← This Day in AI Podcast
Loading summary
Transcript198 lines
- [00:03]
Michael
So, Chris, this is one of those weeks where I would generally tell people, tune out, there's nothing to see here. I've got a lot of criticism about this at the start of the show, but in fact, it might seem like a quiet week in the world of AI, but indeed, I don't actually think it is. I think it could be one of the most meaningful weeks we've ever had in terms of strategic signals from some of the bigger AI businesses out there. One of those being the bigliest OpenAI, who had what felt like a random sort of town hall State of the Union event. It was none other than Sam Altman and the guy who kind of looks like Elon Musk. He's just. He looks exactly like Elon Musk and he's.
- [00:50]
Chris
And he's foreign, so, you know, he's smart. I think they always need someone with a, a non American accent in there to make them smarter.
- [00:57]
Michael
Yeah, it's like this is chief gaggle nerd here that's come out to sort of validate what I'm going to say to you. And really they said at the start of this YouTube presentation, Sam Altman came out and said at the start that it was really about the, the structural updates to the business because they've made some structural changes where Microsoft is now the single largest shareholder. But they didn't really mention that in the, in the update. And what interested me about this most of all is the fact that he sort of said they're gonna pivot like they're changing their entire strategy. And it, it hasn't got a lot of pickup. Like no one's really talked about how meaningful this might actually be. I'll play it two little excerpts and then I want to tell you about some other news that happened this week that is interesting as well.
- [01:49]
Sam Altman
There was a time earlier on in OpenAI where we thought that AGI would be sort of this oracular thing in the sky and it would make all these wonderful things for us. And we now have a sharper view of that, which is we want to create tools and then we want people to use them to create the future. We want to empower people with AI as much as possible and then trust that the process that has been working for human history of people building better and better things with newer and better tools will continue to go on. We can now see a vision where we help build a personal AGI that people can use anywhere with all of these different tools, access to all these different services and systems to help with work and personal Life and their personal life. And as AI gets better and better, as AI can even do things like discover or help discover new science, what people will be able to create with that, to make all of society better and their own lives more fulfilled, we think should be quite incredible. There are three core pillars we think about for OpenAI research, product and infrastructure. We have to succeed at the research required to build AGI. We have to build a platform that makes it easy and powerful to use. And we have to build enough infrastructure such that people can use at a low cost all of this amazing AI that they'd like. Here's a little cartoon of how we think about our world. So at the bottom layer here we have chips, racks and the systems around them, the data centers that these go into in the energy. We'll talk more about the first three today in Energy another time. Then we train models on top of these. Then we have an OpenAI account on top of that, we have a browser now called Atlas. And we have devices coming in the next few years that you'll be able to take AI with you every.
- [03:26]
Michael
All right, that's enough of the presentation. But essentially what was said is that they're moving from this world of them building first party applications on top of the the chips, the racks, the data center, the train models and the open.
- [03:40]
Chris
Useless web browser that nobody needs.
- [03:41]
Michael
Yeah. And so they were building these applications on top and now this huge puzzle piece is going to be the third party applications. And really what it signaled to me is they're saying, hey, we don't want to be out there competing at the application level anymore. We just want all the applications on top of us so we can have the platform, the ecosystem and essentially be like an AI AGI cloud provider thingy. And so, you know, nothing like, I don't think it's like groundbreaking to come out and say this. Of course they want to be a platform and be like the Apple Tax model on top of everything out there in existence. But what I did find interesting about this is Cursor came out during the week and announced cursor 2.0. And I'll unpack a little bit of that in a minute. But why I think it contrasts interestingly to this State of the Union is, is the Cursor announcement said we're announcing our first model called Composer, a proprietary model that from all accounts cursors trained potentially on users interacting with their agent to be a faster and more efficient agent and obviously to eventually make their cogs, their cost of goods sold go down because Right now they're just paying OpenAI and anthropic basically and don't really have a, a fundamental business. But what it says to me is if you build an application that gets really popular and people do a particular workflow on it, it's now becoming possible if you hire the right people to train a model that can do the things that that application is well known for. And so I think that this could be an existential threat to OpenAI because what it says is maybe there's really no mode, you know, in the next couple of years with these models, like you can just go build an app, build it on the frontier models and then learn how people use it and then fine tune for those use cases and improve it based on the very specific workflow or industry niche that you work in. So anyway, there's a lot to unpack here. I just wanted to set it up so everyone's in the loop. But I think this is, this is bigger than we can understand right now.
- [05:51]
Chris
We know, we've known for a while that a lot, large amounts of the demand of these big providers like Anthropic and OpenAI do come from these applications that are incredibly heavy users of it. And I said from the very beginning, OpenAI is not a software company. They're great with the research and the models and that's what really led the way. But if you look at their history of producing software, it's not very good. Like, GPTs was a disaster. Their interface is okay and I know everybody's cloned it, but it's not exactly like setting world records in terms of like usability and that kind of stuff. Really what they did was came out with the best models and kept them there for the longest time. And then people built on that and they had a good brand. So people would use their stuff, but they've never really come out with something that's blown everyone's socks off and made it their, like go to system. I think people de facto use Chat GPT because they know about it, but the people who discover other platforms would prefer to use them. And I actually think them leaning into it is smart be the best and cheapest models out there and just have everyone dependent on you for that. It's a much better business model in terms of, like he says, selling chips and electricity and then the value add, which is their models, rather than trying to compete with every single business that comes along.
- [07:07]
Michael
And I think that's been the challenge. Right. Like you look at ChatGPT Atlas, where I think I Think a lot of people in the know that have used AI browsers are saying that perplexity is comet. And then, dear, the other browsers are just better. Like, they're better at being AI browsers. And look, I'm sure those teams are capable of refining that and making Atlas as a browser better over time. But this is what it feels like. Like they go and build these products and they're half baked. And there's other specialist companies that are just dedicated to this problem. And you know, in your point, your.
- [07:42]
Chris
Point about Cursor is a very interesting one because if you have a business where the primary input cost is, you're transferring a huge amount of the customer spend directly to another company. And it's also something that can really blow out and make an individual customer unprofitable. So let's say someone's paying you a fixed amount of monthly money and you've got to spend a certain amount of that on token usage with OpenAI or anthropic or whoever, then your margin is thin because, like, you've got to give that person value. People generally know what the tokens are worth, so you can't have like a 90% markup on the tokens or they're just going to use another platform. But if you, as cursor can say, look, I can give you 80%, 90% of the functionality, but it's going to cost you half the price. Right. People will take that and their margins will go up in the process. So it totally makes sense for them ultimately to run their own specialized models for this and just bypass the big guys. Like, why would you use them?
- [08:47]
Michael
But then that gets me to the case of why would you build outside of the infrastructure, like the cloud infrastructure on OpenAI stack. Right. Like, they talk about chips, it's like, so what you bought the most Nvidia chips, like, that's really it. Right. And then you built some racks and data centers. Sure. But it's like AWS can do that, Azure can do that. So what's unique there apart from capital, like, they just can raise the most money. That might be one of the things.
- [09:17]
Chris
Yeah. I think when he talks these big numbers, like so many trillion dollars in agreed spend and all that sort of stuff, he's really probably talking to the investors and trying to get more money invested rather than customers. Customers don't really care as long as it works. Right. Like, it doesn't really matter how big their data center is as long as your requests get through.
- [09:36]
Michael
But I think honing in on this Point. Right. So it's like, where's the, you know, where's the platform mode? This is what I don't really understand here because you've got chips, racks, data centers, sure, but anyone can do those things. And the models, I mean, you can run their models on Azure today, unless they're going to restrict their models to their own platform eventually, which I highly doubt. So that's not that relevant either. Also, there's equivalent models. Now some people might prefer, say a GPT5, but I know a lot in the sort of commercial sector or enterprise prefer anthropics models because they're just better at instruction following and tool use. And so there's no real, I don't think there's a real mode around the model anymore.
- [10:25]
Chris
And well, not only is there not a moat, I think there's value in switching and I think we've seen it obviously because our product does that as its main thing. However, even outside of that, I think people like the model flexibility. They want to be able to try whatever the latest one is in their platform of choice. They don't want to go, oh, we're pure open AI and will always be that way. And I think that people don't want to miss out on the ability to use different models. So why would you voluntarily tie yourself into them? There's no reason.
- [10:57]
Michael
This is the thing I frequently am talking to businesses and enterprises and the same thing is said to me over and over again is that their teams maybe have like multiple subscriptions and try different models and they, they're not stupid. Like they know, like, oh, one model's better at this, another model's better at that. And so the idea, especially in such a state of flux or in such a state of chaos of locking into a single model, I don't really see the point. Right, so then the next layer up, OpenAI account, sure, they've got distribution and you can't underestimate that. Like everyone has an account but only a small portion of that total addressable market for them is paying. And sure, like they're making a lot of money, let's not gloss over that. And they have, you know, they're in the zeitgeist. Like every, everyone thinks of AI as ChatGPT. So I think that's their biggest asset.
