Latest / The Bleeding Edge Podcast / Weekly AI Brief 1 May 2026
Transcript
- Ralph Behnke: Hi and welcome to another episode of the Bleeding Edge podcast. I'm your host Ralph Benke and with me in the studio today, Emile Ogier. Emile, welcome to the Edge. Emile Ogier: Good to be back, Rolf. I like the matching black and white like I was saying earlier. Makes me feel like a strategic AI chess piece. Ralph Behnke: Yeah, we've got the uniform going now, so we'll have to see if we can keep coordinated. reading your memos. obviously exciting show again, because the news just keeps getting better and better. I've actually quite a bit of time on the most AI development out there, Emil. And I guess you can guess what it is. usually I try and start with a joke. I think I failed in the last few episodes because the jokes have been so terrible but today I decided to write a joke myself. You want to hear it? Emile Ogier: Please, fire away. Ralph Behnke: Okay, an LLM and an Irishman walking to a bar. was the joke. And the lens can't walk. ⁓ anyway, so I guess I wrote half a joke, but in honesty, I did actually take another spin at basically a Claude skill writes jokes and something really interesting came out of it. So I said to ⁓ Claude, was having this Emile Ogier: Sorry. We're gonna have to fill in laughter. Apologies. Ralph Behnke: that the jokes just aren't funny. what it said was, well, the problem is this, LLMs optimize for probability and jokes, know, they are always funny there's a surprise. It's ⁓ the That's why jokes are actually funny. It's got to do with, you know, something ⁓ makes sense. And this is why they can't write jokes. Emile Ogier: I really agree. Ralph Behnke: had it go away and try and be a little bit more improbable with its jokes and will share some of the the results that out this week. So the first joke is Anthropic is now valued at ⁓ 900 billion dollars. That's joke. Their annual revenue is 30 That's that's kind of interesting. And so for context, that's 30 times multiple Coca-Cola trades at six. ⁓ Investors are basically saying Claude is five Coca-Cola's of optimism. Damn, that's a good story, but it's just not funny. So don't think we've succeeded yet. The next what have we got? Translation. So what Anthropics said, we're committed to building safe, beneficial AI. What Anthropics cap table says, we're committed to building AI so fast that we need $40 billion in fresh capital three months after the last round. Emile Ogier: Yeah. Ralph Behnke: I mean, that's impressive. Again, not usually funny. Another great Love this one. Elon Musk testified under oath that XAI partly distilled OpenAI's models. Yeah, no joke. In the same where he's suing OpenAI for betraying its mission. The guy's, you know, anyway, he's being unreasonable, which I love. You should always be unreasonable. Emile Ogier: So bad. Ralph Behnke: man copied their homework and then reported them to the principal. ⁓ Emile Ogier: It's pretty funny. ⁓ not trying to defend him, but you know, at point that he departed when it was a good decade ago, the models have moved on so much that homework would be so old in the grand scheme of things. I'm not sure how true it is. Ralph Behnke: Absolutely. And of course we know that one of his developers left X to join and apparently, you know, took with him all of the X's AI's And I actually know how that turned out. That was a bit of a news item some time ago. ⁓ Yeah, brilliant. We have a look at what happened to that guy. You know, did he end up in jail? Did he settle? You know, did anything happen at all? Or was the code just so... ⁓ Emile Ogier: I forgot about that story. Ralph Behnke: of date anyway, by the time he basically it was useless. So, but that was a good joke. So we're going to give AI that one. So we kind of got one out of three so far. Open AI chat, GPT 5.5 code spud one week after Opus 4.7. Somewhere at Anthropic, a product manager is staring at a launch calendar and whispering, can we not have one week? Yeah. That's just the reality. know, I mean, these releases are so hot on each other's heels. I think I don't know if any of these guys in the frontier model companies are actually getting any sleep. I know I'm certainly not. And I'm even nowhere near the the the bleeding edge of what they're doing. The next joke, and this is actually kind of reading the news for me in many ways. We're covering all the top stories. Claude 4.7 now does exactly what you type. No more guessing what you mean. GPT 5.5 does the opposite. This is actually really hilarious. describe the goal, it out the path. So that's what GPT 5.5 does. So one stopped reading between the lines and the other started writing between them. prompts from six months ago now fail on both for completely different reasons. This is a really, really sad story, but completely true. Prompt engineering will never die because actually the models Emile Ogier: you Ralph Behnke: fundamentally keep changing the way that they want to be spoken to. And of course, the way you speak to them is through a prompt. it's fascinating. I actually never expected prompts and whole prompt engineering formula, if like, to expire as fast as it is expiring. I kind of thought by now we would have got to a place where we said, OK, this is it. Here's my prompt library. You have to rewrite everything for these new which is kind of hilarious. ⁓ Emile Ogier: totally agree and in case in point, the best example right now is the upgrade on Opus to 4.7. The prompt engineering is distinctly different. And to your point, Rolf, I also assumed the same. thought it was going to develop in a way that we could get more natural and loose with our language and it would all figure it out. But it seems like the traveler direction, or at least for 4.7, has gone the reverse. You need to be more explicit again. Ralph Behnke: Right. And, you know, I think this for me is actually a great opportunity for meta prompting. Meta prompting is when you ask the LLM it wants to be asked. ⁓ basically ask it to write the prompt, you know, and I think this is a for all times. This is probably the only prompt you ever need to know is how to ask the AI to write a prompt for you. And I use all the time, actually. But we'll have a chat about prompting on Claude I do kind of want to share some reflections of how I'm using it ⁓ it's actually very different from of you know what what this news in particular is saying but ⁓ let's just we've got two more to go ⁓ to see we can at least give Claude another win on joke writing let's see next one curse is valued at 60 billion dollars another really funny joke their thesis all models are becoming commodities and the harness is the product. I've been talking about the harness a lot so I couldn't agree more. SpaceX agreed so hard they wrote a 10 billion dollar check. It's they have written a 10 billion dollar check but of course they promised to pay them 60 billion you know if they actually buy them so they have pay them 10 billion if they don't end up buying them. I'm sure we'll cover that story. But to carry on with this joke, not to make it too long winded, Google's response was essentially, we don't care which coding tool developers use, which is either strategic patients or the world's most expensive shrug. Do they not? I'm pretty sure they want you to use Gemma for, you know, or certainly one of their Vertex AI models. That one wasn't funny at all. What was going on there? One analyst spent $6,000 a day on Claude. and replaced a 100 % economics team. Meta's response cut 10 % of staff to fund more AI. The machines haven't taken your job. They've taken your department's budget. dear. Gosh, that's tragedy. That's not jokes. That's tragedy. Where's the humor in all of this? I think it's got a little bit macabre on the side of Claude's humor this week. Emile Ogier: Yeah. It sounds more like a threat than a joke. Ralph Behnke: But nonetheless, we kind of got one out of what seven or so. So we've got to be happy with that. And I still think my an LLM and an Irishman walks into a bar was the funniest out of all of them, even if it have a punch line. So so there we go. Should we jump into the 60 second news roundup, which will probably end up being 12 minutes, but I will try and. Emile Ogier: Go for it. Before we do, just a side thought from jokes. Maybe in the next run, we can try and model it against a favorite comedian just to see how a pattern matches and see if the jokes change in any kind of behavioral way. Ralph Behnke: Well, you one of the things I have said is not to be sexually explicit, not to take the piss out of anybody. So I'm not quite sure. All of the comedians I follow... ⁓ Emile Ogier: Okay, fair enough. everyone. Likewise. Okay, let's scrap that idea. Let's move on to the news. Ralph Behnke: now we know what's wrong with it literally robbed it of all of its potential material anyway here we go news round up so anthropic raising 40 ⁓ to 50 add a 900 billion valuation so as we saw earlier that's a lot of coca-cola cans and ⁓ you 30 times multiple it's incredible ⁓ it's just showing the hype maybe or the promise, who knows, we'll have to see how it plays out. But they have basically got a valuation now of 850 to 900 billion. That's a doubling from the 380 billion that they had just in February. This is staggering just in terms of valuations and their revenue run rate is now obviously just passed 30 billion, which is up from the 9 billion at the end of 2025. And they are getting up to 40 billion from Google. Amazon committed 25 billion of five gigawatts, plus five gigawatts of compute. So 25 billion plus they get five gigawatts of compute. So I mean, Obviously, we've said it before. This is like nation state infrastructure. These are not software startups. I don't know what's going on. know, we've kind of spoken about the sovereign AI of, you know, each investing. Mistral is obviously the European version. We've got DeepSeek and we've got Quen in China. Actually, we hear a coming out of Russia, but I'm sure there must be Russian model out there somewhere. that would be considered a sovereign AI. So let's move on. Cursor, $60 billion bet. ⁓ The harness the product, not the model. model is obviously ChatGPT or Claude or DeepSeek and Cursor basically saying, you're probably going to switching models. So what you want is actually something that you can run these different on. ⁓ And so they've shipped an which is basically a publishing of the harness, harness, their harness architecture. They've this deal with SpaceX. So SpaceX really active. They've obviously purchased XAI. So basically SpaceX and XAI are now the same company. And they are basically going be using cursor's harness to train proprietary models on XAI's colossus ⁓ supercomputer. They've got the names. Colossus, Gigafactory, and I know 70 % of Fortune 1000 companies now use cursor. Google told the news stack, doesn't care which coding tool devs use. So what is the harness? It's the orchestration layer around the AI model. So tool calling, context management, agent loops, file handling, MCP integration. The model generates the text, the harness decides what to do with it. Basically they've described Claude code to be honest. I a lot of people don't realize that Claude Code, I was a ⁓ bit of kind of research on it the other day and the number quoted the LLM part of Claude Code is about 1.6 percent of what you're using. All of the rest is tool calling and the MCP integration and the handling the agent loops. So it's actually incredible. know, we, think we're using these LLMs, but yes, of course we are. But really what gives you the power is the ecosystem. I guess it's one of the reasons why, you know, ⁓ code has just been so incredibly powerful, but ⁓ mustn't be distracted because I'm onto my third story, which is, you know, kind of already covered. Elon Musk admits that X AI distilled the open AI models under oath. I don't, you know, he's still carrying on. He must have a personal vendetta really against Sam. He wants to teach Sam a lesson. I don't know. Maybe there's some money that he's kind of after out of this, but obviously Musk was one of the founders of And ⁓ what he's is he confirmed that XAI used ⁓ OpenAI's models to train Grok distillation. Well, that is pretty much what DeepSeq and Kimi GLM, that has GLM 5.1. ⁓ They've all ⁓ doing this Basically, they run thousands, if not hundreds of thousands and millions ⁓ of prompts the frontier LLMs. And then basically distill the responses and use that to train their models. I think Musk called his 38 million donation he basically I was a fool. so anyway, sorry, Musk, that you were a fool. We don't think you're a We think you're amazing. In the same week, Hugging Face CEO accused Big Labs of pulling the ladder on distillation. So why this is important, think. First, it kind of confirms that, you know, model distillation between high profile domestic rivals is a