Transcript
Will: Okay, well, welcome back to week three of the Business Idiots podcast. Once again, another big week in AI this week. We're going to be talking about some of the advances in open source that happened this week, which is kind of rocking the AI world again. Specifically, that's on the release of Kimi K3. we're also going to be talking about ai regulation uh where it's happening is it happening in the right way uh should we concerned about these things or happy that we're taking a step forward and we're also going to be talking about the cost of ai as these things are changing pretty rapidly as open source changes as costs change as new models come out there's a lot of a lot happening in this space. So I'm once again joined here by Jim Lovell. Welcome, mate.
Jim: How are we?
Will: Good to see you, mate. And also we've got Alex Stenlake. How are you? All right, guys. Well, let's jump straight in this week. We're going to start off talking about some of the advances that we've seen this week in AI technology. So... This week we had Miramarati's Thinking Machines Lab come out with InkLink, which is actually a family of models, but we just saw one model come out here at the start, which has purposefully been released as a generalized model. They specifically said that it's not really designed to beat any one particular benchmark against the frontier labs. It's actually a different approach with the model that they claim is much easier to fine tune for specific tasks and purposes and even the the demos of this showed and being able to prompt it through their harness and saying please fine-tune yourself on this particular task and we go off and do that work and come back and and specifically work in in that way for that task which is a very cool leap forward but i suppose it'd be good to hear from you jim around like this is a this is kind of a new approach so what do you think this this Who would you imagine would be using it?
Jim: Well, again, I love the approach. And again, I think since DeepSeat came out and a mixture of experts as an approach, I think this is then getting out and doing it on a US-backed company and from some alumni that are elite. And so I think they've got a great... I just, it just seems that they went a bit far. I don't know. It's a massive model, like 975 billion parameters. Sure, there's only 41 billion that are active with whatever expert you go. And so is it like, if you think about mixture of experts, it's effectively routes.
Alex: Through cells that can specialize to particular tasks.
Jim: That's the best way to do it. And so then that cell of training then handles the query. And so that's where they talk about 41 billion active.
Alex: But it's still got a 1 million token.
Jim: And again, I thought that was the best part, you know, is that you can have a mixture of experts, a million token context window. It's just that you're going to have to – it locks you into their harness. Sure, it's open weights, and you can fine-tune it on your own data and make it more – useful for your use case or more of an expert on your use case, but you're still going through their harness, their cloud hosting. And so you're into the same scenario you are with the Frontier Labs anyway.
Alex: And their harness bit is kind of one of the... critical shortcomings of this because more and more we're talking about how the harness is the thing that actually matters. The agents are becoming less differentiated. They're kind of converging around behaviors. And increasingly, what you put around the agent in order for it to do its task is what drives its...
Jim: downstream behavior in this performance and and again once you then do the fine-tuning and so once you give it access to all your data your alpha as as alex carp from palantir put it in his well-publicized rant now which again i never thought was that much of a rant but again i'm i've always been a bit of a different duck but i just i thought he sounded really rational but anyway it's more that they and so if you're fine-tuning this inkling so that it has all your data, all your alpha or, you know, like your edge in your industry, what happens, you know, you're effectively giving it all over to thinking machines, which again, you know, they've got a great reputation and, you know, it might... you could say that it's probably the best of the bigger labs. They've probably got one of the better reputations, but it's still a risk. If I'm a small or mid-tier law firm and I'm not that interested in the law-based AI products that are out there and I want to do something for myself on the way we handle contracts, Am I willing to give over my client's data to fine-tune and fine-tunely train this thing? I don't know.
Alex: Yeah, I'm looking at the results they've kind of put together this week and I don't know. I don't feel like this is as big as the press release and certain corners of the Twitterverse are making out to be. Sorry, the experts. Are the results that transformative over existing fine-tuning approaches? Like if you were just to get a copy of Mistral and do some fine-tuning, admittedly, a big Mistral model is going to take some effort to fine-tune. But on the one hand, constraining a model with that many parameters, you kind of need a load of data. And two... Is it that much better than the existing approach?
Will: I suppose I'd ask as well, the companies that do have the resources to do fine-tuning now, are they doing it and are they getting much better results from it? Because I haven't seen a lot of that and I haven't seen a lot of evidence of it. And so I suppose the value proposition of Inkling is that... It'll be easier to do it. And therefore, was ease ever the problem or was it ever just that the value wasn't there?
Jim: No, I think ease was a big problem. Like again, you know, even just getting access to GPUs to fine tune your own model. You know, like you've got to be a big enough player to convince someone to give you enough GPUs to do it. So this could be in the industry. There is actually, and the approach is fantastic. It's just, I just wonder about... Again, I guess that sovereignty element. And because if it's open weights, you think you should be able to run it yourself, but it's so big, you won't be able to do that.
Alex: Yeah. And like that's kind of, if we're talking about lowering the bar, it's got to be size and efficiency.
Jim: I've got to be able to buy an A100 for 25 grand and the servers surrounding it for another 25 and maintain it. in a way that it's then more cost-effective than just paying Claude. And this is where the decision-making, I think, comes down to it, is that is there enough of a benefit from the fine-tuned model? through an external party and take the data risk.
Alex: Yeah, and it really does depend on the use case, right? Like for full-on contract review or medical research or something like that, maybe there's some justification here. But one, do you have enough data in the world to constrain that many parameters? And two, do you trust the output enough of that sort of process that you could realize the ROI? I don't know, maybe cybersecurity or something like that. There's... empirical validation and domain expertise and the requirement for many more parameters to be thrown at this. But I don't know. I'm looking at the use cases that are getting ROI. They're all small models.
Jim: And that's the thing. I think it's just this is going to have a great use case in the future when the cost of infrastructure and compute is far lower. I just think it's a little, and again, they've also said this is their first model, this is their first, you know, the first thing they've put out. And so they're going to keep iterating and keep improving. And so I just think it's for the where everyone's at, they've gone, you know, five years into the future.
Will: Yeah, fair enough. Yeah. I also feel like they were silent for a very long time and what I expected them to come out with is sort of surprised me.
Jim: Yeah. Yeah.
Alex: Look, at the end of the day, it's good to have another player in the space.
Jim: It is. And also to have other options. Like that's the biggest benefit is that, and I like the idea of the small to mid law firm who has a specific way of doing contracts, you know, and enough. and 30 years of data, like sure, it probably still won't be enough, but you can get a benefit of it. Is that the approach you'd take now in your AI strategy and your enterprise engineering approach? Is the way you'd take it now? No. So I think that's where I say is in five years, this is going to be fantastic. And everyone's going to have a fine-tuned model for the way. Yeah, we might look back at this one.
Will: That's it. That was a real turning point, but it made no difference on that day.
Jim: I just think they're a little bit ahead.