- [11:51]
Chris
That's not what we're talking about here. We're talking about what Altman is saying. Their strategy is in the long term. And he more or less said, oh well, originally we were going to make AGI that would control the whole world, so don't bother building anything, guys. And now he's like, oh, actually we're not going to do that. We're going to let everyone have their own personal AGI, which kind of makes no sense. It's either AGI or it isn't. You don't like go, oh my AGI is slightly different to yours or something like that.
- [12:19]
Michael
Also famously said in India, remember that, like, you can compete with us but you'll lose. Like basically just mock them. Like mocked all of India saying, oh, you'll never be able to compete or build a model as good as ours. And like I would argue they possibly could in the future. Now all the tooling's out there so that that entire moat and that entire argument is just evaporated into thin air. And so it's like, what do they really have that's of value left? And it's just that the eyeballs on ChatGPT.
- [12:50]
Chris
Yeah, I agree. I think brand recognition and just the fact that so many people have accounts and use it and that if AI is mentioned in, in media, it is chatgpt that is mentioned. It's not even LLMs or a more general term. It's always chat GPT. I think that's what they have and they need to capitalize on that while they can and lock people in. I just don't think that they really have anything that can lock people in.
- [13:13]
Michael
And this is the challenge, right. As a developer and I'm not sure what's coming, they talk about an enterprise product, but I would suspect the only way of really getting enterprises to adopt something like this is getting, you know, well, first of all, having the security and just trust factor, which I think they lack as a company to be quite well.
- [13:33]
Chris
And they don't even have regional deployments. Like if they, if they were just like, we're going to build a data center in every major jurisdiction, you know, for international business around the world, and it's going to have the best out of the box security and encryption, storage, all that sort of stuff. And you can do private deployments within those data centers. That would be a way to get, to win the enterprise. Right? Because you're like, okay, we can do this. But the truth is now the only people doing that are Google, Microsoft and Amazon. They can do that to some degree. Right, but it's similar to regular cloud computing the way you would have done it in the past. OpenAI simply can't compete on that front.
- [14:13]
Michael
But let's, let's go to the things we know to be true today. Right. Like if you think about, I think the biggest unlock for enterprise and business today with AI is mcp like a hundred percent. Like I think it's the biggest unlock because you can securely connect to data wherever it is. You can unlock the value of having all of that data considered when making decisions where in the past you just wouldn't have had the time or the ability to go and do it. You also can unlock the value from what was once a problem which was all these disparate systems and data stores and instead of these like pie in the sky projects, like one day we'll have it all in a beautiful data center and perfectly coupled together. It doesn't matter anymore because you just with a series of MCPs, link it all together, teach your business. Like put a sort of business workflow as the tools on top of the MCP layer and now you can just query data and make decisions across the. Org and then test those decisions across various models. So to me that is probably the biggest unlock. And then you think about third party applications and where all the data is stored and those companies that would need to unlock it to be out of the box. Right?
- [15:31]
Chris
Yeah.
- [15:31]
Michael
And are they going to do it? Are they stupid enough to do this? That's the real question. Because why would you unlock it and make it so simple and build on top of this platform when you essentially, especially something like Oracle or Salesforce because you just ceding control to, to open air, you're like, you're like, here's our customer base, we offer no value.
- [15:52]
Chris
Yeah, I think that's the thing. I do think that is where all the value is like who can provide a platform that allows enterprises to do what you just described securely. I think that is the key is like how do you get it so the staff can access all of the data and use it to leverage their decision making or take actions or pipe it into other systems through MCP as a platform. Because I think that is, that is actually the crucial part of it that actually adds all the value. And I think even with lesser models you're getting more value by working in that way than you are just being like, well we have one that's closer to AGI by 4% than the other model, but it's hard to use tools with it and it's not.
- [16:37]
Michael
Yeah, I just feel like a lot of the things they're now saying everyone is a bit tired of like these claims about oh don't worry, we're also going to invest compute into solving Cancer and all these other things like it, it's just, it, it's actions speak louder than words. Like you're releasing a browser, you're releasing some doom scrolling video service.
- [16:57]
Chris
Where's.
- [16:58]
Michael
Where's everything else. Whereas I look at anthropic and it's like they're experimenting around with skills, a better way to package up like essentially little applications that the AI has access to. So it's better at instruction following, let's be honest. I mean that's basically what it is. And they're thinking through the problems in the workplace and the problems people have with AI today and trying to make their models better around that and the actions there. If, if the goal is sort of an enterprise platform or these third party applications far outweigh the actions of OpenAI today.
- [17:32]
Chris
That's right. Like you can't talk a big game around enterprise and then take actions that are completely opposing that. Like a browser doesn't help with that. Definitely making TikTok style videos is sort of an antithesis to what you're trying to portray now one week later or whatever it is as your, as your company's main mission. It's just, yeah, it's just all talk and it doesn't really add up. And then the whole like solving like helping people solve diseases and stuff and like a large language model is really the thing that are going to solve the diseases? I wouldn't think so. You know, like even though I have a lot of faith in them, their ability to create new knowledge from existing data is limited. And I feel like a lot of the scientific discoveries are like maths and chemistry based. It's not really something a large language model is great at. I mean I could be wrong, I'm not a scientist.
- [18:24]
Michael
Well, every, every time they've claimed a scientific breakthrough or some novel thing, remember when ChatGPT5 came out, someone claimed some. It solves some novel math problem that was completely discredited like two weeks ago saying, oh no, it existed in the training data, here's the solution, this is completely wrong. And then Greg Brockman and others went and deleted all of the tweets related to this discovery. And I remember reading it at the time thinking, wow, this is a huge moment if this is true. But of course it was not. And so I think, and I have.
- [18:57]
Chris
No doubt that it can help in the science fields, like for example, packaging things into a report, visualizing the data, you know, organizing things, piping things through MCP calls to get things done more efficiently. But it isn't going to on its own solve diseases. It isn't going to be like, okay, one day the people trying to figure out the solution to cancer type in how do I solve cancer? And then one day it just has the answer. That's not really how it works. And he's sort of purporting as if, if they get enough electrons in their data centers that it will.
- [19:33]
Michael
Yeah, I, this whole idea of like, oh, don't Worry Guys, by 20, March of 2028, this is like literally what he said. By March of 2028, they expect to have a meaningful fully automated AI research. Like what A weird specific would be.
- [19:49]
Chris
Bold though to say by March 2028 we will solve cancer with our model. Like, yeah, that'd be pretty cool. It's just like, yeah, no, we're pretty sure it'll be done by then.
- [19:59]
Michael
But like anyone that's using these models today knows, especially researchers who use them all the time, like, of course there's benefits, that's why they're using them. But it's not, yeah, it's not coming up with novel ideas like you say, it's just helping them make connections and making them more efficient in their research.
- [20:19]
Chris
Yeah, yeah, exactly. It's like it is individual productivity leverage based on your own expert knowledge. Like you, the expert, are guiding it through a process that you normally do. And I would argue what MCPS give you the ability to do is take steps that you otherwise might not because it would be too slow. Like for example, say I want to make a company decision. Do I want to take into account market sentiment about my company and our recent sales figures and our historical sales figures and all of these factors where I may not take that into account on a small decision on a day to day basis. Normally because the amount of work required to gather and organize that information is so high. But if my AI system using MCP can do that in say two minutes and then make the decision, then I might actually do it. So I've got more leverage, smarter decision making and things from this, this AI system. But the AI system isn't the one setting the goals. It isn't the one driving the narrative or driving what you're trying to accomplish. And I think that is where we want to get to with the next step. But to act like, okay, suddenly the model is going to just do that stuff is wrong because it's never going to be the model that does it. It's going to be the system around it that does it. Like the real innovation in my opinion is going to be around how we package up agentic behaviors and the environment in which we give an agent to operate so that it can succeed. The model improvements will help, of course, because it'll make less mistakes and be more capable of doing things like vision, for example. But the model itself is never going to suddenly just be like, I am AGI today now and I've done everything. Don't worry about all that stuff you built.
- [22:05]
Michael
Yeah. And so it, the other sort of going, just going back quickly to the, that platform piece. The other thing is into the video he says about two minutes in people are going to be able to build incredible services with AI, starting with our API, with apps and chat, GPT and our new enterprise platform. Right. But like this pitch is basically like we're gonna lock you into our ecosystem. Everyone's gonna use their OpenAI account. I just don't get what the pitch is to the companies and developers who are expecting.
- [22:42]
Chris
It's the people, right? Like you build on our proprietary store, you will make money. Imagine the influences and stuff you're gonna see on, on YouTube. Like, here's how I made $10,000 a day using my fucking, you know, GPT thing I built. And you know, everybody subscribes because it's got keywords in the title and whatever. It's going to be another sort of junk store. My argument is I don't think OpenAI has what it takes in software to accomplish that. I think Apple does and companies like that, but they're smart enough to know that it's not ready for that yet.
- [23:18]
Michael
Just, I know we're, we're hanging on this topic probably a bit too long, but, but like again, sure, the audience, right, but you gotta have a good product. Audiences aren't stupid. Like if you go there and the current proposition is like the hey chatgpt, make me a Spotify playlist and then the Spotify playlist comes up and I can only work with one app at a time, which basically foregoes all the benefits of MCP with context gathering and like action taking and asynchronous tool calling. That product's going to be bad. So there's got to be more to this. Like, are they going to have like basic CRUD interfaces like the Sass sort of create, destroy, edit, update stuff in, in chat?
