Latest / The Bleeding Edge Podcast / Weekly AI Brief 27 March
Transcript
- Ralph Behnke: Hi welcome to the Bleeding Edge podcast. I'm your host Ralph Benke and with me as usual in the studio today, Emile Emile, welcome to the show. Emile Ogier: Thanks, Rolf. Welcome back. Looking forward to hearing about all your stories from Asia. Ralph Behnke: Yeah, I have a few stories to share from Asia and we'll get to that as part of my deep dive today. But first, I am going to do the news in 60 seconds. Now, when I was out there, my client said, I always say I'm going to do the news in 60 seconds and then I take five minutes. It's not my fault. It's just too much news. But today I am going to do it in 60 seconds. Here it goes. Nvidia reportedly gets China approval for H200 sales and restarts manufacturing of the China compliant variant. This is such big news. Big deal for China. I was just in China and let me tell you, they are ready to use those H200s. Emile Ogier: Indeed. Ralph Behnke: Microsoft is unifying its co-pilot leadership. So Mustafa Suleiman is refocusing his own frontier model efforts. Now, the fact that they're only doing it now is a bit of a wonder, but there you go. BMG, the big music group, are actually filing suit against Anthropic for ⁓ allegedly copyrighting I thought that was just for the likes of Suno and Udio, the music generation tools. But there you go. Apparently Anthropic have done the same thing. And I have to say, Claude writes awesome lyrics. has ⁓ basically made OpenClaw part of WeChat. So now you can invite OpenClaw as a contact. And it does stuff for you right there in ⁓ WeChat all over Asia, all over China. ⁓ The EU Parliament has basically supporting delaying some AI rules. So we've heard of the EU AI Act. Well, they're going to delay a few things there, which is no surprise. And it's also not the first time. OpenAI tweaks GPT 5.3. And so ⁓ some interesting use cases there because it's kind of instant behavior. They talk about it being less teaser-ish anyway. ⁓ Nvidia pushes agentic AI plumbing. So some really big news that I will talk about as well. Nvidia isn't just about chips and GPUs anymore. The is temporarily allowing perplexity to play in Amazon's backyard. ⁓ so perplexity a shopping agent. And ⁓ obviously, Amazon appealed they don't want anybody's shopping agent running on their platform. Anthropic is fighting the Pentagon. as you know because they were branded as a supply chain risk and ⁓ denying any wartime kill switch so obviously they've been of that but there you go and then the last one in san francisco there are protesters actually for a frontier ai pause commitment and that actually came out of a single source in the times of India. I guess there are always protesters in San Francisco protesting against AI frontier progress, but there you go. So did I do it in 60 seconds? It felt like two minutes to be honest. Emile Ogier: I was I was compressed. Rolfo was a lot shorter than previously. ⁓ well done. ⁓ But I always enjoy old updates. So that never feels too long to me. Ralph Behnke: you Yeah, so what caught your eye out of that list? Emile Ogier: Yeah, very interesting. I was hoping maybe your memory serves as sufficiently as mine, but didn't Anthropic have to pay out a sum for a lawsuit a year or two ago ⁓ for fair usage on authors' books? I'm if the same thing is going to play out. Obviously, the music industry is quite different in terms of... legal strength they have to fight things like this. ⁓ know the payout was poultry. It only, I think, a few million if memory serves. ⁓ I ⁓ the industry is going to be quite different. Ralph Behnke: Well, course, we've seen deals with the likes of UDO ⁓ ⁓ obviously the Warner Group bought out, I think it was the Warner Group, I'll have to check that in a moment. But they've been doing deals. I think this is the point, can actually, rather than just suing them, you can start to think a bit more broadly in terms of ongoing licensing. Of course, we've seen 11 labs, actually licensing voices of famous people. So you can go and get them, Judy Dench to, you know, read your podcast. certainly Sir Michael Cain is also on there and he would only be happy to read your children's bedtime stories. I think this is really the shift that we're going to start to see where it's more collaborative. and actually setting it up for future royalties because as you say, know, the genie is well and truly out of the bottle. They've already used all of this. this information. They've consumed all of the internet 12 times over already. you've got GROK, you know, ⁓ GROKiPedia, obviously building its own information sets now based on what they have already consumed. And so we've come full circle. And so I think now is the moment really to ⁓ maybe into battle. but use that as an opportunity to leverage kind of co-structure and royalties, frankly, of ongoing content because this is what AI now needs. so we're seeing these music platforms taking ⁓ interests the generation capabilities and partnering with them because they really start to understand that know, there is no way that AI isn't involved in making music. Vel his estate just signed off for him to participate. Obviously, posthumously, he's dead, but he will be the of a movie ⁓ that he desperately wanted to in when he was alive. ⁓ They are now going to make the movie with his AI likeness and so his daughter has signed that off they can go ahead and do this and and so this is what I'm saying, you know I think it's now you have to participate Actually somebody because they train on your data. ⁓ That's and maybe where the small money is but it's a good place to start If you can start to get a bigger piece of the future pie Emile