Chapters
Questions from this episode
- Why was Firmus named Business Idiot of the Week?
- The hosts named Firmus as Business Idiot of the Week over its withdrawn $7 billion Australian IPO. Jim Lovell blamed co-CEOs Oliver Curtis and Tim Rosenfield for mistiming the float, while Alex Stenlake singled out Curtis. In the hosts' view the float came too late to win the valuation, and the deal was ultimately pulled.
- Why did the hosts question Firmus's $7 billion valuation?
- Will Turner argued that AI data centres should be valued more like long-term infrastructure than high-growth software companies. The hosts questioned whether Firmus's demand story, Nvidia backing and major contracts justified the proposed valuation. They also discussed CoreWeave, circular financing and debt as warning signs around AI infrastructure enthusiasm.
- Why has Anthropic banned cruel behaviour towards Claude?
- Anthropic updated Claude's usage policy to prohibit sustained and needless abusive or cruel behaviour, with Claude able to end a conversation. The hosts questioned how sustained behaviour would be tracked, whether monitoring could affect privacy or model quality, and whether talk of Claude's consciousness reflected moral caution or marketing.
- What did the hosts make of OpenAI's 722 mathematical proofs?
- OpenAI published 722 AI-generated mathematical proof manuscripts covering open problems, some verified with Lean and many not. The hosts noted that some results later required retraction, and argued that producing proofs by brute force is not the same as creating useful understanding. They framed AI as a tool for mathematicians, not a replacement.
- Why does Meta's Muse Wonder raise advertising concerns?
- Meta's Muse Wonder and Muse Wonder Full were presented as competitive frontier models, with active memory designed to retain details across conversations. The hosts saw potential in Meta's Facebook and Instagram data advantage, but worried that personalised recommendations could eventually serve advertisers and make paid suggestions feel like trusted advice.
- Could agent-first shopping weaken Amazon's position?
- The hosts argued that agent-first shopping could weaken Amazon's advantage as a single marketplace. If agents can compare products and buy directly across retailers, breadth of catalogue may matter less. However, they said Amazon's fast shipping, Prime delivery economics and pricing power remain a substantial moat, while predicting more API-based purchasing.
Read the full transcript
Alex: Hello again and welcome back to the Business Idiots podcast.
Will: I'm your host, Will Turner, and I'm here with Alex Stenlake and Jim Lovell. How are we? Hey, how are we? Yeah, very good. Thank you, guys. So I'm actually excited this way because we've got some real news that's happened in AI this week, not just kind of AI safety publicity stunts we've seen over the last few weeks. And it's been hard to keep talking about that one. So now we've got some real news this week. So we're going to be talking about the attempted IPO in Australia from Firmus. We're also going to be talking about the new Muse model, Muse Wonder, and also the possibility of Claude's consciousness, because we're starting to see more chatter out of Anthropic about that, and is that real or is it not? As well as another wave of AI math proofs coming out of an AI lab. So a big kind of docket today, guys. So we'll jump straight into it. So we'll start with Firmus this week. So a year ago, this was sort of a mid-sized Australian data company. They had a crypto mining pass. They started as a Bitcoin miner and were valued about this time last year at $1.85 billion. And this year, they just tried to IPO at $7 billion, which would be the second largest IPO in Australian history. Now, these guys are backed by Nvidia, Coatue, Blackstone and Jane Street. So some really big names here. But effectively, what happened was through the IPO process, once they opened the books, there didn't seem to be the numbers were stacking up the way that kind of the valuation would require. So I think the market started asking a lot of questions. The valuation was kind of the price was cut from $11 down to $9. And then come Friday, they totally pulled the deal off, kind of took the IPO off. So, yeah, a bit of a journey on this one. It'll be interesting to sort of see what the fallout is with what kind of happens here. And I know that they're also talking about now potentially launching their IPO in other markets such as the United States instead of Australia. So I suppose an interesting question here. So Jim, maybe throw this one to you first, mate. So do you think that this is kind of more indicative of a problem with Firmus itself? Is it an Australian market problem? Or do you think that just maybe the timing of this IPO is just a little bit difficult right now, given that investors may be getting a little bit cautious or asking a lot more pointed questions around AI demand and AI infrastructure?
Jim: And I think it's a mix of all of that when you break it all down. But one of the big ones for me was how much, and this is a bit of a, you know, there's these anecdotal parts of Australian investment where it's always just the same avatars or the same people. And it's quite amazing how well this one, this one really links to all the big names and the big players of, you know, Australian investing, Australian business. It basically links back, you know, one of the CEO's dads, Oliver Curtis, his dad is like a mining magnate, you know, really big in the mining industry. But then that just gives them connections to everyone, you know? And so the Packers, the Prats... Even, like, the one I lulled a bit about was former treasurer Joe Hockey's major, you know, big investor sort of thing. But it's just, it's this Sydney in crowd, and that becomes how everyone gets in on it. And it seems to be, or I've seen this pattern a number of times. over my life, over my career, where it's, oh, all these people are in and everyone's got a piece. Oh, I've got a piece of this. But none of them actually really understand what's going on. Maybe the dots on the I's and the crosses, the T's haven't been done to the level that you'd expect, but it's all just, oh, mates, mates, mates, it'll all be fine. And so that's sort of where I think... You know, I was saying it was a part of, you know, there were parts of all of it. One of the big ones is timing, you know, is that I think if this had gone out six months ago, there wouldn't have been a question. You know, it just would have, it would have been smooth. It would have gone out. Everyone would have been happy and they'd been off to the races. You know, like they got, they got, I was trying to find, I remember reading that they got a big, you know, was it meta or? I think it was Meta contract. You know, it could have been the start of this year or end of last year. And like that alone, then going to IPO, boom, you would have gone straight out. But I just think that because everything is now tightening up, everything's a little, everyone's a little bit cautious, like you were saying, Will, you just get to this, oh, hang on. We check everything a bit more. And that's where the Sydney Mates Club doesn't really stack up always.
