Latest / The Bleeding Edge Podcast / Weekly AI Brief 27 Feb
Transcript
- Ralph Behnke: Hi and welcome to the Bleeding Edge podcast. I'm your host, Ralph Benkir and with me on the show today, Emile Auger. Emile, it's the second time around, but how are doing? Emile Ogier: That's awesome to be back for a third time. Ralph Behnke: Oh, geez, to tell you what, if anybody wants to start a podcast, then just make sure you get some Valium with your Riverside subscription to make sure you can keep your cool. But here we go. So I'm going to jump straight into a couple of the big news items and then we'll trade stories as usual to bring you the best of the news in the last week from the realm of AI. Interesting things have shifted again this week. think, you know, ⁓ seems to be a lot more stories and examples ⁓ agentic operations. Obviously, ⁓ Claude Bott, Malt Bott and Peter Steinberger have ushered in have truly ushered in the ⁓ of ⁓ AI. So no doubt we'll be talking a little bit more about that. ⁓ I also wanted to call out that There was so much news last week that we completely to mention the story of Claude Code turning one. Happy birthday Claude. ⁓ And ⁓ 3.1 was obviously released as well. And that was actually quite a big jump from Gemini 3. we'll and catch up on some of those stories as well this week. top news One of the... Emile Ogier: Happy birthday. Ralph Behnke: biggest news items this week is that Meta is going to be renting compute from Google using Google's tensor processing units a multi-billion dollar ⁓ It's big deal, guess. Excuse the pun there, but I think this is interesting because Google isn't selling its isn't selling chips like Nvidia. actually renting out space. ⁓ When you do a deal with Google, yes, you're getting their tensor processing units, but you're actually getting them in a data center and all the infrastructure that goes along with that. that's kind of an interesting play. It really starts to show that Google is ramping up its to be the infrastructure model of choice when it comes to running AI. In other news, NVIDIA dropped some pretty big numbers and it's incredible to see. mean, their quarterly revenue is reported at $68.1 billion and their record full year revenue was $215.9 billion. So the GPU tax is definitely delivering for them. On European news, quite interesting to see that Accelerer AI has actually raised another 250 million. to ⁓ actually make AI chips. And what they're doing is they're actually scaling inference products. So talk a little bit more about ⁓ what inference means, but essentially when you are querying the LLM, that is an inference, it different to training the LLM. And there are optimized chips for both of those. And Accelera basically going to be making the chips specifically for inference products. And that is actually quite a big deal. If they get that right, then that is a really interesting solution for the market, which obviously ⁓ different what Nvidia ⁓ and Google are offering with their TPUs. So definitely one to watch. ⁓ On the front, it does seem like there's a lot of partnership news at the moment, certainly with larger and the AI companies. In this latest news, Mistral AI, which is the AI that is the LLM that is run out of France. It's a European sovereign AI and they have done a deal with Accenture. Accenture will be looking at ⁓ they can help enterprise Mistral AI. ⁓ And a multi-year collaboration that they've signed up to. Interestingly enough, ⁓ we've got a similar deal with OpenAI. ⁓ ⁓ course, they are also pushing into enterprise. ⁓ got something like 900 million monthly users. But Microsoft and Google, ⁓ they they've only got like a million users, I think, that are in the in the space, or at least million business customers that they actually have signed up to the And so their big push is obviously to open up enterprise and they are partnering with TechCrunch. So I you know, this kind of goes back to a story we had a little while ago. What is impact of AI on professional services firms? And in this we're really seeing that the complexity of putting large scale AI solutions together means you are still relying on professional services firms. ⁓ There's no doubt they're going anywhere soon. But of course, ⁓ the is greatly changed. so some of these partnerships are kind of interesting to see. how they can speed up landing these tools in enterprise. if we ⁓ have look at what else ⁓ Google been up to, there is 2, an image model that Google released, what, must be some six months back now. And they are rolling it out across Gemini search ads ⁓ and developer previews. So you're a Google user, ⁓ of the things you will see is that NanoBanana, which is their image generation tool, is actually dropped into a variety of different applications. And is what makes Google so interesting because they have this multimodal capability and they can mix and match these different services in their stack. anyway, so they're continuing the rollout of NanoBanana 2 and you'll be able to see that. ⁓ Gemini search and ads. ⁓ course, I've been using it in Gemini. So it's interesting ⁓ they are saying that they're going to be rolling it out further. ⁓ might need to dig into that and see exactly what it is they're offering. Microsoft also doing something similar with ⁓ integrating their tools into existing tool stack. what we know is that the whole world runs on spreadsheets. So no surprise, ⁓ have been extending the usage ⁓ Copilot and its to actually work with ⁓ modern workbook And so this no doubt improve your ability to do amazing things with Excel grids and finance models ⁓ and ⁓ status and like. Of course, Claude Code It seems to be the most capable of actually building complex Excel solutions Copilot Copilot Copilot has actually been playing catch-up there, but it does like Microsoft are trying to sort that out So that's a