Ogier: I think that that movie is an interesting one. It's more tasteful in terms of, you know, creating the footpath for the future on this because apparently Val Kilmer has heritage in what was covered and the director wrote the whole role around him. Unfortunately, he started having all the problems with his throat cancer. know, it's in the spirit of the movie. It was framed around him. ⁓ really looking forward to seeing what they do with it. Ralph Behnke: Yeah, and I think it's a great kind of pioneering use of AI at this point. Of course, we've seen, think probably the most brilliant one, ⁓ I loved reading comments for was a post on X, Iran the war, and it's a movie trailer. Have you seen this thing? ⁓ Basically, movie for Iran the war and ⁓ Liam Neeson ⁓ the... Emile Ogier: Bye. Ralph Behnke: president so he's Trump you've got you've got Gandalf ⁓ as ⁓ Khomeini have got ⁓ Jake Gyllenhaal Khomeini's son and anyway the characters are brilliant ⁓ Zach Galifianakis ⁓ the vice and there you go You know, sometimes the only good reason to go to war is no reason at all. That's the kind of catchphrase. People on X are saying, did these actors really participate in making this? Because it's a blockbuster movie trailer that they have gone and made. If haven't seen it yet, Emil, I will find it and ⁓ share ⁓ Emile Ogier: I'll look it up. Thanks. I'm curious on what they use for post-production. Ralph Behnke: Well, it's very well orchestrated. People on X are saying, did these actors really participate in this? And I just think that kind of just is the world that we're in today. those actors didn't participate in that. It's ridiculous. Yet so real. ⁓ And that brings us to an interesting story, which I'm sure you must be covering. Sora. is disappearing so OpenAI have axed their video platform which means they're no longer in the video game at all is that what's going on? Emile Ogier: Yeah. I don't have a story, but that's essentially it. So, you know, with the explosion of open claw, it's somewhat changed, significantly tilted the tables for all the big AI labs. They're realizing that, you know, the profit margins are in enterprise. a race going forward for all the AI labs to get involved ⁓ enterprise solutions. OpenAI have been involved in a lot of things, lending from Sam's Combinator DNA. ⁓ likes to get involved in a lot of things, try a of things and see what sticks. But ⁓ doing so, they've been hemorrhaging and spending a lot of money in a lot of directions that hasn't been as profitable as they would like. keeping the traction of all the many services they have. I know we were speaking on a channel earlier this week. I was asking the question, which one's going to give? And it seems like that's the one. And makes a lot of sense if you sit back and reflect on it. It takes a lot of compute and inference to run VideoGen. And they could be using all that inference for productivity services for enterprise. So I think it's smart. The smart thing to do for the lab, but I'm sure a lot of people that got involved behind the package and started using it as a service for themselves. woken up the next morning to realize that it's effectively being pulled. Ralph Behnke: Yeah, it's the kind of the first big casualty in all of this. Google Google labs for this very purpose. They're always pulling software. ⁓ And big one that's going to be disappearing in 2027 is Firebase Studio. ⁓ so will but Firebase Studio was essentially the agent coding agent that was integrated Firebase. Very, very good idea. Emile Ogier: ⁓ wow. Ralph Behnke: I could never really get it to work. obviously it's been superseded by anti-gravity. So what's interesting is we're seeing a very kind of short shelf life, you know, for some of the stuff. But as you say, a real cleanup of projects because the focus has to now be on actually gaining traction with enterprise. They've got something like 900 million users at last count, maybe a little bit less after the Mcgeddon. I made that up. I don't know if it was a Claude Mcgeddon, but certainly a lot of people went over Claude ⁓ on the of the ⁓ news with Pentagon. Anyway, and so I lost my train of thought over there. I wanted to go back to the enterprise thing, but Emile Ogier: It works. Ralph Behnke: the enterprise play has been well and truly, I guess, put under pressure with this automation capability with OpenClaw. know, Peter Steinberger really hit on something at the right moment. We're talking a few months back, I think in November when it really started to get going. And he's captured the of what automation could mean. And of course, the thing that really makes it work right now is the fact that there are 500 integrations it will interact with things like Slack and WhatsApp and Telegram. so fact that ⁓ you've Tencent integrating it into WeChat, actually makes ⁓ let's say, pervasive network capable ⁓ of having some form of agent activity, even if that is just something talking to you, know, seemingly autonomously, much multi-book where the agents are having a big agent fest and they're all just chatting to each other and ⁓ humans observing the agents chatting. You know, this is kind of, it's captivated our imagination. And so now the pressure is on enterprise to basically figure out a way of making it work. ⁓ And course the security. implications have been significant and so nobody in enterprise has touched it. yet it is, know, every Joe ⁓ his is trying it out. And of course ⁓ this created very interesting space for Nvidia and so they came up with ⁓ Nemoclaw and is just in the last few weeks it's something like 18 different companies that have all come together to collaborate on this and they've brought Peter Steinberger in because