Alex: When Transformers came out, at the time, it was like, hey, this is really cool. I don't think anyone kind of realized just how big they'd become at that point in time. Maybe. I'm looking at the number of parameters and the amount of data in the world and going, maths, maths. maybe scaling the ambition. Great as a statement of intent, but in terms of practical utility, I'd be saying less so.
Will: It's a great statement of intent.
Jim: Love an extra player, love a different approach. Fantastic.
Will: Great to see these guys get something out into the market.
Jim: That's it.
Will: Okay, so we also saw this week a company called WeCo, We Collaborate, who they take auto research and make it production grade. They were talking about having some evidence of recursive self-improvement. So auto research, first of all, for those who don't know, it's a... I suppose it's a way of using your AI to recursively sort of optimize itself where you give it a target and kind of some criteria and it will take different experimentation approaches, see what works and then double down on the ones that work and continue iterating on it. And they effectively took that program of auto research and then they pointed auto research at it again to see if they could get it to continuously optimize itself. So they've sort of suggested here that this could be recursive self-referencing. understand a bit more about what is RSI and how might it change the game with AI. So, Alex, can you tell us a little bit about, like, what is RSI, what is kind of the promise of it, and is this real evidence of it?
Alex: So, when we're talking about artificial general intelligence, like not AI, the marketing word, but AI, the thing that we've seen in science fiction movies, a key kind of part of that and the whole technological similarity narrative is this idea that eventually we'll get technology to a point where it can improve itself faster than we can improve it. And so certainly, since the early days of ChatGPT at least, there's been this mystique around LLMs that they might be a technology that have the premise to do this. And the rise of things like clone code, et cetera, definitely give us... signal that hey there's something here we can use this machine to make more of itself and and we have seen uh entropic talking about most of the code that goes into their new models being written with claude code now right yeah and and so like everyone's kind of looking at this and waiting for the moment where this technology gets onto this exponential curve where we put in energy like in the form of electricity as opposed to like human brain power to to kind of move things forward in Silicon Valley at the moment. Frankly, that's where these astronomical figures are coming from. Now, that's kind of what recursive self-improvement is. Is that what we see here? We've got an LLM effectively improving the prompt of another LLM. And this is, in some sense, recursive. It's one level of recursion, and it is self-improving. But this is not really what was promised in, like, Isaac Asimov.
Jim: And I think that's where we're a little bit, I guess, detached from it, is that we all have this idea of what recursive self-improvement will be. And it's a lot more that the model will realize something without us even telling us or without even setting up the environment. And then it will go off and spin up an instance and retrain itself so that next time the same scenario comes, it already has itself. And to me, and I guess, you know, that was a little bit of... And sure, this is, and they've even labeled this as the first step or layer one of recursive self-improvement. But to what Alex was saying earlier is that everyone's been doing this already with Chord Code.
Alex: There are frameworks like DSPy, which if you're not in the space, essentially it lets you use chat GPT to generate more chat GPT prompts and then score it and then... take the best ones and recombine them into better ones, right? It's essentially what we're seeing with this WeCo result. It's a DSPy system applied to a more complicated internal object, like the internal research agent.
Will: So what's the missing piece here then? Because it is recursively behaving and it is optimizing. So you mentioned it's not what feels promised by RSI. Does RSI need to have some element of creativity in it as opposed to optimisation?
Jim: Something that the human wouldn't have come up with is that I feel like this was set up, and again, they did 100 iterations and 93 of them didn't find a single thing. They only found seven iterations where there was an improvement. And so it's not a great... Success rate, if you want. But again, this is only the first step. And so I'm being overly critical. To be fair.
Will: I would almost say 0% would be a bad hit rate. 7% proves something.
Jim: It does prove it. But again, it's not what we felt RSI would be.
Will: The feeling of RSI.
Jim: That's it. Well, again, I want to feel like it's autonomous. Whereas if I'm running 100... And again, is that they did it with... Claude Opus, a really big capable reasoning model on Gemini 3 Flash.
Alex: Yeah.
Jim: You know, and so the size difference, like I think it was a little bit set up that no matter what it would find a better way of doing things. And they got a big, you know, they got like 63 down to 34 or something. Yeah. in the size of the prompt or something, whatever they were measuring. Anyway, I think I'm probably being a bit harsh, but it just doesn't, even to start calling it RSI, I think we're a bit early. That's the thing. And I love that they've got a goal and this is where they're moving towards. And they said it was layer one. I just think it's a bit early to call it RSI.
Alex: I think there's kind of two aspects to pull apart here. One is to do with the setup and one is to do with the outcome. And just on the outcome side, 7% hit rate over eight days. If they let this expensive you know perhaps we would have got a better result in time right this is this was a 7% and
Jim: humans have been working on a bunch of very smart people for a long time yeah two years yeah right that's that's not and again like i say i do think i'm being a bit harsh but yeah i just i want to call things what they are you know it doesn't feel like rsi because you still feel like you need to sit there and put ideas into it and that's exactly that's those step and that's sort of where that's
Alex: They had to set it up themselves. They didn't just set the machine up and say, hey, go self-improve yourself. They built something specifically to improve something else. And then we got improvement out of it. And that's not recursive.
Jim: And so, yeah, so to me, it's the same as artificial general intelligence. A lot of people are out there claiming, oh, we already have AGI. Or super intelligence, you know, we've already got ASI. Well, again, is it what I, or when I say artificial super intelligence, it's not something that I have to put inputs into. You know, like I feel like, and just to what you were saying then, you know, Will, about that it shouldn't be... if it was recursive self-improvement, it should be doing it itself in a recursive fashion, not where we've set up this scenario. No bot sitting. That's it. And so I love... To me, the best thing that came out of it was... if you hack the approach. If you're a small or medium-sized business and you've got a whole lot of processes running on AI, you could do the same thing. I think there is some value in paying the freight for Fable once a month or whatever and bringing it in and looking at, is there ways to improve these processes? Here's the results of the processes we've run.
Alex: um here's the prompts that were going in um you know what can you find is better and i think you'd get some results i think i think um a lot of people i've been watching people play around with claude code over the last 12 months or so and and you know some people do sort of go down the loops um rabbit hole and start connecting in custom workflows, custom scripts, et cetera, to kind of optimize their cloud code experience. But the extent to which, one of the takeaways from this for me is that perhaps we should be optimizing our harnesses a lot more from within.
Jim: But again, I think it feels a bit detached for us because the three of us do that. Whereas if you take the average punter, they're not doing it.
Alex: They're using it out of the box.
Jim: And I think that's the...
Will: The forward slash goal was a game changer for many people. And I just said, oh, it's just given me a command to do what my skills already do.
Jim: And I think that's the feeling of detachment we have here.
Will: Because I would also say like Hermes agent can feel very much like recursive self-improvement.
