Latest / The Bleeding Edge Podcast / Weekly AI Brief 23 April 2026
Transcript
- Ralph Behnke: Hi and welcome to another episode of the Bleeding Edge podcast. I'm your host, Ralph Benke and with me on the show today, Emile Ogier. Emile, welcome to the Edge. And today we have a special edition ⁓ Emile's iPhone. So he's ditched the computer, because can't run Riverside. And iPhone can. So there you go. Some pretty interesting stories ever. And think right on the money because yesterday top news for me at any rate is Anthropic has shipped Claude Opus 4.7. So this is quite exciting because they released it at the same price 4.6 or they? So this is quite because The model is definitely number one on VEL's AIs vibe coding benchmark. And you can still process tokens at $5. Well, yeah, kind of like $5 per 25 million tokens. And so that's actually quite high in comparison to many of the other models, certainly 70 to 90 % higher than most of the Chinese models. will cost you. ⁓ But the hidden tax in Opus is that the new tokenizer can actually use 35 % more tokens on the identical prompt. And kind of set to this X high, so extremely high effort level. ⁓ And ⁓ going consume tokens at basically 35 more than what it was just a few days ago. So they may have launched at the same price, but it's definitely gonna get more expensive if you leave it on X high, of course. it doesn't. I actually had a really proud moment in last week. I for first actually ran out of on my Max plan. And that was two people the account, you know, pretty much the clock. over the weekend, ⁓ was a lot going on and we managed to max out and get to 100 % of our available processing. And that's That was great to see. We had to work really, really hard to try and achieve that. It's going to be much easier now with this 35 % more token usage on the identical prompts. So go ahead and if you can max out your Claude Max account. tell you, you I've tried many times to do it ⁓ and amount specifications and thousands of lines of code that I've had to write to actually get to the point where it was maxed out is quite staggering. But there you go. Obviously, it's going to be something that happens more often. Then ⁓ the big that came out was the Anthropic paper on nine Claude agents beating the human alignment researchers. They pretty much is what's called costing you about 22 bucks an hour. If you run a number of these instances in parallel. So to do this analysis that I think they ran four of these Claude instances in parallel. So it came out at about $22 an hour. And basically these Claude Opus 4.6 agents recovered 97%. a week too strong supervision gap in five days and two human researchers got 23 % of the way through that week too strong supervision seven And so the agents obviously, you know, cost a lot less than those two human researchers and covered a huge ⁓ you know, nearly five or four what know, the humans were able to achieve. So they invented four new reward hacking strategies that no human had ever predicted. And, you this really important because alignment. this is kind of where the the goals of AI are aligned to the goals of humans was the one field that said, actually, you can't automate this. And course, now we have some empirical evidence that you can indeed automate it with this 97 % of the supervision gaps covered in five days by the AI. So that's pretty impressive. So, hey, you basically did yourself of a job there. But no doubt we'll still need to figure out, you know, what new strategies the AI has uncovered. presumably we will need somebody to read. all of that research and all those papers. third news item I've got, a interesting story with Sam Altman being firebombed. And there's guy actually been arraigned on that. But what means is you're really starting to see people to AI. And so I think there a recent study from Stanford, was the AI index. published on the 13th of April, and it shows something like only 10 % of Americans are more excited than concerned about AI versus 56 % of experts. So presumably those Americans they were interviewing are not experts. And so I think 73 % of experts versus 23 % of publics, the public agree AI will help. which is quite staggering. I think there is a lot to be excited for. So maybe I'm in the expert camp and not in the Joe public camp, but there you go. This is kind of how people feel. And they are now firebombing Sam Altman's house. So the guy threw a molotov cocktail. And I think Sam actually posted on social media a picture of him and in his family. He's got a young daughter, so him with his partner. Kind of going, hey, don't take me down now. I've got a family. let's see if that keeps him safe. So ⁓ some other news, which is kind of interesting, Maine became the first US state to pass ⁓ 20 megawatt and data center moratorium. So basically they're saying you can't build your enormous data centers here. don't know if that's necessarily a good thing, but clearly they're looking after their energy usage and their water. Then another great news from this week is OpenAI has actually released their AI jobs transition framework. So basically they've analyzed which jobs are going to be over by AI, which jobs are going to significantly be ⁓ changed by AI, obviously what's going to disappear. So this is a bit of research that led by chief economist, Ronnie