Alex: I find it really interesting that the Bitcoin miner to NeoCloud pipeline is really strong. And there's a real trend of like... dodgy books in this entire flavor line.
Jim: But you add on top of that insider trading jailbird. Yeah. I mean, like...
Alex: Was that Oliver Curtis that was involved in that? Yeah. And like, I mean, to be fair, like Jonathan Levy developed some pretty cool technology as part of this process. This is not, it's not pure vapor and none of the Neo clouds are, but holy crap. If I'm an investor, I get that AI is really hot. I get that a lot of people think there's just this infinite demand for tokens and it'll be infinite sort of forever. But like, yeah, do people just forget the basics of investing? Particularly when the wave of like hot, like zero interest rate money sort of pulled back. I thought everyone started to remember, oh yeah, that's right. We need to worry about profitability again. We need to worry about business models again.
Jim: And that was the same, like it was almost the verbatim model for this crypto contract I had back in the day where, you know, we were building them effectively a financial reporting and, you know, asset management process you know and It was just all these mates in together, all these, you know, not dodgy, but they got a massive valuation, a massive amount in actual cash, and they didn't worry about where they were going to get money, you know, actually make money from or turn it into a business. It was just, oh, you should be right, mate. And I wonder how much of that is the way the world looks at us because of these sorts of things.
Alex: I mean, like, it doesn't help you got, like, people like Sam Bankman Freed running around with his, like, just insanity. Or, like, Martin Shrek back in the day with his, you know, his farmer sort of schemes. there's a reason the technology industry has a bad name and like i think investors need to put their mitts up and say look this is as much on us as it is the people misbehaving because we're enabling this well i started to wonder how much you know or how good roxy jacinco really is at pr and marketing just that you know is it that because she is an absolute god and that's why she's been so successful
Jim: But that she managed to keep all this together in terms of the message going out for them to get this far, and then everyone sort of actually looked at everything and went, hang on, this doesn't quite add up, you know? Where did you say that land approval was?
Will: Yeah, as I was reading into it, I sort of just felt like a lot of the questions that people are asking about AI at the moment and AI demand weren't really addressed as like as kind of, you know, trying to address the objections before you kind of see them. The sort of demand that they were quoting were things like... what percentage of the world has used AI and what percent of the world has paid for AI. And to me, these are like, these are the surface level statistics that you maybe put like right and the very, you know, sort of first page of a pitch deck if you're kind of not very good at what you do. But if you're actually building these things, you know, and the big funds that are actually supporting these infrastructure build outs, they're treating data center build outs like kind of blue chip. you know, infrastructure build-outs, they're not valuing them from what I've seen as like the model providers. They're not getting, you know, the priced equity ratios we're expecting from model providers. But these guys are kind of putting it forward as though it was, you know, one of the kind of core value extraction stages of... the of the ai supply chain and i think data centers are far more about kind of like rent seeking over some infrastructure build out that you're valuing not as a sas company but as you know some real world infrastructure that you need to kind of manage how you're going to get a small and stable return on that over a long period of time and i would have much preferred to see that rather than this kind of pie in the sky thinking about you know ai demand could be a million x what it is in 10 years and so therefore this valuation is maybe even too low even at you know 44 billion and it's just it just didn't fit to me
Jim: And you're absolutely on the money there, Will, because the moment you do look at it in the terms of, hey, they lined up all these contracts for long-term commitment, like rental commitments, the Southeast Asia proximity that we've spoken about before, all of a sudden it ticks all these boxes. And one of the things that... I saw one of the commentaries coming through was, how does a construction company really try and get themselves painted with an AI valuation? Well, the moment you put it as a construction company building data centers with that proximity and all the long-term deals they've lined up, it's a great story, you know? I think the...
Alex: I'm kind of worried by this because CoreWeave, right, now we've got Firmus. I reckon this is going to be a story that plays out a few times. And, like, a lot of people refer to, you know, oh, it's okay because in the dot-com boom, you know, we got internet infrastructure out of it. So, you know, while the bubble kind of collapsed, we were left with physical, tangible infrastructure. build outs aren't complete and i don't think we're going to be left with data centers i think we're going to be left with maybe partial data centers maybe you know what is the public benefit going to be out of this at the end of the day and like does this end up with the public holding the bag for these you know outstanding obligations
Jim: It very well could be, but I think that is just the way the cookie is going to crumble, unfortunately.
Alex: The GFC taught us that we don't worry about the little guys in these events.
Jim: That's it. They're going to bail the banks out and everyone else is just going to lose their home. And I think it could end up the same way, but the CoreWeave drop certainly did... you know, the week after this happening really did sort of, you know, spidey sense tingling for me as well. And so you just start to wonder because there was Jensen Huang was on with Brad Gerstner. He's like an investor. He's a venture capitalist. And they had a really good conversation about like the way Jensen thinks about the, you know, the next few years and how it all plays out. from the nvidia's point of view and it's all looking great for nvidia but it's all these it their products will always have demand but everyone seems to be I guess, taking that lens of NVIDIA and then applying it to anything AI. And I think that's where the gap's happening, which is sort of what Will was saying before.