Jim: But that's been constructed. It has a process to go, oh, well, did you have to be corrected? And you can go into the... the repo and you know you can see the code that that gets all this and so i'm all for that being in the place but let's train the model to do that as well and then that's where you start to get to be recursive self-improvement to lean more towards agi on that one so and i think that's where we're going to go and if and if a transformer is going to get us to agi or asi or recursive self-improvement then that's the process it's going to have to go to But again, I think there's so many parameters and it would have to be some sort of mixture of experts anyway, where one expert or one or a cluster of experts is then dedicated to improving the overall performance.
Alex: I think the lack of...
Jim: Hang on, maybe we just landed on something there. Quick get thinking machines on the phone.
Will: AGI might be recursive self-improvement with the model being able to choose where to pay attention and where to go seek novel ideas when it reaches an optimization platform.
Jim: I'm certain that's the approach. That's how they'll be achieving it. It's just, yeah, when it starts to happen, I'm happy for everyone to start calling RSI. Yeah. I just don't think this is it.
Alex: Yeah, I think the big tell here is the disparity between the model that was being trained and the model that was doing the training here. we are still really trying to brute force this problem because we haven't found a substrate for these systems that combines, you know, easy trainability with vector spaces and sort of like actual knowledge representation. Like hallucinations are a mathematical inevitability. The improvement of these systems requires a really good mental model of... how the system actually works. Yeah. And until we can get much closer to a true mental model. Who is it?
Will: Yann LeCun talking about this at the moment?
Jim: Yeah, yeah. Well, but that's why he left. Other than being told to report to someone very less experienced and junior to him. Yeah. Was one of the reasons he said he was leaving Meta was to go and work on world models and actually solve this problem.
Alex: And I think that would be the space to watch that. The word's neuro-symbolic.
Jim: um because there's also there's another uh ex-open ai alum who's doing world models she's an absolute or she spoke at a couple of things and yeah like when you listen to her speak she's fantastic like you just go oh that's right that's where this is going yeah yeah okay well let's talk about the big one this week which was moonshots ai out of china releasing their kimmy k3 model uh kimmy k2 was pretty damn good coding model and um
Will: It was a bit of a game changer at the time. A lot of people were using it for that. Kimi K3 is a huge step forward from this. So the benchmarks are showing Kimi K3 as getting similar results to the world's best public frontier lab models like Fable 5 and GPT 5.6. So this is really quite a huge, it's being touted as a very huge win for software developers and the infrastructure companies and really anyone who's putting AI into their products. But it's also a potentially big threat to the frontier labs like Anthropic and OpenAI if you can go elsewhere and get a similar level of performance, let's say intelligence from an open source model. So... July 2026, is this really an inflection point for us, Jim? Are the Frontier Labs truly threatened by this or are they potentially still going to remain insulated by their products and their ecosystem?
Jim: Well, again, I do think their harnesses are, like, that's where they're truly leading. You know, is it the reason we all love Claude Code and Claude Cowork is because it's by far the best way to utilise those models. Is that, but... A big thing, and particularly with Kimi K2, the previous model, using that model or trying to implement that model is where I really learned about putting the guardrails in and how to build the harness properly so that you can use a cheaper model. is that one of my tasks for this week that I've assigned myself was literally to, okay, how do we now plug in ChemEK3? And what is the additional benefit I'm going to get out of? Because I've gone and learned how to put these guardrails in, and then I'm not as concerned about using an open source model, an open weights model. And so I think, and as, you know, we get a bit repetitive on it, but I do believe that custom harness for your own business or your own department is what's going to evolve out of this for anyone to get the greatest benefit out of AI. So being able to utilize these models and sure, Kimmy K3 is a lot more expensive, and I say a lot more expensive, than all of the other open weights model. Like, you know, it's a lot more towards the dollar per million.
Alex: Yeah, they're really making hay while the sun's shining.
Jim: And I think, and it's a lot more compute intensive. Like, that was one of the things you told me, Will, that I haven't even seen. But you can see how, like, and so you can see the approach they've taken to get the results. But if you're running it on your own local GPU, it's not costing you tokens in your only – or it's not costing you in OpEx tokens. Sure, you've had to invest in some CapEx to allow it to run in your own environment.
Will: You can also assign those benefits to your privacy and control now.
Jim: Absolutely.
Alex: Yeah. Yeah, I think when we look at K3, the – the critical thing here is not so much the model itself but what this represents as part of the ecosystem like we've had a week this week where another major like large language model has hit the market and now we've got now we've got a non-us um frontier model
Jim: Frontier level, that's it. I don't think they'd like you calling it frontier model, but it's frontier level model.
Alex: But it's very bloody cheap by comparison.
Jim: Comparatively, like $50 to 95 cents or whatever it worked out to be, that is game changing.
Alex: Go back to the sort of capital expenditures we were talking about concerning AGI, like what happens when, you know, Let's put aside AGI for just a minute because that really becomes transformative if that ever shows up. Big if. But the... The specifics of token pricing is still that you've got this sort of ability to do work without having a human sitting there doing the work, right? For poorly specified, you know, human-like tasks. This feels like the open shots of a price war.
Jim: I think it's going to have to. Did you see, because again, we're filming this on the end of Free Fable Day, is that they've already extended... you now get a whole lot of extra benefits or 50% more capacity on every model now. So Anthropic have seen this as a, oh, hang on, or whether this specifically, but all of these different things coming out, Anthropic are definitely actively keeping me on their harness.
Alex: Yeah, they're really trying to keep that customer base sticky and they've got an IPO coming up. how initially they were tightening up their limits to try and, to try and, you know, well, juice the goose for the IPO. And now they're on the back, it seems like they're on the back foot every other week.
Jim: Well, but that's, and, but I also think, you know, it's just how much of a benefit getting access to Colossus and Colossus 2 from Uncle Elon was, you know, like, is it, would it have, I, you know, I always like to think about Robert Frost and a road, you know,
Will: two roads diverge yeah what's the other what was the other path there you know was open ai just going to ultimately get it together and and go all the way over the top because anthropic couldn't have had the compute yeah is it yeah well there as well as gpt 5.6 is about half the cost of that that's it to run and that's i think there should be some questions raised about that is it are they just accepting lower margin to be competitive here or do they have some price advantage
Jim: But again, is that a smarter business model? Like do you need to have these massive margins?
Will: Well, at the model layer, perhaps not, but at the harness layer, where we talk about the actual ecosystem lock-in. So I think what's going to be interesting here is there was not really a large incentive even two weeks ago to move off, let's say if you're on Anthropic, to move off using Chlord models and the Chlord code harness. Because the model was the best and the harness was the best. Now where you've got something like Kimi K3 where you can use a truly great model, but the harness isn't good enough, maybe we start to see much more attention being paid to improving these open source harnesses to the point where you're now longer not having to sort of do this trade-off and you can just move off both the models and the harness of the Frontier Labs.