So he's OpenAI's chief economist and he published this framework. Interestingly enough, they did the research with ChatGPT 5.4 and so ChatGPT was essentially scoring its own homework, trying to figure out just how much the world it's going to take over. And they actually did a pretty hefty job looking at 147.9 million jobs. And they found that 18 % are at higher automation risk. 24 % will reorganize and 12 % grow with AI and 46 % see less immediate change. So we'll have to take a look a little bit later at some of the detail around that. And then fifth story, which rounds the news, a really fascinating shift. a company that makes footwear, so they do wool sneakers. They're called and they've just sold their shoe business, I think for about 39 million. And they have put everything into buying GPUs. And they have now built their own data centers and they're reselling compute Interestingly enough their shares jumped 600 to 700 percent in a morning and so, know, this is a B Corp that basically You know B Corp they're obviously Really looking at sustainable ways of doing business and they've literally just gone into You know one of the most energy consuming industries on the planet So I hope they've bought a wind farm along with all those GPUs just fascinating to see people pivoting. This is not the first one that we've seen pivot in that way, but yeah, just interesting that you can go from being a wool sneaker supplier to basically out GPUs as a service and getting a 600 % stock pop ⁓ in process. That's ⁓ probably of the biggest pivots I've seen in a long time. Yeah, it's a bit of a wily dealing happening there. So we'll watch that one, see what they make of it. Clearly, they think that renting compute is a good business. And I think they're right. so let's see. Maybe we'll rent a bit of compute from them and build something great. So what have you got as your first story? Yeah, that's really the promised land you're describing there. And have to say, I've actually been playing around with this because one of the things I find is that, you know, the early days, you really are quite keen to observe and watch and see how it's processing what is actually working on. And so I was actually very happy pushing the buttons and just reviewing each step. Obviously now, as we're kind of going deeper into it and even just trying to be more productive, it's not practical to be with it. And you really want to set up a and let it just run and not just go away and do the build, but actually do build, commit, the code, run. the tests ⁓ can actually use Claude Chrome or you can even use the Playwright Playwright actually one of these from from Microsoft that actually go and browse any website ⁓ it will do everything that know a user will do. And so if you are basically running this on host What it means is Claude can actually debug itself it'll pick up the results of having whatever app, presumably, you know, if it's a web app that you have pulled and it will the issues and fix itself. so at the end of it also has the ability to basically, if you want, you even merge the code and push it to production. And the final piece is I'm also getting it to then write release notes write a communication to the users of the system, as well as build the training material and out support notes. And so is really powerful because what you can start to see is that as as you've got that initial idea and concept well defined. then it should really be possible to hand that over to Claude it'll run, you know, as long as it needs to run to actually build that. so we been this in a number of different ways over the last six months or so. Certainly Tony, who is one of the developers I have been working with is Excellent at this. has a really, really good setup that really can run autonomously. over last week, I've actually been building a protocol Claude code that mimics a lot of what you just described. So it's actually leveraging some of those native Claude features. And what's fascinating about it is I started with some research to say, okay, I'm going to build something that can orchestrate end to end and, know, get a small app live with human in the loop in just kind of key places, but not all the way through like I have been doing it. actually one of the review cycles that I run, it kind of tests ideas. So it tests the logic of what these you're trying to build. And it came back and it said, look, actually you're trying to build what Claude already does. So don't worry about doing that. just use the native orchestration and the Git work that all that capability exists already. So don't go away and build that. We'll just an orchestrator that leverages us. And so that's much where I've to. I actually ran the first test this morning and I haven't actually gone to the results, but there we go. I'll let you know there's plenty of things that I've been trying to get to and build over the last while. And most of it I'm doing on the weekends now because, know, I have work projects and client projects that are taking precedent. So no by ⁓ our podcast, I will either built ⁓ 20 apps have some interesting live, or I'll be crying that it didn't work. But I think this era of is well and truly here. course, we saw that Cowork largely developed autonomously by Claude. That was of the first mainstream