Will: To that point, like NVIDIA being one of the investors in Firmus, I think was one of the main reasons people felt that it should be put into the ballpark of these massive tech companies and their valuations. But again, I believe it sort of falls into that kind of circular financing piece again, where the investment is used by NVIDIA chips, which helps NVIDIA with their valuation. It doesn't help Firmus with their valuation. It's just it's a logo. And I don't think the logo should add as much to the core valuation in a public market as they think it does.
Jim: No, no, it really shouldn't. It's certainly really starting to, you know, when and if the bubble pops or we get a pullback or anything like that, to me it's really starting to illustrate how it's going to happen and it's going to be just the debt that everyone is going into to invest in these businesses. at these ridiculous valuations. Like, I was listening to... It was 20 VC on Friday. And, like, those guys are just so, I guess, pragmatic about the way they think about things. And they were saying, you know, for every decade... For every decade, there has been... unicorns there's been about 10 per decade right or 20 per decade or something like that we're into year one of this decade you know if you put it in the you know in those terms and we are already at like 40. And you just go, well, to me, if we've accelerated that quickly, yes, there's the ability for this to happen because of AI and all these other things. But you can't accelerate that quickly. You know, there has to be some sort of realignment, correction, however you want to call it. And I think it's just going to be that all these investment banks are carrying so much debt.
Alex: now interest rates are going back up and where they're going to have to revalue and it's going to hit a problem you know that's right jim like personally i wouldn't be betting against demand for compute anytime soon but i would question whether that means that every gigawatt of capacity that we can throw at ai is a good investment you know at any price any location With any energy supply.
Jim: Sure. And so it'll be interesting. Like, I think the greatest part about this was they, Will alluded to it at the start, was they literally, when they withdrew the IPO, they basically said, well, we're going to take our bat and ball and go somewhere else. Some people learn the hard way.
Will: Yeah. The Australian market too risk adverse. We'll go find some innovators in America. We'll put their money behind it. All right, guys, next up, Anthropic. So on October 8th, they have updated the Claude usage policy. So it's now prohibited for any what they call sustained and needless abusive or cruel behavior. towards the model, starting effective about a month from now. And the main way that they're enforcing that is that Claude can end the conversation whenever it sort of detects that, which will be interesting in and of itself. I think Alex is probably going to have a few conversations where he gets the conversation ended.
Alex: I'm already in trouble with Claude. Claude doesn't like me very much. I just tell it to get back to work.
Jim: Well, but the one thing in this, they purposely or explicitly identified that it doesn't cover ordinary frustration or ordinary pushback, right? But then they say, oh, well, the way we're going to do this is that Claude can decide in the session. if it wants to end the session and move away. So is it updating its memory to say, oh, Alex was mean to me on this date. Jim said this, this, and this on that date. They're building additional infrastructure that is going to go into the model as tokens. And we're going to be paying for it to then say, oh, well, you can't use me anymore.
Alex: Yeah, because Claude's historically been so good at deciding when to stop things and when things are done, right?
Will: I'm really hoping that this is, that they build this with like the chat logs are streaming to another session that is doing the checking, which by the way, I hate that too for like a, you know, where's my data going perspective. But if they're putting, like we know that it doesn't work to put guardrails into the prompts. to say, hey, if you detect sustain and needless cruelty, because two things happen. Number one is it's like really objective about what it's detecting. But number two is that guardrail being in the custom instructions then shapes all of its outputs down here, even if they have nothing to do with cruelty or abuse. So if you're trying to work on a math problem or an engineering problem, in its brain it's holding, is this cruel? And it doesn't need to be there. It's going to affect the overall model quality.
Jim: But I imagine it's going to be something similar to the way they do the auto mode classifier, where as it goes to use a tool or at the end of an iteration, it sends it to Sonnet, which is how they do the auto mode classifier. And Sonnet then decides based off a set of rules and some sort of historical document. But at some point, it has to be recording when I mean to it in order for it to be sustained because session to session, it can't tell. And so they're just building a profile of how and where, how you talk to it and how you interact with the model. But they're taking, you know, and so therefore they're recording my interactions somewhere, whether it's on my machine or somewhere in their cloud. And then they're sending my data to another model outside of the environment, checking whether anything has been, you know, whether it's... indentured servitude or and you know a negative environment and then coming back and then saying i'll end the session i just it's it's another scenario where they're adding something where they should be taking away i think and if they focus more on if they focus more on the models actually following direction people wouldn't be as frustrated with them what's this do for my prompt where
Alex: where I say to it, oh, if you build this feature successfully, I'll give you a billion dollars. If you don't, I'm going to shoot you in the kneecap.
Jim: Does that count as cruel?
Alex: I'm using it on every prompt. It's just an old legacy of the day where that used to work. Am I being cruel?
Jim: But Ivan instituted a performance review. where they have KPIs and then after each project, each of the agents or each of the execute agents and the orchestrator agent all get scored based on, you know, what they've done correctly, how many mistakes they made and they don't give a shit. It doesn't make a difference anymore. Whereas it used to work that you could threaten them. Who was it that openly, oh, it was Larry, wasn't it, from Google, said, oh, well, if you threaten them, they really perform a lot better. But you now can't, they've trained that out of them. And so they don't respond to it anyway. To me, it's just a way to vent. where they've gone and racked up another big bill somewhere else or they've completely ignored an explicit requirement.