Jim: No, for sure. And that is still, to this day, it's the best recommendation Alex has given me is... was open router two years ago you know is it starting to build around switching a model from the outset that was um that was again i didn't realize it as much at the time but now i mean we literally just get to pick which model based you know based on the use case
Alex: I'd really love to see the development, like all these models that we have at the moment have their harnesses, which seem very baked around the idea of a single developer sitting at their computer. And a lot of them, they have their cloud alternative, right? They'll let you spin up the environment, but it's not, it doesn't look anything like an actual engineering harness. It's all very hand-wavy. Oh, well, we'll figure out what packages you need to install so you end up with all kinds of replication issues. There are standard tricks that we've had for 15 years now. Hey, give me a container and run my code in a container, please.
Jim: That's it.
Alex: I'd really like to see the open models taking the lead on that one and breaking away from the idea that you must have a subscription on your machine and you sit there in that loop actively playing.
Jim: But it also opens up more collaborative... Work environments. 100%. That's what I can't get away from with the single developer, single machine.
Alex: Yeah.
Jim: Is... no one's collaborating properly with it. And so the actual usability of learnings within a team, you know, because Will finds a really great way to prompt in the planning phase, it should be immediately transferred over into other team members.
Will: Yeah, I thought that would be here by now.
Jim: I was expecting that 12 months ago. I'm really surprised that it still hasn't come.
Alex: Yeah, there are ways you can do it. That you can build yourself. Yeah, it's very much home run. Like, okay, I'm going to add an exclusion to this command. Like I'm going to allow this command in my project. It defaults to your local machines path and it defaults to your local copy. You have to specifically say, hey, make sure everyone on my team behaves in this way now. So you end up with just like huge heterogeneity. works on my machine nonsense.
Will: I thought we got rid of this.
Jim: But also the specificity of the permissions. It's, oh, well, you can use the bash command for this specific git command. And that's all stored in the .cloud settings as well. It's quite interesting that they haven't taken that step yet. Wow.
Alex: Real men run in YOLO mode. Dangerously skip commissions.
Will: It's been wearing off on me, Alex, because I used it for the first time this week.
Alex: Oh, it's so good.
Will: And my life just felt so much easier. But I didn't recognize that it was on a completely new project that I was building from scratch with no production users. I would be terrified to run this thing. Because I'd come back to my computer and it would just be like, I've merged three PRs since you went and got lunch. Yeah.
Jim: But it's also, and that's the thing, that's the benefit of doing it on the virtual machines for me and having very different development and production repos. I don't think I would ever let anything go YOLO.
Will: What was the meme I wanted to send you this week, Alex? It was around just, you know, customer-led quality assurance. Yeah, yeah. 98% of our QA expenses and that we just respond when the customers tell us there's a problem.
Alex: So I once saw a joke. They're not complaints tickets. This is customer engagement. That's it. Yeah.
Will: All right, guys, let's talk about the costs here, right? So this is really the big question that's come out of Kimmy K3 is what's the actual cost here? We're talking about the cost per token. But Gavin Baker from Atreides Investment Firm did a post this week where he talked about the difference between the cost per token and the cost per task. And we're actually starting to see this become, I suppose, a bit of a conversation now around with Kimmy K3. the tokens are so much cheaper than the frontier labs however we've actually got some some benchmarks being shared or some some evaluations being done by companies like databricks as well but gavin baker said that part of their analysis they found that kimmy k3 was 50 more expensive on a task than gpt 5.6 through so much more that it came out far more expensive than just running GPT 5.6 on it. So we've also seen Databricks doing something similar here as well. So they're very much recommending that each of these companies should be running their own evaluation suites and doing their own benchmarking. And they should be doing that not just to validate the public benchmarks, but see how that compares against the tasks that they're running inside their business, which, as we know, are very different by industry, by business. So what the CTO, and I'll try and pronounce this correctly, Matei Zaharia of Databricks has noted was that, for example, Sonnet 5 costs far less than Opus 4.8. However, when they compared it on a per-task basis, to complete one of the example tasks that they had, for Sonnet, it cost $2.09 for the task, whereas Opus, it was $1.94. So it came out roughly 10% cheaper to actually use the bigger model. And Anthropic have been saying this for some time now. They've been saying, actually, we really recommend you use, from when 4.6, 4.7, 4.8 came out, so we start to sort of see this this mindset changing from you know cost per token or maybe cost per task alex you always had quite some passionate feelings about this when you've talked to me over the years around you know how we should be evaluating these ar models inside our own businesses versus just trusting what's going on in the public what what's the right way like if you're a cto of a large business at the moment what should how should you be thinking
Alex: Well, we worry about the cost per million tokens and we worry about them a lot. But a lot of businesses seem to forget that salaries of engineers is hideously expensive. I'm going to assume that we're talking about, you know, software development. That's one of the most expensive things that an organization can do. um you know to kind of run that in-house um as opposed to you know the more co-worky type tasks that summarizing notes or putting together a slide deck or something like that the economics don't quite stack up the same way um when we're when we're looking at the cost of tokens rather than thinking just in terms of um the straight line cost at any particular point in time, I think these frameworks are establishing the right way to look at the problem. That is, how much did it cost me to actually achieve some particular outcome? In general, I think that's correct. However, let's dig into those numbers a bit more. They're defining this around cost per task. A $2 task under 4.8, that's a couple of hundred thousand tokens. I barely get through a brainstorming session in a couple of hundred thousand tokens. So what's the actual unit of completion that we're talking about here? I'd be more interested to kind of see what happens... At the level of, okay, what does it take to ship a new feature in an existing code base to a customer? Because, sure, I can see that for certain very vague tasks, 4.6 may work out to be cheaper. I'd say it's probably more of a function in its context window and a blowout due to hallucination. Although, did they specify if it was the 1 million context window equivalent?
Will: Not sure I didn't say that.
Alex: Unfortunate. I think they need to. When I have used these newer models, I find they tend to be far more verbose and frankly piss around a lot more.
Will: but you went through one or two with some QA and you're fine.
Alex: And so the question there is like, does that QA time, the Alex having to jump back in and sort this out, offset the additional bloat that I have from having Fable running for seven and a half hours on whatever the hell it wants to do that isn't doing the thing that I asked it to do?
Jim: And again, I still think it comes down to the measurement. You know, like, is it... Like you were saying, you know, what is a completed task? Yeah. Is it what tasks are we doing here? Is it the Databricks, they assume everyone's filtering into the data lake? Yeah. You know, and they've got all their tasks defined and all the processes are measured and each step of the process is measured. And so you then start to go, okay, well, they're talking about something that is so far ahead of where everyone's at because... no one can actually choose, no one can measure their tasks yet. So how do you choose a better model?
Will: Interestingly, like even pre-AI, Alex was preaching to me about how companies just do not define their tasks or their success criteria. So how do you make comparisons and find where there are issues going on in your business process? And now that we have AI agents able to run some of these processes, we are effectively just scaling up our ability to do the exact same problematic approach that we did in the past.
