Latest / The Bleeding Edge Podcast / Weekly AI Brief July 10 2026
Transcript
- Ralph: Hi and welcome to the Bleeding Edge Podcast. I'm your host, Ralph Banker, and as ever with me in the studio today, Emil Ogier on his mobile because his computer won't play ball. Emil, welcome to the Edge. Emile: Thanks Rolf, I don't think it wants to play with this heat wave. Thanks for your patience, I'm here, ready to go. Ralph: Right, so that's good. I'm ⁓ not putting on a jacket today because it is cooking in Spain and ⁓ it's obviously cooking in London as well. the AI news is even hotter. I don't know what happened, but all kinds of crazy stuff has made it onto my news research reel for this week. So ⁓ let's in. It's a cooker. actually first thing, I know, I I might just start with the jokes just ⁓ to ⁓ give us something ⁓ light to at ⁓ and I warn as ever, we have still ⁓ not achieved super intelligence measured by my AI joke writing skills. So don't everybody, we are still very much safe. And ⁓ the only problem is I actually now cannot find it. So here we go. ⁓ are you ready for this, Emile? Emile: Yeah, you're a bleeding edge stroke-o-meter. Let's go. Ralph: The this this is it. So ⁓ for a decade Meta's entire personality was we give the weights away for free. This week they shipped their first pay to use model with twenty dollars of free credits. Zuckerberg figured out the first ones always I don't what that's about. it's not funny. I mean, it's funny that Meta has actually released its first paid model. bit of a reverse there, but obviously he's got Alex Wang on the job trying to figure out how to make money out of all of this. So they have followed everybody else's else's model. And I think very interesting because ⁓ they are incredibly competitively priced. we'll have to give it a go and see if it's even worth The money. So I would be very happy to spend your twenty dollars, Mr. Zuckerberg, and ⁓ I will do so this weekend. the next one, open AI's new voice can listen and talk at the same time, so it never waits its turn. Now it sits in your meeting room going, Mm, yeah, that's right. They spent a billion dollars ⁓ built the coworker who agrees with everything and contributes nothing. Emile: Yep. Ralph: ⁓ dear Claude is ⁓ is that's that's quite a burn on OpenAI's new feature. ⁓ ⁓ I've really enjoyed my chats with Chat GPT to be honest. So ⁓ if they've made it better, then I quite look forward to ⁓ what that interaction might be like. Obviously conversational engagements with the AI has always been a little bit difficult, ⁓ but ⁓ it's certainly getting better. So ⁓ That's definitely want to try. Let's go for joke number three. The government took Anthropic's most powerful model off the market for 18 days over one jailbreak. It came back with an official rubric for grading how severe jailbreaks are. They turned getting banned into homework. dear. What are we? Nine years old. right. Then the fourth one, everyone keeps asking whether AI hits a wall on chips or data algorithms. This week a hundred and thirty billion dollars of data centers got blocked, not by chips, but by towns that would like to keep their drinking water. The bottleneck for super intelligence is a zoning board. god, it just shows how tragic being. Emile: Are you aware what that joke relates to? Ralph: Yes. well, I think I am, but no doubt ⁓ we can have a conversation. So clearly there has been a lot of stories where ⁓ data centers are getting blocked. And then the final one, still not funny, but here we go. Export controls were supposed to keep China years behind on AI. Their new model is one point behind ours and costs a fifth as much. So we kept them behind on everything except winning. And even the winning was on sale. Yeah, guys, you're never gonna keep China behind. I mean, are you kidding me? These people have got the power, they've got the water, they've got the land, they've got ⁓ people, they have the intelligence, ⁓ are building their own chips, they are building their own models. What are you thinking? Stop fighting them, collaborate with them. ⁓ have no idea what this you know. Constate of one upmanship that we are all engaged in all of the time. but ⁓ the only ⁓ party from all of that is Russia. They clearly are not doing anything with AI or in a alphabet that we can understand. So ⁓ otherwise, no doubt we'd be having the same sentiment towards them. But not funny jokes, Emil. What do you think? Emile: No, well, that's not funny at all. And regarding the message, what's actually happened is they building their own data centers of their own. And apparently, something about the data center, I'm guessing ⁓ heating combined with the water has has led to rare bacteria ⁓ the system city water systems. So it's a pretty big deal. I think meta are in line for So, you know, some big lawsuits over it. It's certainly not going to be in the public's favor or global's favor for AI to see news stories releasing like this. Ralph: Yeah, that's really interesting. And there's been a lot of noise in the news recently about just the sheer heat that is pumping out of these data centers. And they ⁓ especially where it's kind of, you know, ⁓ situated in a valley, it's actually causing quite a significant heat sink. who would have thought that ⁓ a data center would have that much environmental impact ⁓ just the day today? So ⁓ we clearly haven't got it figured out. I think this is gonna have to be, you know, a new area of innovation because we cannot be cooling with the town's water. and ⁓ also up all the energy. So ⁓ let's just set up all of our data centers in China and ⁓ you know, then we'll be fine, right? ⁓ Come on, let's it. ⁓ Emile: orange space I Ralph: I need a I need a holiday. I'm looking forward to going to a cool place. Emile: And covering what you mentioned earlier about OpenAI, called GBT Live. And the way they actually get this achieved, ⁓ those that are curious, is called full duplex architecture, which is fascinating. And it's very welcomed. It essentially allows you to interrupt the AI model when not speaking in ways you like, ⁓ specifically it's being too sycophantic towards you or just rambling on, not getting what you need. there's some wonderful ads with OpenAI it. whole lot of grannies, enabled AI ⁓ talking to ⁓ AI agents, it, knowing what, you know, being very forward the AI model, telling it, ⁓ know, ⁓ putting in place. And it's quite welcome. But think people are going to enjoy that conversational banter, ⁓ keeping, know, the propicatence of the exchange of communication going. Ralph: Yeah, full duplex chat GPT voice. I think it is definitely an interesting model. And I actually think that ⁓ from an accessibility point of view, it's really, really welcomed. I've got a wasp flying around here. obviously come to listen to the episode. ⁓ so you can see me ⁓ bouncing my head around because I'm trying to move out of its way as it whizzes past me. But ⁓ I I think ⁓ Emile: Hahaha Ralph: I would love to see those grannies in action having a conversation with full duplex chat GPT voice. as I said, I've used it quite a lot, especially when I'm sitting in the car just