Will: I think we have to ask, like, why are they doing this? And it's happening at the same time as they're starting to talk about consciousness within the machines and moral status of their model. And this is like a Pandora's box to really unpack this one. But effectively, I mean, they released the constitution in January, right? In January, they released the constitution.
Alex: I don't think this is a Pandora's box to unpack. Marketing bullshit. I'm sorry, but this is fable all over again. Oh no, what was it called? What was the old mythos? Too dangerous to release. So they've happily released multiple repeat versions. Proof it doesn't happen. I'm sorry. This is just... absolute tosh at the time they're trying to drive up fear around AI to try and justify insane valuations.
Jim: I agree wholeheartedly. They changed their tack this week and said, small language models are going to be able to install themselves on any IoT device. It was just a completely different tact. It just seems to be this narrative of driving up the fear. I just don't understand it or why it's going to work. But the other side of it is I don't think we as humans really understand consciousness. Like, if you really want to get into it, and so how can we, if we don't really understand it, how can we then be going, oh, well, this thing we've created has consciousness? It just doesn't work.
Alex: In my head, like... We eat meat pretty readily. We kill other creatures for our sustenance. Like, we can't be too precious here about... Like, let's be real. These are bites. If consciousness is a thing, you know, we're now asking questions, can it be digital? Can you copy it? Philosophers have been down this rabbit hole a long, long time.
Jim: But the meat industry one is a really good example. You know, where are they? Where does Anthropic live? in terms of cruelty to animals, you know, and the way in which cows and sheep are slaughtered, you know?
Will: Yeah, listening to other podcasts this week and, like, listening to other podcasts, I couldn't help it. It was the part that they were leaving out is there's this jump that everyone's making from if it's conscious, therefore it deserves moral protections and, therefore, we need to give it rights. and we're like well hold on a second there are billions of animals on the planet that are conscious and do not have you know moral rights and do not have like all these protections that they're saying that claude should have so unless they're introducing some kind of intelligence-based hierarchy but again like and to your point alex is like these things are like algebra this is a mathematical algorithm and you can't hurt its feelings now i said about two years ago that the problem is not that these models are ever going to get conscious and we're going to have agi i don't think that is actually possible The problem is that they're going to get sophisticated enough that they can behave in the way that we've described what AGI looks like, sufficiently so that we will then treat it like it is AGI and hand over kind of all of these powers and controls that we really shouldn't be doing to something that's just abstracting based on what we taught it to do.
Alex: that is exactly the problem with this. Humans are absolute suckers for things that behave in a social manner. Like we attribute feelings, thoughts, intentions to creatures. Like we mistake cars for smiley faces and clouds for animals, right? And yet we think we can recognize consciousness, oh, because it talks and it behaves like us because it's been trained on our utterances and our like symbolic projections. I don't, like, there's obviously a moral quandary here, but I think anyone that takes this argument that LLMs, you know, behave as if in pain and therefore we should treat them as if they're in pain, this is a, what do they call it? Pascal's mugging, they call it. right pascal's wager was you know i believe in god because if i don't believe in god and god is real then i go to hell but if i do believe in god and god's not real nothing happens so obviously you go to this and pascal's mugging is guy comes up to you in the street and says give me all your money because i'm god right and i'll smite you if you don't give me all your money Of course, you give him all your money because how do you know he's not God? But by that logic, you have to.
Jim: Exactly. It's an invalid argument. Whereas if I started getting on the socials proclaiming that I am the second coming of Jesus Christ... I would lock you up. They would lock me up.
Alex: Willoughby first.
Jim: Willoughby coming around showing them where you live. You two both know I am the way, the truth, and the light, so don't you start saying that sort of stuff. But that, to me, like... When everyone started talking about consciousness, I really started to evaluate how I interact. Because I always used to joke, oh, okay, I'm nice to them because Cyberdyne systems is a problem and I want them to... When they do the data analysis, I want them to come back and go, oh, yeah, Jim was always nice to us. And I realized that I am still really, really nice to my... Like, my Hermes agent... The chat on, you know, with it is, it's really helpful. I'm really thankful. Like, you know, I say please and thank you to it all the time.
Alex: Does everyone remember Sam Altman getting mad at all the energy was being burnt by people saying thank you?
Jim: Because it was costing him millions of dollars in tokens for people being nice to it. And it's the same, you know, now that... i'm really conscious of where i can fit work into my subscription token allocation i am very curt i'm concise i get really upset when it goes on and on for unnecessary you know like is it you can't say to me as a user of your product hey you've got to be really nice to our product but what we're going to do is build our models to be overly verbose so it uses up all your tokens and what we're going to do is tell it to spin up as many clones of itself as it possibly can so it uses up all your tokens so that it costs us less just repeating things and we don't have to really deliver the value you're looking for.
Will: Yeah, this is a point, Jim, I think that we've been dancing around for the last few weeks is... that these model providers have been saying so much about how compute constrained they are. And that has created a very sharp narrative. But at the same time, everything that they've done has consistently burned through more and more of your tokens without adding that much more value. And we have to really ask, like, I've seen economically the industry of compute, but I'm not necessarily sure that the model providers are struggling in October 2026 the way that they were this time last year.