Alex: This is where I think we need to define some unit of work that we recognize. And hell, not every company has software engineers. A lot of companies do not have software engineers in the house. But God, we need some benchmark of what a standard task looks like. Because if Databricks is getting... Spark clusters and that's working out for them. I want to see what their platform team's been doing for the last five years to make that possible.
Jim: But it's also, you know, like I think there's going to be a few CFOs pulling their hair out, you know, in the next 12 months because... Everyone has gotten so, and again, the Frontier Labs have gotten so good at convincing everyone that they have to use the bigger model. And, of course, Anthropic came out and said, oh, yeah, you know, you should have been using Opus the whole time because ultimately it'll come down to $1.95 versus $2.09. Yeah. But... They're just so much further ahead because no one's defined all the tasks properly, defined their processes, documented their processes in most cases. And so then how do you measure when the task is completed so that you know your costs are becoming optimized?
Alex: The amount of fin-offs here to just piece together the information, even know how many tokens you're spending to do something.
Jim: And that's it. Most businesses are still really only AI washing where they're not actually looking at the process and integrating AI where they can properly. They're just buying everyone a cord subscription and being done with it.
Alex: Yeah, and that issue of AI washing, do we want to jump into that now?
Will: Yeah, a conversation coming out of this this week was from the BCG report where we're saying 80% of CEOs were optimistic about their AI investments at the moment. And the blog post was on Substack from the product journey, and they were talking about this number. in and of itself where CEOs needed to claim that they were optimistic because also 50% of them felt that their job depended on the success of that AI implementation. And it's not possible to be a CEO sitting there saying, my job depends on this, but also I don't believe that we're going to be successful. I'm quite cynical about it.
Jim: Well, that's it. I think a lot of people, and that's why I say, you know, the CFOs are going to be pulling their hair out because everyone's going to be running in a direction telling them, oh, yeah, no, this is all working so well, whereas really they're just AI washing everything. And the poor CFO has got to pay for everything because they're then going, oh, well, you know, how can you measure that using the frontier, that you have to have the frontier model set running versus the Kimi K3 because no one's willing. And so they're all willing to, and again, when something goes wrong with the models, they're very quick to blame AI.
Alex: Yeah.
Jim: Whereas...
Will: There's a lot of great excuses for your AI fails at the moment.
Jim: You've got to take a look in the mirror. If you... What's your quote? If you can't measure it, it doesn't exist. I think that's how everyone's got to start looking at these things.
Alex: There's also a lot of naivety around how fungible these things are. When I'm speaking to IT leaders and finance leaders, they're going, oh, well, we can just swap in these. the place of cord code and it's like no you can't yeah i mean hell look at that you know most of us work in normal companies where it maturity is somewhere between the 1990s and about five years ago of the cutting edge right like we're not living in some tech hyperscale at this point in time we struggle with basic things like email migrations and databases going down and and in many cases, moving things off on-prem servers into the cloud. And we're talking about swapping out LLMs to try and get cost advantages. We just don't have the operational muscle memory for this.
Jim: But also it's just a complete lack of understanding is the other side, you know, is that because everyone's... Again, I'm now obsessed with the term, but now because everyone is AI washing and just going, oh, well, the Claude subscription for all my team, oh, we've got AI, they fundamentally then don't understand what it takes to actually implement a Kimi K3 properly.
Alex: This might be something worth unpacking. We've been throwing around the word AI washing for a little bit. We haven't defined it.
Jim: Give us your definition. To me, it's telling everyone or not really looking at how to implement AI properly. Yeah. And just washing everything, washing all the dirty things away and saying, we've implemented AI and it's all fine. But all you've really done is put a chat bot in on your website and it's an off-the-shelf integration into your Shopify site.
Alex: Yeah.
Jim: Or you've bought everyone the Claude Code, and not even Claude Code, Claude Cowork connection, or you've implemented Cowork within Microsoft 365. And you then tell everyone that, oh, well, no, all our team is using AI and all our... all our processes now involve AI.
Will: I think that's okay because that's claims around AI adoption, which I wouldn't call AI washing. I would say it's that very, very last part, which is then attributing things that happen in the business, the amount you're investing, your successes, your failures. to that AI program.
Jim: And that's sort of where I was then going with it because then once you get to the KPI point where everyone's now getting into, okay, oh, well, everyone's had all these. I think it seemed to be a brown March, April this year where everyone was then saying, oh, well, we bought everyone a Claude subscription. Yeah. And that's where I've started joking about it because it's just such a response I hear. Every week I hear someone tell me that same thing. But now they're getting into this point where they're doing their quarterly planning, they're setting their OKRs, and all of a sudden now, oh, well, look how well we've done. Yeah, what do you... Sure, everyone's a little bit more efficient, but what are they doing with the rest of the time? Where's the 30%?
Alex: That's it. Where's the 40%?
Will: That's it. So what's the actual dangers that you see then, Jim, of AI washing? What's the second order effects that we're going to see in businesses, in teams, and in the market?
Jim: Well, but ultimately you have to get an ROI, right? And that's the fundamental thing about anything in business is that there has to be a return on the investment, right? And so if everyone is demanding a token allocation as part of their salary and all these sorts of things, is that what's the benefit to the business? And that's not what's being, none of it is being mapped at the moment. And so, yes, we spent $17,642 on cloud subscriptions for the whole team for the year. Did we get that much more benefit or did it just make everyone's jobs easier?
Alex: Can I propose two alternate facets of AI washing here that I think explains where this is going? The first is AI washing as cover for things you're already going to do. And we've seen some of this already in, for example, layoffs.
Jim: Yeah, layoffs is a great one.
Alex: These were often on the back end of COVID-era hiring splurges or enhanced consumer demand. And... companies scaling stuff in and being able to attribute it to AI, you know, it sounds a lot less heartless than AI.
Jim: That's it.
Alex: And the second thing is AI washing as a coat of paint that justifies budgets. And it's kind of like the slapping on the cord co-work, but I think it's somewhat deeper. Anyone that's ever done a digital transformation should have this come to mind almost immediately in that by not rethinking how your company works or designing a company such that AI can actually drive those force multipliers, of course things are showing up as a lack of ROI, as a lack of moving the needle on KPIs. you're slapping new gauges and new parts into old factory machinery.
Jim: That's a great way of putting it.
Alex: And nothing's changing and you're wondering why. If you want to redesign a factory, you have to redesign a factory. You can't just tinker around.
Jim: And I do think even for companies who are implementing, like, you know, really having a look at the process and trying to automate it or, you know, intelligently automate it with AI.
Alex: Yeah.
