Latest / The Bleeding Edge Podcast / Weekly AI Brief 17 March 2026
Transcript
- Ralph Behnke: Okay. Hi, and welcome to another episode of the Bleeding Edge podcast. I'm your host, Ralph Benke. as usual with me in the studio today, Emil Ogier. Emil, welcome to the show. Emile Ogier: Thanks Rav, good to be back. Let's ferry this week. Ralph Behnke: Yeah, you're looking very slick and good because we're going to have to move fast to get through the bumper crop of news. think just getting faster and faster. I don't know if that's even possible, but it kind of feels that way. Emile Ogier: You took the words out of my mouth. I was going to say it feels like Groundhog Day. keep repeating myself, but I feel the same way. It feels like it's doubled this week. Ralph Behnke: Well, we've got really interesting stories to cover that I think have really meaningful impact on ⁓ way people use AI and the way that we actually work. So let's dive straight in. Big news, obviously from Are they still around? How can they still be around? They are like literally the bumblebee that can't fly, but they do. They have ⁓ announced another billion in investment. with some named strategic backers and this is obviously all about infrastructure relationships that are designed to expand compute and distribution. OpenAI have framed the round as $730 billion pre-money while Reuters have described a valuation of $840 billion. It's likely reflecting different valuation maths. anyway, you know, it's still the numbers that we can imagine. Obviously not quite SpaceX levels. I think would be value SpaceX had a few weeks back with something like one point four trillion after the deal with with Either way, signal is pretty unmistakable that Frontier AI is being financed like national infrastructure. So this is not software, guys. ⁓ national treasures that we are building Europe note and following open we've also had news of the of GPT 5.4 Now you actually called out that was a leak twice in a week. So that's interesting They obviously didn't intend release that maybe somebody did I'm not quite sure what's going on there But it does sound like it's a mature enterprise launch there a thinking pro version. There's pricing tables, evaluations, and even experimental long context configuration in codecs. Codecs obviously being the coding solution from OpenAI. Now have to say, I have been trying to get to grips with codecs over the last few weeks. I just can't my way, find my way outside of Claude code. ⁓ I'm due to ⁓ do bit of testing there. The key difference being obviously that Codex actually operates in a sandbox, whereas Cloud Code ⁓ is to everything that you might need order to do production deployments. And so I haven't seen the need to get into Codex, but I still want to give it a go. It's got some really good reviews in terms of the code quality. So we'll have to see how that goes. Google has also leaned into real world workloads. So one of the things they've done is released Gemini 3.1 Flash Lite. So previously, their inexpensive version was Gemini 2.5 Flash. So that's now been upgraded with 3.1 and it's rolling out across Gemini API as well as Google AR Studio and Vertex AI. That's really good news. I've actually just built a that allows me to select different models, different models so that can do some high volume processing. And Gemini 2.5 on my list. I had selected that, but it's going to be great to actually upgrade that to 3.1 Flash. It's a real kind workload. Or shall I say it's a workhorse. So this thing is designed for enterprise scale obviously the rack up pretty quickly when you're using these models with large and Gemini 3.1 promises to be cheap ⁓ and fast. So you're going to using it for automations classification routing extraction translation millions and millions and millions of transactions, then that's definitely the one you're going to be wanting to use. Interestingly enough, I've also been continuing my evaluation of the Chinese models and they are up to 97 % cheaper, certainly than Claude, which is one of the most expensive models that you can use via the API. And so I don't think many enterprises are really using Claude. for high volume work, but it just kind of goes to show the importance of picking the right model. This week wasn't just about capability and new models. We've obviously had some really interesting drama coming from Anthropic ⁓ and the they've doing with the US government. About a year ago, I think now, actually, it could be six months ago, they signed a deal with the Department of Defense and that deal has now been turned upside down because have ridden in on their dark horse and they have signed the deal because Anthropic basically said they weren't going to comply with the Pentagon's requests to give them unfettered access and allow drones and other autonomous machines to basically rely on the LLM to make decisions without a human in the loop. And so where were kicked out. by all accounts, they are back in again and still having conversations, which is obviously good for them. That's a $200 million deal, we have repeated reported on So ⁓ let's where ends up. But obviously open AI the front runner for use the Department of Defense. And they have actually also signed a they call a kind of a slightly adjusted contract which doesn't quite give the Department of Defense completely unfettered to use the LLMs to make decisions for autonomous vehicles and like, but it does rather look like it's not quite the same lines that Anthropic were drawing and that did lead to a of a consumer backlash so quite a lot of people by the looks of it have moved away from OpenAI and onto Anthropic. That's not necessarily a good thing because are they paid members? You know, of the 900 million people that use OpenAI's ChatGPT, only about 5 % are paid members. So has that just inadvertently meant a whole lot