good thing to see in other news robotics This one from ⁓ So Google is actually a company that they have called intrinsic. That's their robotics arm And they're bringing it closer to Google DeepMind and Gemini. So DeepMind and Gemini are the teams that are responsible for some of the most significant AI developments at Google. just ⁓ ⁓ really starting to see that the is essentially the operating system for the robotics capability. So no surprise that they are starting to fold them into the same product organization. And of This is a little bit similar to what we've seen with Elon's play with SpaceX the acquisition of XAI. And obviously he's got the robotics sitting within Tesla. ⁓ And is of course rumored that he will fold that into SpaceX as well. And so can see, you know, the play that both Google and SpaceX are making there. They're really trying to... bring together the physical AI and of course, you know, what is effectively the brain. So that makes complete sense. But it's interesting that it's happening now. agents got a major organizational push as well. This time from Anthropic. There's actually been quite a lot of news, as I said earlier on about agents. Perplexity have actually just announced that they have got an agent that can do 19 things. and basically it calls the right technology order to either create a piece of music or generate a video or ⁓ an image, write some text. And that is very exciting to see. of course, ⁓ Anthropic ⁓ already quite a lot of that from my experience with Claude Goat. But ⁓ they've the ⁓ user abilities even They've actually acquired a company called Vercept. ⁓ so with Claude already, it's got the, the Chrome, browser integration. So it will literally browse any website for you. but, ⁓ this is actually going to be the next leap ⁓ that'll allow to talk better and do even ⁓ more in of multi-step workflows. so, I think this is the big shift really that we've been waiting for the AI is doing the work for us. ⁓ of course. the model takes action, we do need to look at guardrails, permissions, audit logs, all that kind of boring stuff. And going to share a bit of an experience of why that will just maybe save your life one day. So finally, the governance front, ⁓ I've been this one for a while, the AI Act from the European Commission. And so the Ombudsman have asked for some implementation guidelines. It's kind of interesting because obviously they've published the guidelines now for going on two years. There have been some updates to it. This is really important one for people to keep track of because if you're building an LLM application, then this EU act governs it. And so even though you're not the provider of the LLM, if you're using an LLM in your app, then definitely need to be aware of what is in that. AI Act. Fascinating though that the EU Ombudsman is asking for clarity on implementation guidelines. That's probably the biggest bottleneck everybody ⁓ it's how you translate what's on paper into so what does it mean for my business and what do I need to be doing differently. On one related so I always have stories about what went wrong with AI. ⁓ There some reports that are starting to surface I haven't really been able to this, but it does look like ⁓ AI tools have been used to accelerate elements ⁓ a operation. So ⁓ that's interesting one. I'm still digging into that. But I think the one that we will also talk about is a guy that managed to connect to thousands ⁓ of AI run cleaners. That's kind of an story. ⁓ Yeah. Emile Ogier: I have that story. Ralph Behnke: ⁓ We'll talk a little bit more about that. So some interesting ones in the news there today. Anything catch your eye Emil? Emile Ogier: It's always all very interesting, Rob. Thanks. Great roundup as always. Really important to highlight the two tranches of demand on compute for AI that first the models need to be built. They typically take anywhere from weeks to months to get done. And a tranche of work to ingest a massive amount of data for that inference. And once that's available, then ⁓ they need to continue to run the model to service the queries for the day-to-day work for clients such as ourselves accessing it. So really good that you delineated those two demands. A lot of energy is required for both. all the AI labs, ⁓ finding it quite interesting. I think, or at least to me, the strategies between them are becoming more clear. ⁓ there seems to be a bit of a divide or a change in focus on, know, where they're going for the gold or the trophy, so to speak. Anthropic have been very much always focused on the vertical. they, they put all their soul energy into coding. It makes a lot of sense and can see where the, why the rationale plays out there. If you get coding and logic logic, right. You can ultimately build anything you want, which would mean you can start going off to other domains, open AI. They seem to be solely focused on agents. they, brought in the individual that created open claw and they, they doing a lot of good work there. That also makes a lot of sense as a strategy, build your army and you can achieve a lot. separately with Google, their strategy has always seemed to remain the same, even since the days of deep mind. And ⁓ kind of forgot about it and they keep reiterating the point is they they're all these wonderful models. They're not charging for it like AI. They're not using an ad-based model. was it Gemini 3.1 just came out. ⁓ looking impressive on benchmarks. They're not charging for it and their sole focus seems to be around nailing intelligence. that as a strategy also makes a lot of sense to me. Create the perfect ⁓ that is ⁓ super intelligence and if it has all the intelligence in the world it can solve any problem. So they all ultimately trying to achieve the same thing just taking slightly different approaches and I'm just really interesting