Jim: I think that, you know, the re-looking at the whole way it's done is probably the next step. You know, like it takes a lot to go, oh, well, the humans do it this way. The AI can do it a completely separate way. I understand there's a gap there and that's going to take some work. And so that's where I go, okay, that's the next step. But you still have to have some sort of measurement process, some sort of did the process work, how long did it take, all going into- You can't just retrofit it in blindly. Exactly. And so then once you then measure all these things, then you go into a monthly optimization cadence, you then- we'll start to see how you can do it in a better way.
Alex: I kind of disagree here only because I think the first step should be like not so much trying to improve what you already have, but trying to streamline the process. want to get rid of first um and get rid of here may mean you know putting in an agent to do some stuff it may just mean stop doing the bloody thing that stopped having a reason to do it 10 years ago right like before and the way that we seem to attack these problems in organizations is we look at what we're doing today and we go oh how can we do what we're doing but faster we're doing and going, how would we do this differently if we didn't inherit the history that we have? And then trying to map towards that. But I'm a dreamer.
Jim: No, no, but and also particularly and, you know, for what I see week in, week out, it's how can you actually deliver more value to your customer? 100%. And that's what a lot, you know, I see a lot. And that to me should be the first.
Alex: Most people in a workflow can't even describe how what they do. do fits into the bigger picture of the organization.
Jim: Absolutely.
Alex: This is in small and medium companies. God knows how they solve this problem in bigger companies.
Jim: But again, it's what's going to hold them, I think. 100%. And again, back to Will's original question, what are the second and third order effects we're going to see? I think we're going to see a lot more of what have been – institutional enterprise companies start to really have a look at what they're doing, how have they invested, have they invested in the right way, did they take the shortcut or the quick fix, AI wash everything for 18 months, and then all of a sudden they're sitting at the end of the road.
Will: If they're almost forced, these public companies, into AI washing, If all your competitors are claiming all these benefits from AI and you're not, how that's going to affect your overall share price and your forecast. Some of the realists like us would probably say that it shouldn't affect your forecast because you should be held to account based on actual real measurable results. But if it's really affecting the market that much, you might be forced into rushing some announcements to market before you've actually demonstrated that.
Alex: seeing the benefits already today they're the ones that have been doing the right thing for a long time where narrative has not been the primary factor where they've really been engaged with their customers understood their market sharp leadership with a deep understanding of the vertical in which they operate all that comes down to your organisation's ability to change and if you don't have that ability to change and it's actually the muscle you talked about earlier yeah
Will: then all you can really rely on at that stage is narrative and spreadsheet management.
Jim: But it has to be a bit of leadership. Like that's the thing is I go, my mind, the moment you start talking in that way, my mind immediately goes to like the CEO of Box and even... Mike Cannon-Brooks at Atlassian. They took the hit, you know, and they took the hit badly. Everyone was coming for them. But is it what they did? They went back to first principles and looked at everything, rebuilt, effectively rebuilt both of those products to be AI native. And... Again, Atlassian's sort of on the way up again. I think it comes down to you've got to have that leadership because, yes, there is just going to be a lot of people looking at spreadsheets and going, that's bad. But you've got to – someone has to be the champion and paint an S on their chest and walk out the front and – suffer the consequences if you fail. That's the thing is that we were talking a bit about it last week with the different management styles. And we've created quite a few people who aren't, the majority of middle management, aren't willing to take a risk because it will look bad on their LinkedIn.
Will: The culture doesn't support it.
Jim: That's it. It'll look bad on their LinkedIn for five minutes and they just get ostracized in the culture. And so, yeah, it's interesting.
Will: Okay, well, let's now move to AI regulation. We'll hit a couple of these topics pretty quickly here. So the first one is Google DeepMind CEO Demis Hassabis has endorsed the creation of an independent AI standards body in the United States to regulate these frontier AI models. similar structure to the way FINRA works, which is effectively the industry polices itself with a federal supervision over the top of that one. Gary Marcus, one of our favorite kind of bloggers, has had a bit of a victory lap on this one as he's recommended that in 2023 in his Senate testimony. And really what that looks like is potentially something like this kind of pre-flight testing the standard set, and have to report against those standards before you can release it publicly. And something like that makes it look like what we have been doing is pretty cowboy in the past, right? Just releasing the models straight to public with the old Google Gemini, Black George Washington popping out.
Alex: What do you mean in the past?
Will: It's a new world already. It was recommended, so things are changing. Quick one for you, Jim. Is this the right way to be approaching AI safety?
Jim: The right way is a, maybe not the way to phrase it. I think it's the best way at the moment because, again, you definitely do not want a government department doing it because all that would happen would be, you know, it would start out that there'd be a couple of weeks delay on a model release while they audited it or did their compliance checks. it would very quickly turn into months and months and months. It would just create a backlog and you're not getting the right people doing the checks. You're not getting the best of the best looking at these sorts of things. And so the FINRA model works really well because, again, is it... Everyone's a member or all the financial services companies become members. And so they're accountable to it. They all get votes. They all have a say in what the determinants, the compliance things are. And then their own employees go and do some time at FINRA. And so it's a great approach. And then if anything actually needs to be... investigated or researched, the FCC comes in on top. And so it's a great model. I'm a little concerned about the way it works for AI because the cost of one trader in a financial services firm going over and doing some time in FINRA isn't that much. The cost of an AI engineer... leaving Anthropic at the moment is far too high. So are they going to allow them to go and do the work at the self-regulated organisation? But, yeah, and so I think that it'll take some sweating out, but I think it's the best approach we could possibly hope for.
Alex: Yeah, in general, I'm really torn on this one. I think the idea of something to ensure stability in some of these models as they kind of become critical infrastructure in parts of companies is not inherently... The words non-tariff barrier is coming to mind here. And the EU does this all the time with its sort of restrictions on non-EU trade, right?
Jim: You get the sense that they're pulling the ladder up to the treehouse. Yeah.
Alex: I think there's enough cheap alternatives coming onto the market. Again, if there's not these non-tariff... The whole AI financing model falls apart. They need something to protect them from competition, particularly competition coming out of China, but also coming out of the EU and smaller sovereign efforts. A board that they control that slows down and prohibits the release system is a fantastic barrier to entry to that market. I think that's what we're going to see from it. I don't think this will be consumer friendly at all.
Will: Do you think that there's more likely to be stricter controls if these, if Entropic and OpenAI are public companies or remain private?
Alex: I don't think it'll be about that at all. I think it's geopolitics. It won't be public-private. It'll be down to what country you're from and are you friends with us at the moment. And again, this is historically what we've seen with EU trade policy.
Jim: And it was very clearly defined as American-led.
Will: Yes. I do get a little bit concerned. I don't know if it's related to our regulation, but in the world, once OpenAI and Anthropica are public companies and therefore have joined the key indexes and everyone's got a large portion of their ETFs and the superannuation in the States. 401k. 401K, E, Roth IRAs, everything. Yeah, now being invested in companies that aren't profitable and have these huge price and equity valuations. There starts to be, I think, a lot of movement happening in those sort of political spaces to make sure that there isn't a large amount of risk being created by these companies on the overall market and the overall economy. And in that case, everyone who has a...