OpenClaw is an open source tool and they've said to Peter right we're going to build an environment that will run OpenClaw and it's called Nemoclaw. and it's what they call a reference stack, an open source reference stack that helps you run OpenClaw always on AI agents more safely by putting them inside a controlled runtime called NVIDIA OpenShell. And this is absolutely gold, I would say. course, it does take a little bit of setup time, bit longer than you would set up OpenClaw, just let's say vanilla, but much for the first time, think, you know, here, enterprise can now experiment anybody that has a business and is serious about not having these agents run away amok with their data and everything else. But actually, obviously, are still building an agent that is going to interact with the world. And so this is the place to experiment. It's called NemoClaw. installs OpenClaw and OpenShell in one ⁓ and you can basically run it on this Nvidia platform. ⁓ Basically there's interesting terminology. NEMO is Nvidia's framework for building AI so I'm going to come back to that. NEMOTRON is Nvidia's model of AI, it's a big family, the big NemoTron 3 super, 120 billion parameters. And I think that means you kind of get like 10 % of that when you're running. So 12 billion, that's a pretty smart model. That's NemoTron and then NemoClaw, the agent stack that runs OpenClaw. And so these three things together point to something unbelievably exciting for me because right now, We've really thought of Nvidia as a hardware business, GPUs, CPUs, but here we are, we're actually seeing them as now a build platform. Emile Ogier: Yeah, I think really interesting, Rolf, for several reasons. It's slightly strategy to Anthropic and OpenAI. They essentially, as you pointed out, they're creating a middle layer, integration layer between ⁓ OpenClaw. and the developers. And I think it's quite smart several reasons. One is essentially he's saying we trust the developers to figure this out. We know this is useful. We're going to provide a safe environment that you can work in at an enterprise level. So taking away all that friction while still for that capability and flexibility that the market responded to the pulse that got hit with OpenClaw. And in addition to that, it keeps Nvidia in the game, right? Because the more people that adopt it, takes that inference edge ⁓ local compute to So the more ClueBots we got, the more of that ⁓ support that grows around NeoClaw, the more ⁓ Nvidia ⁓ and more... enterprise gains and the more platform expands in terms of richness and capability as far as skills and services are combined. So ⁓ looking forward to seeing where this goes. Ralph Behnke: Yeah, and of course, Peter Steinberger is at OpenAI. they've said is that OpenAI is going to work on bringing agents to everyone. Agents that even my mum can use, apparently was what said. I'm not quite sure whose mum they were talking about. And I think an on things safely with access to the best models. OpenClaw, Sam has actually said, will move to become a foundation and remain open with OpenAI supporting and sponsoring it. And so now we're really seeing, know, Nvidia basically creating the infrastructure to run it safely. So it's something that's here to stay. And I just think we've kind of hit this wonderful moment where suddenly... a technology that has captivated people's imagination it works. It's all of the stuff. I think they will figure out how to make that useful. And we figure how to make that useful. And so this ⁓ starts to become everyday tool in a way that maybe actually starts to rival. what we do with our messaging platforms, because initially I think a lot of people will be using it for email and communication. so incredibly, ⁓ OpenClaw almost becomes the vehicle that really takes ⁓ mainstream into automation. It's something that you can invite as a contact. ⁓ And we've seen that from Tencent and their WeChat tool. But... Yeah, it's a big, big interesting moment. So why don't we jump into your first news story? Emile Ogier: Yeah, of course. Not a new story. thought bring another lesson or some strategy advice picking the right agents. There was ⁓ interesting episode I followed with Jones. He shared recent insights for evaluating and picking AI agents that are best suited for you. As you know, there's a lot out there. The market's moving fast. There's a lot of noise. Everybody's telling you they've got the best thing. It's really hard to discern, you know, what's best for you. And I think this advice goes quite a long way. The core premises around open core phenomenon, which we've been speaking about, which as you know, Ralph has become a category defining moments in AI for AI agents, similar to what Linux and Android did for operating systems. triggering an avalanche of copycats, open source forks, and corporate responses from giants like Meta, Anthropic, and Perplexity. As a result, it's easy to get distracted in the daily hype or the security concerns which we've been covering. What's important for builders and operators to understand right now is that every new agent launch represents a distinct strategic bet. And understanding these bets is the only way we can make informed decisions over which is going to be best suited for our needs. Nate offers a pragmatic three-step framework for evaluating any agents. So next time any new agent drops, which is likely even happening right now as we record this episode, Try ignore all the marketing hype and rather ask yourself these three granular questions. Where does the agent run? The infrastructure question. Is it local on your machine, cloud-based or hybrid? Why this matters. This directly determines your data privacy