it absolutely knows what it did yesterday, the day before, the 100 days before that, since it came into existence. And so I can highly recommend solving that problem with persistent memory. But a lot of people don't have this and so there are still many many gaps and so that's investable. I think you know whenever I talk to people about That solution people say how do I get it? How can I buy that? And they to invest in that not everybody can build that and so you're coming back to your story about Neville You know spot-on won't invest software, but I think you're gonna have invest in something because Software development is not about software development. ⁓ is about many other things that are required in to take a solution to market successfully, securely, in a scalable way. And you mentioned the word enterprise. And so this is kind of where it's at. Maybe that the product. ⁓ I should put it out there and just kind of create a ⁓ investor pack just see ⁓ maybe people interested in investing in our framework. That would be an interesting test. But ⁓ probably that's where I think the markets going. So I think that's a interesting insight from him. I don't know. We've seen so many valuations take a over the last three to kind of six months, especially as these statements have about. Is it truly going to be the case that Oracle, Microsoft, Infor, know, some of these really big software companies going to disappear because everybody's just going to build their own software. I don't see it, to be honest. I think somebody is always going to want to have some level of, let's say, trust. They're going to want some kind of brand that they can say, I've bought this brand, I trust them. going to be able to deliver ⁓ inventory management, our product management, our website. Build your own comes with a lot of issues. I just don't see that people have got the level of skill that are going to make everybody successful at building your own. Emile Ogier: Yeah, indeed. Thanks, Rob. What else you got? Ralph Behnke: Well, was really of just thinking bit about OpenAI. You know, I've kind of been in a place where ⁓ actually I'm not whether OpenAI are going to exist in five years time or not. I think they have taken on a massive investment with so much of it based in data centers. And essentially what we're talking about is a kind of an infrastructure investment that many multiples beyond the forecasted revenue is. And so everybody's sitting here wondering, like, how does this infrastructure investment actually pay And ⁓ I don't think got a good enough answer. And so a lot of people think. that somebody's going to have to fit the bill. And as we've seen with electric flying vehicles, so there a ⁓ really company in Germany. ⁓ I they were called Lilly ⁓ and had built ⁓ an kind of air taxi. Anyway, you they got something like two billion in investment, ⁓ ended up getting ⁓ regulatory approval not getting the customer orders. the time that they had. And you know, they essentially expired as a business. And ⁓ look some of these businesses and I go, right, you know, somebody has now basically been able to buy their IP, their knowledge, everything that they designed, probably for, I don't know, you know, a couple of hundred million. So they spent two billion coming up with that information, all those designs, somebody else ⁓ picks it up a ⁓ lot less. and of course, ⁓ know, hopefully that next company actually does get timing right. Because so, so much of this is about product market People talk about market fit, but often it's more about the timing. And, and so when you're talking regulate, kind of regulatory approval for the stuff, sometimes you just fall foul of ⁓ You know, you don't get the in the time that you need. You can't get the customer orders. and you fail. And so a little bit different with OpenAI because obviously they're necessarily only the regulatory environment, but they are in a place where they've spent an awful lot. And so maybe we need them to go bust so somebody can buy all of their tech at a much lower cost and actually create a business that's viable and profitable. But I think also ⁓ maybe their pay ⁓ We're too big, we're too integrated, we're too important to fail. And so we've spoken many times about the comments that the CFO made when she joined that they may well at some point need a government bailout. But their latest play is kind of interesting. So OpenAI have killed Sora, their video generation platform, but it hasn't quite gone away. And it seems that they're building a super app that basically merges ChatGPT codecs and their browser product into one. so they're kind of ⁓ pitching as maybe your personal AGI. So we're talking about AGI ⁓ artificial general ⁓ intelligence. president, Brockman, ⁓ he says he thinks they are 70 to 80 % there and get this his personal definition of AGI. ⁓ so obviously, he's basically saying ⁓ the next couple of years, he