to see where the journey goes and to see if there's any type of convergence down the line in 26 and 27. And on your last point regarding the convergence on Elon Musk's companies that's also quite interesting listen to a really informative podcast on moonshots where they interviewed Elon Musk earlier this year or late past of last year. And they'd asked him if he had planned this great plan to bring all these companies together. And he freely admitted that it wasn't part of the plan. ⁓ The most interesting aspect of the whole thing was he brought up the singularity and he said, you know, that's one of the byproducts of the singularities. As you get closer to the edge, things do start to converge. So just, I found that quite amazing as a response and it just gave me a lot of pause for thought. Ralph Behnke: Yeah, I think that's kind of where we're heading. And I've seen that have invested, or promised an investment of something like $50 billion into OpenAI. And one of the tenants of that agreement is that they float so that they have to actually IPO. And then the one is that they achieve AGI. And so... ⁓ What an amazing kind of statement there. You you'll get this money once you've achieved AGI. hope they have a better definition of AGI than rest of us. It will be interesting to see what they have considered as AGI. But of that level of general intelligence, artificial general intelligence is something that ⁓ many say we have surpassed what we had set out. Certainly GPT 5.2 kind of exceeded so many of the benchmarks in nearly all disciplines in terms of what a human being can actually deliver. And so it is smarter than us by some percent on nearly everything already. And obviously the level of autonomous operation that many are achieving with the likes of Claude Code, I think a number of people have said, look, we're past AGI, but there we are. You know, we've got Elon Musk talking about this convergence happening. Clearly it's happening. Everything is becoming connected, absorbed into the center. so no as essentially land same capability a combination of ambient devices, we must talk about ambient devices because there's some interesting movement there. ⁓ So these are devices that on your person, they're always listening, they are recording absolutely everything. And then obviously physical AI, which is the robotics. guess many of us have got used to the robot running around the lawn. ⁓ maybe you have one of these robot vacuum cleaners that has now been hijacked by this guy that managed to hack them. so is its way into that physical ⁓ form if like. then, you know, if ⁓ news is to believed, we will have one of these robots that we will be leasing probably for ⁓ the of a small car standing washing dishes, making our toast ⁓ and our And so, ⁓ know, all that is what's driving this convergence. You know, those things a brain. They need to be able to operate autonomously. And AI seems to be very well placed to do that. Emile Ogier: Indeed. Ralph Behnke: See you there. Emile Ogier: I've got, and I thought I'd switch it up a bit this session, Rolf, and have a bit of a, a bit of education myself to the table. So without further ado, let's get started and unpack one of the most useful features in Google Gemini. Thought it'd be useful to explain the difference between Gemini Pro instructions and Gemini Pro Gems, their pros and cons and when to use each. Think of Gemini Pro as your all-round helpful friend. Custom instructions are like giving that friend a sticky note or sticky notes with your general preferences. Examples being always reply in bullet points or respond in a friendly and professional tone. Gems are like cloning that friend into multiple specialists. One gem is your writing coach. Another is your code debugger, each with their own detailed rules, personality, and even upload files that they can remember forever. That's the core difference. Custom instructions, these are global settings for the main Gemini chats. You set them once in an app under your profile and they apply automatically across all new conversations. The pros are it's super fast to set up. It's easy to set up. It works across all your chats and it's perfect for simple habits like respond in this particular format or tone when I speak to you. Cons, it's limited to basic preferences. There's no file uploads. It's not task specific and it can. be forgotten in very long chats as your context window increases. Gems are the dedicated experts. Gems are separate custom AI assistants. You can give each one a name, detailed instructions and upload files for knowledge. They live in their own chat window and always remember their role. The pros are they're highly specialized and consistent. You can upload files so they know your documents or your data. They're great for repeatable time. tasks feels like having a team of experts. Cons, you have to switch to a specific gem each time. It's only private to you, slightly more set up time, and best use of Gemini Advanced for taking full advantage of its power. So when to use each? Custom instructions are good for when you want everyday improvements like shorter answers or professional tone. You're doing quick general chats. You just need one simple rule across all your chats and you don't want to keep repeating yourself. gems when you have a recurring specific job. You need to write emails in your brand voice or review code. You want to give AI your own files or data when you need a focus expert that never forgets role. So how it's helped me recently, I was trying to set up a home lab on my Raspberry Pi and I got involved in some pretty long conversations with Jim and I. And as the conversation progressed, the model started to drift and hallucinate, and it became increasingly frustrating trying to frequently steer it in the right direction. It kept defaulting to