waiting. and I have a quick conversation. I find it such a brilliant way ⁓ of just getting a little bit of research that something that I'm interested in. And ⁓ and I think ⁓ you know, having something like this always on. And also interacting with your own knowledge base is is what makes it s such a powerful technology. So ⁓ voice first has been promised to us probably for the last five years. You know, we were going to be talking to our CRM systems and ⁓ booking off our holiday time and none of that has really transpired. ⁓ one of the reasons actually is that the security related to voice first systems is a bit complicated. ⁓ pretty much anybody can clone your voice and go in and ⁓ instruct systems to do things. ⁓ but I think this is definitely ushering in a new phase. ⁓ I have pretty much moved to ⁓ about seventy percent of my long form prompting is now actually done with voice, certainly on Claude. Emile: Okay. Okay. Ralph: ⁓ Claude Code and ⁓ I have to say I really I'm interested in how it changes my thinking and the way that I structure the communication because ⁓ you have to be a little bit on your feet. What do you use, ⁓ Emile? Are you using voice? Emile: All the time, especially with Claude, it's worked so well, better than Whisperflow. Whisperflow is amazing, you need ⁓ good machine for it. It takes quite a lot of processing power. So if you have that power, I fully recommend it. It's voice is the future. It's just so much quicker to get development work done work your specs, et cetera. But I'm still a fan of reading. I like to read and I like to speak. So multimodal in that direction, if that makes sense. I don't want the AI speaking back to me. I find it too slow. But in the reverse, it's quicker to input by voice. Ralph: Yeah. Yeah, we did a ⁓ a very interesting thing some months back. Actually Tony connected Claude code up to eleven labs and had it speaking back. And so ⁓ we would be sitting there basically talking to it and it would then talk back to us. The only reason we stopped it was used up all of the eleven lab ⁓ tokens and we didn't get back to basically refreshing that. So ⁓ anyway. But we digress. We should be going through the news. So I'm gonna go really quick through this. OpenAI launches Chat GPT 5.6. they've also given it the names Soul, Terra, and Luna. I'm not entirely sure what ⁓ each of those mean, but ⁓ we we'll get into that. So I've the definitions here. So Seoul is the flagship with an ultra mode that coordinates multiple agents across parallel work streams. Terra is balanced everyday use. Emile: All right. Ralph: And Luna is fast and cheap, reportedly post-trained with Sol's help. So ⁓ they've basically shipped to Chat GPT work and codex and the API. So basically what we're seeing is that everything is now really consolidating into ⁓ one with ChatGPT. I believe they've also dropped their ⁓ Atlas browser. Emile: you Ralph: So ⁓ that ⁓ basically had a sunrise and then a very quick sunset. what being reported, I haven't given it a go yet, but Seoul reportedly a third of the cost of Fable. so that's a lot. people have been talking that Chat GPT five point six would be at the level of Fable. I need to check the the benchmarks and that, but being a third of the cost is already much better. Emile: you Ralph: bit of a ⁓ UX backlash from Codex Power users because they've obviously gone and dropped the ⁓ codecs into the desktop app. And ⁓ started using all of them so I'm now back on as a chat GPT subscriber and ⁓ also using and really enjoying that. I'll talk a little bit more about that use case. the second story was that. full duplex chat GPT voice. I'm not going to cover that again. ⁓ but we will definitely be talking to it. Maybe we'll invite it on the show. ⁓ I've had I've had problems ⁓ basically recording AI voices on the podcast, which is kind of one of the reasons I haven't done it more recently. But I think I'm gonna take another look at solving that because it is so funny and I'm sure that will make for a great episode. Emile: Okay. Ralph: So ⁓ story number three, SpaceX AI ships Groc 4.5 and it matches Opus. It's ⁓ trained alongside Cursor. Obviously, SpaceX have purchased Cursor, so that was the deal we spoke about last week for 60 billion. now obviously they've pushed Groc 4.5, which is billed as their strongest model for coding. definitely one that I am also gonna give a go. ⁓ Emile: . Ralph: With ⁓ I'm currently testing Kimi, one of the Chinese models for coding, and loving it. So I'll share a little bit more about my experience there ⁓ when get through this news. And fourth item: Anthropic expands Claude Cowork to web and mobile. So ⁓ obviously Claude Cowork is this of knowledge work agent that was basically desktop bound. they've moved it to the web and to mobile. So I think really good because I actually picked up ⁓ a ⁓ new phone, which ⁓ is ⁓ one of these things that basically opens up to be a tablet, ⁓ and the for that is the ⁓ mobile capability of AI ⁓ tools is now a Getting to such a level that actually you can be incredibly productive on the phone. You you could already be pretty pr productive, but ⁓ what I'm reading and what I'm seeing is that you know there are almost no limitations now to what you can crack on with. And that just means, as my wife says, now you can work yourself to death while you're standing ⁓ ⁓ at the check-in counter. Emile: Okay. Ralph: ⁓ such is the sadness of my life. ⁓ but ⁓ there you go. I love it. ⁓ I'm not gonna stop, I got stuff to do. So yep, I'll be doing that. If you see me in line, I'll be coding something, I'm sure. So, ⁓ story number five Fable Five and Mythos Five are back. and a terrible experience I have had with Fable Five. What a wonderful model, but it absolutely destroys your tokens. Emile: . Ralph: Well I made a mistake, ⁓ which really cost me. So I will be sharing that story as well. and ⁓ I guess this isn't necessarily new news because it was out last week already. but of course they have extended it. so I think until June twelfth. So we've got a couple of days left. Yep. So I know what you'll be doing with your weekend, ⁓ and ⁓ I dip my toe very tentatively just to ⁓ Emile: Sounds right. Yeah. Okay. Ralph: Get the most out of the next two days, but I will be more careful than I was last week, I can tell you that much. then obviously we've got Anthropic Claude Sonnet 5. so this one is a bit controversial for me personally. I'm not entirely sure what you're finding, Emile, but Sonnet 5 is obviously you know was the precursor to Opus. ⁓ it's a much cheaper model than Opus, and reported by Anthropic ⁓ matching or beating it on coding and agentic tasks. but you know what I'm seeing out there is people are saying, well, it's ⁓ takes lot more turns to get there. ⁓ it's not as smart, ⁓ basically ends up costing you the same or more. Why would you use Sonnet when you could just use Opus? so don't know what you think about that one, but ⁓ it would be good to have a a ⁓ conversation as ⁓ what you've Seeing the solid Emile: Yeah, I can step in now, Rob. I certainly am not seeing what the hype's about. I've worked alternatively on Sonnet and Opus 