- [24:04]
Chris
I mean, I reckon it's almost certainly going to be like a dynamic UI producing style platform. So it'll be tooling around. You can use their models to build something that's going to have dynamic input UI and dynamic output UI that is customized for the thing. And I think if that's what it is, it's totally misguided. That is not what people want. People want to be able to automate common processes at certain times on a regular basis, proactively, reactively, that kind of thing. And they want the AI to work within little frameworks where they're able to start to accomplish things automatically. Like the example I gave before, where it gives you great leverage in decision making and, and getting through your own workday. I think the next step is, okay, I'm commonly doing this task, but it requires some intelligence in order to get through the task. I couldn't just give it to like, let's say you have a junior employee. I couldn't just give it to them and say, go do this. You need to train them. Right. And I think these trained skills where an AI can accomplish the task that say a reg, like a trained employee could do on a regular basis when required, is what everyone wants to get to. Like, that is the next step in this journey. We're all on with the AI technology and I think anything that doesn't do that is a misstep by them. Like if they were the first ones to come out with, here's an agentic framework that is going to allow you to automate common tasks and train a system on doing that. That for me would be an advancement, but anything other than that I don't think is right.
- [25:41]
Michael
Yeah, I think the real benefit today, or the part that's lacking with all, all of this stuff, right, is just people figuring out reproducible workflows, skills, context, actions, compiling them together as like essentially AI apps that other people can utilize. Like I mentioned it last week or the week before on the show around, someone could totally build a, an MCP that fully replaces. I mean, you'd need like a companion app for sure, like somewhere where you set up your inbox and a few settings and stuff. But outside of that, I think today it's very capable, especially with like MCP UI as proposed by them. You could totally build like a full Zendesk replacement as an mcp. Right. And I think that is a good example of an application you could potentially build on a platform like this and get good distribution for. But I would still argue the downsides in a large enterprise are you still want the interface, you still want to be able to go and access the data, and there's still workflows and approvals and permissions and very specific things to different industries and markets, and you're going to then go, okay, well like I still need my core app is just a small interface into the AI model, but then I can deliver that similar to Cursor in a very specialist way in my own applications. With like, I just, I. There's no example in my mind unless someone's listening and, and thinks I'm incorrect. Where this is worked before, like, look at, in our era when we were in, in San Francisco, you had Facebook and you had that, that gaming company, I think it was Zigna Games or whatever, built a whole business like at one point valued at billions, based on making games on the Facebook sort of SDK like this. It's the exact same playbook, right? Like they. There was a bunch of like farmville and all those things on that Facebook ecosystem and that was dubbed as the future. Like build on this ecosystem. And what happened? Like it doesn't even exist anymore. And all of those companies, I mean they rose up and some did quite well, but then everything completely disappeared and it sort of goes against the whole open web as well.
- [28:05]
Chris
What strikes me as crazy though, is either OpenAI is not capable of making the most of their technology when it comes to the AI systems around it, or they think their audience is stupid and wouldn't get it in terms of what it's there for. If you look at the way Anthropic is packaging up skills, for example, that is a real meaningful step towards a better way of working with the AI. It may not hold, it may not last, but it's a step in that direction. Right? Like it's actually a useful thing that would get us closer to what I just described earlier with the automated tasks. Whereas OpenAI sort of is, is a few steps behind. Like he's talking about next year providing the abilities for enterprises to say, build apps on there. But enterprises are building apps now, like with, with any model. Like it doesn't even have to be open AI's model. I just feel like they're kind of behind and I just don't really see how that helps them catch up. Like, again, I don't understand the economics of big business. Like maybe they're just waiting and they want to get KPMG as a customer or Coke or whatever they said in the early days. And that doesn't matter because they'll make so many billions. But in terms of like the forefront of technology, they're just not at it. Like, they're just not even close right now.
- [29:22]
Michael
I think if you're losing $8 billion, I think that's what it says A year and you've got to invest trillions in infrastructure. Someone's got to pick up the tab, right? It can't be investors long term.
- [29:34]
Chris
I remember, I remember when we had committed to Amazon to spend some insane amount of money and you're looking at it being like, oh my God, like we owe these guys so much money. I better be nice to them so I can get out of this and imagine owing trillions. Like, you know, like our business basically has this massive anchor that's dragging us to the bottom of the sea and we have to invent something to overcome that and everyone's kind of lost interest in.
- [30:00]
Michael
What concerns me is Sam Altman single handedly has set up probably, let's be honest, the next financial bubble. I don't think AI is a bubble in and of itself. Like I think it has so much practical use and if adopted and set up right in organizations is totally transformative. Right. And as you say, I just don't think it's being done very well and. But yeah, all the hype, all the sort of like forward investment, the sort of we invest in you, you invest in us and we pay you in GPUs and you pay us in like tokens. Like the whole thing is a house of cards.
- [30:37]
Chris
And like when Enron started selling bandwidth at night or whatever, it was like we're going to sell broadband and trade it on the web.
- [30:45]
Michael
Yeah, it's starting to get really bubbly and we'll get to the Neo robot later. But the, the fact that they're losing so much money, they're committing fake money like 1.4 trillion. He's like, We've got to spend 1.4 trillion. It's like they have a lot of billies, don't get me wrong, but not that many billies. Like that's a, that's a trilly, not a billy. Like these are different, these are different.
- [31:08]
Chris
When we get to numbers that high. I quit. It's just not worth it. Like let's just get gold chains and retire.
- [31:14]
Michael
It's just so fake. We, I said we'd cover the, the cursor update and like normally we don't touch on these individual tools but I do want to talk about it and sort of contrast it to what we just said. Right. So here's a company that sure has built on the back of, let's be honest, Claude 3.5 sonnet. Like they just exploited the fact it was such a good coding model or Visual Studio, which is the interface that most developers were using at the time. And then slowly iterated and then rapidly iterated to build, you know, coding agents better autocomplete and just like, let's be honest, just better tools for developers. And despite a lot of people using Claude code and then Codex, which is the OpenAI competitor, you know, Curses hold held strong and a lot of developers I know, you know, still use it. So a release cursor 2 and it has some interesting features, but I don't think anything groundbreaking like now it can go off and use all models at the same time to solve a problem and then let you decide which approach, especially for large code bases you would like to adopt. So this is the people that are wearing gold chains in large companies and just don't care. They're and, and this sort of validates our point of wanting access to all the models. They're like, just run it across all the best models, we'll pick the best variant and it makes us so much more productive. It's worth the spend, right? They don't care. It's not like people trying to conserve tokens, you know, it's completely different.
- [32:48]
Chris
Think about the cost of a developer, like for a senior developer, at least In Australia it's $200,000 a year, right? Like front end, back end. That's about the price. I mean it might have gone down a bit lately, but you know, around that. So what's that in American like 140,000, something like that. So if you can replace them with a piece of software and you got to spend a bit of money on tokens, it's extremely worth it.
- [33:14]
Michael
It's not replacing, it's just speeding up. Like one dev will get you the output of two.
- [33:19]
Chris
Well, I mean, okay, well it's effectively replacing them then because you don't need to hire them. Think about you trying to pump out three different features at once in your product. You would have to have two different small, like two or three different small teams of devs. If one guy can do what one team used to do, that is replacing people. Yeah, okay, Pyres you wouldn't make.
- [33:39]
Michael
But anyway, my point around it basically is they, they added that feature multi file editing, things like deeper code base understanding, like improvements to semantic search, things that I don't think are that relevant to this show. But what they did also say is, you know, they've announced this composer tool. They say it achieves frontier coding results with generation speed four times faster than similar models. And they mean similar frontier models Sonnet 4.5 and GPT5. They also say that GPT5 and Sonnet 4.5 offer a slightly higher level of intelligence. But this is really a fast frontier option. And you and I have talked about this before. It's why at the moment I'm addicted to Sonnet. Sorry, yeah, haiku. Sorry. 3.5. As I know a lot of our listeners who use Sim Theory are also addicted to for mcps because it's just so fast and it's intelligent enough, it. It doesn't really matter anymore. Yeah. And so I think they've realized the same thing is when you're in this agent interaction, speed's critical. Like, test something, try it. It doesn't work. Try it again. And they know that these, even the frontier models need to do like six iterations sometimes to get it right. So really it's a speed game of like going back and forth, playing.
- [34:56]
Chris
And it might, it might sound silly, but at least for me, my personal productivity, excuse me, is made better by having a faster model. Because I don't get distracted. I'm not like waiting for it to do whatever the hell it's doing and going off and doing something else. I'm getting it straight away so I can move on to the next step. Like, it's a far more frantic pace when you work with a faster model like that.