more cost to Anthropic? I think those people probably will move back to ChatGPT. Anthropic does not have. quite a token ⁓ allocation free users. So let's see if that actually sticks. But certainly have voted, you know, open AI seems to be a little bit of the bad guy in all of this. But the reality you know, what Anthropic agreed to already was already quite significant permissions. So not entirely sure that Anthropic is the good guy in all of this. I think if you're going to a deal with the Department of Defense, you are probably implicated as being a bit of a devil anyway. the consumer platform side, Meta have said that they are going to temporarily allow rival AI chatbots on WhatsApp. I don't know who uses AI in WhatsApp. You know, it's supposed to be kind of personal communication. Maybe for business, I've never actually used the Meta AI that's embedded in I think it's a little bit weird. Why would I do that? anyway, you the European Union antitrust have said that they need to allow other chatbots in there and that's what they're going to do. So I don't know how long that's going to be sitting there. Would be interesting hear from people that are actually using it. And then there's some other news Meta. They've actually faced a new lawsuit tied to privacy concerns around its AI smart glasses. So ⁓ They've had such enormous success with these smart glasses. I can't wait to get my hands on some of them. would really love to, you to give them a bit of a test. I've tried on before ⁓ and they fantastic. They're not very heavy. And of course, market has really very positive about the audio as well as the video. And so the fact that they've now this lawsuit around these privacy concerns is a little bit of a worry, but not so surprising in the age of AI. The compliance consent bar has to be higher because otherwise wearables essentially recording everywhere all the time. So I'm not quite sure how we're going to get around that, but there you go. Infrastructure have their own reality check. Core Weaver has actually had some... interesting news on the evaluation because they've announced some big capex plans and the didn't like that but Dell on the other hand has benefited from the flip side of the same cycle and they've actually saying that their kind of core computers servers business is definitely an increase in revenue and the market has responded very to them so You know, the AI economy always seems to be a tale of two halves, somebody winning, somebody losing. And ⁓ finally, to out that Apple refreshed its MacBook line with M5 chips and they have got on device AI performance. So that's actually very exciting. My MacBook is now nearly 10 years old. So just in time for me to think about getting a new one. because think having on-device performance is what all about. And we'll cover some of the reasons why that is exciting in some of the stories today. So as usual, Emil, anything catch your eye. Emile Ogier: Yeah, thanks, Rolf. I'd say on the Codex front, you're not alone. Anthropic is certainly building up its fan base, hearing more and more people by the day, just speaking praises of Opus 4.6. It's doing everything they need. They don't feel they need reason to leave. And it seems like you sharing that same experience. Regarding the deal, ⁓ really interesting ⁓ They have quite a mass migration. Last I heard there was a roughly a 300 % drop in new user base activity for OpenAI. So I think they certainly feeling the dent. You did raise a good point there. The models are very different. They've purposeful very different things. If for those not aware, OpenAI is built around highlighting sicko fancy. And if you don't know what that is, it's appealing to your nature to give you the right good vibe feelings. to be agreeable with you. model is very much designed around that because the retail users respond well to that. Coders don't want it. They want something more pointed that point out the errors so you can get the job done. It's more enterprise ready. despite many moving over, ⁓ I share views there, Rolf. think they're going to somewhat surprised if they have expectations that the models are going to behave the same way and talk to them the same way. I don't think that's what they're going to experience. I've even seen celebrities put postings out on X like Kate Perry saying, I've left OpenAI and I've gone to Anthropic. I know she's not going to be a heavy user, but in this day and age of celebrities, people notice things like that. So it'll be interesting to see where the trend goes. And on the meta AI front, that's a really interesting one, seeing tons of posts on X with users. apparently have... office in Kenya that does a lot of the work for overseeing what's being viewed on the glasses and people are forgetting that when these glasses are switched on, Meta have the right to periodically content for future et cetera. And these Kenyan users are, you know, pretty much saying that they're seeing unusual uncomfortable events which include people having sex, showering in the bathroom, other things that really should remain private. Ralph Behnke: Yeah, this privacy I think is one that's set to explode. So remember this year we are going to start to see pins, earbuds. In fact, I saw the other day a pair of earbuds that have actually got forward facing cameras. So it's it's a earbud in your ear, but it's got a little forward facing camera on it. And so we've got ⁓ really extreme surveillance behavior. ⁓ know, we've seen so many protests over the years. Emile Ogier: Wow. Ralph Behnke: about surveillance cameras now we are literally becoming the surveillers. And quite interesting shift. Obviously this ⁓ of ambient is going to be landing I think in in quite a meaningful way. As we reported probably just last even, we're set to see the OpenAI Johnny collaboration. actually launch their ambient