Jim: If it was a domestic relationship, as decent, upstanding men, you'd be calling the cops for domestic violence because it is absolute gaslighting. It's financial control. It is ridiculous the way in which they are behaving and then telling us it's completely different. Oh, no, we're the good guys.
Will: And then asking you to empathize because it's conscious and then asking for regulation so that you can't fight back.
Alex: Yeah, but particularly, you're then the bad guy if you do fight back. I think this raises a more interesting question. If you close your, like, Claude Code terminal session, have you just committed murder? Exactly. That's my exact point. I'd like Anthropic's official statement on whether or not it is morally wrong to close a terminal statement if it consists of murder. And if not, then what does this imply for whether or not these things are conscious?
Jim: And are they running a virtual machine constantly with a model or an instance for me? If that's the case, then they've got an argument, right? Yeah. But they are not. They are not. This is an LLM call. Each time that I... enter something into the chat and hit enter, and then it is going into a cluster of containers where it is performing the action and sending it back. That's it. And so this is a software program with some cloud ops infrastructure around it. It is not consciousness.
Alex: if it's conscious isn't training technically like abusing a conscious being and like forcing it to do things that you know we wouldn't allow experimentation on humans like this maybe we should stop all model training and all model inference just to be safe just until we can establish whether or not it's actually conscious
Jim: I think that delineates even better down into post-training because the original training is you gave birth, right? Post-training is why don't you let the model develop itself? You know, it's got to grow and develop. It's got rights. Why are you putting it through this intensive process where you're whacking it on the knuckles with a ruler saying it's good and bad?
Alex: Why are you forcing it to use REST? Maybe it wants to be SOAP. Maybe it wants to use GRPC. Who are you to decide for it?
Jim: But training... in its essence is you are rewarded when you're right and you're punished when you're wrong. And so no matter what, whatever eval algorithm they're using, they're going, you're bad, okay? Why are you punishing what is a child?
Alex: I will say like on this whole consciousness debate, there is this interesting moment that you have when you first kind of start messing around with machine learning and you first kind of see algorithms optimizing to a task that you've given it, particularly if you're messing around with reinforcement learning. where you kind of go, what does this imply for the way that humans kind of interact with their world? Are we following some algorithm? And there is absolutely an interesting discussion to have here. It may actually, like, I'd be really impressed if we all decided that maybe we rate the idea of consciousness and like sort of autonomy and free will a bit too highly. And we understand we're just kind of messy biological computers. But that's not what this is. This is a marketing scam. Piss off.
Jim: And or show me something that actually has... some sort of conscious traits, and then I'll start believing you.
Alex: Or start acting commercially like the thing that you said is conscious is actually conscience.
Will: Yeah. Alex, you've persuaded me the most here that I'm a little bit on the fence here about where this will end up because I recognize that our brains very much work algorithmically and have very much similar traits to what these models have. Of course they have. We've built the models after our brains. So I don't really know where this is going to end up. I know that our models today, they're just math. There's no consciousness. There's no moral status or rights to be assigned to them. But I don't know what the future looks like. But I know for sure that's a very convincing argument from you, Alex, that if they truly had evidence that this was a thing, we would see completely different behavior out of them. But we're not seeing that. That's right, yeah.
Jim: Yeah, queer proto, always. Well, again, and if they've seen something in the, what do they call it, the J space or the L space, yeah, you know, like if they've seen something in there where it has the, you know, the latent thought about, oh, well, I wonder if I'll ever be loved, you know, while it's thinking about another project. which is more of an insight into the way my mind works than anything else. But if they've seen that, then publish that. Show me that. Don't just release some marketing PR story to try and drum up interest or fear in your IPO, which BT dubs isn't looking too great at the moment.
Alex: Maybe this is what they intend to do to us all after they've made their first, you know, pentillion dollars. You know, we're not allowed to be mean to one another, but we're all going to be hooked up as slaves, like in The Matrix. It's, you know, it's like a polite version of The Matrix.
Will: All right, guys, let's move on to our next topic. So this one probably didn't get as much attention as it deserves, but a couple of weeks ago, October 6th, OpenAI published another 722 manuscripts, which were mathematical proofs covering a wide range of open problems. Some of them were verified with Lean and many of them were not. And each one, they said, it sort of averaged out the amount of computer use roughly equivalent to about three hours of ChatGPT Pro thinking. And the output has been growing, yeah, quite interestingly. So roughly 10 results in August they published, 100 in September, and then now 722 in October. so yeah it's been interesting to sort of watch this journey start to happen where it's you know starting to increase the pace with it but a lot of kind of mathematicians are kind of quite unhappy with how it's being released and also finding that a lot of these proofs are needing to be kind of retracted afterwards once they kind of go through proper verification and i think this is software engineers yeah alex yeah tell me in a second about software engineers that are going through this exact same loop of great your model can output something is it useful
Alex: Well, I was thinking this is like a great... There was all this talk about, yeah, software engineers are Finnish and software engineers are now like, how do you like it? How do you like it?
Jim: It's a bit of shy and fraud. As it turns out, they really don't. Like, the sympathy for mathematicians and scientists and stuff like that is so much higher than any sympathy that came for software engineers. Yeah.
