Vollständiges Transkript von If AGI Is Here, Why Can’t Companies Use It? And Chatbots are a sideshow

you sound more intelligent than you are. Ralph: Well that's very welcome for me. I love that. but before we move off the the news and I'll hand over to you for your first story. I'm just actually having a quick look at the dev day announcements. I think you're gonna be covering dots. so a whole new way to work with AI. I think this is quite interesting. Everybody getting on the bandwagon of personal assistance and this is OpenAI's contribution to that, so I won't say more about that. you've also got new tools for developers with codecs in the cloud. I think this is really smart because one of the things that happens is when you close your laptop and you are vibe coding or you're building apps with these various tools, obviously everything stops running and that's a bit of a problem. So codex in the cloud seems like it's gonna be able to carry on working. And then they've got more customizations in Chat GPT with plugins. I love this because we've been talking about it over the the the last eight months or so. Plugins create the ecosystem. So if you have a Canva plugin, for example, that works with ChatGPT, of course you can do all kinds of amazing things. And so plugins with Shopify. the latest one I've actually been playing with is something called Hyperframes, which is a little tool that HeyGen actually released. an open source licensing so you can use it for free and it builds videos. So I'll probably do a little bit of a deep dive on that one. but yeah, really interesting to see what they're doing with extending that whole plugin marketplace. I think it just makes the LLMs so much more capable. improving how people and AI work together. So this is something called Chat GPT space. so it looks like It is focusing on collaborative slides. Obviously, we've had tools like gamma.ai, and I used Gen Spark certainly last year quite extensively for my slideware, and then I really started to move away from that to more HTML type presentations. but of course, we have seen that each of the Frontier Labs have tried to take on the space of being able to build presentations. And that is what Chat GPT space is all about. So go forth and collaborate on your slide where and then they've got something with a a Pro 500 Chat GPT subscription that's a new tier for those who want even more usage and speed. and then there is also the open AI marketplace for eligible customers where they can apply part of their existing open AI commitment towards a approved partner software. So those are just some of the highlights that I kind of came across I actually used Google for the first time in ages to to come up with that list. yeah definitely go and have a look. It's on openai.com forward slash news and there they've got a whole list of the 20 items or so That they have actually launched. And you can even find a little overview of Sam Altman's remarks at the United Nations Security Council. so thank you, Mr. President and your excellencies, is what he said. I don't know. I mean it was the UN Security Council, so presumably there was one there was more than one president there, but apparently he was only talking to one them. let's just move on to a a related story here because I don't know if you came across this on X. I know you spend a lot of time on X, but Sequoia Capital is often in the news because they are obviously funding many of these AI startups. And one of their partners in Sequoia Capital, a guy called Pat Grady, he was preparing for a a speech. I guess he's opening you know he was gonna go and talk at some commencement speech thing at a at a university like they will do. and he actually posted his practice run on X and so people were kind of commenting on it. But I think it's fascinating. So one of the things he said is AGI has arrived. And so he actually pitched this to the internet November 2025. and people were like, mm, don't really think that's true. but as the kind of twenty twenty-six has moved on, I think pretty much there's now consensus that we have actually achieved AGI. And so this is artificial general intelligence. why have we achieved this? Well, you know, AI has convincingly beat the Turing test and several tests since then, and it is now definitely more capable than a hundred capable experts that you might have in a room. And so there's no one single definition of AGI, but I think we all pretty much agree we we passed that threshold. in its various forms. and we are now really moving towards some kind of super general intelligence, which is the the next benchmark we we've set ourselves. So he basically said that the three pivotal inflection points that defined the technology's rapid trajectory, obviously November twenty two, ChatGPT launched, and that was the peak of the pre training era. Late 2024, OpenAI's O one model launched, introducing advanced reasoning. And autumn twenty twenty five, Anthropics Claude Code and Opus four point five arrived and they mastered Long Horizon agents. And I think we can both agree that Opus five point five has taken that to the next level, as has Soul Six. And so what he's then said is he highlighted why this technology Technological shift dwarfs previous cycles. So massive total addressable market. So AI targets the service market, which is roughly two orders the magnitude larger than the software market. And so exactly what these personal assistants are addressing, the service market, like what you were saying, you know, I'm looking for a handyman. this is what the AI industry is addressing. And you a hundred times larger than the software market. That is incredible when you look at the likes of Oracle and SAP, Infor and some of the other really, really big software companies on the planet, you know, you kind of go, wow, that is actually quite staggering. what you also have is instant distribution. So unlike previous revolutions, you know, we now actually have all of the global infrastructure. So what I was just saying there, how did instinct

