- Meta is exploring selling its spare AI compute and model access, a move the hosts see as more pragmatic than most of its recent AI bets because demand for compute currently outstrips supply.
- OpenAI unveiled Jalapeno, its first inference chip built with Broadcom, planning to deploy it by the end of 2026 as part of a broader push toward owning its own infrastructure.
- A German court issued a preliminary ruling that Google is liable for defamatory claims made in its AI Overview search results, rejecting Google's argument that a disclaimer was sufficient protection.
- Anthropic said Alibaba ran a distillation attack using roughly 25,000 accounts and around 16 million exchanges to copy its models' capabilities, a claim the hosts tie to Anthropic's own super weapon marketing for Mythos.
- The hosts warn that regulating AI models as though they are uniquely dangerous risks repeating the mistake of treating open technology like Linux as a threat, which could damage the wider tech industry.
Meta is exploring selling its spare AI compute and model access, a pragmatic pivot the hosts say makes sense because demand for compute still outstrips supply. OpenAI has unveiled Jalapeno, its first inference chip built with Broadcom. A German court has issued a preliminary ruling that Google is liable for defamatory claims made in its AI Overview results. Anthropic says Alibaba ran a large scale distillation attack against its models. In the first episode of Business Idiots, Will Turner, Jim Lovell and Alex Stenlake work through what all four stories say about who actually controls the direction of AI, and close the show by naming the first Business Idiot.
Why is Meta selling its spare AI compute?
Bloomberg reported this week that Meta is exploring ways to sell some of its excess AI compute and model access, a move that would push it toward becoming a cloud provider alongside Amazon Web Services, Microsoft Azure and Google Cloud. Alex Stenlake's first reaction was scepticism, given Meta's history of expensive, poorly justified bets since the Metaverse. But he came around: Meta has comprehensively lost the frontier AI race, doesn't have the single-customer dependency that leaves a company like CoreWeave exposed, and is sitting on a pile of compute the market badly wants. Will Turner made the same point from a different angle: this is one of the more pragmatic moves Meta has made in years, because right now there is far more demand for compute than there is supply, and Meta already holds supply. Jim Lovell wondered whether the timing is also connected to the AI talent exodus toward Anthropic, since a company that has stopped chasing frontier models has less need to hoard compute for training.
What is OpenAI's Jalapeno chip, and why does it matter?
In the same week, OpenAI unveiled Jalapeno, its first inference accelerator, built in partnership with Broadcom. It is already running lab workloads, with a full deployment planned by the end of 2026. Jim noted that OpenAI's Sam Altman has openly said the company sees itself becoming a token business rather than purely a model business, which makes owning inference hardware a logical next step. Will pointed to a concrete signal of why this matters: CoreWeave's share price was down 29 percent that month, a business that Alex noted draws roughly two thirds of its revenue from OpenAI alone. Meta's move avoids that single-customer trap. Jim added that specialised inference chips matter because demand is shifting from a handful of enterprise deployments toward many more small and medium businesses that don't need frontier-grade models, just something cheap and reliable enough to run.
Does a smarter AI model always work better?
The hosts spent time on where "smarter" AI models actually help. Will's view is that a model embedded inside a product only needs to be smart enough for its specific task, whereas for software development work, a smarter model is usually better, with the asterisk that Anthropic's newer, more verbose models have been generating far more code without a proportional drop in bugs. Alex made the same complaint about recent Claude models: more output per fix, not fewer fixes. Jim disagreed in part, arguing the bigger models earn their keep in planning and specification rather than in the code itself, where a well-scoped brief handed to a smaller model like Sonnet gets the job done efficiently. It was Will's comment defending smarter-is-better that prompted him to ask, half joking, whether that made him the Business Idiot of the week, a line that returns at the very end of the episode.
Why did a German court hold Google liable for AI Overview outputs?
The conversation shifted to a German court ruling that found Google liable, at least at a preliminary stage, for content generated in its AI Overview search results. The case involved two publishing companies whose AI Overview summary associated them with a scam, which the companies argued was defamatory. Google's defence, that it discloses the content is AI-generated and users should verify it, wasn't enough for the court. Jim explained the key distinction: ten blue links point users toward "the answer is somewhere in here," while an AI Overview states a conclusion in an authoritative tone, effectively making a new claim rather than indexing existing ones. Alex pushed back a little, arguing few users read disclaimers of any kind and that the ruling might be Google being held to account for marketing AI Overviews as intelligent rather than as a statistical summarising tool. Jim's practical read was that the ruling is likely to stay specific to Google's search summaries rather than extend to a general AI service being liable for whatever a user does with its answers.
Who decides when a new AI model gets released?
The hosts turned to the pattern of frontier labs routing new model releases through government process. Anthropic's Mythos model was walked back shortly after release and then quietly reinstated following a US government approval gate, and OpenAI has said it will voluntarily submit GPT 5.6's release process to US government review. Alex's read is that Anthropic's own "super weapon" framing of Mythos backfired: once you tell politicians and bureaucrats something is extremely dangerous, they act on it, especially after reports that an early tester found a jailbreak. Will connected that to Anthropic's Department of Defense ambitions and its distillation attack claims, suggesting the US government may be deliberately restricting access to a model Anthropic itself has called a super weapon while simultaneously saying that weapon can be copied. Jim's take on OpenAI's approach was blunter: it is good marketing, buying engineering time while the US government effectively vouches for a model the hosts weren't convinced was as strong as its rivals.