applications that we've seen wholly built by AI itself. And I think ran about four weeks. So we know that Boris and his cohorts there at have figured this out. Us mere mortals are still trying to play catch up. But on what you just described, there really advanced and mature features that are now putting that kind of capability in the hands of everybody. So, do want to go with the story or you got something else there? The next time I run out of tokens, just go shopping on GitHub. Hopefully I can a good API key to use, but ⁓ think that's really solid advice. I run quite a lot of these scans and one of the automations that I've included in my automation is, you know, quite an advanced security scan. So I think that's the key thing. And yeah, mean, just have to have that part of your workflow. So here's something quite interesting that I've seen in the recent news, Emil, and that is that ⁓ Mike actually, he's just resigned from Figma's board as Anthropic is prepping to build a competing design tool. So I didn't that they were planning to do this, but you know, it just goes to show. why the SARS apocalypse coming because so much is obviously moving into these LLMs and the is forever shrinking. So Figma obviously is ⁓ a design application. It does really great storyboards, wireframes and you design front ends. ⁓ It integrates with a of different code development tools that are using AI. Sphigmas slid something like 6 % on the first reports earlier in the week. Adobe fell like 2.7%. Wix, Israeli website platform, they fell 4.7%. So again, can kind of see that this ripples across the market. that has got a front end design tool going to feel it. So I think fighting for relevance is going to be the defining term of 2026 because a lot of these companies are really just going to see that AI actually does a much better job. And, know, I ⁓ the is this. What's kind of interesting on reflection in the last couple of weeks, I've really been looking at Why do we need applications at all? like an LLM will just do the job for you. But of course, you know, most people don't necessarily know how to use an LLM in the same way and to actually get the same level of consistency and quality that we have in applications. And this is why applications exist, even though, you know, something that literally just has a window in it. Now, if you think about most enterprise applications that people are using, there are probably some hundreds, not thousands of pages, all of those pages meticulously put together by a developer. each of them have got the specific fields that you need in order to do a certain aspect of the job. You know, with AI, you can do the same activity of ⁓ chat And so you just kind of. Lose the need for all of those pages and all of the API calls and all the integration that's required to actually, you know, make an application run in that way. Obviously you still need the integration. You still need the data and, and that doesn't go away, but the kind of orchestration, the complex orchestration that happens really just disappears. And we've probably had this conversation four or five times, but, yeah, it just, I just reflect on it. And I think where I've got to is that actually applications probably won't disappear overnight because the fact that people just want simplicity. They don't want to figure out how an AI can necessarily do the job, even though it can. And so I the market reacting, ⁓ you on with Figma of dropping 6 % on the first release this news. Fair enough. I'm pretty sure Figma and companies like that, if they innovate now, there's no doubt that they'll keep their market. They've got a really big user base. They just have to find the value. And think ⁓ really the now for SaaS because this kind of services as a software is definitely the way to And, uh, and maybe it is, you know, development as a service even. this DOS, uh, you know, essentially we're just, uh, building for your very unique personalized requirements. And even if it's a big company like Figma, uh, why not, why not build a custom and, uh, personalized capability, you know, for, for the clients. So, uh, anyway, that's my backstory. Yeah, yeah, I like that. So you heard it here first services as a service and Emil Ogier, ladies and gentlemen, your pioneer. So, you know, there we go. Your of fame, Emil, we'll have to get that on LinkedIn immediately. Yeah, I think, you know, a bit of stress in the system is actually always good. just in humans, you know, we sit up and listen when there's bit of stress in the system. And it's of the things that actually helps us commit to long term memory because we take ⁓ deeper, you know, ⁓ awareness what's going around us. And I think the same actually true for the LLMs. And there's been a number of these types of tricks over the, you know, that I've picked up over the last two years or so. And one in particular, is if you actually, you know, include in your prompts, stop and a breath. and, again, it kind of puts it into a somewhat deeper thinking mode, which is, is kind of interesting. but I really ⁓ love this idea of, putting that into the, the MD file for ⁓ Claude to read, know, just as it's about to set off. There it is. Your enemy, you know, number one, your arch nemesis codex, and it's going to be reviewing your work. So be on your best behavior. And