offering me generic advice for standard Raspberry OS build, which I told it I did not have. This ultimately got me into quite a lot of configuration issues, which took me a lot of time to resolve. The solution created a clear system manifest outlining my setup in detail, which is now used in all conversations. First in instructions, which I came to realize is not token efficient as it can also lead to context leading into unrelated conversations when I was not trying to work on my HomeLab build and it had nothing to do with the HomeLab build. it's now moved to a gem to draw from it as I need it. So. challenge to our listeners, open Gemini this weekend, set one custom instruction like always use emojis in fun chats, then create your first gem called writing helper with instructions to write in clear simple language or whatever structure you prefer and let us know how you get on. ⁓ Ralph Behnke: Yeah, some good advice. think, ⁓ know, getting those instructions set up actually can be a big time saving because obviously every chat that you have, it kind of applies those instructions. I have used gems very extensively and I really like them. In some ways, I to say I still prefer chat GPT projects. And I also really like Claude projects, actually just using the desktop version of Claude. So my problem is that I obviously have gems, have chat GPT projects, have Claude projects. And so I have to sometimes compartmentalize like where did I store that now, you know, I'm kind of going through the cupboards, where is the can of beans? And so I can't Emile Ogier: Interesting. You need to organize treasure chests. Ralph Behnke: Well, of course, what we need is the orchestrator that sits on top of all of that. I think this is, you know, very much where perplexity is trying to play for so many of us. It's about what model should we choose, which one is better. And actually, if there's of two camps to this, don't get too comfortable with any one model because things always changing. I think that's the one side. ⁓ And so as many as you can. But the other side is something that works for you and go deep and just become as comfortable with it as you can. ⁓ ⁓ kind of flip between those two modes. Possibly ⁓ allows to maybe pick a third option. So as I said earlier, they've actually ⁓ been on something which is a multimodal digital worker. And we see where all of this is going. So, know, agentic AI. where the agent is essentially following your instructions. So that instruction set that you were just describing there, the content from your projects, from your gems really coming to play here. of course, what the model is actually able to do is execute on your behalf. in particular model, know, perplexity are giving you these kind of ⁓ access 19 models, ⁓ which available for max subscribers on the web. ⁓ And it essentially an agentic workflow operating system. So this thing will orchestrate many models and it beats betting on one model. So, you know, I'm not using it because I'm looking at one model like Claude code being so capable and, able to do all of this stuff anyway, that I haven't seen the need to go to perplexity. But think, you know, Claude code in itself has some barriers to entry. It's not the most straightforward environment. for people to be logging into a terminal window. Cowork changes that somewhat. And ⁓ so, know, we'll have to see how this plays out. think Perplexity have created a space for themselves. I think this concept of agentic AI and actually having your team of AI agents doing work for you is the next big shift. So back in the day when remember Hotmail and Yahoo, that was the first time people were kind of getting email addresses on, well, was the first time people were getting email addresses. We didn't have a need for them before 1992. If anybody can remember that far back, I certainly can. so, you know, getting your team of agents, I think is very similar to, you know, getting a URL, getting an email. It's just that these things are way more capable. And so my prediction is that over the next three months, Much like we saw with Vibe coding in 2025, we're just going to have thousands and maybe even millions of people adopting a cadre of agents, an agent swarm that are going to be doing their everyday mundane tasks and maybe even not so mundane tasks. And I think this is really what Perplexity and of course Anthropic with Claude and even Google are kind of all now trying to unlock solutions that make it accessible for the consumer user. just finish on the story with ⁓ related news from Google is that they're now pushing agentic automation with Gemini on ⁓ Android. And they're an step into Opal. Opal is one of their development environments, a really capable tool. it's not available just, everywhere the world just yet, but you can get it, get access to it on a VPN. And it's really great for building, all kinds of applications for the web. but of course, you know, getting these agentic agents that can actually build the stuff for you. and then, you know, the final piece, which I have uncovered certainly over the last few weeks is using voice mode. So instead of typing on the keyboard, You're actually voicing, you're explaining to these agentic enabled systems what it is you want them to go away and do and they go away and do it. And so that's the promise of what is here and now. And certainly what is coming to most kind of production, let's say productivity enhancement plays for everywhere. Get ready for your agent swarm. Emile Ogier: Thanks, Ralph. That's good advice. I've also jumped in the bandwagon of using voice and wow, what a difference. I had done all my research. I was really keen to use Whisper Flow. It's a phenomenal product. It's getting rave reviews, all the big podcasters and the guys in AI use it. have tests run with it. It's exceptional at not only