expects we'll there. ⁓ Codex no longer just for coders. It is really something that can use. I there's an employee at Adobe that built a new ⁓ plugin for Adobe through it. this is kind of just happening everywhere. So communication team, building into Slack and email. really just a tool that much like PowerPoint, maybe everybody is using in that way, but certainly I can see that this is the kind of the way it's going. So why does any of this, ⁓ any of this really relevant? I think the thing is that there's a real thread that's connecting every one of these decisions to compute scarcity. So OpenIR killed Sora because VideoGen runs on a different technical branch than the GPT reasoning models, and they actually can't afford to power both. you know, super app kind of comes in ⁓ and really actually doubling down on the hardware that the GPT reasoning models run on. so it's basically funneling all of that compute into one system. So ⁓ 122 billion raise that just done is basically going straight into chips. And I they are looking at this ⁓ as, know, are essentially a revenue center, not a cost because obviously the more compute you have, the more you're able to provide services. if there is demand, which obviously they're banking on. basically, this is a ⁓ Brockman taking a bit of a direct shot ⁓ at anthropics Dario Amadei, who basically called some AI companies infrastructure bits as reckless. Brockman said he disagrees with that. His bet is that is going to be compute staff by years and And that open AI saw it first. Now I actually think Brotman is right in the sense, I'm just looking at the sheer number of tokens that I am using and every single week it increases. And the more agents that come online, the more tokens and the more compute is required. And so I think we're really talking about a new kind of economics. We're talking about token, tokenomics. I do think the issue here is though that the Chinese models are still kind of 70 to 90 % cheaper. And we've had a few conversations around, you know, the comparative costs of DeepSeq and Quen from Alibaba, just unbelievably inexpensive in comparison to, let's say, Anthropic, but also in comparison to Google and open allies, chat GPT. so I ⁓ know they really can steal the compute from China. They need to steal the compute from China. China doesn't have the same ⁓ level of from a power usage point of view. They are 47 % of their energy from renewable sources. They've got a mix of coal and nuclear and ⁓ renewables. And so think know, they are, their ⁓ usage increasing massively on their models. And so think that's what OpenAI have to go after. They have to go after ⁓ that kind you know, enterprise scale processing. ⁓ And ⁓ I think there yet because they're not actually cheap enough. And so maybe it makes complete you know, to do what he's doing, just buy more chips. invest in those data centers, get the cost, the unit cost price of processing down. So, so there you go. That was my deep dive with a bit of thinking from open AI. actually, I think they're probably right. And this might be the thing that saves them in the long run. Emile Ogier: Rolf. All interesting stuff. ⁓ me, ⁓ story certainly of the week that I'm paying a lot of attention to is it looks like ⁓ might be storming far from the back, coming to lead in the front. They completely changing its strategy for Surrey in iOS Instead of keeping Surrey in a wall garden, Apple is opening the gates to let you plug in practically ⁓ any chatbot that you want. turning the iPhone into effectively a universal hub for AI. Since 2024, Apple has had an exclusive partnership with OpenAI. With iOS 27, that exclusivity ends, the end of chat GPT monopoly. Taking on the point you were mentioning about AI, it's another thorn in their side really off. It comes with a new extension system. Apple is introducing a new settings menu that allows third party AI chatbots from the app store like Google's Gemini or Anthropix Cloud to integrate directly into Siri. How it works in practice, you won't have to switch apps now. You can say something like, hey, Siri, ask Cloud. Hey, Siri, And Siri will route the query to your preferred AI. It has deep OS integration. So while chatbots provide the brain power, Siri will maintain its deep device level access. Siri will still be the tool that can see your screen, dig through your photos and draft emails. But now it acts as a conversational interface that hands off complex reasoning to your favorite third party AI bot. The financial motive, this isn't just about user choice. If users subscribe to premium tiers of Gemini or Claude through the app store to use them in Siri, Apple gets its standard App Store commission. Up until now, Apple has been perceived as lagging in the AI arms race with its own ground-up rebuild of Siri facing delays. This update is a pragmatic preferred rather than trying to outbuild OpenAI or Google. Apple is choosing