posture, your security surface area, and who is legally responsible if the agent makes a destructive error like deleting your inbox. Question number two, who orchestrates the intelligence? The model question. Does the agent rely on a single proprietary model? Does it have multimodal harness? Why this matters? This dictates your operating costs, the quality of your agent's work, and your level of vendor lock-in. Question number three, is the interface contract? The UX question. What to look for? How do you actually talk to it? Does it hijack your desktop? Do you message it via WhatsApp? Why this matter? This dictates whether the tool actually stick for you. Be honest about your own habits. Pick an agent that integrates into the message platforms you already use. So currently there are around five major strategic agent plays. There is OpenClaw, the sovereignty play. How this runs? It runs locally, fully modular. and you bring your own LLM. The trade-off, you get maximum data sovereignty and flexibility, but you take on maximum technical complexity and severe security risks, e.g. supply chain attacks on plugins. It's best for developers and highly technical power users who demand total control over the infrastructure. Next, we have Plexity Computer. the delegation play, how it works. It's a fully cloud-based virtual box execution environment. The trade-off is you trade away control and you pay a premium for it, circa $200 a month. So you don't have to worry about security or infrastructure. You just decide the outcome and they handle the subtasks. It's best for enterprise and non-technical knowledge workers who want professional-grade outcomes without managing the backend. Meta's Mantis, the distribution play, how this works. It's a highly accessible agent designed to serve billions of users at scale and keep eyeballs inside of the Meta's ecosystem. The trade-off is you get convenience and an easy UX, but you're trading away your data privacy to Mark Zuckerberg. It's best for everyday consumers and small businesses who want agentic capability without touching the terminal. Anthropix Dispatch, which is new, the safe and secure play. How this works, it allows you to drive Claude on your desktop directly from your phone. It's single threaded, secure connection. The trade off, it lacks complex multimodal routing flexibility of OpenClaw, assumes you are a super fan who only wants to use Claude. It's best for non-technical professionals who prioritize safety and just want a reliable mobile-friendly way to execute tasks like calendar management or building PowerPoints. Finally, lovable, the relentless simplification play. How this works, pivoting from a simple Viacobing website builder to a general purpose agent executor. The trade-off, they are attempting to walk the tightrope of keeping the interface incredibly simple, low technical complexity, while giving the user high control over complex agentic workflows. So the ultimate takeaway for 2026 is that relentless simplification is compressing the software layer. Every vertical tool is collapsing into a conversational agents and where the AR tool is stuck in the middle, not deep enough to be specialized experts and not broad enough to be a general delegator, it will die. So to round up, choose your agents based on where you're willing to trade technical control for convenience. Ralph Behnke: Yeah, so our agents are like consumables. not... I want them all though. Emile Ogier: fail. Open your wallet. Ralph Behnke: There you go. That's right. And I'm actually starting to see, you know, where the costs are hitting. So we've got quite a few automations and different tools that we've been running lately. And I guess I've been reflecting somewhat on ⁓ tools we spend most time on. And I have to say, definitely, ⁓ you Claude code ⁓ and even Claude desktop. I use a fair amount I think I wonder about all these agents and what people are really getting benefits from. And so are my top picks just of what I've been finding quite useful. Practical ways to use AI. So first all, with Claude, actually ⁓ just connecting it your ⁓ it summarize what's going on in your email. Some amazing insights. Building write like me file where ⁓ basically able to analyze your voice, ⁓ your tone voice and your writing style. then using voice and you can just dictate ⁓ a message and say the outline. ⁓ write like me skill, you bought a little skill and it will then reformat that email in your voice and it'll just send it off. Wonderful you're doing summaries ⁓ about ⁓ and describing requirement or you're providing an update because it actually just fills in all the gaps. then actually ⁓ using tools like Notebook LM. I've spent a lot of time in China training out that let's just say the principle of Notebook LM. Of course Notebook LM is a Google product. Google doesn't work in China. It might work for a few days, but then they always get you. I tried everything with Coat figure out how Google had figured out that I was in China and had started blocking. the use of my apps even though I was using a VPN. Those guys are sneaky. But you know there are other tools like it like Lily's AI. These are research tools. You just load in all kinds of PDFs. You load in all kinds of YouTube videos. You can load in research from tools like illicit over 200 million research articles. you can load in customer feedback and insight from tools like Reddit into Notebook LM and ⁓ Lily's And immediately you can then build unbelievably detailed research and also publish ⁓ your ⁓ in ⁓ that like Excel. And so these tools are so accessible and we use them so much just over the last two weeks. It was actually ⁓ really to not only be stuck in a kind of terminal