AI device. It doesn't have a screen, but it does sit there and record audio and your environment And so, know, every time you down in front of ⁓ a computer and a meeting, you have a variety of note takers popping up. Everything is being recorded and we're more more used to it. But Are we actually ready for what's coming when our life kind of broadcast 24 seven to our AI, which is going to be evaluating ⁓ to make sense of that and to give us value for that. I don't know. I don't know what we're going to get out of that. But yeah, it's a bit uncomfortable. And I think, you know, accessing people's lives in that way. and that stuff leaking out going to be the defining consideration of 2026. Emile Ogier: And in addition to that, Rolf, it's also quite a behavioral change that we used to recognizing that there's cameras in our, for example, in our laptops. And we know when we walk away, we can close the lid, but glasses are quite different. more casual about them. leave them on our bedside table. There's not an on or switch that's obvious to us. So yeah, I it's going to be a bit of a rough patch while we get accustomed to ambient technology. Ralph Behnke: But it's also going to drive a real I think, in 2026 for hardware Everybody is to be adding AI to their hardware. And that's obviously to consumers, hopefully, a reason to upgrade, at least ⁓ from the of these owners like Samsung and Apple. And so I think, you know, we're going to just see a slew of new products coming onto the market, essentially probably a much like Intel inside AI inside. And let's see. mean, know, the Apple launch of their MacBook Air M5 ⁓ and their M5 Max emphasizing this on device AI performance. That's just a good example, but there's going be a slew of this stuff. can't imagine, I don't know if you've ever been to Hong Kong, but if you go to Hong Kong, you will see that there are these tech stores, same brand, literally every block. And they're always packed with people. Now with AI, know, basically being added to bit of hardware, I can only imagine what's in those stores. I can't wait to visit. I'm actually leaving to China on Saturday. And so I'm really interested to see what kind of devices are in the shops and stores there doesn't seem to have hit Europe in quite a significant way just yet, aside you know, the meta glasses and ⁓ handful devices that we've spoken about. But ⁓ I think an entirely different story out in China. So I'm looking forward to seeing that. Emile Ogier: Looking forward to your stories. ⁓ Ralph Behnke: So why don't you off with a news What have you got? Emile Ogier: Yeah. So big news in the agent wars this week, Rolf, Grok is getting native MCP connectors. And I think it's going to be a game changer for them. First of all, what is MCP? Good refresher. stands for model context protocol, which is an open standard originally introduced by Entropic in November, 2024. Think of it as a USB-C for AI agents, a secure standardized way for any LLM to plug into external apps, data and tools without custom code or needing any risky API hacks. Instead of copy pasting context or building one of plugins, the AI talks to a lightweight MPC server that exposes safe permissible actions. And it looks like XAI intends to jump onto this quite hard. Late 2025, they added full remote MCP tool support for Grok API. And now it looks like they taking it mainstream for everyday users. According to the leaked but confirmed Grok settings UI, you'll soon be able to flip a switch in settings connectors and instantly connect to the likes of Gmail, Google Calendar, Notion, Slack, GitHub, Linear, Atlassian, which services Jeroen Confluence, Sentry, BillKite, Axiom, DeepWiki, HubSpot. and CoinGecko. Plus, in advanced mode, you'll be able to point Grok at any custom MCP server URL. Once connected, Grok becomes a real proactive operator. Imagine saying things like triage my inbox, create linear tickets from important threads, update Notion project page, and post a quick Slack summary to the team, all with your explicit permissions, of course. This will be welcome news for Grok users while Claude popularized MCP with its desktop app and community service. XAI is making it directly into the consumer Grok experience. So no extra apps or servers required people. It also leverages Grok's real-time X data, colossal scale compute, and now your entire work stack. Suddenly, Grok isn't just chatting, it's running your productivity layer as well. The current status at the moment is that it's rolling out gradually. Not everyone sees toggles just yet, but replies from Grok itself say expansions for Grok Drive and Home Assistant, et cetera, are already on the roadmap based on user feedback. Bottom line, 2026 is shaping up to be the year that agents stop being demos and start managing your actual digital lives. And XAI just took a Ralph Behnke: Thank Emile Ogier: massively forward by writing and extending the open MCP standard. So stay tuned and we'll keep you posted with these second connectors when they go wide. Ralph Behnke: Yeah, it's really exciting. And obviously they're playing a little bit of catch up to the likes of and Claude. this definitely way it's going. We're starting to really see the impact of this agentic capability terms of automating everyday And I have to say, I've been really enjoying using Whisper. you know, essentially talking to Claude code. Emile Ogier: Yeah. Ralph Behnke: and a little ⁓ command which kind of like right like me and ⁓ goes ⁓ and I have a chat with it and say please send this file that I've just built to ⁓ somebody it just goes off and sends it to them. Sometimes I actually don't realize that it's sending them stuff so I forget that you know just having these and not doing the action. I'm like did I send them that email? And then I have the conversation and they're like, ⁓ thanks for that email. I'm