- [35:20]
Michael
And as you know, like, I'm working on a new feature for Sim Theory, like Vibe Doc Editing, which is a lot different to, I think, what's out there today. And when using it, it's just so critical to have a fast model because you just want to, you know, quickly iterate. But then sometimes you want it to go and do longer things like go off in the background and research something and then add that chunk into the document at the bottom while I'm working on the top bit. Right. And that's okay, but if I'm like working in line with it and need quick answers, I don't want to sit there waiting for this document to think and update and change. And so anyway, I think the speed is important and the reactions on. On X, just from the quick assessment I did, is that people are really impressed with the speed they're saying tasks are completing in under 30 seconds. The parallel, like being able to put agents in parallel is much faster, yada yada. Anyway, I don't really want this to be some sort of free promotion for that tool, but what I would say about this is I think that this is an example of a successful product that's been built on the backs of anthropic and OpenAI's models, you know, and they can call it ecosystem and platform or whatever they want, but what they've done is they've gone out, figured how AI can be helpful to developers honed in on those use cases and workflows. Now they've gotten to a point where they're like, okay, now we can just train our own model because we have all the data for this specialist thing. Thing, and that's what they've gone and done. And yeah, okay, it's not the best, but it will get better, won't it? Like, exactly.
- [37:00]
Chris
But think about the brilliance of the economics of that. You've basically used an available tool at a high cost, knowing that it's a high cost and just accepting that burn, right, that you're spending more than you're making probably on the tokens for a while. Then they would have reached some level of economies of scale where they're like 50% of anthropic usage or something like that. The reason none of us could get quota for a while, but then they've negotiated a better rate, then they're like, hey, how about we just cut you guys out completely run our own stuff. All we need then is the GPUs themselves. And now we have a fixed cost because it's just the same ongoing GPU spend, no marginal cost for tokens, and we just bank money. Like, it's a brilliant, brilliant business strategy. Like, in terms of lowering your input costs but gaining the market share while the technology's evolving. Like it's, it's, it is not become.
- [37:56]
Michael
The playbook for everyone now. So, like, imagine you're in the, you're building like the cursor for finance or financial, like the new Bloomberg terminal that everyone seems to get up hard on over. Wait, I meant to beat that out.
- [38:09]
Chris
You beat the wrong bit.
- [38:10]
Michael
Yeah, but you know what I mean. So, like, they go and they, you know, they, they have some maybe MCPs connecting to all the financial data. You can buy it out with this thing and make investments and that becomes the new sort of investment platform, call it. I'm just making this up. Maybe it exists. Maybe someone's listening and going to build it. So this product comes out, everyone starts to use it because it's just better for investing, say. And then what's the next step? Okay, well, they go train their own model. And so you're really just benefiting on the backs of the R and D that OpenAI and Anthropic have poured billions into.
- [38:46]
Chris
Like, when you think about it like that, if you were A major institution. Why the hell would you ever sign a deal directly with one of the major providers like OpenAI or Anthropic or who's the other one? Google? Because something better and cheaper might come along. You might want to develop your own, you might want to run your own like this. You need to keep that flexibility there because the locking yourself into one means that you're guaranteed to pay the maximum and not necessarily have the best model in the long run. So I think that it presents a major challenge to those people as platforms in the sense that you've got this very, very healthy competition. And like you say, there's pretty much nothing stopping except for money. Someone coming along with their almost frontier level model. I mean, X XAI proof that, right? They just came along, they're like, let's throw some money at this and just make our own. And then they got one that's almost as good as the others.
- [39:45]
Michael
It's probably obvious to OpenAI too that chat GBT wallets in the lead. Now, like us Jeeves was at one point, just kidding. But you know, Vista, this lead can be eroded quite quickly, right? Like if Google ever get their act together and release a decent assistant, or Apple get their act together and release a decent assistant and put it everywhere, it just starts to erode the like. Yet maybe it's not going to kill their lead immediately, right? But it starts to erode it because people are like, oh, AI is here, AIs there, AI is everywhere. Like, I can get this chatbot experience literally anywhere. And it's, it's good enough. Like the image editing, the image creation, the video creation, like whatever it is, it's all the same. And so it gets to that point where it starts to get eroded and then you think, okay, well what, what doesn't get eroded? It's specialist workflows and applications. These do not get eroded like they, they never will be. And he acknowledges that in the excerpt I played at the top of the podcast where he says, you know, humans have always been building better tools and you know, things to do stuff mine gold. And so he's now saying, oh, we admit that that's the future. It's not some magical AGI box that we fantasized about. And so they've realized that it's these specialist applications and their pitch is like, build them on top of us. But I don't get the pitch. The pitch doesn't make sense.
- [41:08]
Chris
I think, like forgetting OpenAI for a second. I think that people in companies who are now building these AI Systems around processes in their organizations or exposing data are basically taking the steps needed to get to some level of automated intelligence in the future. Like these are not throwaway things that will need to be replaced. Like if you can teach the AI to do a fundamental component of what is done in your industry with intelligence, then that can be a building block later of a wider AI and smarter AI system. You're building towards it. Having a smarter chatbot is not going to accomplish that. Like, it's the, it's the stuff around it, it's, it's the, the arrangement of the data. And I think you had a topic we were going to talk about today about building smarter MCPs and I think that that is a big part of it is like, it's actually the architecture of how you arrange these skills and tools that will lead to, you know, better intelligence later.
- [42:13]
Michael
Yeah. And to, to really just sum up that whole conversation and why I think it is a big week in AI news, or at least what will impact the future, future of everything, is that this cursor model is a big blow to OpenAI and anthropic and Google and whoever else. Because what it says is there's no moat. Like we can basically build products on top of your APIs, get to economies of scale and then make our own model. And look, I'm not sure how this plays out if people will eventually just prefer this composer model or not, but I would think maybe they will. And then you think about Mira Moretti and her Thinking Labs company that allegedly is building it. So you can just train your own models, right? Like you can just build your own custom models for your apps. If a company comes along like that and offers a service to future app developers where they're like, hey, you're currently using OpenAI and Anthropic. We can study the workflows that your users use and then train a custom model for you and then it's yours. Like that's our service and we maintain it for you. Yeah.
- [43:22]
Chris
And I, to put that into perspective, let's say your business is spending 30 to 50,000 US dollars a month on the big players models. And someone could come along and be like, well we can actually reduce that down to like 5,000amonth running on a couple of GPUs. Like, that's very appealing for a business. And it's like, and you're going to get most of the performance. And because it's specialized to what you do, it'll actually be faster and more accurate. I mean, it's a very appealing proposition and I don't really see how the big players defend against that except for with good marketing talking about how hold us back from from taking over the world.
- [43:57]
Michael
One thing that I I questioned is where was Greggy boy? Where's Greg Brockman? Co founder? The only remaining co founder left at OpenAI other than Sam. Sam's bringing out the the AI Elon clone to the table. No Greg and and he's up there talking about company structure vision. Anyway, there'll be a sad song at the end of the episode about Greg if you want to listen.
- [44:24]
Chris
I haven't heard it. That'll be good.
- [44:26]
Michael
It's gonna be good. All right, so next next topic. As you alluded to, mine's a bit more clickbaity than what you said, which is does MCP need to die? And I based this whole whole thing on on a tweet by, you know, informer Sig Kitten over on X. He said MCP just needs to die. Anthropic's already abandoned it and he's showing in the Anthropic UI trying to add a custom MCP connector and their interface just crashes. Like it doesn't. It doesn't work. And then he said, you don't want to see the issues with the MCP and Chat gbt. And he says in Chat gbt, you need to enable Developer mode from the obscure place and settings. If your account type even has it. Then you have to go to the connector screen and not miss the weirdly placed Create button. I agree with him. It's very weirdly placed. And then he says, then have your server running so it can maybe connect to it. And I get and guess what? If you restart your server, you need to come back in here, disconnect, connect manually or your chats are going to fail trying to connect in Claude. It still sucks, but at least the MCP connection is done in the browser. So you can just refresh the page, then make sure you select it each time from the Overload Overloaded plus button or Chat Todd won't know about it, so you've got to focus on the MCP you're using or Chat GBT has no idea. And then he says, oh and my favorite part, when you click disconnect, if you want to do anything that closes the dialog, it's gone. Your MCB connection disappears and you have to use it.
- [46:04]
Chris
Wow. I mean this makes sense. This actually increases my self esteem quite a bit because I feel like in Sim theory we're like 100 times better than that. I, I can't believe that that's the state of it in there.
- [46:15]
Michael
Yeah. He's like, literally no one's going to go through and use this shit. And I think this is like the bit with mcps, right. Is it's so hard, so inaccessible. These companies have completely failed to make them work well that, that everyone blames MCP. Like it's MCP's fault as a protocol or an app.
- [46:36]
Chris
Yeah. And I think this is the thing. Every time I show someone how easy it is to set up their own and how well it can work and quickly they're blown away. Maybe that's why, because even if you're setting it up for, say, Claude Desktop, there's a lot of steps. And I think the other thing is that because people inconsistently adhere to the protocols, that there's a lot of fiddling around on a per MCP basis. Like, you hear one announce, you're like, that's great, I'd love to add that. And then you realize that there's all these different things you have to go through to actually get it going. It's not as plug and play as it appears on the surface, even though it should be. But I don't think that dismissing the whole thing because of a poor implementation like that or a poor way of working with it is right. Obviously. Because I think that if you work with it well, we've shown like the ability to have parallel tool calls, one click install and nice oauth connection processes on them. The power is massive. Like of mcp. It's just that unfortunately they're being presented poorly in these major providers. Yeah.