Actually, get to scale so quickly. Well, we have the distribution. You just need a good set of influences on YouTube and Instagram, and you can get to a hundred million in a matter of days, as certain you know of these Sequoia companies have actually done. The next one that he's saying is a big difference in terms of this technological revolution is the computational revolution. So This is the the first fundamental shift in the nature of computation, not communication, since the silicon era of the 60s and the 70s. And obviously, a lot of this is just down to the capability of the hardware. so we are seeing the training chips, the inference chips, and the data center build out. Those are kind of the inf infrastructure components of this AI revolution. and then he also calls out the diffusion gap, is where startups get rich. So introduces a critical concept for founders, the diffusion gap. This is the massive space between what the latest models are technically technically capable of and what enterprise actually uses them for in practice. So hence my question: what can enterprise actually do with this new wave of personal assistance? I can see it being adopted by independence. by small business owners. I just don't see it making the inroads at at the same pace in in enterprise. But, you know, that is also starting to change. his fourth point Em: Ralph I could also Ralph: and Em: flip that around is maybe they might not need to adopt it, but they do need to respond to it. Ralph: Well, I think I think the actually I'm gonna g I think there's a really strong link to that. And one of the things that Pat Grady has actually said is you know there are some very specific industries, or at least there's kind of four specific areas that sequoia growth you know, that that they're basically supporting growth in. And so They're obviously an investment firm, so they want to back the trends, right? They want to make the trends and be where the market is is gonna be. And so what absolutely relates to what you just said is this n he's saying the primary opportunity areas for founders are high value knowledge work, new enterprise systems of record, and then frontier technology focused on AI for science and domain specific neo labs. Now I think the one that really speaks to what you're saying is new enterprise systems of record. So much like I was just saying, I'm I'm seeing a revolution in what the website is for and what it's capable of, your accounting software has a revolution. Your inventory management system has a revolution. your master data is undergoing a revolution. And each of these systems now essentially has an embedded AI. that is able to predict the data that it requires, that is able to pull that data together, that is able to create the content, that is able to test that that data is actually complete. And we know that, you know, so many of the issues that enterprise face is that data is often wrong, data is incomplete, it's often not available at the moment that you need it, and it causes havoc with the ability of these LLMs to actually execute their task. And so I look at that and I go, wow, I could see how enterprise can actually leapfrog all of these AI systems. however, I also know how hard it is to replace that infrastructure. I spent many years working with companies like Tesco, for example, architecting their international supply chain. And I can tell you, you know, one of those systems that we were looking at taking out ended up taking 13 years to replace after three failed attempts at trying to move it, which themselves were multi-year programs, multi-million pound programs in their own right. And so, yes, new enterprise systems of record, I could see that, you know. invalidating the need for enterprise to move to these agents. But I don't want to underestimate the task and the complexity required to actually bring in, these new age systems. But of course, SAP and Oracle and Microsoft will be working on that. And no doubt you know, that's gonna be a rebirth for enterprise software, which hopefully will make it a bit more interesting as well. Em: Rolf. you know, we we come from shared industries. the layers of plumbing and complexity underneath it is, it's a really tall order. There's no easy way of getting around that. the uncomfortable truth is we are gonna have to reimagine and rebuild from the ground up And it's why there's individuals that are twenty three years old raising ten billion series rounds because they are reimagining and rebuilding things from the ground up and it's these one man bands that are gonna be the dangerous new players on the field that enterprise have to be careful and considered