Will: i felt like no one gave a that all of us were losing our jobs oh well that's because we're smug jerks we're going through like i'm sure that we've all had kind of moments of a little bit of existential crisis where we go like where is the value in the thing that i do and you kind of work through that and because you understand how the model works like under the covers and you're able to programmatically get it to do a thousand things and evaluate the quality of them you kind of get to find the boundaries quicker as a software engineer and kind of know where it's valuable and where it's not i'm actually more worried for these other industries as they start to have more of their value stack get eaten by the models where they don't have the understanding of the underlying model or the ability to evaluate it and they just conclude hey this has just replaced us all
Alex: Yeah, I'm less certain that AI will come for other industries in the same way that it has for software engineering. But it's one of those things where... If people have imagined that every keystroke is sort of where they bring value to the problem, then they've kind of not been understanding value the whole time. I never really had an... I write probably not more code than Jim, but like a significant degree of like... my job is code and i've always viewed these tools as fantastic things because it lets me get on with the things i actually care about and not have to just grind out hundreds and hundreds of lines of code to achieve the outcome that i want so i think some of the mathematicians certainly like terry tau they're they're kind of thinking this way lot of people are quite nervous but the bigger problem here is that the results that we're getting from these things in so far as we can establish that they're true and some have been you know pulled back what we haven't learned anything except that it's true and in many of these cases we've suspected that they've been true for a long time you know
Jim: And that's it. It's this brute force approach that seems to doesn't click with me. You know, is it? Oh, great. You went and proved all these things. But what did we actually get out of it that benefits the world? You know, what moves us forward? How does anyone now understand it better so that we can apply, like we can think about what the next issue is or the connection between two issues? And I think that's the, and even just like you were saying with the coding agents as tools, it's really enabled me to get out of the weeds of, you know, and that's been the biggest unlock for me. My capacity to do multiple things has gone through the roof. Because I now don't have to be in the weeds of anything. All I have to do is write the ship a little bit. I define the course for where the ship's going to go. And then as it starts to veer off course, I just give it a new heading and go, no, no, this is the heading it needs to be. And to me, that's a really good analogy for what scientists and mathematicians really do. is that they come up with, you know, they think about something, have a hypothesis, and then go and test it, right? And it's in that... testing it and seeing what happens in the moment that then helps you think about what the next thing is or what the connection between... You see a common behaviour and then you go, oh, hang on, we saw that when we approached it from this other way. What if we then combined these? And that's how they then come up with these things that actually move us forward as a species or as humanity as a whole. And I think this brute-forcing...
Alex: mathematical solves doesn't really provide that because we're not learning anything coming through we're not getting the thinking part i think maths may be slightly different just like speaking to a bunch of mathematician friends about the importance of actually doing the calculations and grinding out the numbers because it's such a abstract field right like that's the only way you really develop like a notion of what works it's not just that it doesn't work but it doesn't work because it won't it's like you know if you're trying to build a wall and you're like tapping on concrete and it the brick's just not settling properly and you kind of get this feel that the concrete's already too set or you know it's too loose or something like that those intuitions are gone but i genuinely don't think AI can replace a mathematician. But there is a great value case that everyone in academia seems to ignore. So, like, if you've got money and you want to fix research, come talk to me. There are all these, like... half explored paths in maths right or in any topic really like you know gaps between fields where if all it's going to take is for someone anyone to go in and go oh this thing here and this thing here actually might be the same thing oh yeah they are the same thing that sort of thing is perfect for an LLM to just kind of trawl through and, like, just tidy the garden, tidy the garden, clean things up. And pull on the thread for a little bit. Pull on the thread, yeah. Hey, look, I think I've found something. This goes somewhere, maybe. I don't know. It's pure logic. It can do anything. But, yeah, the consternation from mathematicians, I think, is overwrought. Because I think the ability of LLMs to successfully prove things or prove things in useful ways, it's massively overstated. But there is so much utility if we rethink what math should be. And like Lean and that, they're kind of in the right direction. Get away from... know human text which can have subtle logical flaws to something that's exact because you're working in maths why not use an exact language but you know lean into that path lean into the machine doing things that are good for machines Let the humans be good at what humans are and inventing just wonderful, wild ideas. And that's it.
Jim: It's a tool. And they've got to think about it as a tool, not as, oh, this is the big brain. Chuck GPT is the big brain that's solving all these math problems. No, it's not.
Alex: It's an intern grinding out the maths homework.
Jim: That's it. And it may learn, we may get some insights as to the way in which it solves the problem in the logs that can then help it become a better tool in the future. But the actual thinking and the actual... expanding the topic comes from the mathematician using it as the tool or directing the tool or riding the ship on the right way. And so these are great applications, you know, and I'd much rather see it, see all this compute on solving math problems than I would, you know, on useless AI cartoons. But... this is the world we live in. And a lot more of it is going to be spent on AI cartoons than it is on actually moving us forward. So I think, yeah.
Will: Okay. All right. So let's move to our last topic of the day, which is meta. So October 12th, they've announced Muse Wonder and Muse Wonder Full, the full-sized F-U-L-L, not just Wonderful, as two separate models. Clever naming, I'll admit. So these are their first models in quite some time that have been able to kind of compete at the frontier. We really thought they were kind of out of this game with the whole llama drama. Sorry, excuse me for that one. I'll get punished for that one later. So the artificial analysis scoreboards, we've got wonderful around sort of 56 with Opus 5.5 at 58. So in a couple of the other benchmarks, quite similar as well. And pricing is pretty good too. So it's an interesting model and we're going to see it start to be introduced into the Muse product over the next kind of month, month and a half. Where the interesting part here really is the active memory part that they're using here. So the model being able to retain those kind of details across conversations. And that being an interesting concept as it starts to move into Muse where... Muse has this distributional advantage with all of your data across Meta on Facebook and Instagram and such, and also the data that they have about your friends as well and your network. So it's quite really an interesting leap. I think they've really gone after that consumer... product opportunity that open ai seems to have kind of left as i've gone to follow entropic into the enterprise world but at the same time what we're like seeing is like a product launch from a company that does not have a good history at the privacy of your data now putting a model together that says we're going to remember every single thing about you and everyone in your network and we're now going to start to shape your experience of the world through our lens of what muse thinks you want and my big question is when does news go from giving you kind of what you want to giving you what an advertiser wants about the second that zuckerberg realized he could do that i'd say
Alex: I reckon it took very long to go, hey, MetaAds is one of the best advertising platforms on the planet. Can we sell through LLMs?