Alex: stick where they can wield around risk is going to start wielding the stick well i mean we've seen the same thing in australia with the property market right like there's so much wealth tied up in that that the government hasn't been able to make proven decisions for the better part of a decade.
Jim: Probably longer than that, yeah.
Alex: Yeah, certainly back to the 90s. There's just too much political risk and then financial risk and then institutional risk. The super funds getting involved, much like in the US. has the potential to detonate retirement, which would just melt the political order.
Will: Which totally then changes the political position that they're in, right? And we will absolutely see that once these
Jim: Yeah, I think they have to now. Particularly open AI, it has to, the way in which it's gone about, just the way it's organised everything.
Alex: But who buys that first IPO round? Because if it is the large public pensions, that's going to be disastrous for the US. But the IPO price cannot be sustained.
Jim: But if every – yes, if everything goes wrong. But it's still – that's the thing. That's where I think is that a lot of it will turn back to that product and that harness and that. And so this is where I think, yes, governing the models is probably a good thing. It just – I never want it to – and it's back to last week about semi-Altman, you know, voluntarily giving 5%. You know, is it – you've got to be careful about the fox in the hen house. Like what's actually going to happen. You don't want it to be the government deciding these things because all it's going to do is slow everything down and then restrict it in a way that only the little guy gets hurt.
Will: Wow. Talking about that, Jim. The Australian government has announced a national AI framework, which they are aiming to have part of law by early 2027. This includes a couple of things such as the creation of a new AI office, some new rules for approvals around data centres, which seems to be one of the large components of this piece, which is things such as the data centre companies paying their own cost to connect to the grid. They need to be able to show efficiencies for how they're going to use energy and water. We've also got some copyright components in here just around restating around AI training on copyrighted materials. And then lastly, some guidelines around the safety of consumer AI and the frontier AI.
Jim: Jim, is this a good step forward for our country? Yes, in that there needs to be some sort of AI, you know, regulation or AI supervision. Yes. However, the approach that the Australian government takes always just hits me in the pit of my stomach because, again, it seems like they've had this great announcement which changes the media cycle. So we're now talking about something else, elbows out there on four corners, talking about something else other than what he said on a podcast. whereas they haven't properly thought it all through yet. And yes, I agree. We need some regulation for data centers wanting to come in. Is it, you know, there's some advantages for Australia? having data centres to serve a lot of Asia because, again, it's sort of you can get to billions of people in 200 milliseconds and there's a lot of space here, as it turns out, you know, and the access to a lot of energy. However... But all they do is just – all the Australian government does is say, oh, well, you have to do these things. They don't define it in any way.
Alex: Yeah.
Jim: And so no matter what, in the way in which they announced it and the way it is in the announcement, no matter what, those costs get passed through. And so, yes, the investment funds are going to create data centres in Australia, but the costs are going to get passed through. Same as with the consumer protections compliance stuff is that – Yes, meeting compliance requirements in Australia is an important thing to do when you're a big company offering an AI, consumer-facing AI product. However, one of the big advantages to AI is that the one-man shop on the corner can also integrate AI to get some efficiencies. If it's based on the way the announcement was made, if it's consumer facing, they have to meet all the same compliance. And so it's just very, very difficult for the small business to do that compared to the big SaaS style. AI product.
Will: So you might want to see something like the way the ESG regulations are rolled out in stages for the size of the company.
Alex: Slavery provisions. There's like some market capitalization.
Jim: And the big one for me is with the younger than 16 social media. It wasn't we want you to definitely do it in these ways or this is the way it has to be compliant. They just said, oh, well, the big AI companies and the big social media companies, they've got all the infrastructure, they've got all the technology, they'll work it out. No, you should have gone and engaged with them and gotten them to say, okay, well, these are the ways we can do it and then created law to say you have to do it that way. You know, and so this is sort of where I go, the fact that Australia, the Australian government doesn't seem to think it all out before they do the big announcement really concerns me because the ambiguity in meeting the compliance if it's consumer facing is all just going to fall to the SMBs can't afford to implement AI.
Alex: Yeah, I think a lot of previous legislation doesn't give great hope that this will be done well. That's it, yeah. For those of you who have never met Jim, when he gets a bottle of wine into him and starts talking about the government and small business, it gets spicy very quickly. But in this particular instance, I'm 100% on board with you on this one. I'm not going to argue any point you've raised. I will say we need to watch that office that they create.
Jim: No, no, and that's the thing. I just want it to be that I think it needs to happen. We need to have something like this, but I would much prefer a lot more engagement. It wasn't Four Corners, but it was on the ABC, and he was literally – and he's a Labor – Prime Minister, supposedly for the unions and the people.
Alex: Yeah.
Jim: And he's on there saying, oh, well, workers are just going to have to accept the fact that AI is coming. That's not what I want to hear from him if I'm a member of his party. And that's the thing is that I think I just, I remember before he was elected, he gave an absolutely amazing speech at the National Press Club. And I went, okay, that's it. He's absolutely going to be prime minister. It was a fantastic speech. But it was – from that point, I just think he's compromised and compromised and compromised just to try and stay in the good books.
Alex: Yeah.
Jim: Whereas back to my – if you've got good leadership, you can actually win scenario. I think his speech in the press club was absolute leadership. He's now got to, oh, well, how are we possibly going to get elected next?
Alex: I'm just worried about this becoming another cudgel where the government can essentially say, to do X, Y, Z. And this is why I say watch the office because if the office has no enforcement powers, but it does have powers as a loudspeaker, then it's a political weapon.
Jim: All it can do is make noise. And that's the thing.
Will: But I'm also worried about enforcement when their policies are ambiguous.
Jim: And it's the ambiguity that really gets me because I just get iron-rimed. flashbacks you know the fountainhead is it all of a sudden it's just oh well like you say that there's these things that you can do but you can't do them and there's no definition in it there's no proportionality that's it no like conducting assessment figure out how much of this actually applies to
Alex: If you found a material in breach, obviously it will come for you. But conduct an assessment, do some benchmarking, here's some information.
Jim: That's it.
Alex: Reasonable. And I don't think anyone would complain with that sort of legislation. Everyone knows you can't legislate for every circumstance.
Jim: No one's asking for that. And that's sort of where I come with all these sorts of things, is that GDPR data... are actually quite easy to meet, right? And I don't know why people just don't do it from the outset. Why would you ever have to validate that you meet GDPR? Because you should just be doing it anyway. But it was clearly defined what you can and can't do with consumer data. So why not just set those sorts of things out in terms of AI? And that's what I want to see.