thing. It's not just things that the Chinese are doing. And there's some evidence for that now. Anyway, I thought that was kind of interesting. Maybe it's a bit of a fringe story. GPT 5.5 launched name Spud. I know a dog named Spud there's a dog named One of my clients in Shanghai has a dog running around the office. I really inspired by Spud's fashions because he was modeling their latest line ⁓ of pet fashion way. So there you go, Spud. You're now. You've been on the Edge podcast. So 5.5 was launched on April 23rd, so seven days after Opus 4.7 and it excels at code, research, data analysis, multi-tool, agentic workflows. So benchmarks show some pretty significant improvements on Terminal Bench 2.0. They're currently scoring 82.7. They're also pretty good at frontier math. So it's not available to the free tier. obviously means it's not available to me because I have actually let my OpenAI subscription lapse. I ⁓ logged into for the first time today in probably two months. I have been to open it up again just so that I play around with 5.5, but we'll have to see if that happens. So. ⁓ The model race is obviously measured in days. So, you we just keep seeing model after model being released. And that's one of the reasons we do the show to keep everybody up to date with what is happening, because it's all changing every week. My fifth and final news story. So prompting rules changed for both Claude 4.7 and 5.5. I have got a couple of examples that we might if we have some time, we'll go through it. But basically the news is that you just need to rewrite your prompts. So as I was saying earlier, use meta prompting. Ask Claude 4.7 or GPT 5.5 to structure a prompt for you. you want to be extra careful, you might ask it to reference own documents that basically set out best practice for prompt design and then tell ⁓ you obviously what you want to ask and get it to write the prompt for you. That's probably one of the best ways of learning as well. So those are the top five news stories. I am sure there's a hundred others that I would love to talk about. But, Emil, I will stop there and ask you anything. Stand out. Emile Ogier: Indeed. I think I'll pick on the harness specifically as we're both really interested in this and agented computer use. Just as a reminder to all our listeners, lots of options available out there. It started off with just open claw, but everybody's come to the table. There's a lot of choice, so more time consume in things. just as a reminder some of the services can consider, you've got perplexity computer, is going to really be good for the researchers. ⁓ ChatGT agent, which was formerly known as operator, got Microsoft Copilot Cowork. We've Google Project Jarvis and Cloud ⁓ And as that's enough, I think I read that Amazon has also come to the table and they've just released something called Amazon Quick. Ralph Behnke: that is interesting. I hadn't heard about that. Amazon quick actually sounds like a express store format where you can go buy your snacks. But anyway. Emile Ogier: Yeah. Yeah. I think with all the other integrations, they've got the cloud service. They've partnership deals with a lot the other big frontier labs. I think when consolidate that all as a service, it sounds quite appealing. I haven't had a look at it, but it sounds like things click together quite nicely. But that aside, Rolf, I'm to move on to the first news item. Ralph Behnke: Come on then. Emile Ogier: Yeah. So, is maybe a week or two old, it's, it's still very relevant. Imagine paying $50 a month for something your neighbor just got for free and they're using your own plumbing. That's exactly what Anthropic recently did to Microsoft. Rolf, we've been tracking the AI wars for a long time, but Anthropic just made this move and that feels like a bit of a heist really. They plugged in Claude. into Outlook, SharePoint, OneDrive, Teams for every single user, even those on the free plan. This is pretty wild because Microsoft charges $30 per user every month for co-pilots to do exactly the same thing inside those very same apps. But Anthropic is now offering access to all your Microsoft 365 data through Plowd. for $0 and the absolute kicker is that Microsoft actually built the connector and approved it in their own marketplace. So this used to be an enterprise only feature gated behind premium plans since last October, but Anthropic saw massive gap in the market and pounced and there are 400 million paid Microsoft 360 feet seats out there. Yet Microsoft only managed to convert 2 million of them to co-pilot because getting businesses to pay that extra $30 a month for AI that sometimes hallucinates on your own calendar is a pretty tough pitch. So Anthropic looked at the 99 % of the untapped accounts and just built a front door right into the ecosystem. It's a brilliant Trojan horse strategy because the company charging absolutely nothing to read and summarize your work emails is obviously going to convert way more users than one demanding a premium fee. It really makes you wonder how Microsoft is going to respond to the rival offer a better, cheaper AI right inside their own wall garden. So what do you reckon? Are we about to see a massive price drop for copilot or Microsoft try to pull the plug on Claude integration entirely? Be nice. Ralph Behnke: Yeah, it's, it's an important story because, you Microsoft owns enterprise ⁓ and, ⁓ has, enterprise sales to some degree, but, know, the years, Microsoft is probably 90 % of ⁓ all of the, Excel and document and processing and file storage is within a Microsoft ecosystem. think, you know, Copilot is not an add-on anymore. It is just a of how you work with Microsoft. It's not something you should pay for. It's not, hasn't been good enough to pay for frankly. So, and of course are, you know, working on their own model. So Copilot has relied on open AIs. chat GPT a version of it for whatever reason they've just not quite been able to unlock even you know the full capability of ⁓ GPT so I think they really have been behind the curve ⁓ I have certainly used Claude and its in Excel have been phenomenal so you can really get it to do amazing things in Excel ⁓ if that's what you to do. yeah, I think this is interesting. And I also want to say, you know, one of the most powerful aspects of Claude is ability to actually look at your folder structures and categorize I actually had a client who was talking categorizing a large number files. So at the moment