- [47:40]
Michael
Which is like two of the biggest companies. Yeah. And Anthropic, who invented it. And you see this reply I had up on the screen earlier, someone replied, looks fine to me. And it's a Screenshot of like 17 errors that come up based on MCPS not connecting correctly when they load up. Claude. So it. Yeah, it's a mess. And like, quite frankly. And I think to say that like, we as users of Sim 3 will know it's taken us like a good two months to refine the experience to get it now where I think it's mostly reliable and sometimes it's not like you're relying on so many providers and so many different variables that it can be quite difficult. But I think we're at a point now where it's reliable enough that I use it all the time and don't really think about it that much anymore. Especially if you refine Assistants with a limited MCP tool set rather than just trying to smash them all out.
- [48:32]
Chris
But I mean, some of that has been us having to rebuild the mcps ourselves. Like, we're like, okay, that's a poor implementation. Let's make our own version of it, that that is better. So it's taken a bit of that, But I feel strongly enough about it that it's worth it. Like, it's worth getting them right, because the productivity gains from people using them is so high and quite frankly, they're not that hard to build properly. It's. It's something that, it's more about the way you work with it that's important than the thing itself. However, I think I mentioned earlier the other thing that's bad about mcps is that I think the people making them in some instances don't really understand the way they're being used in that they're building it like it's just an API with a series of functions and sort of hoping that the AI will figure out, okay, well, I can get this parameter from this tool call, and then I'll pipe it into this tool call. And then after that, after both of those are done, I'll call the third one. But the problem is that isn't a really nice way to work with the model because it's going to have to go through three or four iterations to actually get to the thing you asked it to do. My opinion is the tools need to be more like individual procedures or like accomplish a sort of range of things to get a task done, and then the AI can then feed them all the information they get to use that. So, like an example of that is, let's say it's your Outlook calendar and you say, make a meeting with Mike next week at 9am if you think about the steps involved, it needs to check that it can access your calendar and that you're free. Check which calendar it's referring to in case you've got multiples, connect multiples connected. Look up the contact of the person they're referring to to check that it has it and it can add it. You know, that's three or four steps. Now, if you just have an MCP that has those as individual things, that's going to take a while. And there's. Each time it takes a step, there's a chance that the AI will make a mistake, right? Or pause or do something dumb. If you can combine that as one thing, it's like, this is how you make an appointment. You do the following four things. It's much more efficient and much more likely to succeed. And I think that the MCP design needs to get more intelligent in that way in order for it to be perceived as better.
- [50:53]
Michael
And it's interesting you say the design because there was this post on X by David Kramer. He says, I'm so tired of would be AI influencers. That's us shilling GitHub as an example of why MCV is bad and thus why CLI is better. You have all done a disservice to the community by miseducating people. GitHub is a really poorly designed MCP. The other thing he says that I'll call out because we can explain why it's bad is so just because GitHub and some others have built bad implementations on top of MCP doesn't mean the need for tools and the metadata required to operate said tools is the problem. We are still early in the lifecycle and there's a good chance we'll be fine tuning agents on top of our tools to improve upon this. But any version of training is still lossy compression and will be more prone, error prone. So please stop spewing MCP is bad because some random dude on the Internet who doesn't work for a living told you so.
- [51:48]
Chris
Yeah, so the problem with GitHub is it has like a hundred tools in it alone. So if you enable GitHub as an MCP, your model has to contend with going through the definitions of a hundred tools to decide if it should call one or not. And then even then there are tools like I described, where it's got to like, look up the repository name from scratch every time it's got to look up something else there's no idea of. Like, okay, we're working in this context. So this is the set of tools I should be working with. We have that in SIM theory. I filter the tool calls to give it a more curated selection of things that here's, you know, here, sir, here is your tray of whiskeys or tray of cigars. Pick, pick your favorite. Rather than like, there's a wall of books and games and garbage behind me. Go read them all and decide which one you're going to use. Like, just straight out of the box. Enabling the GitHub MCP is going to make your experience working with AI worse. Like there's no other way to describe it. You're muddying the waters. It's going to cost more and it doesn't even work that great when you use it. So it's A, it's, it's a canonical example of a poor mcp and it's basically useless for that reason.
- [53:01]
Michael
Absolutely useless. And it's, it's one where we're like, okay, we'll have to rebuild that. Like, we'll have to do their job for them because they've done it so poorly.
- [53:09]
Chris
Like, do you really want, every time you type a command to your AI, model it to decide if it wants to comment on a public issue on GitHub because it has that ability, you know, like, do you really want it to create a gist of what you're talking about?
- [53:25]
Michael
People that don't understand coding lingo but understand like this sort of, you know, AI chat style interface. Imagine every time you, like, like the classic example of hello, saying hello without any form of filtering would need to do the following if you had the GitHub installed. You're talking about a prompt that guides the tool use alone and the tools and like the, all the examples and stuff in them of pasting like a full blown essay in when you say hello above that there's an essay of like, hey, here's all the stuff you can do with this tool and then multiply that by like 20 or 10. And this is why I think obviously OpenAI have just restricted it to one. And with anthropic, I think it's like a couple is because you're putting essays in every single time. And so that bit is just so dumb. But you can fix it so easily. Like one of the experiments I did during the week is if you have a tool which is like get schema or get info or get more tools, so you just have the most critical tool listed. Like for say image editing, you might have create image, you know, or edit image. Sorry, one tool listed and then have another one more image tools. And so then the AI can decide to go and discover more tools.
- [54:45]
Chris
And you've basically just described anthropic skills.
- [54:48]
Michael
Yeah.
- [54:49]
Chris
Which is, which is exactly the way that works. Like they give a sort of top level, here's the thing it can do and then all the additional prompting and mechanisms for running that are encapsulated within that skill but not included in the primary prompt. And so yeah, you've made that but in toolkit form.
- [55:08]
Michael
And I think this is why people are excited about Claude skills, because you're, you're encapsulating or you're teaching people and you're getting the help of the AI if you build it through the AI to, to encapsulate those things. That people would be building. So instead of me as the developer with my mcp, trying to figure out, like, your workflow or everyone's workflow, you now can then take. Well, at least in the implementation, we want to do a series of MCPs and instructions and then tune your own skills so that you're teaching it how you want it to operate it.
- [55:46]
Chris
Yeah. And then keep those tools within the context of that skill so you're not presenting them to your main AI model every time you do anything. Like, it doesn't make sense to do that. It makes sense that they would be clustered and working in a certain domain rather than just this global thing where you have all these tools. I mean, we have people with literally 500 tools enabled. The AI won't even take that. I think the limit's like 125 on anthropic, for example, so there has to be a level of filtering there. But I also just don't think that filtering is the future. I think it's what you're describing, which is these curated skills that contain sets of tools within those to accomplish specific procedures or specific decisions or research or whatever it is.
- [56:30]
Michael
It's this evolution and iteration where I think people are quick to come out and say, oh, MCPs are terrible. But I think it's just the mainstream experience with it. I've seen in the SIM theory community firsthand, people starting to say, like, hey, the. The MCPs are good in here now. Like. Like, it's running well. We're doing asynchronous stuff. We're getting stuff done with these. And I think that sort of validates all the work we've been doing figuring out, like, how do people want to use these? Listening to, like. Like, quite frankly, just constant criticism and responding to it and just taking it on the chin and going, okay, we agree. Like, let's make it better and better. And then now I'm at a point where, like, you know, I. I'm using it again for literally everything. Like, you know, when we do anything in the overall system or respond to a support ticket now, you can just say, go off and figure all of it out, come back to me and I'll approve it. And I'm using it. Like, I shared an example with you earlier in the week. I'll try and bring it up on the screen for everyone that watches all two of you. And basically what I was able to do is. Let me just bring it up here. Oh, wrong screen. Hang on. Yeah. So basically what I was able to do is I had a power bill that was sent to me by my energy provider I've talked about on the show.
- [58:03]
Chris
You can tell we have in our country that both of our main examples of huge expenses in life is our power bill.
- [58:10]
Michael
Yeah.
- [58:10]
Chris
And so couldn't afford to run trillions of GPUs. That's for sure.
- [58:14]
Michael
No. Even though we have all the energy in the world. But anyway, so they, I quit this energy provider. They sent me a bill for what would be less than a single month. And normally my bill might be like say 200 bucks a month. And the bill, the bill, the bill was $732.
- [58:33]
Chris
Oh, you have solar.