obviously, you know, we've seen that when the knows it's being reviewed, it actually behaves differently. This has been proven a number of times, especially they're doing the evals. So gonna add this one in. I'll let you know if I ⁓ actually any differences there. But yeah, that's a great top tip. And just us, which file are actually putting that in? And that's like in the Claude MD file. So it reads it, you know, on startup. Yeah, yeah. Yeah, amazing. Really, really good stuff and simple as well. So that's one to take And even if it's not true, of course, I think people find quite a little bit of an issue when they're actually ⁓ reviewing between multiple models. I've some people a bit of success with it. I've tried it in the past, but have to say now with of the custom skills that I'm using, I rarely tend to move between models over the last weeks, especially, I've shared this in a previous podcast. pretty much now moved ⁓ all of my activity Claude I pretty much use that 90 % of the time. Sometimes I'll go into Claude desktop, pretty much everything that I'm running within Claude code. And I very rarely need to go to another model. So the last ⁓ weeks, know, it's taken over from GenSpark. So GenSpark, a very, very capable AI tool actually basically gives access multiple and it picks the right one for the job. It's got lots and lots of features with image generation and video generation. It's got its own version of OpenClaw. It's also got a really, really capable builder. All of as capable as Genspark ⁓ is, is surpassed now by what I can get out of Claude Code. And so certainly been able to sweat that Claude way more. just, yeah, there's been no need actually to engage in any other applications. And more importantly, I've been able do some pretty advanced ⁓ using applications built Claude And just recently, you know, I put a gathering and research agent a retail client. ⁓ honestly, the amount of work that this app does to go and do product level research is quite staggering. so, yeah, there's ⁓ nothing else that close to it. And of course, the latest version of Claude code Claude has something called routines. routines are supposed to replace in a 10 and Zapier even routines, I you know, I have had very little need to actually run any automation through an 18. I'm actually now converting most of ⁓ n 18 into just into ⁓ kind of code Python that are orchestrated ⁓ via cron jobs. and you can kind of schedule them to run, but now Claude Code Routines will that even easier. I haven't used it yet ⁓ and I you were that the latest feature, I think you were alluding to, this Claude Code Routines is riddled with bugs. So just been released and essentially it gives you a no code, low code automation. platform and there's no need now for things like N8n and Zapier but we've heard that before and we might have to wait a little while for this routines to become a bit more mature and less buggy which is unfortunately the case when you Vibecode everything. Yeah, that's a good one. actually did try to access routines in Claude code just with an automation that I was setting up for this product and for this new news research. yeah, it didn't quite work the way that I want to do. So I ended up with a cron job, but I will give it another go and see if we can get that to work. I mean, you know, it just goes to show. These tools really consolidating at a rate and they're eating entire segments of the software industry. And so, you know, I know what it takes to have a strong moat these days. think unless you have some kind of specific hardware you just a critical mass of optimized features in a particular vertical, you know, an industry that that the LLM is just can't quite trump with their more generic feature set. I think it's very, very difficult to see how a lot of these tools are actually going to survive. in fact, know, Canva, which is a tool a lot of people use as a design platform, they've just launched 2.0. They are testing a valuation at $42 billion for their IPO, ⁓ Again, that's quite staggering really. And so what can do with their latest AI is actually describe your project in plain English and ⁓ can editable, iterable design with persistent memory. And so that persistent memory obviously has been largely missing from the AI experience, over multiple sessions. Certainly Claude has got memory file, which improves it. Obviously you can use... ⁓ chat GPT projects and this type of thing to get around that. But I think a lot of people will find that Canva AI probably will be quite helpful. I used Canva quite a lot, probably about a year ago. I actually found it just bloated with features, complicated, difficult to use at times. And so I ended up not using it, especially around the time when Nano Banana came out and so, you know, Canva is by a lot of people, certainly amateur designers or at least, know, solo entrepreneurs definitely have relied on it. But it be interesting to see, you know, where they go with their AI 2.0 launch. Do you use Canva Emil? Yeah, this is it. It's like a real consolidation. And I say, I think, you know, the the moat is maybe a dream now, a distant dream for so many software companies. So I expect the market to change quite significantly. And, you know, interestingly enough, I think if you have got a piece of hardware that probably is, you the biggest