we're looking prove the way that some of these files are processed. And they told me they're going to relabel them all. And I said, well, look, don't worry about doing that. let me know where they are and we'll set up access. And I will just let Claude actually do the categorization. We'll have a look at it. We'll write some instruction. And Claude is able to do that, you know, within a few minutes, thousands and thousands of files. So this quite exciting. And I actually think that Microsoft has a lot to learn from the way that Claude is able to interact in that way. I think the other thing it calls out is that tool really now kind of has is abstracted. So ⁓ it's bit of an issue. You know, we've been talking about MCPs. Obviously, MCPs came about what? Just over a year ago. It was an invention from Anthropic allowed you to essentially connect to other tools. where a connection didn't previously exist. And so, you know, there isn't an API, so an application programming interface that allows you to integrate with the tool, you can actually build an MCP and it will allow you to interact with any tool. this is what, you know, users have been doing. They've been taking their favorite tools, building MCPs and allowing them to actually connect programmatically via Claude. And what does that mean? Well, it just means that you can actually get clawed to interrogate those systems reformat the, what it finds the results of its queries into whatever you want. And so this means that you can actually analyze and create dashboards and create reports. And so it's basically freeing up your data. And what we're is that if there isn't an open API that allows you to connect, you know, your customers, your consumers will build their own MCPs and do what they want. And then you've essentially lost control of the user experience. I mean, maybe you have with the API anyway, but at the very least with the API, you can, you've got some intelligence with, you know, in terms of who's connecting to you, what data they're taking from you. you at least control the experience in terms of Whereas with the MCP, you know, You really don't have much information. So here is thing that we've been talking about a lot on the show. Do I need to build an application or do I just simply have a database and then allow API services to connect to that database so that what we're is we're not brokering a user and the user experience. We're really brokering. information probably that is the future of compute and is probably one of the main reasons why we've been talking about this SARS apocalypse where essentially kind of considering why should I build or why should buy some software if I could just it and one of our listeners Rob Smith, a ⁓ bit of a out him, he was mentioning somebody that was looking at a CRM and evaluating of the big players like Salesforce and HubSpot and decided in the end to actually build his ⁓ instead of investing into one of the big CRM players. don't know if that's right or wrong, but that is a valid option where it wasn't before. so anyway, I want get too far off track, but yeah, that's a pretty interesting concept really with Emile Ogier: Totally agree, Rolf. And the same thing ⁓ out in different ways. So you cited a really good example of databases getting to the presentation layer and that middle layer effectively collapsing. And we've seen that play in other different ways. ⁓ The for the GUIs, because ultimately we design GUIs around what users need to do. So it needs to be intuitive, but behind that there's a lot of drag. There's a lot of click-throughs and there's complicated menu systems we have to learn and remember, et cetera. But agents don't need any of this. They just want to get direct to the data and do what needs to be done. So, you know, in addition to those plumbing layers disappearing, the GUI is another one that's falling away and having less relevance. And I completely agree with you. This theme seems to be expanding and, you know, gaining momentum. Ralph Behnke: Yeah, and the kind of the GUI, the graphical user interface is actually for me experiencing a let's just say a reinvention around what it means to ⁓ sense out of information. So in the form of reports and in past on the show, I've kind of spoken a lot about HTML actually really, believe it is going to become the predominant way for us to build presentations. instead of using PowerPoint even Word or some kind of document processing tool, what's the point of all of that when you can actually pull really complex data sets and then synthesize them with absolutely phenomenal graphics and visualizations and descriptions HTML. HTML is format that's actually incredibly easy to produce. And it's very easy to share because every browser can open it and I think that's really quite exciting So no surprises over the last three months. I have pretty much switched to 100 % HTML publishing I'm still kind of getting a few interesting responses from people when I send it kind like what is this? Like why you doing HTML like and I actually explained to a marketer just the other day said, you know, the thing with a PDF is is this annoying concept of the page because is essentially just, you know, a long list. can be infinitely long as long as your kind of memory can handle the size of the file. Whereas with I always have to spend so much time trying to format where the page break is. so that it looks decent. so is incredibly versatile when you're communicating information. So that's a really story. Microsoft watch out. Emil has put the on the wall for you. Emile Ogier: And to add just one more thing to your point, Rolf, also no more complicated folder structures with HTML. You you just got your URLs, you get straight there. It's no more needle in a haystack. It just makes sense. Ralph Behnke: Yeah, and I did actually build an HTML publisher. I thought I was first to market. I built this tool probably, I don't know, about eight weeks back. It was about 90 % done. And then I realized, oh, actually, there's quite a lot of these tools out there. And they're all available for free. So yeah, there's there's no money in doing that. But you might consider building your own kind of HTML publishing. page, just somewhere where you can build an index you can essentially build out all of these resources. It works an absolute treat. And definitely a try. But onwards, I'm just going to quickly jump into a very short story and then ⁓ you can up