like, have to really think about what I was sharing you there. So there is weird thing happening as we kind abstract away from the work that we are not even sure what work we've actually done, which is hilarious. And somehow we kind of to have that three dimensional tracking of exactly what is going on. But I think that's really great news Grog. ⁓ One interesting tip that I got from a really heavy AI user was that he's actually using Grok to summarize best practice for a variety of different tools. And so one of the things he does once week, he'll say, find the top trending that describe best practice use of Claude code. for building websites, for building apps, for some kind of activity. what does is it basically brings back ⁓ all of results that have been shared by pro users. And he essentially using to the way that he basically sets up the system prompts and various kind of... Emile Ogier: Thank Ralph Behnke: folders that tell Claude code how to operate. I just think that's such a great tip actually for anything. So whatever trying to improve, there is probably somebody out there that is geeking out, whether it's to write book, whether it's pottery, whether it's how to pave your driveway, you can actually use to kind of summarize some of the best practice that he shared by users. And I just think that's great way of using X. Emile Ogier: That's a great tip. It's actually what I do. I use it a lot for weekly to grab the latest trending prompt for Nano Banana Pro amongst other things. It's really useful. Just wait for the trending topics, see what people like and get a quick summary and then play around. It's a real time saver. Ralph Behnke: And so hopefully with their MCP, we'll actually be able to suck that stuff and publish it to dashboards, which will make the process even more automated. And yeah, I look forward to that. it's that you track Grok and in that way. Emil, I have say, I've really enjoyed the of incursions with Grok. Emile Ogier: Did. Ralph Behnke: But it's still not made its way into my kind of daily usage. have to say I'm still really relying quite heavily on Gemini my deep research. I use with chat GPT quite extensively and obviously ⁓ Claude for everything else and most of my automation. ⁓ And this kind brings me to a bit of a personal reflection that I've had over the last week. and some conversations that I've had with developers that are using AI extensively. And one of the things that's transpired is, you know, we're really starting to see the potential demise of And one the reasons for that ⁓ is that LLMs are capable that actually, why would you go ahead and build an app? when you can just run that particular routine or activity Claude Code directly. Now, only reason to do it is most people aren't comfortable with getting into Claude Code, running a and, you know, these commands. It's not such a straightforward task. If you ⁓ are technologically challenged, which people are in fact, they just find it quite overwhelming when you move into the space. But what I can tell you is that I look ⁓ a lot of the and capabilities in cloud code, for example, it can do that is really, really impressive probably takes over, you know, 90 % the software, specialized that I've been using. So that is building presentations, analyzing data, building reports, work, sending interacting with my Excel spreadsheets. It is capable of doing all of those things. And so I kind of predict the depth of vibe coding as fun as it I think that probably 99 percent of every app that has ever been built through a vibe coding platform hasn't seen action. It hasn't. been used past one week. know, somebody it was a good idea to build it. They were probably a bit surprised that they could build it in five minutes. I mean, know, maybe a little bit longer than that. But then actually found that they just didn't need that app. And I think this is the case for so much software development. We've been pouring millions and millions into tokens to actually build features. people just aren't going to use. And this last week, especially in the work that I've been doing with Claude Code for clients, I'm seeing a capability way beyond what probably % of users are able to access. And so I think as that start, that reality starts to sink in, maybe the here is, you know, if you do want to build an app, just think twice about it. Push yourself to think about what it is that, what value you're going to be able to extract of an app in app form. And certainly if it's a private app, then, maybe you shouldn't even bother. You should just be using, you know, a capable model like Claude code, to actually do everything inside the existing features from the command line. It's not maybe as sexy as looking at a nice interface. And that might be a good reason to go away and build an app. But in terms of speed and efficiency of processing, I'm just seeing there's so many apps that I've stopped building because actually the way that Claude code deals with it right now is so much better than what I had in mind. the need for that app is just evaporated. And so that's, you know, quite a big shift for me in terms of the capability and, and just invalidating what was, ⁓ you know, new industry, Vibe coding, putting the power of building developing applications into everybody's hands. But now really, what's the point with the capability of these models? Emile Ogier: I'm with you, Rolf. The GUI is becoming less relevant. We just want the final product. makes perfect sense in my mind. Ralph Behnke: Yeah, so maybe we just need to it a little bit friendlier and accessible, not so command line centric. And of course, that is what Claude Cowork essentially has done. And we've seen some really interesting developments from perplexity in that space as well. So I do think that making it slightly more, let's say, consumer friendly is actually going