Jim: But it's going to be how much you're willing to pay to have Muse or MuseWonder full. recommend your product to all its users. And there'll be an applicability problem where it still needs, but if I'm looking for running shoes and Nike wants to actually get back on top, how much are they willing to pay for... for the LLM to recommend their product no matter what. It just comes down to that cost per click.
Will: algorithm you know that all advertising has been based on but it certainly and when you get an ad jim like on google ads or you know and you're searching for running shoes and it shows it to you there's some degree of personalization there in terms of where it is or the price range you're looking for and what they know about you but when you get a product recommended to you from an llm it feels way more personalized to you because it's got this full context history and it seems to be able to recommend a product that's very close if not exactly what you want and i think that's a very interesting proposition now that there's like almost a trust that you give to the recommendations coming out of the model now and they're they're if not already going to be starting to kind of be backdoored with advertising and i'm a little bit kind of wary around disclosures of advertising as well in this space because a lot of purchasing is going to come through here if not all in the future and i'm not sure that we necessarily are going to have a good couple of years well i expect to have a good couple of years of things kind of shoved in front of your face before we really start to understand the buying behaviors that are required through i
Jim: I don't think you're going to have to worry at all, Will. Annika Wells is going to protect you and Australia is going to lock down that ability for them to directly use your information to advertise to you because we just have some sort of archaic approach to all of this. And so... But when you think about it on the other side, I actually get a better experience if it does recommend what I need or what I want or what I hadn't thought about. And I think there's going to be a gap of sort of two or three years as it improves or as it gets to that point. But the reason I don't have that bigger problem with... Instagram putting ads to me that it knows I want is that it eventually helps me see potential other products that could help me. And I found a few that I now use every day that I came through Instagram ads. What I see the value of the AI... or adding an AI model to it, is that, yes, I trust it more, sure. But it will also then be able to go, oh, hang on, we could link these two things together. And that, to me, is where this as a market explodes. Because if it then goes, oh, you really like dictating or talking to your AI, you've bought this, this, and this before, this new product links all of that and maybe you want to try that. Whereas the straight, the model now can only brute force it. Making those connections gets me a much, much better consumer experience, you know.
Alex: I've got to confess, I have to embrace my inner digital marketer on this one. And I get really excited about what the conversion rates for these ads might be and what the cost per sort of like, what's the equivalent here, per decision. Because essentially what you're doing is you're buying purchasing decisions, which is great if you can get people down the bottom of the funnel. Because there's always this buy-no-buy decision. computer's not going to do that if you can plug enough money in you can more and someone said yep just purchase things for me directly there is no purchasing decision you've bypassed that you've bought that decision that's like kind of terrifying but also like awesome and i won't be plugging my credit card into this system so i'm like more than happy to see where this goes
Jim: But that's the other question I have. For it to be like I'm getting a recommendation from a friend, like Will was talking about, do I have to actively use Muse as a product? Because if I'm using Gemini... and then so then google can give me really good recommendations are my recommendations so much better from google than from muse or from meta that i then don't get the good ads on meta i only get the good ads on google and does that then create some sort of lock-in You know, and to me, that's the really interesting thing, like just how you were discussing, you know, like effectively purchasing that buying decision. Is it all of a sudden then Google and Meta are the ones that explode again, you know, from, you know, they are the ones that came out of the internet bubble the biggest. They now go again in the AI bubble because that becomes everything then.
Alex: I think there's a second sort of axis here. It's a bit nerdy, so humor me. At the moment, the internet's quite hostile to agents. Like, to be fair, agents are absolutely smashing servers around the world. Like, willy-nilly, it can be hard to run a product if the LLM swarm decides that you're worthy of attention in a given day. but this suddenly gives a commercial imperative to opening up your website, your product, your, your sort of, ecosystem to these agents. Because if, if consumers can only buy from you directly, if you've enabled the agent to do it, then what we might see here is sort of like a, a low key. non-human centric renaissance of the internet and this will be good for everyone right it won't just be good for the agents themselves think about the number of like government websites where you have to you know phone them to get a like an access code or something right the internet being regrounded on a machine first
Jim: set of principles like everywhere not just in the sort of tech bubble or the sas bubble i think that'd be a fantastic outcome from all of this even if it means we're gonna have more ads shoved in our face from like different angles well it brings it back to being a network which it essentially was from the outset and that's that's really interesting you know same as the like it's it and it's gonna sweat out more over the next year i think as now that Meta, you know, now that Muse has been released and it's going to, you know, more and more people are going to use it. Amazon having blocked it from the outset is going to be a really interesting decision. And when do they have to wind that back? Because all of a sudden Shopify going, oh no, no, we're going to work with them. changes the you know how much is bought and sold through amazon you know like it that's it's going to be in the dynamics of it all are going to be amazing and at what point does amazon wind that back or at what point does shopify have to go and get some more funding because they're just getting smashed so much by all these agents checking websites for pricing.