Alex: To come back to the EU, right? While the EU can be... a nightmare on the train front, a lot of the good it's done has been setting reasonable, concrete consumer standards that if you want to engage with the EU, you must comply with. And that's spread through a lot of, like, other consumer products. The most famous one is standardisation of phone chargers, right? Remember when every phone had its own charger style? But, you know, that sort of normative effect on... you know, taking the lead and ensuring certain minimum standards are met Anyone in Australian politics or technology or investment would complain about them.
Jim: But I think it then costs Australia in terms of investment, you know, like particularly financial services. Like, again, the financial services reform of whatever year that was, was, you know, like something that needed to happen in Australia. But they went so far and it's so ambiguous that most of the big investment products in the world are not available to Australia because they do not want to meet the compliance. And so it's the same thing is that if we start to get left out of AI products and have access to these models because they do not want to go through the rigmarole of meeting the compliance, then I understand their use case, but it really, really diminishes the capacity for Australian businesses to benefit for it.
Alex: Yeah, and to sort of play devil's advocate to my own point, I do have friends that have built companies in the data space around helping people do data engineering tasks that will not engage with the EU because the specifics of EU law don't really work for third-party data processes that are there to automate parts of your data engineering function. And that's sort of like... Lack of tech awareness or legislation pinned at a point of time where technology keeps moving does have that potential to kind of like hold us back and prevent innovation in the country. It could be a really good thing, particularly around copyright, but I don't call me particularly optimistic on this one. No.
Will: All right, guys, well, let's wrap up this week with our lightning round. I want to get some quick opinions from you guys. So first one is Macquarie Technologies announced they're building a $3 billion data centre in Sydney.
Jim: Jim? No, good, and they had the land picked out for a while. Is it they executed the option, paid the money? Is it they've seen the opportunity? And again... Good on Macquarie Technology. It's been a really good fund. They've done some great things. I think they'll keep doing it.
Alex: Yeah, as long as they protect the water supplies, yeah, no problem with this.
Jim: Oh, yeah, no, that's it because it's right next to that wetland thing.
Alex: Yeah, and, you know, like we don't have quite so many rogue operators in Australia as you get in the US. I think a lot of the fear-mongering, well, it's not fear-mongering. There is legitimate fear around environmental hazards coming from these centres. We've had data centers in Australia for a while. We tend not to have too many.
Jim: But again, Australian owned, like again, every, you know, well, again, I guess Macquarie investment is very much a boomer, highly successful Gen X millennial sort of thing.
Alex: Yeah.
Jim: But it's still Australian owned data centers that aren't, that are governed by Australia, not by offshore interests. Yeah. Fantastic.
Will: I want our data centre industry here to thrive. So I'm just hoping that these build out something at the moment are all quite successfully and are good for both the economy of the country and local communities so we can continue to thrive.
Alex: I do worry that in the past, like some of the data centres we've seen here in Queensland have not been economically viable.
Jim: But it's before. I think it's also the potential scale now, particularly of this one in Sydney. And again, is it... In Australia in general is that we can serve 2 billion people in 200 milliseconds.
Alex: In a geographically stable region.
Jim: Geographically and geopolitically stable region that has a lot of energy and a lot of space. So I think it really should be exactly what you're saying, Will. We want our data centres to thrive. We want it to be smooth and efficient for that investment to happen.
Alex: Just don't drink too much water because that's the one thing we don't have in the world.
Will: Okay. Apple is suing OpenAI for theft of top secret information.
Alex: This one's fun. I think the OpenAI is running out of work here. It has been... making a lot of promises and cutting a lot of corners. And the fact that Apple is taking this one along is more than really fun.
Jim: That's the one that got me is that Apple has been particularly not litigious over its whole existence. for them to actually go, we're coming for you, they know they've got something.
Alex: Relatively hot off the press, Oracle. And someone else who OpenAI does quite a bit of business and there's been some gray areas there. Oracle's just found itself in hot water. And they are notoriously litigious. So let's see what comes out of this.
Jim: Again, it will be very interesting, but particularly with the Apple one is that it's just... I never understand. When you're leaving a company, just leave the company. You can't ever take anything with you. And for the tweet or the text message...
Will: The evidence I saw was that OpenAI asked them.
Jim: No, no, but that's the thing. Coached them, I think. Coached. But it's the ex-VP of product at Apple is now at OpenAI and they've hired 400... people from you know from apple and there's just somewhat clear um or from what all the reporting like i haven't actually read the lawsuit or anything but from all the reporting it seems pretty clear that they were coaching it and then the engineer that was like the text messages came out where he goes oh lol i've still got access to the network drive You just go, buddy, come on.
Alex: The one thing I'll say there is what was Apple IT doing? Like, come on, just take their laptop from them.
Jim: But it's also, even if you do want to, you know, take a copy of everything, like for your own, keep it for your personal records. You can't ever, you can only go to a new job with what's in your head. But what's in your head is yours. Just stick with that. I think, I think, If Apple have sued them, Apple have got them. And it's another headache for OpenAI going into an IPO. Another headache. And you just start to go, how complex do you want this investment to be? Back to what Alex was saying before about who's going to invest in that first round.
Will: I don't know how a Wall Street analyst is meant to assess all these different ongoing risks and the unseen ones. Last topic Alex touched on, so the S&P cut Oracle's credit rating to near junk status, citing OpenAI as the key risk.
Alex: Well, Larry has a fantastic law firm with an attached IT department. This ends in very interesting ways. I'm hoping this goes to discovery and we get some look at the open AI financials because I've seen a lot of hand-waving and a lot of smoke and not a lot of fire.
Jim: Again, I think open AI can't afford to have too many – you can't fight a war on too many fronts. Napoleon – taught us that really clearly. And so I think they go, hey, Uncle Larry, we'll sort it out. Sorry. What can we produce to improve this for you? We'll just raise another round and sort you out. And speaking of another round, Uncle Larry, can we please have some more money?
Will: In the form of some circular equity, I'm sure. However it comes, however it comes. All right, guys. I feel like maybe Business City of the Week might be an easy one to pick this week. Gentlemen, what are we thinking? Again, it's... I reckon Jim might come back to the Australian government on this one.
Jim: Well, again, I would like to say, but again, I'm yet to be shown whether the Australian government is... Because it's Prime Minister and Cabinet as well, so it is actually Albo. And so I'm yet to see whether... It comes out as idiotic as it's more just my internal fears that it's going to be. At one point, the Australian government will be the business idiot, but I'm pretty sure it's the head of EP, the ex-head of EP of product from Apple who is coaching people to bring all their hardware, bring examples of hardware over and do a show and tell.
Will: Yeah, I'm agreeing with you on that one. That's got to be my business idiot of the week.
Alex: I'm tagging that one engineer who sent a text message. You reckon it's the engineer, not the head. The VP was at least sneaky about it, but sending a text message in plain, come on. And then touching a network drive. Sorry, that's a lot.
Will: That's business idiot. No, no, fantastic. Excellent. Thank you. That's another week. We'll see you next week. Cheers, guys. Bye-bye.