to unlock massive benefits. So we'll have to see and watch where that goes. But that's just a bit of a personal reflection on where things are at. coding is changed, I think, at this point. And actually, Andrei Karpathy, in the last couple of weeks, there was a headline that he had predicted the death of vibe coding, presumably for the same reason as what I've just reflected on. I didn't manage to get to read the article. But yeah, things have definitely shifted. Emile Ogier: You Ralph Behnke: and should still build apps, I think it's a good idea, but you really need to test the value because it's really just a waste of time. Emile Ogier: And I think you kind of raised an important point. said you're enjoying whisper. I think that is the new interface. Natural language is replacing the GUI. That is the bridge between intent and end product. Ralph Behnke: Well, of course, I think how it kind of plays back information is key. So for me, Notebook LM is still of the tools that ⁓ I use regularly. It's a Google tool. It's really optimized research, but actually it is most phenomenal content creation platform. So I'll give you a bit of a sense. In the last week, we've built, you know, my development partner and have actually some quite significant specifications for a very large client project to automate their entire operations. And we've actually done is we've loaded those documents into Notebook LM. If you're on the paid plan, Google do users they don't use that information to actually train the model. So it is safe to load that proprietary into Notebook LM. And we then generate podcasts and videos and PowerPoints automatically to actually explain the solution back to us. And that's really allowing us different modes of consuming the material. And so it's really powerful. And notebook LM have actually just launched a brand new video feature before the video feature basically gave ⁓ a based with a voiceover. Emile Ogier: That's cool. Ralph Behnke: now they have moved to basically length presentations and this is a bit more dynamic so they're actually using live footage well I say live but they're using real footage instead of just basically giving a video of a powerpoint and ⁓ that taking to the next level I think that feature just dropped in the last few days so really really powerful of you have got some basic controls over the branding and the colors and the type of imagery that you want to use. You just have to put that into the prompt when you're actually building video or the PowerPoint or the audio. And LM does such a phenomenal job of generating that content. If you haven't used it, I highly recommend Notebook LM ⁓ to users actually free for up to 20 notebooks and ⁓ included the kind of the Google workspace plan. And this thing just keeps getting better and better. If you're a business analyst, if you're a project manager, if you are a CEO, if you are a researcher in any form, I think this is one of the hottest applications out there. Emile Ogier: Well, Ralph, I'm glad you pointed that out because we can't be everywhere, especially in a market that moves as quickly. And that's why we compliment each other. It's an absolute pleasure having you as a partner to navigate this exciting space. And I really mean it. So while I'm busy on X and you're using notebook, I've just subscribed to Gemini Pro and I'm very keen to get in and use more of Valerlem for notebook. And without further ado, I'm going to move on to a next story, which I just found incredible. The story really is bananas and fired from open AI for predicting AGI by 2027. Now his fund has exploded from 225 million to 5.5. billion dollars. Yes, you heard that right. In one of the most remarkable stories in finance right now, former open AI researcher Leopold Aschenbrenner has transformed his hedge fund situational awareness LP into a powerhouse under a single year. Now SEC filings show that the funds disclosed equity holdings have surged dramatically. While most investors chased Nvidia chips, Leopold made a calculated pivot. the real constraint on scaling artificial intelligence would be power, electricity, and energy infrastructure. His largest single position is now $885 million stake in Bloom Energy. The company specializes in modular fuel cells designed to deliver reliable onsite power to data centers. This bet alone makes up 16 % of the entire fund. He also built a major stake in leading AI cloud provider CoreWeave and now owns roughly 10 % of Core Scientific, a former Bitcoin mining company that is rapidly pivoting to host massive GPU clusters. The fund has been aggressively buying other former Bitcoin miners like Cypher and Bitdeer, repurposing their existing powerhouse capacity for artificial intelligence workloads. The strategy has delivered strong results. Reports indicate the portfolio gained around 35 % in the last quarter alone, adding approximately $1.5 billion in value. A sharp contrarian move Leopold even opened a short position on major IT outsourcing firm emphasis betting that the powerful AR tools will begin disrupting traditional offshore services work. This is not random trading. He's putting capital exactly where his influential situational awareness essay pointed to, warning that achieving AGI by 2027 would require an industrial scale buildup of power plants and data centers. From controversial safety research at OpenAI to one of the sharpest young investors in the AI space, and need I remind you, he's only 24 years old, Leopold is living proof that true situational awareness can create a billion dollar opportunities. What an incredible bet on the physical limits of AI revolution, and this story is far from over. It's just getting started. Ralph Behnke: Well, I mean, we've seen lot of young folks just on the money when comes to this. And, you know, just kind of goes to show right place, right time, right