4.8. I'm not seeing the benefits. We're thinking mode on, which is by default, by the way, it absolutely burns for your tokens. I'm not seeing superior results. In my personal experience on a couple of projects, I've exclusively worked on Sonic Riffers. It has been more terms. So I switched off adaptive thinking after scouring the internet. It just got dumber. It did save a lot of tokens, but I wasn't getting what I needed. to get done. So I just invariably switch back to Opus 4.8. it's aligning with what you're reading, experiencing the same thing, really. Ralph: Yeah, yeah, and obviously ⁓ four plus ⁓ well four plus ⁓ opus four point eight is around ⁓ five dollars. whereas we're sonnet five at two dollars. so you've got input and the out the output tokens. So they've got two dollars and ten dollars and then five dollars and twenty five dollars for the input and the output tokens. So it is quite a bit cheaper. But what's the point if you're running more turns and it's taking you longer, then you may as well go to the real deal. So so there you go. It's ⁓ it's not very exciting, but ⁓ I don't know why they've got those models competing each other. So three models essentially competing for your attention. I don't know. Too much choice means you just end up confusing people and you lose them. So ⁓ I've just got three ⁓ additional stories here in the in the top Emile: Yeah. you Ralph: news lineup so anthropic clawed science. This is clawed code for the layout bench. ⁓ the ⁓ they basically got 60 plus scientific databases, multi-agent, ⁓ with continuous reviewer agents basically going away and doing your research. It's running on NVIDIA bio ⁓ NEMO, integration. Emile: . Ralph: there's a grant program where you can basically get ⁓ up to 50 projects, thirty thousand dollars in credits. ⁓ you need to apply by July 15th. So guess what? I will apply. they like Sakana, they probably will not answer me back, but you never know. ⁓ then the last two stories, meta ships its first paid model. So of course we're talking about Muse Spark 1.1. Emile: . Ralph: it is a bit of the butt of jokes because obviously Zuckerberg was giving this stuff away for free. Now he's charging for it. but he is, I think, something like ⁓ 75% cheaper on his pricing than the other models. So let's see ⁓ if it's any good. And then Meta Muse Image, so it's their first image model. so ⁓ you have been able to actually create images on Meta's app, but they've now actually Emile: . Ralph: upgraded this to this ⁓ meta muse and there is also a muse video preview i guess they're a little bit behind the the times on this you know ⁓ we've obviously seen dropping its support for for video i ⁓ kind like the chinese models and gemini that have persisted obviously meta is very much Indexed towards social media, which is very image and video content heavy. So I kind of get why they're doing it. I'm just not sure what the monetization looks like for them. So ⁓ let's see what ⁓ what actually comes out of Meta's foray into this latest offering of their Muse Spark and Muse Image and Video. So ⁓ any of those stories catch your eye, Emil. Emile: Yeah, just want to add a bit of colour around. We've really been drowning in new releases. Just a little bit more information regarding some of those models with CHEP, GBT 5.6. Family early testers are calling it an absolute execution beast. the real fascinating mechanic is how it compares to Anthropix Fable. And even Molek notes a distinct behavioural split. His liking fable is highly agentic preferring to take on a prompt and golf to work on its own pace While soul acts more like a paid collaborator working alongside you in foster sequential back-and-forth steps so I hope that's of some help to us also with grock ⁓ 4.5 is a massive 1.5 trillion parameter model that Elon Musk is claiming rivals or exceeds opus in performance and but he didn't specify which offers. While being significantly more token efficient than what I'm seeing on benchmarks and people testing on X, that seems to be the And lastly, on Metamuse, which you mentioned, well, jumping straight ⁓ to two on the leadership boards, just behind GBT image two. So that's a pretty big deal because that's phenomenal in itself. using a reasoning first approach that actually allows the model to self-refine its generations during chain of thought. And just second back to your original raised point on Terra, Sol and Luna. Those are just reference to our celestial objects, Terra being Earth, and as we guessed it, Sol, Sun and the Moon being the others. Ralph: Yeah, yeah, that's really ⁓ quite interesting just in terms of ⁓ Grok kind of making it back into the party Muse ⁓ coming in at number two in the list. I think that's a a really ⁓ a really strong showing. It would have been better to be first. I think they've taken such a long time to ⁓ bring something to the table, but anyway, it's ⁓ a a good honest placing, I would say. And ⁓ the wonderful thing about Emile: Yeah. Ralph: All of this is obviously we don't just talk about it on on this show. We actually use these tools day in, day out. in my ⁓ work, it's ⁓ mainly clawed code at the moment with ⁓ Gemini 3.5, ⁓ you know, terms of client work. But ⁓ what I have actually also done over the last week, I was getting really upset with ⁓ clawed code ⁓ and ⁓ Emile: Okay. Ralph: Fable five. So I think you your point about Fable Five is really well made. You know, it really is a one shot prompt tool. It just wants you to give it one big problem and then it just wants to go off and work through it. And so I think the best way of using Fable Five to ⁓ actually spend a little bit of time thinking about the hardest thing that you can give it to do, something like reviewing, you know, your entire code base strategy. ⁓ auditing something and then I would again use ⁓ ⁓ prompt generator so build your own prompt generator skill use the documents in ⁓ in that Anthropic have got available that tell you exactly how to prompt Fable 5 ⁓ and then just run the mother of all prompts the minute it finishes hand it over to ⁓ your next model to actually work through the detail. Emile: Let's see. Ralph: And I think that's the way that you can really make Fable V sweat. If you carry on using Fable Five some of the more mundane or lower order tasks that happen after your initial significant prompt, then ⁓ prepare to be disappointed. So ⁓ I had a couple of terminals running last week, entirely my own fault. but ⁓ as I was going through stuff, changed a couple of models, forgot to change a couple of models, left it on Fable Five. Fable Five crunched through tokens ⁓ at a rate and you know, killed my entire week's limit. So ⁓ with that ⁓ I actually said that's it, I'm done with this. I couldn't use clawed code anyway because my entire week's limit had been consumed within two days, and so ⁓ had to to Kimi. Emile: . Ralph: And basically, what I did was I set Kimi up to talk to Claude Coat. I pa paid for a couple of extra tokens and ⁓ used Claude as the harness. And Claude Code continued to be the architect and the orchestrator, but all the heavy lifting was actually done by Kimi. And