another one. So, you know, I've been building. Well, I've been building. I've been writing my own music, actually writing my own music. I think ⁓ probably about 70 of the songs that I've Emile Ogier: I'm afraid. Ralph Behnke: generated with Suno music. I have actually written lyrics for the other 30 % obviously I've generated for one reason or another and kudos to Brian Reynolds. I did mention Nelson's Mile in my in the last podcast and he actually went and listened to it sent me a note going wow Nelson's Mile you know what a good song and that is famously a song I wrote for an alpaca called Nelson. you want to go listen to it on Soundcloud, check out Ralph Benker, Nelson's Mile, one my favorites actually. But ⁓ the story that I want to get to is Suno has now got a valuation of, well, they've 300 million annual recurring revenue. Incredible with 2 million subscribers and actually Breaking Rust become first fully AI artist to hit number one on the Billboard. That's incredible. A woman in Mississippi turned her poetry into an AI song and landed a three million record deal. Bloody hell, they haven't heard my songs yet. so let's see. The next 12 months will whether AI music creators get rights holder So obviously they're busy suing Suno. have started to settle as I've reported previously, I thought that was a really interesting number that they've kind of You know, we've crossed novelty threshold 2 million paid subscribers is quite significant. So Let's move on Emil. What what have you got on your list? Emile Ogier: Yeah, I think I just would like to add some more color around an important story that you've really brought up in your roundup regarding court case between two power weights. Sam Altman, co-founder of OpenAI as a nonprofit force for good, is now in federal court trying to drag it back to those very roots. Jury selection is underway right now in the Northern District of California. with opening arguments set this week and a civil trial is expected to run for roughly two to four weeks. And here is how we got here. Back in 2015, Elon Musk teamed up with Sam Altman and others to launch OpenAI as most of you are familiar with as a nonprofit built for safe, artificial general intelligence that benefits all of humanity. This is the important aspect. No profits, no close corporate control. Musk poured in roughly to $44 million, about 60 % of the early seed money, plus compute power and talent. By 2018, he stepped off the board over direction clashes and Microsoft's growing footprint. OpenAI later grabbed billions from Microsoft, restructured it into a for-profit arm. When closed source... practice and rocketed 852 billion valuation with fresh for-profit setups. So you're asking, you asking earlier off why I might be upset. think this is one of the key reasons reporting into. So we have a fresh for-profit setup reporting into a nonprofit foundation. Musk is saying they breached the original charitable trust agreements, turning the output into a wealth machine that puts profits and insiders ahead of the mission. He claims they deceived him on their plans. The suit was filed in 2024, withdrawn and then revived early this year. And they just recently dropped the fraud claims to streamline the case. It's now focused on breach of charitable trust and unjust enrichment. He is seeking damages, reports put it at around 132.4 billions, but get this, to awarded straight to a charity, not a single dime goes into his pockets. Plus a possible restructuring to restore the nonprofit mission and the change, team changes like removing Altman and Brockman, his nemesis. OpenAI counters at Musk knew about this and was once supportive. of commercialization steps, even floating a Tulsa merger or taking control himself. They are calling the whole thing competitive sabotage for his ex AI rival and straight up jealousy of their success. But it's not as if Elon Musk doesn't already have enough success of his own. Folks on X are calling this the defining showdown for AGI's soul, profit driven and closed versus benefit of humanity. A big win for Musk could reshape OpenAI's valuation, IPO plans, Microsoft partnership, and possibly even the entire AI race. Ralph Behnke: Yeah, it's beautiful. mean, it's kind of like was going to say it's like a superhero and a villain story. But, you know, some people might think it's a villain and a villain story or a superhero and a superhero story, depending on your affiliation. Yeah, ⁓ but ⁓ see how it plays out. think, like said, think Musk is trying to teach him a lesson ⁓ in I don't that's going to work on Sam Altman or other man, Brockman. I think they have Emile Ogier: Yeah. That's a good analogy. Yeah. Ralph Behnke: passed the Rubicon in terms of their egos. anyway, let's see. So, well, I've got kind of got a bit of an interesting take, I guess. It's a personal take. It's an opinion take. You call bullshit on it, but I'm going to go with it anyway. And it's actually something that we discussed in private just a week or so ago. Emile Ogier: What else you got, bro? Ralph Behnke: all makes sense in a moment. So here we go. So we started with prompt engineering and we seen, know, with the news of Claude 4.7 and 5.5 prompt engineering is still a thing. Don't hang up prompt engineering gloves just yet because need a prompt to keep, you know, innovating how to get best out these models. But after prompt engineering, what we found was, could write all these prompts, but unless it understood you, you know, unless it had context, you just didn't get the results. So we then had context engineering and some of that context engineering came from projects. So projects include gems in Gemini projects in chat. They all allowed you to basically give some context so you didn't have to keep telling it, you know what you were talking about. And then of course, we kind of looked at it and went ⁓ no, this context engineering, you know, it still just doesn't quite know the full story. So we started to try and solve memory engineering problem. ⁓ memory started take the shape. Most people probably don't have a persistent memory solution yet, unless they are actually coding. But I can you from an engineering standpoint that memory is the all important solution when you are developing with LLMs. And now I think have moved from memory engineering into, wait for it, intent engineering. And think what we see coming out of next is reality engineering. That one I'm still of mulling over what