- [58:35]
Michael
Yeah. Anyway, but this is not my point. So. So there's a. This huge disparity. Right. 500 bucks. Now I am so busy in my life, I normally would just pay it and just accept it and be like, oh, they couldn't have got this wrong. There's no way. But I was curious and so I asked the AI with MCPS to go off and research it. I cut and paste in my solar data, like, which actually tracks my energy consumption. And the AI is like, it doesn't match. Like, this doesn't match up. Like, your energy usage tracked in this application does not match what they're saying. And then they sent me a few days later an email saying, you know, we got it wrong, we apologize. We've basically stuffed up your bill. And I thought, oh, good, you know, they've corrected this mistake. I'm glad I hadn't paid it. And it said, we owe you an apology. And then they came back and reduced the bill by $4 and said, we stuffed up on the solid credit. So anyway, I'm like, that's it. I'm going all in. So I fire up the AI lawyer. I have an assistant called Lawyer, which says, you're a really, really, really good lawyer. I think that's seriously the problem.
- [59:46]
Chris
And there's a prompt engineering guide for you.
- [59:49]
Michael
And I'm like, here's all your tools. You got access to my email. So it went through my email, it found. And in the new Gmail that you built, thankfully it can read attachments. So it went through. Read all the historical power bills. I think it, in a couple of seconds, it pulled up like 20 or 30 of them. It was able to go through those PDFs, read them all, put together a case that, quite frankly, I was like this, you know, maybe it's hallucinating, like, maybe it's not true. But I sent it to their. I, you know, I sent it. It also looked at the legislation and my rights as a consumer. So I emailed them the most legally summary and breakdown of like every problem on earth and was like, you know, I could basically take you to this tribe funeral and get $100,000 payout. And then it listed all the case examples of where this particular company had lost when this had occurred in the past.
- [60:43]
Chris
Wow.
- [60:44]
Michael
Anyway, then they came back to this email, not realizing how good my really, really, really good AI lawyer, how much.
- [60:52]
Chris
How much money it would cost a real lawyer to do that research and draft that letter for you. It would be thousands of dollars.
- [60:58]
Michael
It would never happen, let's be honest. So anyway, they come back and they're like, you know, as a gesture of goodwill, like, we haven't done anything wrong here, but as a gesture of goodwill, we'll give you 50 off the bill if you pay it and end this matter. So I wrote back, Rahul, your 50% offer is rejected. And then, you know, whatever the AI lawyer wrote, I didn't write any of these emails. I don't have time to do this. I was just like, respond.
- [61:24]
Chris
You're just like a spectator in your AI agent doing your dirty work.
- [61:28]
Michael
My wife was like, oh, you probably should just settle this. Like, you're wasting a lot of time on it. And I'm like, no, I'm not. I'm just telling you what it's wasting a lot of time on. And so anyway, it replies and sure enough, it gave them a deadline. It decided that it was this Wednesday at 5pm And I was like, wow, that's really aggressive.
- [61:45]
Chris
That is a really, really, really good lawyer.
- [61:48]
Michael
Great lawyer. Really, really, really good lawyer. And so then I get back. Dear Michael, thanks for your email. I have waived the debt of $732.13 as a once off resolution for you. Based on the above, I've now closed this matter and it is resolved.
- [62:04]
Chris
Yeah, once off, like don't use this as a precedent for other people.
- [62:09]
Michael
So my point is around this is the power of working MCPs. This isn't saying like, obviously AI is the, the is the equalizer, the great equalizer for the consumer where these companies could basically take you for a ride in the past because they had all the resources, you didn't. Now you have all the resources, they do and more.
- [62:31]
Chris
And without bragging, I think that SIM theory actually made that possible with the way we implement the mcp. Because if you think about it, let's say you were using ChatGPT. You would have had to do way More copying and pasting. You would have had to go, go into every email and get the contents from the PDF or upload the PDF or whatever. Right. Or you would have had to be directing it to call each MCP individually, like, look up the laws, you know, analyze this PDF. You know, it would have been you just basically guiding it through the process. You probably, again, even that probably would have been too much effort to actually go to, I would imagine.
- [63:13]
Michael
Yeah, I mean, there's just no way I would have invested the time otherwise. And I guess this is what I mean, because our experience of MCP is this and everyone else's. Like, the 99.999% experience has been like, how do I even access it? What even is it like? I, I would argue if you're just a regular ChatGPT user, you, you would have no idea what it even is.
- [63:37]
Chris
And even in anthropic Chicago for you.
- [63:40]
Michael
But even in anthropic, like, sure, you can connect to email, can't read attachments. And I'm not saying, like, that's some killer feature. I'm just saying these are the workflows and things that people actually want to do with AI, right? And just expect to work. And this isn't to sort of, I get, like, this isn't. I know it sounds very promotional, but it's really not. I'm just saying, like, why haven't these larger companies with far more resources than us, and quite frankly a lot of talent realized that this is, this is the benefit. Like, this is, this is what the.
- [64:15]
Chris
One thing I wonder about it is. I mean, the whole reason we started this podcast was because we wanted to actually use this technology, and therefore we're talking about it. And I think a lot of what's good about the MCPS and SIM theory has come from us, and by extension our community actually using them. And like you say, constant criticism like, this didn't work, this was crap, the output was garbled, it did something dumb. And then us just going back and be like, well, we know the model is capable of better than that, and going back and making it better, or where necessary, completely replacing the MCPs. Like just this week, for example, I replaced the YouTube one to be much better at searching and much better at getting transcripts and analyzing them because people weren't using the previous one because it sucked. And so I think that you can't just go, oh, well, that's the YouTube MCP and just end it there. I think that we have to see it as very Very foundling technology that needs some. Some attention on it to get better. But the. But people need to look at the possibilities. Like, you really need to think, like, what if I gave it a series of tools that could do this for my company or whatever system I want to work with, Would it be able to do good stuff with it? I think in most cases the answer is yes. And therefore, to dismiss it outright and say, oh, MCPs are dead, is ridiculous and wrong. That's the Sim Theory promotional song.
- [65:56]
Michael
So it was a promotion after all, disguised as a. As content. Do you want to know something really interesting? I always am alluding to the fact that the tracks are now all on Spotify. We've actually risen in the rankings. People are getting songs like Love Rat. Like, you know, when they do, like, Spotify dj.
- [66:19]
Chris
Yeah.
- [66:19]
Michael
They are actually getting like, Love Rat, just like. And for those that, you know, don't know the. The beauty of that song.
- [66:28]
Chris
But.
- [66:28]
Michael
So they're listening. They're on their commute home from work. Here's the play. You know, they're on the train.
- [66:33]
Chris
You know, if I'm ever at a cafe and I hear that song Come on, I'll know we've made it.
- [66:39]
Michael
They're on the train, they're commuting home, and all of a sudden they're listening. And this song sounds sort of, you know, like a mainstream track. And then they're hearing stuff about, like, Jeff, I had it on in the.
- [66:50]
Chris
Car the other day, and everyone was just, like, enjoying it. And then it said something about like, GPT4 or something. And then the kid's like, is this AI? Yeah, like, yep.
- [67:03]
Michael
Anyway, all the songs available in the album AI Diss track collection, and then average tracks from the show, you know, wherever you get your music.
- [67:11]
Chris
We should do like a triple j hottest 100, but for AI songs, and let everyone submit them. And wait, like, no voting. We pick.
- [67:21]
Michael
You know, the other day, someone reminded me about. Remember that original? It's. It's on one of the shows, like, from long ago. That song. It was like that first country one we heard about AI that really gave us. It spawned the idea for all these songs, I think.
- [67:36]
Chris
Oh, yeah, yeah. Have you got it?
- [67:38]
Michael
I remember that. I can't. I'm not that good. I'll. I'll try and pull it up before.
- [67:41]
Chris
It's like, sorry, but I won't do that. Yeah, as an AI language model, I can't do that.
- [67:47]
Michael
You know, back when rejections were a thing. Wait, they still are. If using anthropic all right, moving on, moving on. Because this has gone on way longer than I thought for a quiet week.
- [67:59]
Chris
So at least you didn't tell everyone to skip it.
- [68:02]
Michael
Well, I sort of, I mean, I gave them a little warning. So in the future, Chris, in the future there's going to be very muted colors. I'm thinking like, you know, tans, whites, all that.
- [68:15]
Chris
Like non, non confrontational colors.
- [68:20]
Michael
Not yet non confrontational, sort of like beta colors is how I would describe this. So this company, 1X, they have investors, including OpenAI is one of their investors, of course, because they've got all the money. And so they released the, or allegedly are releasing in a year from now the Neo home robot, which you can order today or subscribe to for $500 reduce a month. And I'll mute it because the audience Australians do it.
- [68:52]
Chris
Or is it just America as usual?
- [68:55]
Michael
I don't know, maybe we should. Maybe we should.
- [68:57]
Chris
I think we need a company one like get on that.
- [69:01]
Michael
No, I'll get on. I'm sure someone, one of our listeners, will accept the shipping of the robot to their house in the US if we can't. But anyway, it's this like whimsical sort of old fashiony video. We're in San Francisco and this robot's. Yeah, like doing stuff with them. Like, you know, it gets a carpet, opens the door to the house and they allege that it'll be able to like fold clothes, like tidy the living room when you're out, pack the dishwasher, unpack the dishwasher. And don't get me wrong, I want to live in a world where all of this is true. But in the demo video, you know, there's like two scenes, like one where it takes a cup off the guy and another one where I think it opens the door where it's actually using the AI model to operate the robot. The rest of it is tele operated with a guy literally using a VR headset in another room just acting like he's the robot. And I get the whole fake it till you make it tele operator thing. But what I believe they're proposing, and it's all speculation on my part, is that you could in the future for some tasks have like a tele operated robot in your house. So there's some guy in India on like, yeah, I don't want to do the action, but he's just probably looking.