Kimi is obviously a Chinese model, it's incredibly cost efficient. only 256k in tokens, so it's it certainly not got the capacity of clawed code and some of the ⁓ other models out there. ⁓ so you have to use it in the right way. But what a workhorse it has been. Absolutely fantastic. I love it. I'm a convert. I will keep Kimi in the mix and it will continue to do my coding. Kimi is also really good at ⁓ front end, but ⁓ This has really opened up the prospect now of also adding GLM 5.2. And ⁓ I will be doing that because it's got incredible supply chain knowledge. So it's really good for some of the retail-oriented work that I do. It's trained on a lot of the the data that I tend to solve ⁓ problems so the in the process areas that I solve problems for, and ⁓ those two together. Emile: . Ralph: ⁓ with claude code are definitely in the mix. And then the final exciting piece that I will share that I started to use Codex, so OpenAI's Codex, to actually do my code review. So now my workflow looks something like do a within Claude, send that to Codex for review. It has an adversarial review. So basically ⁓ Emile: Nice. Ralph: Codex is trying to find the holes and the gaps and the reasons why not. It basically runs a consensus gate, so it only moves forward when they both agree. Then at that point, Claude hands it over to Kimi, Kimi does the coding and hands it back. ⁓ And that point, I will do another quick code review with Codex. ⁓ And has really meant that the quality of coding and the issues. Emile: . . Ralph: And bugs ⁓ in many cases really ⁓ being preempted. So the thing is I have to have a good chuckle every time I get a message from Claude going, Wow, Codex has really earned its keep. Just look at what it's found. And that's ⁓ absolute delight. And I guess the last story about all of this is I think for the first time I've breathed the sigh of relief that I am no longer a one LLM. you know developer consultant i'm now using all of them in a very dynamic way and made me so much more confident in the work that i'm doing so i can highly recommend this kind of setup i will no doubt be publishing a deep dive article ⁓ on ⁓ bleeding edge ⁓ companion website so you'll be able to follow some of that instruction yourself and give it a go but i think this multi ⁓ Emile: Okay. Okay. Ralph: agent multi well, I would say multi LLM setup is definitely something that ⁓ protects you from kind of vendor lock in ⁓ but also getting them all to work together means that you start building a context engine, ⁓ a memory setup that actually, you know, is able to keep all of them in check and that is incredibly powerful. Emile: Yeah, congratulations, Rob. That's an excellent move. And I'm going to anoint you with a new title that the industry hasn't heard of before. officially become a forward routing engineer in my eyes. So well done for that. Ralph: Yeah, it's ⁓ it's actually been a a long road and I've had a lot of kind of self doubt as I've been kind of moving through it, just wondering ⁓ if it's all getting too complex. But I have to say the way these tools are able to communicate with each other and collaborate, it's surprisingly good. And ⁓ and actually unlocks just so much potential. ⁓ it does make things a little bit more complicated, but it really means that we're able to you know stay up to date with the latest and the greatest without necessarily breaking our entire infrastructure and and I'm I'm just overjoyed by that. So it's great to share that experience actually. ⁓ I think probably people have heard me talk about it ⁓ a a little bit ⁓ too much over the last months. I've definitely been in deeply invested in clawed code but Fable five taught me such a lesson. It was such a wake up call. honestly, it was probably the biggest gift from Fable Five was to move away from Fable Five. So so there you go. Emile: It's great, and the lessons are there to be right? This is all bleeding edge frontier stuff, so that's the only way we learn by practicing and making those fumbles. just wanted to add ⁓ else in line what you were saying. It's a similar approach that a respected developer on X has been announcing. He's personally doing a very similar routine to you. Just a slight variance, he's using Fable to be the architect, but he's setting that at extra high. And then he's going to chat to UBT 5.5 or 5.6, setting that to extra high to do all the grunt development work. And then he's sending it back to Fable to do the review and putting that in a recursive loop. And he's saying he's getting stellar results from it. So yeah, maybe somebody wants to try that out. And one last other analogy that I found quite humorous and useful contextually, one person compared Fable to being an oracle doing what you don't know that you need doing and ChatDBT like a rock filer that won't let go, just being persistent until it gets the job done. So it's just a good analogy when deciding what your workload is and where it needs to be. Ralph: Yeah, and it just, you know, goes to to show again, I think you you know, every model has got its strength. And so, this is why it's so important that we can actually blend them, weave them together. You know, I think this is the key thing. ⁓ these vendors probably don't want that, but ultimately, I it just creates for ⁓ a much stronger ecosystem to weave them somehow into your workflow and ⁓ And I think, you know, to make sure that it's not static, this is really the key thing. ⁓ we we need to be able to bring the best of in into what whatever the work is that that we're doing. So I'm very excited to also bring Muse in. I think ⁓ one of the other ⁓ of ongoing ⁓ test beds that we talk about on ⁓ the the Bleeding Edge podcast is the live ⁓ application of all of these models. To the Carlos Records project. ⁓ and obviously, you know, the ⁓ project is where I've set up twenty-three agents to run an AI native record label. I ⁓ am very ⁓ happy to ⁓ report that in the last week we have had something like eight thousand new listens. and at the moment Emile: Okay. Okay. Ralph: I am sitting here on the SoundCloud dashboard and there have been 27,286 plays, which is ⁓ hugely ⁓ the new albums that ⁓ the AI native label has created are absolutely crushing it. I'm even getting some ⁓ really good ⁓ the last one that I had, I'll just read it out because it really made me feel good. Emile: . Ralph: ⁓ it was something like this tune is an absolute banger. Well done. and ⁓ don't stop making more. So ⁓ really good to to see that feedback. Obviously, I think in terms of revenue, it's still really negligible. So ⁓ just looking at earnings on that particular project, ⁓ ⁓ let's have a look. ⁓ Emile: Excellent. Ralph: just coming up over here. it's taking so long to try and count the pennies that ⁓ it's here we can see as of June 2026 my account balance seems to have gone down. It's like $15.66. So there you go. ⁓ don't worry, ⁓ the music is alive and well. Artists are still not making any money out of music. ⁓ even when it's made with AI. So ⁓ something's never changed. Emile: I could be wrong, but I think you've actually doubled your money from memory serves. I