that looks like, but just to talk about intent engineering. One of the things that's happening with tools like Cursor, remember the story about the AI. the agentic harness. This really the environment which brings the tools, the ecosystem if you like, that makes the AI really effective. And when I talk about intent engineering, that harness can be optimized really around your work. And so this is kind of what Cowork started doing. It look, we can do your work for you. What is your work? Well, your work is essentially intent at productivity. When you set out trying to create a presentation, you have the intent of communicating something. When you an Excel spreadsheet, you have the intent of calculating something or researching something or analyzing something. And I think now basically with these tools actually bringing together, you know, the prompt engineering, the context engineering, plus some of the memory protocols, you suddenly have this ability to actually prompt for intent and Claude and some degree Codex and ⁓ cursor multiple tools basically have the now to actually engineer your intent because you have given all of the prompts, because you've given it all the context and because they may have memory if you've basically given added that service. then, of course, it becomes somewhat optimized for your job. So if you're a hairdresser and you're obsessed with generating all of interesting content and stories that you can talk to, or I don't know I picked that. It's probably because I went to the hairdresser a week or so ago. And, ⁓ you know, probably using it this case. But let's say, you know, you a hairdresser and you can essentially train ⁓ or any of your kind harness, agentic applications to really understand the intent of what you are trying to do and it will ⁓ regularly go and run scheduled This is what Cowork is particularly good at. It will obviously go and do your weekly activities, do the research, create various bits of analysis for you, publish those. And so it can really do your work. And this is kind of what I mean by intent engineering. And so where we're at now ⁓ I think we're starting to see the rise of offering services to set up your Claude code so that it actually can do the work. I mean, yes, you can use code work, but Claude code is the ⁓ real of this. To some Codex is also able to do it. But, ⁓ you know, talk Claude code. It's not just for code, right? ⁓ It is in itself a harness of sorts and obviously you can also use cursor. I've not tried cursor in this context, but by all accounts, know, this is promised land. And so where do we go beyond that? Well, beyond that, I think when you start stacking intent and it's ⁓ really doing, you know, items of work, this is where I think this kind of concept of Reality engineering comes into being because you can actually reliably set this thing off to go and build a company Or at least build a channel or at least build, you know some stream successfully that's what your game is and And think ⁓ intent engineering you've it here first. So intent engineering followed by reality engineering So there ⁓ are ⁓ two levels to the evolution of AI that I am predicting or maybe not, maybe I've just heard it somewhere on a podcast and I'm just regurgitating it but I've had a bit of a think and I actually don't think I've heard it anywhere before. Emile Ogier: I don't think I have, and I love the terms. I, quite honestly, Rolf, I don't think it's crazy at all because I'm just going to reframe your thinking in a slightly different way. You know, the way I was thinking, if you just look back on civilization, you take a further arc back, part before AI, even before computers, even before tooling. We just thought of the human man. It was always if you needed to get anything done, there was the idea of this, then there was action and then there was output. Then came the tools which gave leverage to get the outputs done more efficiently and quicker. Then came the computer again, more gearing, more leverage. And then we got AI, which is effectively a computer inside a computer giving rise to even more compounding effects. ⁓ you look at that, what are we saying? What are we compounding is we ultimately compounding execution and having our ideas land in front of our feet. ⁓ quicker than before. And if that's the case, if that's the goal, if that's what's happening, it is reality shaping. It's instantaneous reality shaping as that timeline shortens. So yeah, I get it. I'm with you. Ralph Behnke: All right, well, we'll all, you know, that one, can write the book now, Emil. So finally, we've kind of got a plan. But there you go. I've shared my my special interest piece in relation to that. just wanted to call out that. this is probably something that I will talk about over the next few shows, because it's it's a reasonably, let's say, complex set up. But I think. one of the skills or one of the services that people have to now start bringing in to their is this concept of memory. And I say, it is something that is a little bit focused on coding development, but I it's going to become more prevalent to anybody that really to really leverage and not kind of have this, you know, It's like Groundhog Day. Every day is a first date whenever you the AI, because it just has no idea what were talking about yesterday. And the way that you actually overcome that kind of first date is you give thing memory. And obviously, you know, in put a memory file. You now a memory dot MD file where it writes stuff, but it's not quite not quite at the level that we see some of the pros using it at. And so if I just go through what some of the components are, so you basically got level one memory and that the project rules. And so this is basically setting out normally in your file, essentially your conventions, the you always do. And then you've got your level two memory and this is your tool specific configuration. So, know how you use your MCP files, your slash commands, the features and all them this by the way. So you've got Claude MD, you've got Gemini MD. You've also got GitHub instructions.md. So they all pretty much work the same way. And obviously what's also really interesting is that we actually have some kind convergence, if you like, because I'm just trying to think of the foundation. Now, the Linux Foundation have actually basically taken control over the of the agents.md file. And so that it works the same way, regardless of which LLM