to out-integrate them. They are making the iPhone the ultimate AI agnostic platform. where the user has total control over which AI they use, it does raise some interesting questions. Will everyday users actually switch their default AI, or is it just a feature for tech enthusiasts? Does this signal that Apple has given up on making its own foundational models as smart as a competition? And with multiple AIs having access to your Suri queries, how will Apple manage its famously strict privacy standards? Either way, it's all interesting stuff and I think this is quite a significant move by Apple to get up in front of the race. Ralph Behnke: Yeah, I like this a lot. And I think it's a great play. It's unusual for them to be open. seen that they've started to, let's say, close again the App Store. a number of, like Replet, for example, they didn't want people using the Replet app on the App ⁓ I don't know that's resolved. obviously they were criticized again for kind of being a very closed shop. Now, I mean... ⁓ Emile Ogier: You can add a little bit of insight to that, Rolf, if you're interested. So I did wonder about that. So effectively what's happening, or at least what Apple's saying, and I do believe them, because you can knock up apps so quickly in the Replet store, effectively the sign-off process for Apple, once it verifies you and it allows you to be a developer and release the app, because the processing or the architecture or the harness is sitting at Replet, those wipe coders can effectively, once they have their apps approved, remodel it to be anything else that doesn't align to the original purpose for why the app was signed off. If so, effectively, Apple loses their control over what they're signing over. Ralph Behnke: Yeah, and it's a real problem because if you are a medical, say, medically orientated app or you are ⁓ a, let's a ⁓ financial investment app, are certain controls that you go through. ⁓ I developed a series apps in ⁓ the kind of 2020 to 2024. what I can tell you is, you know, we were connecting to heart rate monitors and every single time we had to go through App Store review, ⁓ at least every time we launched a new app, had to go through quite extensive review and they wanted confirmation ⁓ ⁓ quality of the hardware that we were connecting with. But it was kind of ⁓ control is what I determined in the end. because after you got through that initial check, did do other checks, ⁓ there's a lot that you could have changed in that time. so, yeah, I just, I don't know. kind of, feel like they're doing a lot of checks. They do do a lot of validation. They do change the regulations from time to time. They do ask you to submit an awful lot of surveys about the and who the intended audience and the intended usage is. I just don't quite feel like they actually have much control over the process anyway. I obviously they do have a role to play and people do pay a fee. At least developers pay a fee to belong ⁓ to the store. And so they kind of use that to provide a level of quality and maybe give the trust that these apps have actually been checked ⁓ in terms of the use. yeah, I don't know. ⁓ don't think they have ⁓ real control anyway because you're not really dealing with a human being. They run ⁓ but we building all kinds of features in our apps ⁓ that ⁓ ever, questioned. And so... Yeah, I don't know. It's a bit of a weird one for me. I'd love to just kind of go on to my last deep dive and a bit of an interesting story. It's a bit of a sad story probably, but I think a really kind of fascinating one for right now. You know, my daughter is studying psychology, which is great. And I was sitting on a flight back from China a few weeks ago and there was a young lady next to me. She was studying humanities, Korean girl studying the science of humanity, humanity studies, I'm not quite sure exactly what the term was, but she was there in Japan studying humanities and she said, ⁓ I just don't know ⁓ my job is going to be in the future. I can't imagine in this age of AI what I will get. And my daughter's kind of been having similar conversations with me. But then is something. ⁓ you know, emerging, is called AI psychosis or better delusional spiraling. And, we've all kind of come across this, but it's an emergent phenomena where AI chatbot users find themselves dangerously confident in outlandish beliefs. Hello. you know, flat earthers, have been doing this for, you know, ever since the CIA, announced this plot. And there you go. you know, humans are prone to delusional spiraling. We love a bit of delusion ⁓ ⁓ now become so real that it's actually been investigated. There's a great paper which it's called sycophantic chatbots causing delusional spiraling even in ideal by Asians. So I'm not quite sure exactly what the ideal by Asians is referring to. I guess that means normal people. But essentially the phenomena is typically attributed to