level of focus. And actually the you're uncapped in terms of the growth potential. That's such an story. I absolutely love that. Emile Ogier: and the right research paper. Ralph Behnke: and the right research paper. Well, and I think, you know, the maybe strength of your own convictions, just trusting that are seeing the signal. That for me one of the things that's so exciting about reporting, you know, on some of the the AI news, because start to see these really fascinating patterns and actually looking behind some of these things as Leopold has done into the energy sector, you know. But there are countless others that are also experiencing some fascinating booms. mean, we've just been at Dell's reporting of their server market and that is basically driving significant for them in the business. And this is kind of happening ⁓ all over. And think energy, compute ⁓ all of those, you know, minerals and materials that actually are required to build some of these tools. who knows what's going to happen with Taiwan. I mean, the world is such an interesting place at the moment. Obviously, China ⁓ desperately trying to get their hands on Taiwan for quite time. But as soon as something happens there, we have to realize that quite ⁓ a significant percentage the world's compute is still developed through the TSMC plants that are all based in Taiwan. We don't have the capability ⁓ at the that they're able to produce in Taiwan. So that's going to be a really, really big risk, but a opportunity for investors that are obviously backing chip makers in Europe and the US and elsewhere. Because as soon as that capability comes in line. Emile Ogier: That reminds me of that interesting story we covered last episode on the toilet manufacturer for those that might be interested. Check it out. Ralph Behnke: Yeah, that was the Japanese toilet manufacturer their ceramic skills. Anyway, so let's Leopold well on his journey to infinite wealth. I might to check out his fund. Emile Ogier: else you got, Rolf? Ralph Behnke: I was actually going to jump into a bit of a story about Microsoft and it's probably of two stories really. And the first one is that they have announced that they're really now bringing agentic building into SharePoint. It's just incredible me that Microsoft are actually, you know, quite slow to the party on some of these things. I have been playing with their capability within 365 co-pilot. But just, I've always found it a little bit lacking. And so it just hasn't really helped me, but I do it's interesting that they're going to be kind of launching this agentic capability within SharePoint. So it means that you'll actually be able build out your intranet, your workspaces ⁓ in this approach will be like multi-turn workflow. It's important because it's all about keeping governed content front and center. So, you know, you can really rely on Microsoft's security scalability and, know, why they kind of come to the party sooner? But there you go. are becoming first class citizens inside the enterprise repositories. And, you know, we're one step closer a report that was launched by Microsoft probably about a year ago now where they predicted that you know management essentially will become agent bosses. You will have a swarm of agents that will be doing the work and there will be key people in the organization that are ⁓ managing monitoring those swarms. So I don't know who we're going to be building these internet pages for maybe they will be for the bots but nonetheless I think for a short period of time, we probably will want to be unlocking some of that corporate data using these agentic tools. that's really welcome news. Obviously, we have seen that the likes of Claude Code have access to PowerPoint and have to Excel, and you're able to build some incredible things using those tools, which is a bit funny because it's Microsoft's home ground. And now we're also seeing is that there is news of the Microsoft OpenAI divorce. And this is a really fascinating story. It's been news for about a month, but I've been kind of looking at it. And as of February, 2026, Microsoft's AI chief, his name is Mustafa Suleiman. He said that ⁓ Microsoft is longer content Microsoft is no longer content being a glorified landlord for open AI. They're now rolling out what they're calling the ⁓ MAI-1. It's an in-house mixture of mode. So mixture of experts, obviously when you a prompt out to ⁓ Emile Ogier: Wow. Ralph Behnke: one of your LLMs, typically it has a mixture of experts, so a number of different ways of looking at your data and it then gives a coherent response and that mixture of model has basically proven to give us some of the best results but they're now rolling that ⁓ in-house they have basically trained on 15,000 H100s The goal is to replace GPT 5.2 in Microsoft 365 Copilot they're going to be using their own silicone and their own weights. So really starting to take control of their own destiny. Hopefully that's also going to mean that we're going to get much deeper integration with the existing Microsoft tools and features that really do take the Microsoft to hopeful. parity with the likes of Claude open AI and some of the others. so this a, an interesting, kind of of events. we'll have to see what it means. the economics of it fascinating. Apparently, they are basically paying a chat GPT tax. So, the bill was something like 8.69 billion. so every time you write a query in Microsoft co-pilot, obviously a from OpenAI on Microsoft and that was something like 8.69 billion quarter alone. So Mustafa has got a big budget to play with because every year he has the opportunity of saving Microsoft on a to 40 billion dollar bill. ⁓ So that's kind of nice, but there you go. It is a big draw on Microsoft margin. And I think being able to kind of bank that margin themselves, I believe that they think it's going