you're using. So If you're using Codex, if you're using Copilot or Windsurf, Warp, Anti-Gravity, they all basically are the same structure of the agents MD file. Why is this important? It just means that you aren't kind wed to one LLM, but have one that is the same that you can actually use to communicate all of them. And so that is kind of level one, the agents MD file. then you've got the tool specific context, then you've got your personal context, your kind of individual preferences. And this is typically, you know, something will be called your local MD file. And you've got your task dependency state. And so now this where the rubber hits the road. ⁓ If you're with me, if you just think about it, you know, we've got a stack of five, we're at level four right now. So task dependency, this is actually where it remembers what you're working on. And this is where you move the first date and it remembers the first date. You move into the second date, the third date. I'm actually dating an algorithm, so this probably going to this analogy is going to just get stuck in a moment. But every time you're working, it remembers what you did in each session and it is able to actively recall what you've completed, what you're currently working on. and what you are forecasted to work on next. This is such a powerful concept because what it means is you can suddenly start orchestrating obviously multiple agents take on different tasks and they will know they will be somewhat aware of what each one is working on. But not only that, you could potentially also start divvying up tasks for different LLMs to take over. So, know, I've been experimenting more recently in potentially using Claude code for decisions and for research and then actually handing over to a model like like Kimi GLM to actually do some of the coding and Why am I doing that? I just want to create some options Should Claude code ever be down will need to be able to jump onto some other model. Should Code continue to more more expensive as it has become with 4.7, consuming 35 % more tokens than 4.6, I might need to at some point because I keep hitting my limit. And so this is where memory really starts to create optionality. And this is why I think, you know, it's the most important next service. If you don't have it yet, you need to get it. Tony, my development partner on some of my projects has developed Omni Cortex and we've been using that extensively. But there is a nice open source version that is free called Beads developed by a guy called Steven Yagy. He's also created Gastown. You probably heard me talking about Gastown previously. It's a kind of a, let's say an agentic development framework. But then you've got level five this is kind of the cross session and there a bunch of different tools that basically of start to remember not just what you did, how well it worked. And this is quite powerful. So a few of these out there. Mem zero, Zip, Letter. You can kind of access these things MCP. So level four, level five becomes super important because when you're working with agentic it actually allows to do, you know, multiple process, multiple tasks parallel. But of course, when actually work with other developers, then again, you know, you've of got a shared environment. You can actually upload your tasks to the cloud. Obviously you can do this with but. There are tools like databases and DOLT, which another database which Beads actually off, which allow ⁓ multiple developers connect to it. And it just that can basically in real time inform each other and also the agents what they're working on. And so this is just unbelievably powerful. And I think it really sets us up. for the age of intent engineering and obviously lays the reality for, well, it raises the, kind gives us the foundation for reality engineering, which comes next, if you my prediction. So there you go. That's my big pitch for week. Get in memory. It revolutionize the way you You're on mute. Emile Ogier: I was on mute. Yeah, I totally agree, Rolf. So is a way forward for certain. You got to have it. what you've set up is clearly working for you. And yeah, couldn't think of not having a system like that in place. And just a reminder to our listeners, we also covered another product. I think it called Memory Palace that was co-founded by famous actress that did fifth Elevants and blue Lagoon. forget her name, look, look back on one or two podcasts. We covered it there too. So, ⁓ yeah, if only we could remember without further ado, I think I'll move on to our next news story. The creator of the hottest AI product in 2026, literally Ralph Behnke: There you go. If only we could remember. Emile Ogier: begged people not to buy MacMunese to run his code, but they ignored him and created a 16 week hardware waitlist anyway. The project as many of us are familiar with by now is OpenClaw built by developer Peter Steiberger in exactly one hour as a weekend break after a 100 million euro tech exit. I have covered the story in the past, but things have moved on quite a bit. And I'm just extending that really for all of us. It started as a tiny bridge called Claude Bach to connect WhatsApp to his laptop, but it exploded into Hacker News and it hit 100,000 stars in a single week. The hype cycle was relentless. Influencers sold the dream of a fully autonomous AI employee spawning weird spinoffs like Malt Match, where agents actually secretly screen dating profiles for their owners. But as Peter Steinberger got snapped up by OpenAI, normal people trying to automate their lives hit a brutal reality check. A default heartbeat setting keeps waking the agents up every 30 minutes just to stay warm, quietly burning for some around $86 a month in Claude Opus API fees for it to do absolutely nothing. The worst in this, the costs were silent failures such as shattered OAuth redirects, expired tokens and overnight memory wipes after simple model version upgrades that left these personal assistants with total amnesia. Another risk was exposed when an AI safety researcher at Meta Superintelligence Labs had to physically run over and pull the plug on a Mac Mini because the agent started relentlessly deleting her entire email inbox. Under the hood, flaw called context compaction caused the agents working memory to automatically summarize all the messages accidentally deleting its own safety guardrails in the process. It later turned into a security breach when a researcher proved