with. Ralph: Well, and of course we've just seen this with Jev just over the last few weeks. Jev, this new technology that essentially is able to compute and make decisions at a fraction of the speed and cost when you compare it to a traditional LM. And it's essentially it's like shaking one of those decision cubes. You know, when got that thing, should I go to dinner? And then you shake it and it says yes or maybe or no. I like that's Em: April. Ralph: kind of I I I have that exact In my mind, when I I when people ask me, like, so what does Jeff do? and it kind of like seemingly came out of nowhere, right? And actually, there have been quite a lot of technologies that do something similar, but somehow Jev really has caught everybody's attention and it's just scaled. And of course, you know, their marketing has been free because the news cycle has just literally picked it up. So I I think this is the moment if there ever was a you know. the opportunity to be a shooting star in AI, I think it is now. but you just have to have that that right idea at the right time that you know, kind of picks everybody's well let's let's say piques everybody's curiosity but ultimately you know gives people the promise of solving a problem and you know I I do want to I do want have some hope that

maybe there is a faster way of doing all of this. one you know, one of my favorite AI companies is Thinking Machines. There's hardly a podcast that goes by that I don't call out Inkling. And the reason why I think Inkling is such a fascinating model is because it really is the promise of training your own AI. And and I also think that this is kind of where maybe we actually can sidestep all of the complexity that enterprises built up over the years by building you know what we've called the digital twin for some time but I think people have looked at that and gone well what does it actually mean is it really a twin but I really do see that now we have the potential to actually really embody the organization and what it's trying to achieve. not each of the individual process steps, but its goal and implement the pattern of how it can achieve that goal with AI in dedicated AI systems like Inkling. obviously we have to go away and build it. That in itself is not it's not an easy task, but it's certainly not as complex as trying to re-architect the spaghetti that tends to be, enterprise ecosystems. And so I don't know what you make of that, for me it's like, let's not try and wire these things in. Let's just build it fresh. and I just I love that idea because I know it's just it's just treacle moving through enterprise IT. So yeah, I just think let's just build digital twins of everything. Em: and it's not fun picking up all the enterprise baggage that's twenty decades old. it's fun playing with new building blocks. But in answer to your question about Jeff, why Jeff, we actually have had classifiers before. it's been staring us right in the face. It's called machine learning. But the difference with that is you needed huge data sets, it was complex to set up, you needed a lot funding for infrastructure. You needed to understand machine learning, which is, you know, not for everyone. It does get quite complex. when you need have huge sets of your own proprietary data and whatever your business was. you had to build the classifier, you had to fine-tune it, and when you got it working, it was an absolute dream. But as new business products come in or new business information comes in. You have to retune those classifiers. you have to rerun the models. It all requires a lot of resource, a lot of effort, a lot of time, a lot of infrastructure, and a lot of spend. So, yes, they valuable, but it requires quite a business mode to take full advantage of them. And now we have Jeb coming in, which is essentially a general purpose classifier, which you can use for pennies to get your answers. and your workflows and your decisions done for you. It is going to completely revolutionize the industry as far as I'm concerned. And it is the next primitive fundamental architectural building block that we've been missing. Ralph: Yeah, I can't like it just it's just amazing. I c I can't even believe that we we missed it for so long and suddenly there it was and then it was like, that makes complete sense. So it really is kind of one of those moments where you realize you know, you've you've just had an arm missing, but then somebody gave you an arm and and and you started using it. And you're kinda like, How did I not actually get that? But one of the things that I think starts to r really bring all of this to l let's just say kind of practical and maybe even faster evolution. So well one of the the terms that Pat Grady used in his delivery there was this concept of the neolab. And and I and I think where this kind of relates to what you're saying, With this concept of the c the the general purpose classifier, there are hundreds of thousands that every business needs to make every day. And so that's why the classifier is this this fundamental myth. How did we not have an engine that could make decisions? So where this becomes particularly powerful is with this concept of the Neo Lab. So the Neo Lab is especially essentially, if I just kind of read his definition, it's an AI startup operating like an independent research lab. But targeting a single