view where I'm using code code. Although I have to say, know, ⁓ for your buck. That is definitely the most effective tool because it you're spinning up multiple agents. ⁓ all going away doing tasks. You're building your own, you ⁓ your agents with their own skills. And so there's all this talk of agents and there's all this talk of kind of picking the right platform. You've just given a bunch of different options. Yeah, there it is. Claude code is just happily already doing that. Having done that for months, know, it's kind of fascinating for that we even have the conversation because ⁓ there it It's what most developers have known probably for nearly the last 12 months that Claude code has had this kind of capability. even Cowork was really a commercial, let's say, ⁓ it commercially accessible. And so we really this multi-speed AI where we've got these waves of adoption. And yes, people do need to make a choice, but maybe what they just need to do is choose Claude code and forget about the rest of the note. the noise and you know just walk off into the sunset. Emile Ogier: Well, for me, I'm partly where you are. For me, the decision is to stick with Gemini and Claude. I think the pair is a compliment. You get all the video gen and then you get ecosystem of Google, which is huge. And there's very wide enterprise adoption in that. as mentioned, Claude is doing very good work on connectors to the likes of Gmail, et cetera. So maybe you don't need it. It depends on your use case, really. But interesting stuff, but I do have a couple of questions regarding your experience in China. the same measure, was Claude also not being blocked as Gemini or ⁓ Notebook LLM was? Ralph Behnke: So Google was particularly aggressive. I was able to access my Gmail, but definitely Notebook LM. They started to block off for a few days. And Claude fine. ⁓ I to ⁓ use via VPN. And interestingly, asked the manager, said, right, if you were going to ⁓ Emile Ogier: All right. Ralph Behnke: develop something, if you were going to build something agentically, like what would you use? What are your tools of choice? He looked at me in horror. He said, I don't code. I said, don't worry. You will be soon. And he he doesn't know what tools they use. So I asked him he could find out. And he contacted many of his peers and he took a little mini poll and what transpired was absolute gold because ⁓ it's Emile Ogier: you Ralph Behnke: basically seems that anybody that is serious about coding is using Claude code even in China and ⁓ they are using a VPN. Gemini was also on the list but of course DeepSeek is also an option in China and so if they could they were trying to use Claude there are so many You know options and guess we will start to see that that changing but I ⁓ do it I absolutely get it. I have for absolute brute ⁓ simplicity and thrown Best at it with access whatever I need to do, I do. ⁓ don't worry about it. And ⁓ getting really, really good results. And so ⁓ team was amazed at the of some of improvements that we could make. And I have to say, you know, again, we talk about agents and ⁓ we talk about of levels of and One of the real interesting eye-openers for me was just how fast it is to write code with these coding agents that will do an activity. So it just doesn't mean that you're actually deploying AI. You're just deploying and ⁓ especially ⁓ Python that are able to do something very specific. And normally these things are really simple, like copy a bunch of information from one Excel spreadsheet, merge it with a variety of variables and upload it to another spreadsheet, which then gets uploaded somewhere else. And so, much of what I've seen is just basic application of using the AI tools to solve admin. It doesn't have to be as sexy as end-to-end automation. it's really just interacting with 20 spreadsheets ⁓ Emile Ogier: You're just distilling why open-claw has become so popular for that very reason. Rolf, that's really interesting. I'll tell you why I was curious is just because the founder ⁓ Anthropic has quite a strong view on not wanting to share those NVIDIA chips with China because of the AI race. Ralph Behnke: Right, exactly that. Emile Ogier: So I thought for that reason, they must probably have a very strong view on blocking Claude. So that's really interesting. And my second. Ralph Behnke: Well, they definitely block, but I think, you know, as these things are, a lot of people are running VPNs. It's definitely problematic, but if you get it right, it can work. interesting is I was reading the local news and ⁓ has made such an impact on the local economy that they have run out of Mac minis. this is just because OpenClaw runs really well on Mac minis. ⁓ had a variety of companies that have offered open days where you can come and bring your Mac mini and they will set up OpenClaw for you and configure it safely and securely or as safely and securely as they can. And so it's a movement over there. I saw a number of councils providing funding for sole entrepreneurs who want to become a one-man show using the of OpenClaw to automate their business. They are absolutely the participation in the AI economy and trying to enable it and support it. Having said all of that, the reality on the ground is nobody knows anything about pretty much everybody is just as much in the dark as they are in the East, West, North, South around the world. Most people are in the same boat. Top 1 % are using these tools at an advanced level. The 99 % have no clue. and I can't remember the exact stat. It's something like 60 % of people have never touched AI anyway, you so we are still at the very early stages, but I see that it's a very practical application of AI that is