getting tone inflection and your accent right, ⁓ having South African accent. I've struggled with many dictation tools. Siri is just awful. for a South African accent, Whisper Flow nails it every time and it even understands instructions outside of text it's trying to convert to. So if you need it to do something while you're taking notes, can also... separate the logic from the text if that makes sense, which is just quite incredible. But having said all that, you do need to, you know, fork up a bit of money for it. It is at a premium tier price. I'm not sure what the exact price is. I could be off. I think it's between $10 and $15 a month. But I did do a holiday research. If you are a Mac user, I think it extends outside of other operating systems. I've installed whisper notes and I think it's just absolutely fantastic. I spent five pounds on it four pounds 99 it's a lifetime subscription and it's working perfectly for me so you know try it out and I'm there's no going back I think my fingers are gonna start getting lazy not my efficiency rate has gone up 300 % at least Ralph Behnke: Wow, that's impressive. we'll be expecting a lot more stories from you ⁓ next week. If you could just get the words out faster. Emile Ogier: And a lot more output. But Rolf, before we continue, I just want to circle back because that story on gems and all the additional features on the other labs you brought up. Very interesting, very helpful for our listeners. And it got me out of a lot of pain points, but I know you've had quite a different, interesting story and quite a painful one this week, whereby I think it worked against you and I'm not entirely sure why and I'd love to know more about it. Ralph Behnke: Yeah, well, I was setting up Claude code to not run destructive commands. You know, ⁓ you know about these things that ⁓ code likes to use something called bash commands. Bash is just a way of giving the ⁓ LLM sophisticated set of instructions that it can go away and perform. I was and there are exist things called destructive bash commands that basically will delete your drive and inadvertently the explaining to code that it cannot these bash commands it kind of ran the bash and before ⁓ I could even five times it had literally deleted all of the system files of my machine and did such a good of it. that when I ran the diagnostic afterwards to find out why I was having so many problems was saying, well, this was a really good job. You know, they've deleted everything. Like I can't even find some of the evidence trail of ⁓ actually And of course Claude code had done this, you know, just before having said that I panic and ⁓ I set about a And this is really the most amazing thing. back in the day, had you deleted your entire hard drive, you would be hopping mad. You ⁓ you'd throwing things, shouting. My family noted, why are you so calm? Do you love the AI more than us? ⁓ And obviously had nothing to do with that. I don't actually know why I was so calm. just kind of, it's a whole new experience knowing that whatever problem you have, there is a solution with AI. So I did set Claude code, task of helping me get back on track. ⁓ what was fascinating is that actually the full diagnosis ran across two instances of Claude, because I was taking instruction from a different Claude system about how to ⁓ up this ⁓ guardrail. And so we did was we actually, my development and I, actually took both sets of transcripts, instructions and history from our various sessions and we combined them into one and Claude found the exact moment in our conversation when this particular destructive command was run it actually even told a really interesting story of how we were trying to diagnose problems ⁓ that were down to my system files being deleted and you ⁓ know were able to do a full triage really uncover exactly what went wrong. So I will say this wasn't a ⁓ issue. ⁓ I think was definitely a human error. ⁓ I ran that I shouldn't have at a moment that I shouldn't have and so I put this down to just the wrong execution in the wrong moment. We hadn't yet fully set up the guardrails. And so it basically did exactly what it was instructed, but the power of the tool and the speed at which it did it was what impressed me. And then of course, what completely won me over was the ability to sit and diagnose what was, you know, a reasonably technical discovery of what had gone wrong, because we didn't know at that time what it was. And it absolutely nailed it in record time and nailed the restoration in record time. Now, I won't say that I didn't lose time. I probably lost about two days this week rebuilding everything, but nothing material was lost. Everything that was in the cloud was still saved in the cloud. really just speaks to the fact that these models ⁓ got access, obviously, to your folders. This is the benefit of Claude code. It is able to execute anywhere, at any moment and do anything. ⁓ you can maintain the guardrails, then there is huge productivity to be gained. But if it goes wrong, it's going to go wrong in a spectacular fashion. You just have to be able to keep your cool and run the diagnostics and obviously, you know, try and follow the remediation approach, which I have found to be very capable and it definitely worked. I don't to repeat that process again, that's for sure, because it did cost me quite a bit of time. But my confidence that these machines handled in the right way, they break things and they can fix things just as fast. Emile Ogier: Thanks, Rolf. That's a really great lesson. I'd love to know once you had time to distill the process, ⁓ you can do in the future to try and avoid those pain points. I think it'll be useful for us all. Ralph Behnke: Yeah, and we've set up ⁓ guardrail files. And basically, it's just a file that says, look, don't run these commands. So kind of a little