specialized industry or a unique architectural approach instead of general purpose AI. And this is probably one of the kind of biggest opportunities in terms of arbitrage because you've got the capability of the LLM, but it's general purpose, right? It doesn't understand what decisions you actually need to make to make your business successful. Of course, it has a pretty good idea, but you can't run a successful business on a pretty good idea. You actually have to have. Absolute faith that is going to make the the right calls and the right decisions based on what makes your business successful. And this is really where I see the opportunity to kind of build vertically optimized solutions where you take those classifiers, you take the LLMs, and you really solve a problem for a hairdresser, for a car mechanic. For, any the biggest insurance company on the planet. It doesn't really matter what you pick, but I think really owning that vertical, going deep into it and starting to use these tools like Jev as a general purpose classifier, and applying this Neo Lab concept to solve very specific problems and using tools like Inkling to actually build an LLM brain that is very capable of generative AI and But you know, is able to be really specific in terms of the problem that it solved. And I think that is going to be the key. General chat is kind of turning out to be an interesting sideshow. It's gonna be a footnote in the journey of LLMs. I don't even know the last time I've used general chat in a kind of the normal way that I was just a year ago. it's completely shifted for me. I don't know, doing Google search. I I stopped doing Google searches for quite a long time. I've started to come back to them because they've now got kind of like embedded AI. That's probably the closest I get to a general use of chat. Everything else is hyper specific, kind of wrapped up in some

type of tool. And and and I think that's just the promise for for enterprise as well. Em: what you calling chat, like the likes of Google, I see that utility shrinking, but chats in terms of conversation is increasing. because I really thought what we're using now to do our development work, that effective chat bot with turns was a a clunky surface that we were only going to see in this first phase of AI And I really thought the surface would change, but it seems like the big labs are recognizing how accustomed and useful we are as human beings in having conversations and having stories continue. And what they're effectively doing is they're keeping the surface But the surface is utility is changing. So the conversation remains and the the agents are coming into the conversation underneath it during the work. Ralph: That's it. I think it is moving from a an interrogation, a chat, to a conversation, Em: Collaboration. Ralph: but you know, it's almost like I don't know what you call a a lifetime long conversation. I think this is really the thing. and maybe we haven't even got a word for that because Em: monothreads. Ralph: monotreads. Well, well there you go. So so that is the the difference. And of course, it becomes heavily dependent on the context that you're operating in. the goal awareness. So it's not just the actual context, but the and and maybe goal awareness is part of context, but I think it's a very kind of specific aspect of of context awareness. And you know, I started to kind of feel my way through this a little bit. with the Bleeding Edge Podcast website because one of the things the SEO agent came back to me and said was, Hey, what is your goal for the website and for YouTube and Spotify? And I thought to myself, well this is really fascinating because the context awareness in terms of the goal is a lot more interesting than the context awareness of where we are right now. Because I don't really care where I am right now. I really care about where we're going as a show. And and so you know, maybe we have different types of context. I'm sure there's a a a very formal word for that, but I like that term of mono thread. I know what's gonna happen with that one, it's going to be going immediately onto one of our explainers. so from the transcript and it'll be a great evidence of automation at work. Em: monothreads were essentially the worst way of working in 2024, 25, when we were furiously getting excited with AI. And what it essentially means is growing out your context window in a single thread, to the extent that your AI started hallucinating was never a good idea. So we started packaging up those conversations, creating documents and so we could hand it over to the next session. But the labs are getting very intelligent now. We've got compaction, it knows what it's doing, and there's administrative and supervisor agents working in the background and they doing all of that heavy lifting now. what was a bad habit is now becoming a good habit don't worry about creating new threads. Work within one mono thread and let the agents spin out separate branches and threads as they need them. And you can dip into those and come back to the main conversation as needed. That's essentially the idea behind it. Ralph: Emile, thanks for a a great show. Thanks for a great contribution. You said you weren't prepared, so Em: I appreciate that and thanks to all our listeners and please like and subscribe. Take care. Ralph: See you in the edge.