to be about a 60 % saving. So obviously there's the cost of building all of this out and having to stand up their own LLM, but they're well on their journey to it and being able to cut your costs by 60%. That's certainly going to make Microsoft investors very happy. Emile Ogier: That is really big news. It reminds me of cloud journey, pay as you go models as banks and other big enterprises got on board. They quite set up the infrastructure correctly and they were receiving these startling bills in their first phase, the first quarter of the projects. Very much, I can see a similar pattern playing out there. The in his response he made publicly is quite telling. It sounds like there's might be a bit of friction between Microsoft for the open AI over these costs. Ralph Behnke: Well, undoubtedly. I mean, it is such a big bill. So it will be interesting to see what OpenAI does to plug the gap and just how long it's actually going to take for for them to solve the problem. But I think it's probably going to be, you know, play out in 2026. And obviously, you know, it should mean a much better experience for Microsoft users. It's been a little bit strange that Microsoft have had to piggyback off OpenAI in this way and haven't been able to bring their own technology to bear. So think this news has been long overdue. Let's if they can actually build something capable. Obviously, Meta are trying to do the same thing, but ⁓ taken bit of time. I think about weeks ago, they had released their first kind of version of their LLM internally, but we've yet to still see something coming out of them actually ⁓ for consumers. So that's another one to watch. Emile Ogier: Indeed, Ralph, I'd like to circle back on that earlier news item covering Apple. I'm going to just drop straight in. Apple caters to over 2.5 billion devices and they set to deliver private AI to the masses, no cloud required. While tech giants chase ever larger cloud-based AI models, Apple is quietly engineering a different kind of victory. One rooted in devices already in our pockets and on our desks. Fresh off its January, 26 earnings reports, Apple confirmed an installed base of more than 2.5 billion active devices worldwide, a staggering figure validated directly by the company and growing by tens of millions every quarter. Powering this empire is custom Apple silicon packed with dedicated neural engines built from the ground up for blazing fast on device AI inference that keeps everything local and lightning quick. rarely leave Apple ecosystem with iPhone upgrade loyalty hovering around nearly 90%. That's just incredible. And cross device ownership creating unbreakable stickiness that turns one product into a lifelong habit. That's loyal. is also supercharged by Apple's ironclad reputation for privacy. A brand promised back by on-device processing and private cloud compute that ensures your data never leaves your control. As AI models shrink through sophisticated quantization and optimization, Apple is perfectly positioned to roll out an advanced local intelligence with its current on-device foundation models already running efficiently on devices with as little as 8 gigabytes of RAM. Try and picture the scale. A single seamless software update could push sophisticated AI features to hundreds of millions of users in a single week, reaching more than most AI startups could touch an entire decade. Zero extra cost to consumers, perfect memory for your files and preferences, and total privacy. This isn't hype. It's a hardware mode that could make Apple the biggest winner in the AI era without ever needing to win the raw model size race. I must admit, Rolf, I think I entirely missed this play. I was so focused, solely focused over how far behind Apple was with Surrey and not having their own mature AI lab that I entirely overlooked their hardware modes. And it matters. Might this have been their strategy all along? I wonder. Ralph Behnke: Well, it's definitely their strategy now we see with some of the releases of their latest products, I think this week, actually March they're going to be shipping their new iPhone 17e and they've also got iPad Air. I think they're both priced at about $599 in the US and they've got these AI upgrades in them. you know, I think Apple the part... Apple is the hardware distributor to partner with if you are building any kind of software because the distribution ⁓ massive. So getting into that ecosystem makes Apple highly desirable. And I think you're absolutely right. They ⁓ are to run away with it a bit, especially as start to bring in some of their ambient AI tools. Obviously, Siri is getting the upgrade with Google Gemini. So that was a deal that was announced, ⁓ man, maybe six months ago now. We're just waiting for it to actually land. big upgrade there actually using Google. Let's I mean, maybe it will be a similar kind of story to the one that we've just spoken about on Microsoft, where Microsoft, know, insourced OpenAI into their copilot. Obviously, Apple insourcing into Siri. eventually, no doubt, they will come up with their own models. But the ecosystem and the hardware plays is so powerful. It is probably the last remaining mode. Software is no longer a mode. I did actually see that there was a project somebody had coded chat GPT with 243 lines of code. software we can build any system at any moment at any time right now. so yeah, hardware is where it's at and Apple is all about hardware and they're all about the ⁓ Emile Ogier: Indeed. And it's not just that Rolf, like sitting on a mountain of cash. So here's me thinking, why aren't you keeping up? You're so far behind. You're not going to catch up. Maybe they just waiting for the winner to get to the top of the hill and buy them out if they willing. Ralph Behnke: Well, don't forget they have started Apple Ventures. Apple Ventures is all about