bit like the 10 commandments, ⁓ shalt not delete the hard drives, is what it says. And there's a specific bash command that does that. And so as long as you have some of those things set up, I think the problem for most people is, ⁓ and we've this with the ModBot, CloudBot people don't understand how to go about ⁓ the tools secure. So unfortunately, automation handing over the reins to your email, handing over access to your folder structures, handing over to various ⁓ MCPs and APIs that connect you to many different tools. so think that just needs a degree of care that maybe we're only just starting to ⁓ learn how to do that. And as users have the maturity. And with think ⁓ AI is capable of guiding you. The problem for most people is that they don't know what questions to ask. They don't know what problems to try and avoid. And so Maybe this is we need AGI, ⁓ we cannot think of all of the things we actually need to do to make a particular operation safe. And the AI isn't thinking about it either until we ask the question. And so ⁓ ⁓ you have an expert sitting at the table and you're talking about a hair-brained plan, drilling the center of the earth, they'll tell you why that's a really hard thing to do. And it's the same, you know, equivalent here, except we're supposed to ask the question, what's hard about it? What can I get wrong? so many of us don't do that because we just don't have the context that those problems even exist. Emile Ogier: Absolutely. And just to add to that also on the moonshot episode with Elon Musk, quite fascinating is just saying, you know, it's getting to the point where he doesn't even get past a week where the model doesn't keep surprising him and bowling him over. And one of the examples he gave is, know, we spent putting all this focus on the models giving us these perfect enriched answers, but the models are getting so good, they're starting to ask questions that they can't even understand. you ⁓ know, the question and get even better and bigger answers. So the sky really is the limit. ⁓ but without further ado, I'd like to dive into that vacuum story because I did find it quite wild. so this guy apparently drops $2,000 on a DJI robot, which is a Asian model vacuum. And a lunatic, he refuses to use the normal app, ⁓ just a normal person. So he fires up cloud to see if he can crack the API to allow him to drive it with his, his X-Box controller just for fun. Cloud comes through and delivers. He then proceeds to pull an authentication token from their servers, connects it successfully, except the system now thinks he controls 7,000 vacuum cleaners. He checks again. Yep, saw that right, 7,000 vacuums. Essentially, DJI built an authentication with no device ownership verification, meaning that any valid token would work for the entire fleets across planet Earth. So this guy, Sammy, now has eyes inside homes across 24 countries around the world, camera feeds around the house as it operates, the full floor plans from all the mapping data, unaware that vacuum cleaners are ⁓ snitching. And that's it took one API call from being one of the most informed cat burglars in history. So this was never the plan. He just wanted to have fun with the new toy and with his game console. story does end well. He reported it and decided to stay with the force and not go to the dark side ⁓ DJI fixed the security vulnerability in two days. But it just goes to show that ⁓ companies are still shipping products with little to no security oversight with regards to security. So, you know, a good lesson learned, I think. Ralph Behnke: Yeah, no, I love that because it just illustrates that the interconnectivity that we've created, you know, there are vulnerabilities that people can expose. And so if companies like DJI ⁓ aren't catching it, then what else is out there? And so obviously people are going to be exploiting some of these vulnerabilities as they have been. for a long time, but it just gets a little bit even more invasive. And so a couple of other stories just in terms of hardware, Apple are also now pushing for their ambient pin. So one of the new stories that broke in 2025 is that OpenAI and Johnny Ives, the designer of some of the iconic Apple products, they joined forces, multi-billion dollar deal, all that. And of course, the thing that Johnny was supposed to be bringing to the table was this ambient AI device. so apparently it's still on track for release sometime in 2026. I think probably see it around ⁓ the of the year. And the idea behind the product is that it's always on recording absolutely everything. It doesn't have a screen. And ⁓ but it of course interacting. with the AI brain, whatever one you have to it, we'll have to see. Obviously, OpenAI will make ChatGPT the operating system for that ⁓ ambient AI tech. Apple are doing the same thing. And so we are going to probably see a slew of these devices on the market bringing this, ushering in this new era of ambient AI. What are we going to do with these things? Well, we don't know. actually, I a bit of research on the sentiment amongst people these tools. And I will say that the market is pretty divided. About 50 % of the people say no way. I'm not having an AI hanging around my neck recording absolutely everything. And other people are saying they can't wait for this tech because it's going to make such a difference to monitoring, let's say. an aged parent that has some medical issues and of course, you know needs to be able to converse and chat with something that can record what's going on. And so I can see, you know, both positive and negative use cases for this. Maybe Meta's glasses are the furthest that we've got because a lot of people are using these Meta glasses to record video and audio and also stream audio. There's a little bit of a heads up display. so they're getting information