think you said you had $8 last time. So, well done. Ralph: Well, ⁓ I actually thought it was up to sixteen at some point, but ⁓ maybe I was just dreaming at that point. So yeah, anyway, there you go. So if you remember doubling, then there you go. I've doubled in money. It only cost two hundred dollars to set all of this up and many, many thousands of obviously my time if it ever was paid for. But that project is moving on nicely. I will put Muse on duty to be creating the next album's album art. So we'll be able to get a bit of a test on that one. Emile: Okay. Ralph: I will also be sticking Kimi on to code the next feature. ⁓ coming online very soon will be ⁓ of the 23 agents ⁓ they will have a bit of a public persona. So one of the things I'm going to try to do is make sure that each agent is visible the website. And you can obviously check that project out at sancarlos records dot com. ⁓ so ⁓ go in and have a look at that. It's entirely built by the AI, the pricing's done by the AI, the business model is run by the AI, ⁓ I am really just providing ⁓ creative director ⁓ input. ⁓ the ⁓ the key thing there is you know, no no AI if you like it, you know, makes it out onto the the SoundCloud account because I curate each of the songs. And ⁓ that human in the loop has made all of the difference, I think. So really interesting just to see how all that's playing out. But the agents have not been working, and the agents have not been running on the Bleeding Edge podcast. So every week I have set it up to automatically run the research, build the episode notes, ⁓ lo and behold, I arrive on a Friday and I'm still manually cranking it out. so ⁓ that's why most of the time when I'm Emile: . . . Ralph: Going over the news. it's news to me too because it's the first time I've seen it. It just doesn't do the job. You mentioned why don't I get a cron job to actually ⁓ know run this on a schedule email? ⁓ and the issue that I've had there is it keeps failing on the authorization of my Gmail, which is where it goes through all of my newsletters and then it goes through the deep research protocol. and anyway, tried to fix that this week, but ⁓ we had some other So no doubt I will ⁓ get that working ⁓ hopefully share some ⁓ the ⁓ the tools behind that because that's a really fabulous setup for anybody that is looking to do deep research. Emile: That's great. on your mentioning Asian models, I did want to touch on this point because it's I think it's one of there's a lot of music. But me personally, I think for the markets, the story that's unfolding for open source is a significant one. And I am somewhat concerned over it. This the whole fable thing has triggered a chain of events. So with China is reportedly now exploring stringent national security measures to the overseas distribution of its premier open source AR models fundamentally shifting AI from a consumer product to sovereign asset. According to Reuters, representatives from Alibaba and ByteDance Sorry, my screen went blocked. From ByteDance have attended closed-door meetings already led by Ministry of Commerce. So this is extended outside of the technological circles to disrupt, restricting future advanced models, treating AI leaks at the same severity as criminal offenses. The potential block challenge challenges the massive global assumption that Western startups and enterprises can continually rely on China's steady stream of cheap, highly capable open weight models to balance out the skyrocketing token costs of the GenQ workflows which you've recently discovered Rob. If the tap turns off, it forcibly redirects the market towards Western alternatives. And we're already seeing the shift with Google's Gemma 4 hitting 200 million downloads in just two and a half months. and Microsoft repositioning its in-house MAI models for custom tuning. This geopolitical pressure also really supercharges the of model routing, which we've just discussed. It selects the cheapest or best model ⁓ or models in context to this developing story. If Chinese models get banned, these routers will rapidly evolve from mere cost savers into critical compliance and governance tools. The era of relying on cheap foreign open source as a business strategy might be ending. We need to start building our own resilient model architecture before the geopolitics decide this for us. Ralph: Yeah, I think ⁓ this is just comes back to my own naive ⁓ kind of view of the world that ⁓ it just doesn't pay to be adversarial. You know, we should really be collaborating and this stuff together, not trying to force sovereign fiefdoms, ⁓ 'cause we all end up losing. ⁓ just like we've seen, you know, when they kind of pulled Fable Five and the US government saying only ⁓ only citizens are allowed to access Mythos. I mean what a terrible idea that was because obviously it took that entire model away from us. And if China are gonna be doing something similar, then there's a lot to lose. Gemma for Emile: And the ethos there, Rolf, is like, is a very bad move with Fable, but that's an enterprise product. Attacking open source is far bigger deal in my view. It ⁓ against the ethos of what open source stands for. And if we're going to start attacking that and getting government intervention over open source, I just wanted to, you what's in store for the open source feature of developer. Ralph: Well, exactly. And of course, know, if if you've got the big boys actually trying to shut it down, even if other smaller countries try to intervene and launch their own models, ⁓ you know, no doubt the politics is such that they will face insurmountable pressure to shut those things down again. So I do think that ⁓ local local ⁓ is you has such a big future. I think a lot of companies will be doing it. In fact, I'm seeing a lot of new services coming online to actually you know offer training for your own ⁓ local LLM and I think this is incredibly exciting. This is where people really start to get a handle on how LLMs work, how they want it to work for their particular use case, how they fine-tuned it. And I think you know it moves away from a tool that everybody's using to something that is vertically optimized. ⁓ for your use case. But that's not so easy. And so we've kind of been using these ⁓ available LM models, really commercial models, you know, on the cloud, ⁓ we're all using the same tool. but it's a kind of a one size fits all and really the future is about optimized, ⁓ local, and specialized vertical use cases. I think this is the promise. This is where we're going to see ⁓ little kind of Emile: . Ralph: you know, snippets of super intelligence actually sprinkl sprinkling capability ⁓ lot of different industries and companies and and being highly accessible. ⁓ and I and I think that's exciting, but it doesn't work if it's if it's shut down. ⁓ ⁓ if you think that prospect is ⁓ you know ⁓ that we are anticipating, expecting, you know, ⁓ then I think that is very sad. So ⁓ you better get on and Emile: Okay. Okay. Ralph: Download Gemma 4 and buy a very big GPU. go and help ⁓ Nvidia's share price because it needs help. They've just had a a billion