moat you can have as a technology company now. And so that's definitely the way to go. But one interesting that came across in the last or I think last week we spoke about Anthropix and Mythos. Mythos ⁓ of course the tool is able to do the cyber security analysis. ⁓ And so powerful that it keeps vulnerabilities in almost every bit of code ⁓ in code that we're using every day. the House is preparing to give federal agencies access to Anthropic's mythos. Why is this a funny story? Well, of course, they actually deemed Anthropic as public enemy number one when Anthropic basically allow to use all of the of Claude. their military hardware. obviously, know, you go, once again, Anthropic at absolute cutting edge and the White House ⁓ having to go and hey guys, you know, can we talk? maybe not all has been There was obviously a $200 million deal that OpenAI basically won as a result of Anthropic not signing away full rights to use tool in defense. so, you know, with Methos, it just goes to show how powerful this tool actually is because Trump's regime definitely they were never talking to these guys again. So we'll have to see how that one plays out. ⁓ And I don't the White House that kind of power, you know, that's quite a scary thing given all the weird things that they to be doing these days, but there you go. Yeah, talking about providing services to agents. I don't know if you know this company, Andon Labs, very interesting AI first company that has actually been looking at business models where you can run autonomous businesses with AI. And famously, they have been running the vending machine test. So probably about two months ago, they ran the second vending machine test, which Claude actually won. this is where they give each of the different AIs a vending machine and they them a budget and they them to make as much money as possible ⁓ selling KitKats bottles of water out of the vending machine. And the first was a bit disastrous. This was ⁓ back in... think late 2024 and 2025, the backend of they ran it again, got an entirely different experience because the models had improved somewhat. And now they've ⁓ and actually leased a San Francisco store they've handed to an named Luna. basically Luna, the first job that Luna did was it and hired two human employees over the phone. I'm not entirely sure exactly what those human employees are going to do. Probably stock the vending machine you it's got a three year lease and it's basically going to be to run a store. So there you go. And on labs, really fascinating business. Definitely want to check out. can have a look at what they're doing at and on labs dot com. In fact, out of all of the evals. Uh, this is kind of like the benchmarks for which AI is doing the best. I love the vending machine the most. And that's the one that I really look at because there are so many factors at play when you hand over a budget and a profit target to an AI. And, uh, obviously, you know, Claude, Gemini, uh, chat Upt, those were the top performing. Um, and I think they did about 20 or 25 different AIs. Most of them made at least a small profit, but they were about 25 % that actually lost money. as say, I think this is where the rubber hits the road. That's the real proof that we want to see. So I'll be watching that one very carefully and maybe even on the next show or something, we'll do a bit of a deep dive on exactly what that and on market launch is all about and which is they're actually going to be putting through the mill. So there you go, know, AI as your retailer. Definitely a positive signal. know, I think this is why the ⁓ recent by OpenAI's chief, what is he, research guy, basically, you know, you look at what they up with higher automation risk including you know, bookkeeping, Clark's data, intricate, customer service reps, sales reps, administrative roles and, yep, you know, higher automation risk. mean, I would say it should be possible more or less to entirely automate those, features. but no doubt, you know, if you are a bookkeeper with particular knowledge about tax codes, then they may still be ⁓ a much kind of order. that you can actually provide. mean, data entry key is probably not, but as a data entry Clark, you you may well become somewhat of an automation specialist and actually look at how you can automate the entry of data. And that's not really that hard or that technical. And I do think that even folks are at that kind of cold face of admin have the opportunity still use their skills, which is what you were saying earlier. In terms of jobs that will be reorganized, they're talking about lawyers, software developers, obviously, we're kind of moving from in terms of software development, very much moving from the actual coding to really thinking and describing in systems. And not such an easy thing to do. If you ever kind of sit down front of a computer with a blank page in front of you and say, right, you're going to build an app. know, good luck. I the art of really thinking in systems is something is available to but certainly software developers do think in that way. And so I think they're well placed to take on a much kind of higher order level of thinking and output. so definitely a big reorganization there. with AI, so these roles include software developers, graphic designers, and there's just