starting to take shape in China and... I think they'll go far with it and these are going to be quite interesting because they are, ⁓ in fact, for data centers, they're accelerator chips. They're gaming cards and they're mainly used to run AI. and heavy scientific computing ⁓ faster than normal CPUs. And so, you this stuff is going to be part of the infrastructure. So no doubt be able to use it for training large language models and running inference at scale. You mentioned inference a few times. It's, you know, it's a big deal basically to be able to get these requests ⁓ and come with answers. very large models. So I think long context, this is one of the things obviously we've seen Anthropic or Claude code moved to 1 million token context which has really changed development for me. I've seen a big shift in how I can work. And then obviously, you know, high performance computing, climate, weather, simulation, physics, chemistry, genomics, you know, what's next? What's next is AI is picking the research problems and then solving the research problems. And so think China will ⁓ in front. They are excelling in that front. And so I think, ⁓ I know, should we be worried? I mean, this is the part of the problem I see. You know, we're all concerned. ⁓ We to move the many as one. ⁓ And the world is so divided. that that's never going to happen. But it's going to happen anyway. I think it's unavoidable and it's a big economy and don't kind of for a minute think it's easy for them. I well, I predict I have experienced that there is a cognitive barrier to ⁓ onboarding at scale. ⁓ And a very alien experience for many people. Emile Ogier: Yeah. Ralph Behnke: in the enterprise. Training them is very difficult in the US US trainers. I haven't heard this firsthand, but certainly there seems to be from trainers that are training in enterprise. One of the first questions people ask them is, do I need this AI training? I'm in three years. And so, you know, that's the sentiment, then you are somewhat lost. Emile Ogier: And in addition to that, Ralph, it is really interesting because I know a lot of companies now have moved from the training phase to, you will be measured by this. Your year end review will be measured on what your uptake on your AI utilization is. And that's all well and good. But as you say, if there's this, this, this barrier to the psychological barrier to, to absorbing AI, that, that psychological hit, how that's going to impact things even when they have the best intentions to take up this new tooling. There's only so much cognitive load you can put on people, especially when things move as fast as they are. Ralph Behnke: This is our biggest blocker and our biggest challenge. And this is why, you know, I think getting to grips with some of the basics, actually using it to have a bit of a crawl over your inbox, you know, actually using it to build a write like me skill and then dictating something to it. Nearly everything has got voice enabled now. Certainly ChatGPT has a wonderful voice mode. Claude has voice mode, Claude code has voice mode. don't have to use whisper flow anymore. so ⁓ tools are really making it accessible. And so much of our development is just having a good conversation with Claude and Claude and letting it get on with the task. And so things have really shifted. I saw a fascinating stats of Boris Cherny, co-founder of Claude Code, and he posts something like or pushes something like 6,000 commits ⁓ per So a commit is when you commit some code to GitHub, you know, you're committing it to 6,000 a day. And so that really is the level of perform at 6000 anything a day is a lot. So imagine. Yeah, well, well, it's an amazing stat. And so, but this is of course, you know, I think Andre Karpathy also Emile Ogier: I think I shared that with you. It's incredible. Ralph Behnke: I was talking about the kind of the FOMO, the fear of missing out on his token spend, his token allocation, token budget. And I had to deal with that myself because every night I go to bed and something's not running overnight, I feel like a complete fraud. ⁓ And I have to get up and just invent something, which is ridiculous. I don't actually do that ⁓ very often, I have done it. Emile Ogier: Yeah. Yeah. Ralph Behnke: You know, we want to use the capacity because the tools are so capable. Yep. It's a big moment and things definitely ⁓ moving fast in terms of automation. But so tools remain ⁓ underutilized. Who's to miss Sora? I don't know. Emile Ogier: Yeah, interesting. Ralph, just because we've been covering Anthropic quite a bit and all the tool use, especially on the more mundane workflows, et cetera, I thought it'd be quite useful because it's new just to cover Anthropic's Cowork Dispatch. And what it is, what it does, for those that are interested, it's a new feature from Cloud Cowork called Dispatch. It offers persistent local execution, which allows Claw Desktop to run autonomously in a sandbox environment. on your local computer. Remote delegation is possible by messaging Cloud and assigning a task directly from your smartphone while you're away at your desk. Asynchronous work is possible because conversation is persistent. It runs on your machine so Cloud can execute the work autonomously and in the background. You can assign a task remotely and simply come back to your computer later to find the finished work waiting for you. In essence, dispatch acts as a remote control to your desktop AI agent, allowing you to trigger complex local tasks on your computer using just your phone. This essentially enhances Anthropic service offering, bringing it one step closer to Clawbot capability set, but without all the security concerns. So definitely