- [70:23]
Chris
For crypto where you store your crypto keys like that.
- [70:28]
Michael
So like, anyway, why would anyone on earth let this thing into your house? If it can be, if it can be tele operated. But it, you know, it's.
- [70:39]
Chris
Honestly I, I thought because I read it that it was a paid add on to do some stuff. I didn't realize the majority of the stuff was done in that.
- [70:46]
Michael
Well, I think they're claiming that it'll be able to fold laundry, vacuum, vacuum and carry groceries. But apparently to fold it took. So demos showing it taking two minutes to fold a single shirt while struggling with balance.
- [71:03]
Chris
So it's sort of like a drunk elderly person.
- [71:06]
Michael
Yeah, pretty much. And it says critics are calling it a modern day Mechanical Turk. The 90, the 1700s, chess playing automation that was secretly a human inside. And demos show it taking yeah, two minutes to fold laundry. The privacy, you know, the, the privacy concerns. People are calling it creepy, a privacy nightmare and questioning why anyone would pay $20,000 for what's essentially a remote controlled robot operated by strangers who can see inside your home.
- [71:36]
Chris
I bet you it sells out though.
- [71:39]
Michael
But so humanoid robotics is a category companies. This is outside of Tesla's robot and, and some of the like internal project. 3.2 billion in 2025 alone.
- [71:55]
Chris
This is the first robot I've seen where the main demo is. Isn't someone just kicking? Seems like all of the main robots, their main job is to just take a punch and take a kick.
- [72:08]
Michael
I just like, I believe, I believe this, this will happen. I don't believe this company is.
- [72:13]
Chris
Look at what happens to this robot if I hit it in the head with a cricket bat. It responds perfectly. So it's.
- [72:21]
Michael
You can have your very own next year apparently for $20,000 or 500amonth. The one thing I'll say about this is I got to experience last weekend the Australian version of fsd, right. Because my car's too poor to get that. So software and so a friend of mine, he had it, we were away for the weekend and he was like, you got to check this out. So we get in the car and I'm thinking, you know, my expectations were very low. Not because I didn't think it would be good. I just thought, oh maybe it's like one of those things like you watch the videos on YouTube of people in the cars and you, and, and you're like, wow, that's amazing. But then when you experience in real life there's all these quirks and problems with it and you're like it's not that great.
- [73:06]
Chris
Yeah. Anyway, yeah, the door keeps falling off, but we're working on that.
- [73:09]
Michael
Yeah. Here's my genuine experience with It, I had an existential crisis. It drove around roundabouts, it drove us to Woolworths. In this remote town, there was like road work where you had to cross onto the other side of the lane, like people with stop signs signaling. It drove us all the way there comfortably. More comfortably than a human drives into Woolworths, which is like a Safeway in Australia. Goes into the parking lot and finds a park and then reverse parks into the park. And I, I don't know, I would equate it to like losing your virginity. Like that is, that was the experience I felt after I was like, hang on, what? I was like, this is like, I'm sorry, like a late adopter here, but I'm like, this is going to change everything. Like, like every other manufacturer. Like if no one else can copy this or get to this standard, eventually, once everyone experiences this and they perfect it, which I'm, I'm certain will not be that long, maybe five, ten years. And I, like, I know people might disagree with my timeline, but I reckon it'll still take another decade before it's, you know, you can go to sleep. And also let me clarify, go to sleep with your family in the car on a long trip and young children.
- [74:24]
Chris
And trust that you're safe. Yeah.
- [74:27]
Michael
So I was like, this is going to change economics. Like how, how much could this company be worth? Like trillions? Because if, if this, if they truly are ahead, right. And this thing is this software can be applied to like other vehicles, like trucks and transport, everything big one, right.
- [74:43]
Chris
Like if you can do road freight with no drivers, like, that's going to be pretty incredible.
- [74:47]
Michael
So I'm a huge believer in this stuff and I think the humanoid thing sure will eventually get there. But think how long it's taken Tesla to get to this point. It's been like over, I think over a decade or maybe just under a decade. And there's still edge cases from what I understand where like if, you know, it can freak people out or make mistakes sometimes. And in fact it did make one mistake. I'll call out where it caught fire.
- [75:12]
Chris
Crash.
- [75:15]
Michael
There was a left turn and my, my friend, he, he does like admit he never washed his car ever. Like, he, he does, he has, he, he has a certain color choice in cars. These, both, these cars are the same color so they can remain dirty and no one knows like that sort of like dark gray color where it just looks dirty naturally. And anyway, we were driving into the sun, it's making a left hand turn and it panics and slightly goes Onto the what? You would, I mean, the wrong side of the road, but in the sense it's one of those like estates with a really narrow road and it, it just sort of goes over to the right when it should have gone to the left and around this like sort of hub thing in the road. The wrong way anyway. Like, if it's something a drunk driver would do, basically. So outside of that, like one thing, everything else was incredible.
- [76:04]
Chris
May have just like killed someone or a pedestrian.
- [76:06]
Michael
I don't think so. I don't think it would have killed anyone. But yeah, so, so that one thing. But I think similar to what we experience with LLMs, right, there's moments of brilliance, like 98% brilliance and then that other 2% where it fails and you're like, God, this thing's dumb and will never get better. And I think the self driving car thing, that's sort of where it's at right now. And overcoming that last sort of 2% is going to be the hardest challenge ever. Similar to these robots, they're, they're at like the 1%, so they got to get to the 98% before they're, they're useful. I just think this stuff is still probably like 15 years away. If I had to put a number on it.
- [76:47]
Chris
Yeah, this is like that, what was that rabbit. Rabbit AI thing? It's like the equivalent, but in robotics.
- [76:54]
Michael
Right, the scams are just getting bigger.
- [76:57]
Chris
Yeah, it's like we know that people want this so we can sell something that kind of looks like it does it. But actually, because I just think about things like in the house that, you know, they're not easy. Like our door lock, right. Is a bit stiff, it's kind of hard to use. And I'm like, is the robot just going to persist with trying to open the door for like 45 minutes, trying different combinations, like, oh, sorry, I appear to be having an issue with the door and then like, just keep trying. Like, I like the idea of one that's just going to escalate things. Like I'm going to go get a hammer and that will open the door or something like that. The other problem I noticed with it, this sort of fabric look, it has, I can imagine it getting like wet and dirty. Like if I do the dishes, I spill water on myself and I just wonder if, like, if it gets damp, it's going to stink. Is that the image? I'm.
- [77:50]
Michael
None of this matters because this is just a pump and dump scam. That's all it is. Like this thing I Just if this. If this is it, if. If this is it, if this is the big breakthrough, this is the robot. I mean, I. I can't go on because I. I wonder if, like, for.
- [78:08]
Chris
The people controlling it, you could have, like, celebrity celebrities or, like, you know, twitch stars, like, doing your I'll be your house husband for a day as your robot. And that's, like, a perk they can offer people.
- [78:20]
Michael
So it comes in three colors. They're interesting color choices. There's a. A white sort of tan one, which they used in all the. The demo videos. Yeah. There's a black one. I. I don't know why that, like, given, you know, slavery and having slaves in homes, it just.
- [78:36]
Chris
It'd be a. It'd be an interesting choice if that was the popular one.
- [78:40]
Michael
Yeah. Like. Like, who's ordering the black one? I just. Whoa. And then there's like sort of a slightly, sort of, like, I would say the middle shade between.
- [78:50]
Chris
It's not brown, Grace.
- [78:52]
Michael
It's not brown. It's like a gray. A gray color.
- [78:57]
Chris
So, I mean, I think the thing is, like, part of me, even though I recognize that it's probably, in practicality, not there yet, the idea that it might work, it's just so appealing. I just think that there'd be enough people out there willing to spend 20 grand even if it can just do one or two things. Like, if it could just fold the towels or it could just unload the dishwasher. Like, if you could unload the dishwasher. That's pretty useful.
- [79:23]
Michael
This is similar to the thing Apple released, like the Vision Pro, where, like, you use it a few times. You're like, wow, it's a glimpse of the future. And then you put it in a drawer. Like, I can imagine this thing being put in a body bag up in the garage. What else are you going to do with it?
- [79:44]
Chris
I guess you could talk to it. The whole companionship for old people thing is pretty cool. Like, if you could sit it with someone, you know, who's got dementia or something, and it'll chat with them and remember their stories. Freaking told me this already. I know it better than you do. But. But, yeah, like, that kind of use case is actually kind of cool. Like, even, like, you think about an elderly person who. Who wants to live independently at home. Right. One of the big risks is they have a fall. And nobody.
- [80:20]
Michael
This seems to have a fall. Not the elderly person.