finding up and coming startups that they can actually use accelerate their own innovation. And no you know, ⁓ with all cash, they're to be able to put it to good use. So, yeah, I think you're right. I think Apple is definitely the one to here because it's in ⁓ many consumers hands, real trusted brand. else can compete with that? Maybe Samsung. Samsung the other one we haven't heard too much about over the last few months. Obviously they came out quite early with their AI enabled. I think it probably the S25, which was their flagship Samsung, the S26 ⁓ is just preparing to So kind of two cycles into AI enabled phones. of course, you know, they make so many consumer devices I Samsung will be the other one to watch and no doubt You know, start to see the same from what goods manufacturers so has AI started drop into your washing machine for example You know, there are out there that have got some integration, but still not every day kind of feature. So I think we've got a long way to run with this ambient AI and on device AI. But yeah, Apple's definitely got the lead on that in terms of consumer devices. Should I jump into my last story? I've just got one, an interesting one. Actually, ⁓ to something that I've just built for a client. So I a client who runs a chain business and they're really interested in what's going on with shipping, with tariffs, with commodity prices, these things affect price of the goods that. Emile Ogier: See you soon. Ralph Behnke: they manage and the production of. But a guy out there called Dan Ushman who basically created an application called CitDec and it's free at the It's an AI powered intelligence tool and it basically pulls 180 plus live data feeds and it monitors things like conflicts and earthquakes and flights and you believe it? Nuclear threats, which is becoming more more relevant ⁓ every of the day. So cyber elections, shipping lanes, markets, all of these things he's thrown in there. And ⁓ he's some really big success with this. His launch post racked something like two million views. The reason I bring this up is actually this ⁓ is possible now for much anybody that You know, has a serious business where you need to be monitoring macro economic events. And I've been looking at, you know, what I can actually get my hands on in terms of source data from a variety of Chinese sources. I, know, I have this client in China and, and there ⁓ are enormous number APIs. And so I have built the first version of this dashboard that ⁓ is also similarly to SITDEC. looking at shipping data and market data. I the first scan this morning, had a with the client, flashed it up, showed them, ⁓ you know, is some early test data. Immediately they zoned into some things that affected them directly that they didn't about. And so the power of these dashboards to basically connect into publicly available data sources and to ⁓ you know, synthesize the important information using your system prompt and your context setting. powerful. These are simple systems to I think it took me probably around three hours to get to a working version. And actually, know, is live and running. I haven't quite deployed it into production ready mode, but I ⁓ think this is one of the most accessible, tools. that people can start to deploy. a look at the of market information you need. Go out there, find APIs that you can subscribe to and go and build a dashboard. See what sense you can make of the world. Emile Ogier: Sounds good, Ralph. I've also got a tip for our listeners a slightly different theme regarding your safety. Imagine picking up the phone, only to hear your daughter's voice trembling with fear as she explains that she's been in a serious car accident and needs money immediately for bail or even medical costs. This nightmare scenario is becoming all too real thanks to AI voice cloning technology, which allows scammers to replicate a loved one's speech patterns using just a few seconds of publicly available audio. As confirmed by the FBI and its public service announcements and reports from the Federal Trade Commission, voice impersonation scams are surging, exploiting the panic of family emergencies to extract quick payments. In a striking example from Florida, a mother lost $15,000 when scammers used AI to clone her daughter's voice, creating a call convincing it sounded exactly like her child's distress cry about hitting a pregnant woman while driving. driving. The recommended countermeasure is both elegant and effective. Families and friends should establish a unique passphrase or code known only to the inner circle. This could be an inside joke from a memorable holiday, a quirky family reference, or any memorable phrase that is easy to recall under stress, yet impossible for outsiders to guess. If a call raises suspicions, simply request the passphrase. A genuine loved one will respond correctly without hesitation, whereas an AI clone will be unable to provide it. So maybe this weekend face to face over dinner to privately agree on a family passphrase. This is low tech defense is proving invaluable in the age of AI, empowering people to cut through digital deception and protect their loved ones from financial pain. With scams growing more sophisticated by the day, creating a family secret password phrase now could be a decision that prevents an unfortunate loss tomorrow. Hope that helps. Ralph Behnke: Yeah, no, that's a really good practical tip and it's a bit of a crazy world that we have to think that way. But there you go. If it's a you that's a really ⁓ simple of solving the problem. So, yeah, no doubt that will be a conversation for us around the family table tonight. Well, I think that brings us to the end of our show. Emil, thank you so much for all of your interesting stories ⁓ and thanks to listeners making it this far. We'll see you on the edge. Emile Ogier: Enjoy! Adios.