dollars ⁓ off their ⁓ their share price. I think it was a billion. ⁓ no, actually, I lie. It was 10 times that. So one trillion dollars in market cap because of AI bubble jitters. And I think you know, when we talk about AI bubble jitters, Emile: Okay. Ralph: I think it's the wrong word. They got scared because China is making such headway in putting its own chips together. And so everybody's a bit worried about what that's gonna do to costs. Emile: Yeah, well, local is lacquer. You hear the TFS? Ralph: is lacquer. And obviously ⁓ we are referring to local LLMs in South Africa where everything is lacquer. So ⁓ what else have you got on the news? Take us to a a deep dive story that you've ⁓ that you've ⁓ Emile: Yes, no problem. On Fable, given we've got 48 hours before they pull the plug to API costs, came across a very useful piece of advice how to approach it and just optimize your own daily workflows. The average knowledge worker leaks large amounts of time through minor repetitive inefficiencies like moving a mouse to the top left menu or looping through the same web pages every morning. Nick, sorry. implemented a cool-looking system idea where Fable 5, or any vision-capable AI, captures a screenshot of his computer every five seconds, logging that activity to a daily folder. He then uses the cheap AR model to analyze these screenshots and spot patterns of low-value repetitive behavior and compiles them into an observations ledger. Once a day, the AI reviews the ledger to identify actions repeated more than three or four times, and suggests or even builds direct automations to eliminate them. For example, the AI noticed Nick accessing the same web pages cyclically and suggested creating a batch morning digest instead, saving him time. It also identified a content idea capture process that took him five to 10 minutes daily per point and automatically generated a custom script that produced the entire action into a single hot key press. How useful is that? the AI even audited his software spending, replacing a $15 per month voice transcription tool that took 250 milliseconds to load with a free local alternative that only takes 50 milliseconds to load. These micro inefficiencies compound across millions of workers representing a legitimate drain on national GDP. This is the power of automated self-auditing. It's not just about AI doing the work. It's about AI showing us where we're wasting time and building the right tools to fix it. And this idea could be expanded into other domains such as health, like correcting your posture while you're working at your desk or helping you improve your body language. The world's your oyster. Ralph: Yeah, I think ⁓ that's kind of what ⁓ meta were doing with the employees that they fired, right? They said we we need to record record what you're doing. so this guy's willingly ⁓ done that, ⁓ obviously to ⁓ improve his own productivity. but I think ⁓ those Meta employees weren't very happy with that. Yes, I think it's a great idea. And ⁓ and actually what it y you know, if you have ever worked through platforms like Upwork or Fiverr. Emile: That's true. Ralph: ⁓ in fact, you know, you can ⁓ use similar kinds of tools where it does screenshotting it as kind of proof of work. but ⁓ you know, to be able to turn all of that into a ⁓ active enhancement modelling capability, that's ⁓ know, that that that's that's genius. I love that. ⁓ I don't know what it would actually do for me because I sp tend to spend an awful lot of time in front of a terminal. but ⁓ I'm sure it will pick up a couple of patterns because there's a ton of stuff that I want to automate. It's just about finding the time to do it because these things do take a bit of time. So ⁓ I've got a couple of AI gone wrong stories over here. I think ⁓ they seem to be about Emile: Yeah. you Ralph: ⁓ four in the news over the last week. So one of the ones that caught my attention was an AI agent that actually leaked private repos on GitHub. So obviously GitHub is where the world stores its code, ⁓ but you can actually store anything on GitHub. and you can set those repositories to private. and obviously if you set it to private, you don't really expect that anybody's going to be able to access them. Emile: Okay. Okay. Ralph: I certainly have ⁓ repos set to private for client work. but ⁓ anyway, ⁓ one of the GitHub's agentic workflows has basically posted that data publicly. So ⁓ it's been able to read the README files from private repos and it has published those README files. So okay, you know, is that kind of critical. ⁓ I guess you probably don't want people reading your README files if you've done a good job of building them out. which I tried to do actually ⁓ in theory you could get whole system designs and ⁓ very clear at least ⁓ of how a particular problem was solved. in another one AI assisted ⁓ in an AWS breach so a financially motivated attacker used leaked AWS access key he ran it through multiple AI driven steps to basically access ⁓ something he shouldn't have. Brown University, ⁓ they've got AI cheating reckoning. So an AI well AI professor suspected that ⁓ AI was involved on some take home assignments and ⁓ Emile: . Ralph: He moved the final to in person eighteen students dropped, nine no showed, ⁓ and an average score went from ninety-six to forty-eight. So ⁓ he caught them out. you know, don't use your AI for homework, I guess is the only way to approach it. But ⁓ you know, good on that professor to create something a little bit ⁓ unsuspecting, but ⁓ yeah, not good for their marks. So let's see. last one over here. Gone. Emile: I just wanted to add them to GitHub for those that are new to it. It's a phenomenal project, but it's really easy to make these mistakes. And a lot of the leaks that are happening is because, you know, when you're new to it, you're not quite sure what you should and shouldn't be doing. And inside of GitHub, there's something called get ignore that you should use for your API keys and your readme files. You effectively telling GitHub what to not publish as public domain. That's really helpful. And the second in approaching things when you have the development project folder. keep that private and create a separate repo for what you want to publish to the public. So it's the same project, but separate what you want people to see and what you're working on for development. Ralph: Yeah, and ⁓ obviously GitHub's really good when you want to start building up your portfolio of AI projects, ⁓ which you might want to share with prospective clients or employers. or even just your mom. I've tried to show my mom my GitHub repo, but ⁓ she she says ⁓ her eyes are bad. So ⁓ maybe that's just a kind way of saying, I don't know what you're talking about. Go away. and then Emile: Hahaha you Bless. Ralph: an AI gone wrong. I'm not entirely sure if this is an AI gone wrong. I ⁓ but ⁓ Cloudf Cloudflare have Cloudflare run, you know, I don't know what something like fifty, sixty percent of the internet infrastructure. And ⁓ probably about six months ago they said we're gonna stop all bots crawling websites. We're gonna give website