- [80:23]
Chris
Well, yeah, the robot. Well, vice versa. They can. They can have each other's backs. But, like, I, like, I Reckon there'd be a segment of people out there who'd be like, look, I don't want to put them in a home, but if I chuck this robot in, it's going to, you know, make their eggs in the morning, it's going to have a chat with them and it's going to alert me if anything goes wrong. Like that's quite a useful function.
- [80:44]
Michael
I 100 I'm with you. I get the appeal. But I think what you're saying is 10 years away before, like before we get anything like that, like I don't see. I don't know, I just get the feeling that it'll happen and when it does we can't fathom the economic impact. Like, like my experience in the, in the Tesla. But I just, I, I think this is peak AI bubble in terms of raise Billy's Pump and dump, get a bunch of orders in, raise more cash out and you know, and then like the, the, the bubble burst. People realize this, this stuff's a long way away.
- [81:22]
Chris
I'm very much looking forward to the fail videos from these things when people start to get them.
- [81:28]
Michael
Anyone will ever get it.
- [81:29]
Chris
Like the real funded business.
- [81:32]
Michael
No, it's, I think it's, I think it's like reasonably legitimate. Like they've got a decently designed website, although that doesn't, you know, the, the one thing on their website is you're right about this old person companion. It's this old guy hugging the robot. I mean what him and the robot are up to, I'm not sure but.
- [81:47]
Chris
It'S like he's, he wants to kiss it.
- [81:49]
Michael
Yeah, it really, there's nothing to kiss. Yeah. And that nearly all the examples are like the, the elderly couple are playing cards and these robots carrying stuff around. But you could actually see this going another way where they realize the AI is just simply not good enough, which I think they kind of have already. And instead it's just tele operated, you know, slaves.
- [82:10]
Chris
What if you bought, what if you bought like 10 of them and have them act out of play for you? That'd be fun.
- [82:15]
Michael
Yeah, we got to get our hands on one of these. Even if it is tele operated I.
- [82:19]
Chris
Really have to like it's worth it.
- [82:21]
Michael
For content and this is probably why it'll be successful.
- [82:24]
Chris
Did it take a piss in the toilet?
- [82:26]
Michael
No, it was, it was pretending to clean. I don't know why. This has been the biggest lull ever.
- [82:36]
Chris
So this is when some dude is remote operating, is that right?
- [82:40]
Michael
Yeah. So all these ones are remote Operated. Yeah. The laundry. Even the laundry example.
- [82:46]
Chris
That is so sad that someone sitting there in a $1,000 VR headset and he's painstakingly putting your dishes away over the Internet.
- [82:58]
Michael
Yeah. Like, because you don't want them in your house. But it's fine if they're through the robot, but then it's like they're seeing everything like it's a real person.
- [83:07]
Chris
If you can throw away 20k on a toy, then maybe you could just afford a full time maid. Like, I don't understand the benefit of this right now.
- [83:19]
Michael
I just don't believe it'll ever ship or work or they'll ship it. We'll get a lot more lulls and then it'll be like one of those, like, you know, we're sorry videos maybe. Or is this like juice era, you know, where the juice was already, you know, ready and then the machine just squeezed it? Is this juice area in the sense of like they're just going to tell the. Operate these things and you pay $500 a month and they give 100 bucks a month to some company in India to operate.
- [83:46]
Chris
Yeah, but I mean, I guess it still has the like, dexterity and motion and stability to, to perform those tasks, which seems pretty good.
- [83:53]
Michael
Dexterity. I don't know. Watch the, watch the Jonas video. It's okay. I think the examples I've seen from one other robot company and then obviously the Tesla robot seems just so far ahead. I mean, these, the Tesla robots, there's video of them being used on the production line to assemble things. I mean, they're trained on those tasks, obviously, but yeah. Anyway, I want one, don't get me wrong. And I want this to be true and like, I want to believe.
- [84:23]
Chris
Yeah, it's sort of, it's sort of like, I guess that's what we always talk about on the podcast. Like when we used to make, remember in the early days making the images and like people would have, you know, broken hands and weird stuff, but you just want it. It's like the same with this. It's like, I get it, I understand it's a scam and it probably won't work, but I still want one quite badly.
- [84:44]
Michael
All right, any final thoughts on the. The big happenings in the AI market? OpenAI becoming the platform of choice, apparently. The state of MCP and the neon robot.
- [84:58]
Chris
I think Sam Altman's just further diluting any sort of genius level sway that he has by doing these constant events that are not cohesive. I really think what they need to do is like, like Willy Wonka style. Like shut down the factory, go hide for a while, don't say anything to anyone, and then come out with like the golden tickets or some, some massive event. But I'm talking about this from a pure PR play. I think that for me, what would sway me is these guys leaning into being like huge platforms for like MCP hosting and coming out with the best tools out there to build on and staying far ahead. But as we discussed, I just don't know if that is. Is possible.
- [85:45]
Michael
Yeah, I think that's it, right? Is like go away, focus on real use cases and instead of just talking about it, deliver this future. Like deliver a really good platform that developers like us and others want to build on that are excited about it and get benefit out of it. And then if they can deliver on that vision, then, you know, maybe that can be the future of the business.
- [86:06]
Chris
One way a massive data center is imagine if GPT5 was the fastest model. Like imagine if it was just so fast. Like it was just blazingly grok level with Grok with a Q level, fast and reliable. And it was the best model and they made it like a tenth of the price. Like to me, then why would you use anything else? Like you would only use them and that's how they could possibly win. And then because if presumably that comes from economies of scale, it's just so hard for the others to compete. Like that is how they could win on the front that he's talking about.
- [86:42]
Michael
Yeah, I think that. And also just like figure out these real use cases that, that businesses, you know, want to do. The, the, the chat GPT thing is so confusing to me. Like they don't mention that core product. They're just like, it's like one of our products on top of this huge stack. But it is the product. And the reality is like also maybe just go all in on that. Like, I don't understand, like it this seemingly this threat to Google and it makes me question like, why don't they go all in on it? And I'm like, it's probably because the cogs aren't there. Right. You can't make any money off it. Like even at 20 bucks a month or 200amonth, I would argue they're probably still losing money.
- [87:21]
Chris
They definitely are. I mean we, I mean their cost base would be different, but still like those plans are too cheap to make any real money.
- [87:29]
Michael
Yeah. And also I think they've in some ways done a disservice by just devaluing the, like, the benefit of these models, like, I. I think I would be willing to pay a lot more to just retain access. I don't want them to obviously do that, but I think, like, if they need to make a profit, that's okay. Like, yeah, you know. All right, we will see you next week. Stick around if you want to hear the Greg Brockman sad song for not being included in the YouTube video. We'll see if it's a banger, who knows. All right, we'll see you next week. Goodbye.
- [88:07]
Greg Brockman (poem narrator)
I ironed the cuffs of a suit I didn't wear Set out a chair by a screen that stayed there the countdown flickered then face is in light Three on the stage and I'm left with the night I built cathedrals out of code, out of time we dreamed of truth that could reach reason could rhyme but somewhere between the chips and the sky they found a new climb that left me behind Is this the price of building the dawn.
- [89:04]
Michael
To watch.
- [89:05]
Greg Brockman (poem narrator)
It arrive with your name withdrawn A billion lines, a single co cut A door I open quietly shut and I am not in the frame tonight not in the questions, not in the light I see the future dressed in gold but I feel the weight of what we hold if all our engines roar and rise who keeps the compass for our skies? I love the spark, I fear the blaze don't, don't let the world be set ablaze you set platforms set infrastructure trust silos of power silicon dust a trillion promises long steel lines but who will guard the hearts and minds? We taught a mind to think like fire to stretch a thought, to climb it higher but will it learn to care, to stay when all we do is pave the way you measure progress in gay Watson grass But I hear echoes in the aftermath the model's cheaper, the cost is down what's precious now that thought is found and I am not in the frame tonight not in the credits, not in the light I see the future split in twos the things we make, the things we lose if every every workflow wins the war who tends the soul at the core? We'll cross the sea, we always do but who will anchor something true? There's a kind of silence only founders know the clock that ticks behind applause the hands that shake when you let go we said we'd make a lantern for the dark but lanterns cast a shadow too if we forget to hold the heart the brightest path won't carry through Teach it to reason, teach it to wait Teach it to doubt when high the stakes Teach it to Listen more than let to know the lines we want and need not just alignment on a slide but vows we keep when markets rise A promise etched in human hands to build what serves not just expands and I am not in the frame tonight not in the crest it's not in the light I see the future split in twos the things we make, the things we lose if every workflow wins the war who tends the soul at the core? We'll cross the sea we always do but who will anchor something true, True I wanted wonder not a crown to lift us up not press us down if I'm not there to say the rest Let every word still mean our best and I am not in the frame tonight But I am holding to the light A patient future carefully grown where wisdom weighs what power sows if we must build an endless sky Let mercy be the reason why.
- [92:59]
Michael
And if.
- [93:00]
Greg Brockman (poem narrator)
I'm gone when credits roll Remember guard the human soul Turn down the glare Let conscience in not just to race but learn to win I'll take my bow behind the screen and pray we earn the world we dream.