want to check out for those that are keen to get on board with Clawbot. Ralph Behnke: Yeah, and the game changer here is the fact that you can really be productive on the road. And so the really big shift in terms of the capability on the phone. One of the best features that I've been using recently ⁓ voice mode, both ChatGPT and Claude. Actually, you can have a really good conversation with ChatGPT, but does a really good job ⁓ of gathering requirements and away and doing work. And you can have that conversation just on the way to the lift while you're waiting in the lobby or while you're sitting in your car or while you're actually, you know, at the bus station or on your lunch break. And I think when I'm out and about, this is the kind of work that I'm doing, actually just recording the work to be done. ⁓ a phenomenal way of actually interacting with your AI because ⁓ goes off and ⁓ talk about practical things. You know, it's sending emails. It's actually ⁓ creating It's building and formatting a document. And you have to write anything because you can actually speak these particular instructions. They can be very sophisticated instructions because you can of course invoke skills skills essentially tell Claude how to do certain repetitive tasks ⁓ if you've got a good skill then it does that really really well and that skill might be creating a document, creating a spreadsheet, doing piece of research, writing some code, building an HTML file. So many these things ⁓ are really useful in of the everyday know, content that we create. And so I really see a paradigm shift in how people can participate and continue to be productive when they're on the road. Emile Ogier: Absolutely. And what I'm seeing, I'm seeing tons and tons of developers and managers writing on X saying, we've been discussing saying this is the of orchestration, but it ⁓ takes while without you running on the ground and actually using the tool and seeing how it changes your day-to-day workflow, how it essentially is a real paradigm shift. And they saying essentially, the skills used to have is not what they ultimately are. should be good at anymore. They really are becoming orchestrators, they're directors and they're letting the agents do the work and it's more about how well they speak to the agents and how clear their thinking is rather than doing the actual task work at hand. that is quite a, you know, it's a huge fundamental shift to to terms with. Ralph Behnke: Yeah, and it's obviously also really shone a very interesting light on this concept of memory. memory is important because ⁓ especially you're working across your phone and your desktop and maybe even have different places that you're using Claude, you've got Claude code, you've got Claude you've got Claude in the browser. each one of these you can run in parallel at the same time, know, so maybe you can be interacting with it in four different places. And of ⁓ these tools to provide ⁓ a common across all of those form factors and use cases. And ⁓ I this what starts to make it so powerful because It does that over a period of time. the of setting context starts to melt away for a lot of people as ⁓ it's building up its own internal knowledge base. And it does in a very simple way with these MD files. It just writes memories to these files. I think it's incredibly in the way that you've got the you know, the various levels of Claude files where you're giving it system instructions, project instructions, all the way down to even instructions for sub agents. So the way that you can actually explain to it what you wanted to do at each of those levels really... does two things. It means you're very specific, but it also saves a massive amount of token usage and improves performance because you have the information to guide Claude in way that you want it to process at the point of use. So you can be very sophisticated. And if are methodical and deliberate about the setup, I think people can get a really good result. with the agent capabilities for everyday work and not be terrified and also more importantly not be completely technologically overwhelmed. But I will say, I will come back to the barriers are real for most people that are not participating in these tools. Everything feels alien. The questions that they can ask are the biggest barrier because they can't even imagine what they could ask. And so think there is the big learning curve and that was my big learning moment as were training, even making really practical and relatable to the task that people were doing. You know, there is still a barrier to even knowing that can ask for those things. ⁓ For example, give me brand guidelines for my client based on their website. This is something you can ask. It will generate a very long document of every font, every color, well as potentially even the brand voice if you ask it nicely. Emile Ogier: Yeah, to echo your point, Rolf, I know this whole AI explosion is incredibly intimidating. It's moving fast and just note to all our listeners is, you know, if you're feeling ⁓ you're not alone. We are too, you all the leading experts, including Caparte is that they feeling this pressure, but need to remember that the LLMs, they're infinitely patient and we don't need those special to networks to have access to ⁓ people domains with and knowledge. We can just ask the model and it can... it be infinitely patient with us until we understand it. So sword cuts both ways, know, the access is there for you. So just take a deep breath and work through it and be confident that, you know, you'll find what you're looking for. Ralph Behnke: think that is probably the single most important piece of advice and you really don't