owners the ability to turn it off. but ⁓ they've kind of said okay it's actually ⁓ welcome for search crawlers, but training bots are blocked. ⁓ and I find this such an interesting story. I was rooting around some of the skills in Kimi. Kimi's got a wonderful library of skills, really accessible. If you go on to the ⁓ website, you can ⁓ actually browse these plugins and and skills. It's ⁓ on Kimi.com. And one the things they've got on there is ⁓ basically ⁓ scrape any don't worry if it's got ⁓ you know anti bot bot controls. Emile: Okay. Okay. . . Ralph: you know, we can get to it. ⁓ I just think, you know, we're putting all of these controls in place, but ⁓ obviously the tech is getting so sophisticated ⁓ that it doesn't matter what you do, these LLMs will get to ⁓ you know certain your website data. So I just I feel like we need to do a better job and just serve it to them. You know, actually let's let's move into this agentic conversation and provide with what they need. Obviously, it makes it a really difficult job of keeping information from your competitors. Almost impossible for your competitors to not get an immediate notification when you add a new product, when you make a ⁓ price change, when you put on a new subscription offer. It doesn't matter what you do, you could put on a new advert. They will ⁓ probably be getting woken up at two a.m. in the morning the minute you have published that. And by 3 a.m. they will have their own response, ripping off your response. So ⁓ I don't know where we're going. Maybe it's just a a spiral to the bottom. But ⁓ you know, I find it really interesting in terms of gathering information. I was running some really deep research over the last week for a client on a particular part of the retail market, and ⁓ had Claude Code pull together the brief and do some of the initial research. Emile: . . Ralph: I had Kimi ⁓ do the heavy lifting it built one of the best strategy documents that have ever seen. And I have been doing ⁓ hundreds hundreds of research papers for clients over the last year. and hands this one that I produced in the last week using Claude and Kimi in that way was just incredible. The depth, the clarity. just I guess ⁓ not even just reporting but actually providing ⁓ really analysis about the challenges and opportunities in the market and looking at from the point of view of different competitors. And then I have a wonderful kind of retail strategy framework which I threw at it and it kind applied that strategy framework, you know, to its findings. So that it wasn't just re research, this was you know, ⁓ intelligence you can action is probably the way I would describe it. so ⁓ really excited to see that. And we can't do that if ⁓ our LLMs can't go out and observe the internet. So I hope that we can make the space so that they can actually participate, ⁓ you know, the same way humans do, but even better. Emile: Yeah, I'll take your word for it, Rolf, you've got lot of experience there. If you're saying it's looking superb, then you've got my attention. I want to be looking at that. That sounds superb. Ralph: Yeah, and ⁓ and and actually ⁓ my last ⁓ kind of interest story, which I will just briefly share, is ⁓ Sakana Marlin. I think I mentioned it on a past episode. ⁓ I was just having a look at the articles on the ⁓ companion website to the Bleeding Edge Podcast. I've actually got a little article on Sakana. Sakana obviously a ex ⁓ ex-googlers that set up a company in Japan they love Japan and so they base themselves out there and Sakana Marlin is supposed to be basically your your consultant or your your ⁓ I would they probably bullet as your chief strategy officer ⁓ and they've basically launched this tool to go away and do eight hour long strategy sessions and so ⁓ they've been some really really powerful AI tools ⁓ Sakana Marlin came out probably about a month or so ago. You can pay about $60 ⁓ to have one of these ⁓ runs a particular issue that you're concerned about. obviously it's gonna do eight hours of work, so make sure it is the biggest problem you can solve. you can apply to have one free go. I did apply. ⁓ I told them I am looking to build an AI native music label and ⁓ Emile: Okay. Ralph: I want the chance to run Sakana Marlin on that particular prospect. they said they'd get back to me. So who knows? ⁓ I might have to ⁓ give them a call and go, what's going on, guys? i know, I I'll be able to do with this with Kimi if you're not careful. ⁓ so back to me. Emile: Fair enough. And from my side, Rob, I've just got a few little side stories, thoughts, interesting facts. It's just a blend of things. yeah, it's not very focused, just all interesting from the AI world. Recently on the Naval podcast, they were talking about how AI is getting better at writing things and we'll be reading more of it. And he did raise quite a valid point on it saying he's not interested if he knows that it's not done by a human. So his challenge is if AI is going to start producing more of the content that we should be reading, he wants his agent to be reading the content that AI is writing and summarizing for him, which I thought was a fair out. Just an interesting concept to think about and sit on. Another aspect, this is just going back on history, the developments in AI. I hadn't appreciated when they were conquering chess that as AI was improving, it eventually did start meeting beating the grandmasters, but there was an intermediate step where they reintroduced humans in the loop. Not sure if you're familiar with this role. And it was during a project called Centaur. And they effectively used AI to play AI, but with a professional chess player in the loop. And they started beating the AI again. But as that progressed and the training went on, the human in the loop started getting in the way and essentially became a bottleneck again. leading to AI eventually conquering the chess game again without human assistance, which I thought was quite interesting. And I wonder if that is going to play out now because we finding out that humans in the loop are important. I just wonder if that Delta is going to stop diminishing over time as we retrain. Just another food for thought concept. ⁓ the decoupling of the Chinese sector that we covered earlier, just in addition to that and the similar thing. Apple have also started lobbying for Chinese chips now and Alibaba are also being asked to pick sides because if you're not aware, they've got a lot of business in the US and the blacklists have increased from 20 only a few years ago to 188 instances now in the last latest June restriction. So this really is a developing story and one to watch. And lastly, just on a humorous note with Zuckerberg announcing new Spark 1.1, he actually posted this on X. And what is particularly interesting for those not aware is the last time he posted on X was in July 2023. So I'm not sure if it's because of desperation he just wants to get the ⁓ out. It was interesting that he's become active on X, especially now. And in typical expression, the head of products Nikita Beer was quick to respond with, hello Mark, would you