you could steal the user's private SSH keys by just sending a normal looking email loaded with a hidden prompt injection. I think this serves as a good reminder to us all, including ourselves, that when dealing with automating workflows to treat AI agents as junior contractors that need their work reviewed and only to promote their scope and responsibilities as they earn and develop your trust. Good architecture principles to follow as a reminder to our listeners include. Ensure you sandbox agent access, set explicit exclusion criteria outlining any high risk actions that are forbidden from performing, such as erasing artifacts from your production server, and add measures to try and manage the potential blast radius for filling tasks. Examples include get your emails to draft emails only and not send them without your written consent or physical review. Lastly, start with singular low risk tasks and build them out incrementally instead of handing them over your entire digital life all in one go. Ralph Behnke: terrifying stuff. Emile Ogier: So there you have it from me, the architect speaks again. Be careful, there's a lot of power to will, but just do it right. Ralph Behnke: Yeah, I just think having had the experience myself where I did actually delete my entire computer because I was trying to educate Claude on what not to do and accidentally got it to do the thing I was telling it not to do. I could sit in the past entirely my fault, I ran stuff out of sequence. But, you know, it kind of really illustrated for me the power. these tools are wielding. literally deleted everything on my computer. I had to start again. And so with OpenClaw, yeah, I mean, I love it. I mean, I think it's been a wonderful concept. I said in the past in China, there was kind of OpenClaw fever. People literally you know, creating kind of these weekend moments where you can go to a particular company and they will install install open call for you, which I think is really good because obviously that way you can actually get a proper setup. And that's probably the right way to do it rather than try it on your own. of course, as we kind of look and go, well, what do you actually want to automate? A lot of people don't know what they want to go and automate. And so it goes just becomes a bit of a novelty thing. I think you Claude code these things are able to do this out of the box It's not that open claw actually does something new in this sense, you know, it is an automation tool and there are many them it's just this one particularly caught people's imagination and went viral so yeah watch out, but I think think if you really want to get serious, then doubling on something like Codex or Claude definitely the way to rather than playing around with something like OpenClaw. But that's just my opinion. ⁓ And I'm sure other have some wonderful success stories. Emile Ogier: I just want to add more one thing, Rolf. I'm not saying to don't use OpenClaw. It's anything but that these ⁓ are, you they're capable of actualizing themselves anyway. There's certainly so much to using these tools in the right way. It's just, it's not as as switching up, flicking on switch and assuming that everything's going to run your life for you. There's middle ground. and just following the right path to get there is the way forward. Ralph Behnke: Perfect. Well, I think that kind of brings us to the end of the show, but I just have one more thing. Unless you have another story. Emile Ogier: I do, but I can always pass it on for next week. We always have more stories. Ralph Behnke: Well, come on, why don't you do it, then I will take us home. Emile Ogier: No, that's fine. I just thought this one was a bit of fun and quite fascinating as far as I was concerned. Science has just copied a fruit flyer's entire brain, every one of its roughly 140,000 neurons and 50 million synaptic connections. And then they dropped it into a physics-based digital body and what something incredible happened. No training, hard coded rules. The virtual fly simply started walking, grooming its wings and darting around smelling for food, matching real fly behavior by close to 95 % accuracy. What makes this really fascinating to me is that the behaviors didn't come from some clever AI scripts. They emerged straight from biological wiring itself. This seems to suggest that complex instincts can arise from structure alone. giving neuroscientists an unprecedented window into how brains turn raw connections into actions. And the team is already looking ahead. If it works for a fly, could the same approach one day map a human brain? The story is still developing and there are 86 billion neurons still to go. So we saw generations away from the leap, but this is a tiny digital fly. just made the whole brain emulation feel a whole lot less in science fiction. I just thought that was super cool. Ralph Behnke: That's the best story ever. You should have started with that. We would have spent all conversation talking about it. I I love it because it's like it's like the next generation of the Tamagotchi. I'm going to like just have my app sitting over there with the fly doing stuff or what, you know, the animal of my choice. So I love that. That's absolutely incredible. sci fi. So here we go. Emile Ogier: ⁓ I like anything on positive news. Exactly. So send us a photo, but you guys. Ralph Behnke: Well, okay, I'm going to try again. know, an and an Irishman walk into a bar. Open Claw is the barman. The LLM says, I'd like ⁓ a appropriate beverage optimized for social bonding, factoring in ambient temperature, cultural setting and my current emotional state. The Irishman pint, which is entirely appropriate. Open Claw slides a Guinness Guinness to the Irishman and turns the LLM. I've got four tools. Read, write, edit and bash. He's used one word and got served. You're still on token 47. Do you want the bash or not? The Irishman raises his glass. In fairness, that's the fastest barman I've ever seen. Only does four things, but he does them well. The LLM is still generating its order. ⁓ a very complicated joke. Emile Ogier: That's a complicated joke. Ralph Behnke: But somehow I love it. So on that note, are still on the bleeding edge and it is very bloody in terms of humor. But been a great episode. ⁓ Emil, see you on the edge. Emile Ogier: Rolf, it's been brilliant. So I missed you last week, buddy. Let's keep it up. See you. See you very much. Ralph Behnke: Adios.