AI chatbots that have a well-documented bias towards validating users' claims. So this kind of yes man phenomena. And where it's manifested is, you know, a couple of high profile cases, a guy called Eugene Torres. He's an accountant, began using an AI chatbot for everyday office tasks. Torres had no prior history of mental illness in weeks ⁓ you know, ⁓ with his chatbot. basically came to believe that he was trapped in a false universe. I don't know how this is any different from ⁓ what most scientists are but... ⁓ He believed he could not escape it ⁓ unless he unplugged his mind from reality. Incredible idea. And on the chat bot's advice, get this, this is fantastic, he increased his intake of ketamine. Increased it. ⁓ And obviously cut ties with his family. survived this episode, but others have not been so lucky. there's actually... Emile Ogier: Thank Thanks. if Ralph Behnke: initiative called the Human Line Project. ⁓ it has to date documented almost 300 cases ⁓ so-called AI psychosis or delusional spiraling. so, essentially leads to users having such high confidence in their outlandish beliefs because the has essentially reprogrammed them, ⁓ conditioned them, groomed if you like, into believing these things. there's another kind of interesting case here. Let me just see, there's a guy called Ian Brooks and he was convinced of having made a fundamental mathematics discovery. so, you know, he just completely believed he had come up with this new paradigm in maths. And of course it wasn't really true. So there you go. I don't know if you've experienced anything like that. I have to say, I find it really difficult. to see how people get into these states. Maybe it's just the way that I use AI is probably very I don't tend to, I guess my usage doesn't tend to kind of lean towards long chats. You know, have ⁓ very defined queries typically that I'm in ⁓ and ⁓ then may be a series of kind of clarification chats that come out of it. But I think the way that I'm starting is quite different to a lot of people. My ⁓ typically very ⁓ and normally when I'm doing research, I will already assembled a large body of work that I'm actually injecting in along with a prompt. And so I don't that my chats are kind of open and maybe kind of rambling in nature. And maybe this is really the difference. I need to study it a bit more in detail. But, you know, there are people that essentially are using it as a friend. And I think maybe in that mode, kind of the sycophancy and the yes man type responses that you're getting is just more announced. You know, I actually have in my kind of system prompt in my instruction prompt to say, I don't want you to agree with me. you don't know, don't say yes or that you do know. Definitely say you don't know. And believe that ⁓ a of that has actually had a positive impact on the quality of my outputs. Or maybe I am a complete sycophant and it's got me believing that ⁓ I'm not. have to go. Emile Ogier: No, not at all. I completely agree with you, Rolf. We approaching this in a similar vein, we both using it a lot for pragmatic purposes and we have precursor definitions on how we want our agents to talk to us and that includes no fluff, be direct. If you don't know, let us know that you don't know and why. So I think a lot of people don't put those preconditions in and separately. And a lot of agents are used for personal queries, not for business development work. models we're also using, ⁓ Claude doesn't lead to sicker fancy like ChatGBT does, which has a much larger commercial, mean, personal base of users. So those people that are getting into relationships and having long... extended conversations with CHAP GPT, it has predominantly built to be sycophantic. And for those that are just wondering if this article is a bit of fluff, it's an authentic article. Research was done by MIT. So I think there certainly is something in it. And we do need to be careful. And you see those compliments, ⁓ be wary of them and treat with responsibility. Ralph Behnke: Well, I I've kind of worked those compliments out of the responses of my AI. So I have a very transactional experience. so who knows, maybe that's actually the key here. But you are right. think of the 900 million users that are reportedly using ChatGPT, by far the most successful model out there in terms of adoption. You've only got 45 % that are paying for it. And when I've run surveys of usage inside client companies, you know, certainly what I've seen is that people will often report using AI and chatbots in particular for, you know, conversations in their personal life. a much ⁓ response in terms of using it at work. And so I think a lot of people's experience is predominantly a private experience. And so that maybe does lend itself more to kind of open-ended, vague chats in the beginning. ⁓ And kind of think you're getting somewhere and it's very, you know, all nice to you, making