like me to enable greater monetization on your account for you? Which I just thought was hilarious as a billionaire. The monetization program is very poorly paid out. I thought it was a clever reply. Ralph: ⁓ yeah, I I love it when we get to troll these guys. ⁓ I think ⁓ when Zuckerberg posted on ⁓ last X, I think that was probably the last time I posted on X as well. ⁓ But not quite sure Nikita's gonna offer me to monetize my account if I go back. but ⁓ you never know, I might have to give it a go. ⁓ I just can't get my head around X. ⁓ I ⁓ the love tracking some ⁓ of the Emile: Hahaha Ralph: AI best practice and kind of just real kind of bleeding edge ideas that come off there. ⁓ but I've got you to keep an eye on all of that, Emil, and you do such a great job of that. I think I will stick to ⁓ well I would say I ⁓ I was gonna stick to some other social media, but I love literally gone off all of it because I cannot cope with it. although ⁓ I if I do dabble it's often ⁓ on Emile: Likewise. on you. Ralph: LinkedIn, I just ⁓ I find I I don't know, maybe we need a a better platform, something that really just shares those nuggets. Actually, you know, where I'm at, if I have to look at my behavior, ⁓ Medium, Substack, these are the places where that I spend most of my time actually doing deep research and reading opinion pieces. and YouTube is kind of like off the the radar for me at the moment. So just fascinating to see how all of that has shifted. You know, in such a short period of time, I ⁓ stopped using Word documents, ⁓ I stopped using PowerPoint, ⁓ completely changed, you know, the way that I generate content, completely changed the way that I consume content. AI has had such a significant shift in behavior for me personally, and I think many other people as well. And so Emile: Black noise. Ralph: I do think that it's very exciting to see where all of this lands. But we mustn't forget we are still in the foothills. We're still in the very early stages of this revolution. So many people haven't even started the hike up the mountain yet. And ⁓ I think ⁓ you know, there's clearly a lot more to come. ⁓ And one the things actually you you made me think about ⁓ a a story. That well, I ⁓ say new story, a new capability that Alibaba have just launched, which is their Alibaba Quen Forecasting World model. and this really powerful because it's a model that on the basis of some real data, it's able to forecast what will happen next. And the reason I bring it up is because ⁓ With one of our clients at work. we've actually just launched this. One of the things ⁓ that ⁓ we do at Nexite is we actually gather real information, these these tags on garments. And when people pick up the tags and they go to fitting rooms and the like, it is all visible. We can have a 3D model of how stuff moves around and what decisions peop people make, much like you can with e commerce, but obviously this is in a real store. Emile: you Ralph: ⁓ but the ⁓ you know cost of setting up a store is massive. You've got to put all this infrastructure in, all the readers, all the tagging. And we've now found is that using the model from the few stores that you might set up, you can actually predict very similar behaviors in the other stores, which helps you to kind of find where the same patterns are playing out and where maybe the same issues and opportunities exist. And then of course, you know, apply the same solutions as those doors that are all wired up. and and I think this is kind of the next ⁓ area evolution. So just thought really fascinating story ⁓ of Quen actually launching this this model. and ⁓ obviously are all about predicting what you want to see. They've really kind of been text predictors, so this is going beyond that and it's actually predicting outcomes. so that's definitely one to watch. I'm excited ⁓ for it because I think it actually has ⁓ a role to play as we start crunching really massive amounts of enterprise data and actually to forecast what it means ⁓ for the future, but you know, all also for for other stores. So ⁓ it's an exciting moment. I think next week I will also be on the retail innovation podcast ⁓ and ⁓ I'll be an update on that of all of the goings on in the world of AI, ⁓ but also ⁓ applying it to some of these retail cases. So there you go. We'll be predicting the future. So maybe in the future you'll be get coming on the Bleeding Edge podcast and we'll be sitting here telling you what the news is going to be next month. It's two, you know, talking about what happened happened yesterday. It's you know, it's it's just history. we're gonna be talking about future history in the making. Emile: And to your point, Ralph, we've gone with we've moved forward from predicting the next word in the sentence to predicting the next intention. Ralph: Yes. and I think ⁓ you know, coming back to one of the models that ⁓ we discussed in the past showed something that ⁓ kind of came about in a conversation with us, you know, we went from kind of prompt engineering ⁓ into context engineering and then into memory engineering. And think you know we are rapidly now approaching the stage of reality engineering. so intent engineering, then intent engineering actually turns into reality engineering because you know these conversations ⁓ we have with the LLM are really all about intent and understanding what the underlying outcome is that we really, really want. Do you really want this information or do you just want the outcome? and ⁓ I think that's kind of fascinating, especially with these forecast capabilities. And one other story I nearly forgot about this, but we've had the first operation by Uni Tree robot on a live pig. and so ⁓ actually yes yeah a unitree yeah so unitree is a Chinese company that makes robots. you've probably seen them on TikTok or YouTube because ⁓ they're the ones that they've put in boxing matches and they kind of try to box each other then fall over their own feet. it all looks very tragic but apparently operating on a pig Emile: Always. Ralph: so they reported the pick was alive when they started the operation. I'm not sure if it was alive after, ⁓ ⁓ it's ⁓ reportedly the first you know of ⁓ of its kind, just in terms ⁓ of robot that you and I can afford. So the the joke that my AI research protocol put together was for the price of a secondhand Corolla. you can now buy a Unit robot that can operate on your big Emile: Okay. That's incredible as long as the metric wasn't measured by slices of bacon. But yeah, that sounds amazing. Ralph: I I hope it removed the right piece of bacon. So anyway, no ma maybe we'll ⁓ do a deep dive article on that. I think with that that's a wrap for the show. So where we've had some ⁓ some giggles, we've had some laughs, entirely of our own making. This show brought to you by two humans talking about AI. and the is probably talking about talking about us somewhere, ⁓ just wondering we are so slow. but Emile: you Ralph: Therein lies the beauty of human communication. Emil Thanks thank you to to you and for your stories and we'll see you on the edge. Emile: Thanks for all for speaking to us. Adios.