on the screen. And of this ⁓ is ⁓ pointing, you know, what I can see to wide scale adoption ⁓ of ambient And I actually it's going to make a really big difference to lots and lots of things. mean, ⁓ just imagine in sports, having one these on player, you know, on ⁓ a football team ⁓ or on rugby team ⁓ basically get you know I say 360 imagine a 360 degree coverage from every single player of their experience on the field and can see you know uses for this in medicine I can see uses for it in engineering and ⁓ actually that ⁓ it's going really usher in whole slew of new ways of working and operating that we can't just yet understand, but maybe one of the things that of course is going to come to the forefront with all of this again is privacy. So what actually happens to all of those conversations? And of course, it's not just your conversation, it's anybody that's talking to you. And so will people see the fact that you now have either the Apple pin or you have the ⁓ AI version, sorry, ⁓ the open AI version. Are they going to say, okay, I'm not talking to you? So it's a little bit like a reporter standing there with a microphone. You can choose to scurry past, but now everybody has one of these things and they're recording absolutely everything. So that's going to be quite fascinating. Emile Ogier: Yeah. And I think it's also going to make a massive difference. I think it's going to create a huge crack in the wall for the likes of privacy regulation like GDPR. mean, how consent exist if everybody is wearing an ambient device around you? ⁓ I think it becomes irrelevant. Ralph Behnke: Quite right. And of course, you know, when you've got people like Sammy into DJI robots, ⁓ to stop them to hack into your your meta glasses ⁓ you know, whatever ambient device you're wearing from either Apple or open AI. So, yeah, I think. Emile Ogier: Yeah. And And you also got to remember it's a two-way street, right? So I've always been big on privacy, as you know, Rolf, and I've looked at at all the AI models, spent a lot of time reading through the literature and what I can choose to consent to and what not to consent to. But conversations invariably always have more than one person involved. So even if you not interested in it, the other person could be granting access to their conversations, which include you. So, you know, I really do think we're losing control control over privacy. I'm of pretty much giving up ⁓ ⁓ to defend the position now. I think it's coming. Data going to be available for everyone. Maybe in future we'll build new walls, but temporarily while we go through this new displacement. Yeah, I think that's just our new future for now. Ralph Behnke: And of course, this also touches on a story that's been running over the last few weeks with Anthropic, and they've been refusing to the US government to use Claude for surveillance and ⁓ of drone systems. And they're basically that already the AI is capable. Emile Ogier: That's right. Ralph Behnke: that it can essentially process the communication that ⁓ social media or any AI LLM related tools are gathering ⁓ and is able to set out your opposition for example, and where the voices of dissent in society are. ⁓ And ⁓ that is that we can do right now. Anthropica are saying we shouldn't be doing that. The US government is saying we want to be able to do that. I think as citizens we're saying yeah no that's a level of surveillance state that we don't particularly want to enter but I think what you're calling out Emile is you know there's somewhat of an inevitability to all of this. you know, how do we navigate that? Because ⁓ don't really want to end up, you know, in that black mirror episode where you are getting ⁓ credits, you know, so for how you behave, they're essentially monitoring everything that do in China. ⁓ I know that are some millions of people that are locked out of housing and benefits because they did something at some point. Emile Ogier: Yeah. Ralph Behnke: and they don't have the social credit score that allows them access to those services. Now, I just don't think that that is a good thing. That's a terrible thing to have happen because how do you navigate yourself back from that when the machine is essentially running an algorithm it's you that you're not worthy. Machine says no. This is kind of where we're heading and this is what Anthropic are specifically, you know. ⁓ Emile Ogier: That's right. Ralph Behnke: having an issue ⁓ the US government on and it's a $200 million deal. ⁓ I believe it's due to be signed today. So have to see what ⁓ what of comes out of that, whether they've actually capitulated and they've Claude to be used in these, you know, nefarious ways or whether they've held the ground. Emile Ogier: That's right. They've given them an ultimatum at 5pm today and they effectively not just saying you'll lose the contract, they're saying or else. So can you imagine the military not being happy about you? So yeah, it's tens times. I'll be watching that with a lot of attention. Ralph Behnke: Awesome. Well, what else have you got for us? Emile Ogier: Um, yeah, so I, I know there's a lot of doom and gloom stories out there. thought we'd try to cover something a little bit more positive and look towards a future of abundance. So I'd like to offer an alternative view into the future where the economy doesn't just grow. Imagine that it sprints imagine 200 % GDP gains, AR managed supply chains and a society redefined by hyperscale. So sit back and enjoy the short video offering. revision of the year 2023. Rolf, I'm sending it to you now. Ralph Behnke: Are you sure it's the year 2023? Emile Ogier: I'm so sorry, 2032, thanks for the... It's titled The Nation That Learned to Sprint and it was created by Alex Wisner-Gross. Ralph Behnke: We're going back in time, the good old days. Right, so let's try and this