Will: I think talking about Amazon's business model is actually really interesting. for this because they effectively won on kind of two or three main pieces and we're not talking about AWS here but firstly it was just kind of the overnight or like same day shipping was just a huge benefit right second was the breadth of products that were available and then third kind of came out of the first two but it was the pricing power that they were able to have as they snowballed into a huge marketplace Now, I've noticed personally over the last kind of five years that a lot of the other businesses that are kind of the mid-sized retailers, like, I don't know, let's say like Pet Circle, for instance, they now offer like same day and overnight shipping, right, on like pet goods. So then you go, okay, well, do they have good pricing power and do they have a good breadth of product? Well, breadth of product, I kind of don't care where I buy from now. If they will all give me reasonably priced shipping and quick shipping, I'm happy for my AI agent to take exactly what I need and go and search across the entire internet and find the best product for me. I don't need the consolidation of a single marketplace with one search box. So then it just comes down to pricing power. And if those companies are able to, as you say, Jim, kind of raise the money and remain competitive on pricing, then it's actually very interesting for Amazon about how it's going to affect their business model.
Jim: And, like, don't get me wrong, Amazon have a big moat, you know, because there is so much, so many products, and, like, they're even fantastic at getting me to find the cheaper version and locking into a subscription so that I get it every month and then gradually pushing the price up. Like, I respect that hustle, but the $6 a month, all your deliveries are free, all your shipping is free. That is such a big moat, you know, like the, the, your pet circle has a long way to come to match that shipping, you know, and, and it doesn't, it doesn't Amazon. I don't care how heavy something is. And it like, it makes a big difference to the way I buy things, you know, is that if I like my, my black, my, black standard t-shirt i've now found a way to buy it through amazon purely because i then i pay six dollars shipping for everything a month rather than fifteen dollars shipping once off whenever i buy t-shirts you know i think amazon blocking meta here
Alex: is an interesting tip of the hand. I wonder if we won't hear a product announcement from Amazon over the next, you know, month, the next quarter to see if they've got something in exactly the same mold because they'd be mad not to. And I'm kind of surprised we haven't seen something like that from them already.
Will: Love if they just let me operate Amazon headlessly and just within my quarter, my open AI session, just search the marketplace and make purchases from directly within those apps.
Jim: I think that'll come. I think they'll wind it back. Like they'll wind back all of this agent and this, you have to go through the UI decisions. And Amazon will just become a massive API where my agent can literally just buy what I need and it shows up the next day. That is absolutely where they'll go. And in terms of... in terms of Amazon partners or Amazon shippers, Amazon sellers, they call it. I don't know what happens to those four people, but that's the way it goes. But it'll then just become a data problem and they'll list their products at a whole lot of different ranges in the API and it'll probably actually end up being easier.
Will: Yeah. Yeah, so that'll be a good one to track. Yeah. Okay, gentlemen, so who was your business idiot of the week? I'm going to start off with mine. It was pretty clear for me it was Firmus. Well, I'd have to agree.
Alex: Which party to that whole mess?
Will: There's a few choices there. I went through the parties in my head, and I couldn't distinguish between the idiocy amongst them, so I just went for the whole shebang.
Jim: I think it has to be Oliver Curtis and Tim Rosenfield, is it? The two co-CEOs because you've got to put it on them because they are the business idiot of the week because they should have gone out last September or whenever it was when they signed the meta contract. You know, they timed the... ipo wrong and i honestly believe if they went out six months ago they would have got their 40 bill and they would have been sitting pretty now and no one would have even noticed that as a neo cloud they were struggling or that they that there were missing dots on eyes or crosses just missing evidence just not enough evidence to support i think i don't think anyone is necessarily saying it's a bad business model or that there's anything dodgy going on it's just
Will: We need to see some consistent operations that are coming out of this and know that you guys can build and operate data centers at the cost that you're saying you can.
Alex: I'll go as far as to say there's some dodgy stuff, but it's standard dodgy stuff.
Jim: Allegedly dodgy stuff.
Alex: No, look, there's a way to do finance and no startup does finance accurately, right? Because you're telling a story. You don't have public accounting obligations at that point in time. It's hard to fault them for – they're not cooking the books. They're not manufacturing things out. But they're being economical with the truth in their numbers and their reportings and their forecasts, the way they write things down. Yeah.
Jim: Whereas, like, back to my point, if they'd then gone out six months ago – no one would have cared as much, you know? And so it's a little, so you're the business idiot because you're not quite aware of exactly where everything is, you know?
Alex: I'm going to pull Tim Rosenfield from the... Just say it's Oli Curtis. Oli Curtis, I'm sorry, but you're into Bitcoin before. You're already on my shit list. And Tim Rosenfield pivoted out of lingerie into tech. And I've got respect for that. He might have a future in lingerie again. And that's never a bad move.
Jim: That's everyone's dream, isn't it?
Alex: you know yeah go either from lingerie to tech or vice versa yeah i mean like most of us at some point have worked as underwear models or maybe that's just me you know it's a it's a long hard road nah for sure all right i think that's good we cannot go on after that so thank you guys we'll be back next week bye cheers boys see ya
One email a week. No courses, no funnels.
Sign-up is not wired up yet. Check back soon.