a high kind of, you know, demand elasticity. I when output gets cheaper, people, you actually end up ⁓ asking for of it. So if I kind of think about music, for example, the fact that you make unbelievable music with something like... uh, you know, uh, Suno music, uh, uh, and it just, uh, you know, on Apple music, I think there is a, uh, completely made up artists with something like 11 songs that are in the charts. I think, uh, this particular artist, AI artist is now number three on, Apple, um, uh, on the Apple music list there. And, uh, I, just, you know, I just think. doesn't necessarily mean that artists won't a future. I just see the fact that you can actually create music for all kinds of experiences. a restaurant actually launch its own album. A mechanic shop create its own music that it plays for its customers while they're waiting for their oil to be changed. do you want to do these things? Maybe, maybe not. Do they become, you just another dimension for people to experience your brand? Absolutely. Is it now possible for, you know, business to actually have its own soundtrack their Yes. I think this is kind of the case in point, you know, it just drives an increase in demand. But of course, you still need people that know how to make good music, how to write good lyrics. AI can do a lot, but it can't select, can't curate for that as well as we can. so think just changing the nature of how we work, know, this kind of higher order thinking. And we see it everywhere. less immediate change. The jobs included teachers, nurses, plumbers. Firefighters. don't worry, Sam Altman, you'll still get a firefighter to come and put out that molotov cocktail. And health So obviously really there. You know, people that you know, that have palliative needs. AI can help, but it's different. not going to take those, ⁓ physical aspect of the Child care janitors. and, I think, obviously housekeeping and cleaners. So, no doubt, a need there, you know, for people that interact with the physical world. And so, I think there's one interesting one that worth noting a court reporters have a very high AI exposure score. Ranking in the top five in some measures, but in some jurisdictions, a human is still legally required to certify. each transcript and record. And so, you know, as was saying earlier, we're just going to be spending a lot more time reviewing what the has actually captured or ⁓ done analysis So there you go. So it's not all doom and gloom. think ⁓ it's it's going be long, long road yet. know, the future is definitely distributed. And so it's going to take time for potential of AI to really make its way into all jobs. But it's definitely coming. I'm excited by it because I do think it makes work so much more interesting for us. Well, can't agree with you more. At the end of the day, I have written an album for an alpaca farm. And ⁓ it's a fantastic album. I listen to it regularly. It's on SoundCloud. We've got a few more songs to release yet. But of each of the animals has they're actually, know, it's a herd of 20 alpacas and each one is named after a different music artist. And so we've taken that inspiration to write a custom song for each alpaca and know the farm is very well loved here in Ibiza and actually there are some 20 odd people that have adopted an alpaca and so you know hyper it's meant for a million people it's going to be playing to an audience of I don't know maybe a hundred people but you know those hundred people are invested in those alpacas and they love the songs And so this is what I think is so exciting for us. if you are ⁓ songwriter you a producer, you I think you just learn express yourselves in a different way. Ultimately, AI just becomes an instrument. But the creative process necessarily change. still, you know, are really driving and directing it. And that's kind of what's exciting. Of course. mom can do the same thing. So just have to get really good at it really really fast and so I think that probably the end of our show I was Trying again to improve AI joke writing skills in the humor segment. I absolutely Did laugh out today ⁓ the first joke that ⁓ tried to write? and it said, Anthropic shipped Opus 4.7 at the same price as 4.6. And it went on to say, in the same way, a used car dealer is technically a financial advisor. While that in itself wasn't that funny, what was obviously funny was the way that Anthropic had styled the fact that it's the same price, even though it's actually using your tokens 35 % faster on its extremely high thinking mode. Anyway, so on that note, that's a wrap for the show. Emil, thank you very much. Yes, we will have to do a big Ibiza launch. Actually, invented a new sound. It's actually called... It's called ⁓ Ibiza Reggae So, all of them this kind of reggae influence, but there you go. I Googled In fact, I Google it. I on ChatGPT and asked if Ibiza Reggae Lounge exists. And it said no, couldn't find any reference. So, there's another first on the show for you. invented a new fusion sound called Ibiza Reggae Lounge. So you can check it out on SoundCloud if you go in and do a search for Ralph Banker, you will probably find an album called Al Packers in Ibiza and take a listen. I think there's about five songs on there and there's another 20 to come. So on that musical note, we'll see you on the edge.