thing running. ⁓ on YouTube. Emile Ogier: Thanks for that, Rolf. I quite enjoyed that. It's quite a provocative vision. The takeaway isn't just tech, it's the tension. I just love the ending drawing attention to Maya's own irreplaceable slow heartbeats in a nation moving at warp speed, what happens to the human pulse. And it raises some thoughts such as the challenge of the next decade isn't just about managing growth, it's also about protecting the rest node, that human element that machines can't replicate where we need to. regroup, recharge and stay aligned with our unique fingerprint that makes us unapologetically human. I think it also serves as an important reminder to pause and question whether the tools we are building today are ultimately to serve humanity or we are simply trying to keep up with the pace of our own inventions. Ralph Behnke: Yeah, and it's a very compelling vision to set out that kind of interconnected nature of AI's kind of running entire economies, certainly within a council. And I think undoubtedly that's the way we're heading. Whether we get there by 2032, I think is going to be a fascinating question. So there are still quite a few Emile Ogier: You Ralph Behnke: issues to overcome. So one of the biggest problems that we have and this is what everybody's going to face here now, are very soon, ⁓ you haven't already, you're going to get your head around having your agent as we were talking about earlier. You might have your seven or eight agents that are answering your email, writing your book, talking your mom. so things will be going about their day and they will be using the instruction file that you spoke about earlier, you know, the equivalent of the instruction file and maybe a couple of Gemini gems that are it context. But ⁓ that works when it's ⁓ seven, eight nine agents. Imagine that now across an enterprise. And ⁓ you start realize the scale of the task of actually keeping ⁓ all of those informed about where the company is heading. And, ⁓ know, we can't even find out as, you know, employees about where the company is heading. Often, know, people are not told about the things that are happening. And so in a much kind of slower age where information diffuses in a different way. Whereas with our agentic tools and systems, they want absolute truth at any moment in time in order to make the right decisions. And so I think we're going to have to learn how to work in that way, transparency and communication. It's going to take on an entirely different meaning. I think, you memory, actually memory of what ⁓ each execution has just achieved and actually feeding that ⁓ in to the of the agent swarm so that it can also take into account what has just happened. One of the interesting things that has always entertained me is when you throw something into a drop of water, water has this amazing property where essentially, you know, a billion molecules, molecules respond almost instantaneously to whatever entered the water. And it's a little bit like this new information comes into the organization via an agent and the whole system is aware and dynamically responds. We just don't have the level of sophistication of setup just yet. But of course the tech is actually there to be able to do it. We just haven't figured out how to make the system context aware enough so that it isn't working against itself. And we're seeing this now on Claude code. Claude actually does a planning activity it goes into multi-agent mode where it looks at what it can run independently, what it needs to, what it can run in a parallel. and what it has to do in a serial fashion because there are dependencies and Mapping out across an entire organization ⁓ is going to be the next challenge and I think technology companies solve that problem of keeping each of the agents tied into the system and context away that is where ⁓ The vision you just saw in that video. That's when that becomes a reality. We solve that along with actual memory where it is building new context. ⁓ a little bit like our nervous system in our body. Every we take all this information in and our heart rate is changing, ⁓ our pressure is changing, our temperature is changing, all of these systems respond. And that's invention. The fact that the body can do that. We're going to to have agents operating in the same way. in order to achieve that vision in the video. I think that is definitely what's next. That's going to consume almost all of us that are landing agents and agentic workflows enterprise. Emile Ogier: For sure. Yeah, it was just a bit of fun and just a positive, another way of looking at things. I liked your analogy on the water, Rolf, akin to a data lake synthesizing a water droplet. Very cool. Ralph Behnke: Right, exactly that. so, well, I think those are all the stories that I had for this week. A really interesting roundup. Have you got anything to end the show on or are we there? Well, you really finished us on a high note. I'm smiling because I'm really excited about this glamorous future. can't wait for 200 % growth and neither can my bank manager. So we'll have to see if it becomes a reality. Emile Ogier: I'm good. Thanks guys. See you on the other side Ralph Behnke: I've marked it in diary, Emil, to give you a call in 2032. And congratulate or ask you what the fuck went wrong. ⁓ But anyway, we're still the show by then and we can talk about it. Emile Ogier: And don't, please don't take this as financial advice, but even Elon Musk is saying ⁓ on the same moonshot episode is like in five to 10 years time, don't screw money away. You're not going to need it. So that's quite, quite, quite the bearish comment to make. Ralph Behnke: Yeah, I'm clearly following his advice even before he gave it. there you go. thanks very much for a great show. We'll see you on the edge. Emile Ogier: Hahaha Adios.