need another guru, another course. ⁓ helps to be in community is no doubt because you get some inspiration of what to ask, what's possible, what can I really ask for, but everything that you wish for ⁓ could definitely get some way down the road with these tools. Emile Ogier: Yeah. Ralph Behnke: and they are completely self-sufficient. So I think that is probably the biggest gift AI and it's become so capable that it really does blow me away. Emile Ogier: Yeah, so just on topic of what we've just discussed, ⁓ I ⁓ just want to add ⁓ a word of warning to keep things Andrei just explained the scariest happening in software right now. it ⁓ reminds me of the term, great power comes great responsibility. And I think the is quite apt. So which was the poisoning of a Python package that gets 97 million downloads a month, which is really terrible and bad enough as it is. But it's actually far worse than that. The contagion spreads to any project that depends on light LLM in this example, meaning a simple PIP install was enough to exfiltrate SSH keys, AW, SGCP Azure credentials, Kubernetes configs, KIP credentials, environmental variables, all your API keys, your shell history, your crypto wallets, SSL private keys, CICD secrets and database passwords. So it's significant people. And here's the part that should terrify every developer live. The attack was only discovered because the attacker wrote sloppy code. The malware used so much RAM that it crashed someone's computer. If the attacker had been better coding, no one would have noticed for weeks. One developer using Cursor with the MCP plugin had outlights LLM pulled in as a dependency that they didn't even know about. And then the machine crashed and that crash saved thousands of companies from getting the infrastructure stolen. It's a bit of a wake up call. Every time you install any package, you are trusting every single dependency in its tree. and any one of them could be poisoned. Vibe coding saved us this time. The attacker vibe coded the attack and it was too sloppy to work quietly. Next time they may not make the same mistake. So Karpathy's takeaway is that classical software engineering would have you believe that dependencies are a good thing rebuilding pyramids from bricks. But in his opinion, this needs to be reevaluated. And it's wise growing a verse to them to use LLMs to yoink functionality when it's simple enough and possible. So just a good reminder to us all to be careful, but enjoy the power. Ralph Behnke: Yeah, this was a brilliant story. I understand it, the guy that had written this original piece of code was maintaining it for something like 20 years. And then became somewhat mentally and wasn't able to keep patching it. If we're talking about the same story, he wasn't able to keep patching this thing. so somebody offered to ⁓ do maintenance on the code for him and it a bit of a funny name something like Jian Tang can't remember the exact name but they ended up being ⁓ you know essentially a code and as you say they had written this code which was then exposed to be malware the way they infiltrated this single of ⁓ scaffolding like, you know, internet computing is built on and nearly actually inadvertently. Well, I mean, they engineered how to get access to, as you say, the corporate digital keys. this was, know, yeah, like. like LLM as far as I know. Emile Ogier: I didn't know that part of the story. I'm certainly going to look into it. Wow. What a, what a romp of a story. Ralph Behnke: Yeah, if this is the but it sounds like it's the same story, Veritasium is a wonderful channel and ⁓ he a deep dive on this particular story. But I guess the point was that the internet or internet technologies are littered ⁓ by one man shows that have built a small piece of code that all of it runs on. And here, you know, somebody actually went in, this is a foundational unit, and they were able to essentially, you know, grab all of these details. Yeah, it's a beautiful story. Emile Ogier: And while you're doing that, just the thought from Elon Musk, something to ponder over, which really got my head spinning is he's essentially predicting that AI will run out of things to make for humans. Ultimately proposing AI will be able to produce so many things that fulfilling all human needs will get saturated. in his words, AI will literally run out of things to do for us. I just thought that was quite mind blowing stuff. Hopefully he's wrong. Hopefully, you know, as we capable of doing new and more things, we'll just have new interests that never ends, but who knows. Ralph Behnke: Well, on that note, I'm sure I need some stuff. I wish I had an AI that was making unlimited amounts of it outside. But there isn't. So I think we've got a bit more time to play. And no doubt, you know, this tsunami is going to be going around the world five times before we even realize what's really going on. Because it feels like that's what's happening right now. But maybe the embodied AI You know that will be the game changer for everybody if it's all making stuff what I can see is enormous amount of spend on robotics in China I can see that will probably have an of robot ⁓ arms can on nearly any production line but obviously we also seeing these kind of robots that will no doubt roam around the home and they look very much like us. We already have them with wheels vacuuming our house, our lawns, so the future is well and truly here. Emile Ogier: Indeed. Ralph Behnke: So that's a great show. Should we wrap up there? Going once, going twice. Thank you for your time Emil and going for the third time. We'll see you on the edge. Emile Ogier: Sounds good Rob, thanks for your time. Cheers everyone, see you next week.