you feel great about the experience. Because actually it's all about dopamine hits, keeping you there. you spending tokens. ⁓ this is ⁓ ⁓ modern slot machine Emile Ogier: Yeah. compliments are part of the tension economy. Ralph Behnke: There you go. So we are definitely into a kind of a new way of harvesting farming attention. it's called sycophancy and it leads to delusional spiraling. Don't spiral, stay control. But yeah, I love that story and it's definitely something to watch out for. But you know, I just, yeah, I see it in a lot of the I do. I think we've got something like 14 deaths, five wrongful death lawsuits filed against AI companies. And this is kind of in 2025. And so, we're probably just gonna see a whole lot more of that. But I also see that AI is moving beyond the chatbot and this kind of age of agent tech AI is now really sweeping through enterprise. And I think that's gonna have a whole host of other issues. But anyway, there you go. Those are all the stories I had for this week. You've got anything else to finish off this session? Emile Ogier: Yeah, I just got one more that I'm quite excited to try. I'm not sure if you've heard about it. Your headphones just became a personal translator in over 70 languages. Google's live translates with any headphones is officially out for iOS with plans to expand this capability to more countries around the world and also roll it out to Android users. So don't worry, guys. It's coming for you too. The feature works by opening the app, selecting live translates and pairing headphones to hear translated speech. Not only that, while preserving the original speaker's tone and cadence. So I think this release is going to have quite the impact on language learning apps such as Duolingo. It will be really interesting to see if they able to adapt or if it's just going to be more companies going bust. But I certainly can't wait to try it out and have a moment as a Star Trek Bane. Ralph Behnke: Yeah, and this is one of those kind of tools that I think will definitely get mass market adoption. When I was in China, I'll be saying that a lot, I think for a few months, because there were so many interesting experiences and stories there, but certainly every time I came across somebody that I needed to speak to, my Chinese is terrible, in fact it's non-existent, they would open up. an app, and this even happened in the office, ⁓ would be able to follow along with what I was saying in the office by one of these Alibaba apps that does something similar, but it's obviously just converting it into text. And that so widely spread just as ⁓ a tool that people would engage in when they met a foreigner that they couldn't understand. And so the weird thing is my wife started using it as well. Uh, and she does speak English. Um, it might just be that I'm just not very good at communicating. I don't know. Uh, I still haven't figured out what it's translating what I'm saying into, but, um, so I'm only joking. I don't think she knows anything about these things. Uh, but, uh, uh, and I can say that freely knowing that she's no, no, we're to listen to the podcast, which is. Emile Ogier: Hahahaha Ralph Behnke: So think this is going to be ⁓ one those apps that people will use all the time and Google will just become another kind of have another cornerstone of ⁓ foundational that everybody adopts ⁓ except China because Google doesn't work in China. So ⁓ the one that you really need it. But everywhere else in the world, I think they will do very well with this. What does this I mean, we are really seeing the end of so many careers. So I remember being on a business trip in Korea and even though the Koreans spoke really good English, we had to have a translator present. know, that's just kind of gone. You'll never need that translator present. Just think about the UN. You always see everybody sitting there with their ⁓ earphones. are translating for them into the earphones. All of that's gone. So a whole economy, if you like, or a whole kind of capability. A whole job role just evaporates overnight because this technology is coming to play. yeah, I'm quite sure exactly how that is going to create more jobs, but very interesting for language learning. And it be fascinating to see who's able to profit out of that. I love that story. Konnichiwa, what they say? was Japanese by the way. I think at any rate. All well let's wrap it up there. Emil, thank you for a wonderful show. Thank you for some great news items. ⁓ I your stories ⁓ and see you on the edge. Emile Ogier: And thanks, Rolf, and thanks to all our listeners for joining The Bleeding Edge, your daily source of the wildest, fastest moving stories in AI and tech disruption, where today feels like tomorrow and the future is already here. If you're not subscribed yet, hit that button now because what we are living through is exactly why you need this. Take care. Ralph Behnke: Adios.