What did Anthropic claim about Alibaba's distillation attack?
Anthropic reported that Alibaba ran a distillation attack against its models, allegedly using around 25,000 accounts and, in a figure raised later in the discussion, around 16 million exchanges to try to replicate the underlying model's capabilities. Alex questioned how meaningful that scale actually is against a model's full parameter space. Jim raised the hypocrisy angle: frontier labs trained their own models on the world's data, largely without licensing it, and are now unhappy that a competitor tried to learn from their outputs the same way.
Why would regulating open source AI backfire?
Looking ahead, Will raised a scenario where an open source model reaches benchmark parity with a frontier model like Mythos within six months, and politicians respond by treating open source AI as inherently dangerous. Alex's objection was blunt: most of the modern internet, and most cloud infrastructure, already runs on the open source Linux operating system, and nobody proposes regulatory approval for every Linux upgrade. Jim agreed that open source and open-weight models are becoming more important precisely because large enterprises and sovereign environments, including examples like the Anthropic and Nvidia work with Palantir's Nemotron deployment, need models they can run in fully controlled environments rather than send data to someone else's data centre.
Who gets named the first Business Idiot?
Closing out episode one, Jim first nominated the marketing teams behind the week's AI announcements. But the verdict the group actually landed on circled back to Will's earlier joke about whether he was the Business Idiot for insisting smarter models are always better. Alex held him to it, Will accepted, and Business Idiots officially has its first name on the board.
Chapters
Questions from this episode
- Why is Meta exploring selling its spare AI compute?
- Bloomberg reported that Meta is exploring selling excess AI compute and model access, moving it toward becoming a cloud provider like AWS or Google Cloud. The hosts argue this is one of Meta's more pragmatic moves in years: demand for compute vastly outstrips supply, and Meta already holds a large amount of idle capacity to sell into that gap.
- What is OpenAI's Jalapeno chip?
- Jalapeno is OpenAI's first inference accelerator, built with Broadcom, already running lab workloads with a planned deployment by the end of 2026. The hosts frame it alongside Meta's compute move as evidence that frontier AI labs are diversifying away from pure model building toward owning infrastructure, chips and cloud revenue rather than relying solely on one customer's spending.
- Why did a German court rule Google liable for its AI Overview outputs?
- A German court issued a preliminary ruling holding Google liable after its AI Overview summarised two publishing companies as being caught up in a scam, which the companies said was defamatory. Google's disclaimer that AI outputs should be verified wasn't enough for the court, because an AI Overview reads as one authoritative answer, not a list of links.
- What did Anthropic say Alibaba did to its models?
- Anthropic reported a distillation attack it attributes to Alibaba, alleging roughly 25,000 accounts and around 16 million exchanges were used to copy its models' capabilities. The hosts link this to a wider pattern: Anthropic has framed its Mythos model as a super weapon while also needing government protection from a country it says is stealing that same technology.
- Why are AI model releases now going through government approval?
- Anthropic's Mythos model and OpenAI's GPT 5.6 have both gone through slower, government-linked release processes instead of a straight consumer launch, with OpenAI framing its approach as voluntary cooperation with the US government. The hosts read this as labs and regulators reacting to the super weapon style marketing used to hype frontier models before release.
- Why do the hosts think regulating open source AI would be a mistake?
- The hosts argue that much of the modern internet already runs on open technology such as Linux, and treating open source AI the same way would cause similar disruption. They warn that if politicians decide open models are inherently dangerous once they match closed frontier benchmarks, the resulting restrictions could damage far more of the tech industry than intended.
- Who was named the Business Idiot in episode one?
- In the show's first verdict, Will Turner was named the Business Idiot after jokingly nominating himself mid-episode while arguing that smarter AI models are always better for tasks like software development. Co-host Alex Stenlake held him to the self-own at the end of the show, and Will accepted the tongue-in-cheek award to close out the episode.
Read the full transcript
Will: Okay, welcome to our brand new podcast, Business Idiots, where we're going to be discussing AI, business and technology and trying to make sense of it all. We're going to be tackling a bunch of topics sort of across AI across the globe and how it affects us here in Australia. So I'm joined here by my good friends, Alex Stenlake and Jim Lovell. Welcome, Jim. How are you, mate? Very good.
Jim: Very good. A lot better today. I had a rough week. I finally got caught by the bug that everyone's been getting. And I haven't experienced it since I was like in high school or whatever. But like Tuesday night, I was in bed like cold sweats. It was unbelievable. I just couldn't stop it no matter what I did. And so it certainly, I feel a lot better now, but it was a rough couple of days.
Will: It did, isn't it?
Jim: have some ai agents running in the background so you're still productive but again is it because i i'm too cautious as to their permissions i didn't i didn't i wasn't hitting the proof on uh on floor and so they all uh they all sat there for eight hours waiting for me to get back online oh good and alex welcome mate hey how are you very good mate excited for the podcast
Alex: I'm thrilled. It's been a big week, a rollercoaster, particularly with Anthropic graciously allowing us all to play with Fable again. And Jim, cowardice, YOLO mode is there for a reason. Completely untested Frontier models, YOLO mode on corporate devices.
Jim: It's certainly, you know, like... trying to get to my 50% in the week, you know, because again, they gave us to July 7th to use without having to use the API key. And so desperately trying to hit my 50% with Fable and then winding everything back. That's as YOLO for me as it needs to be, I think.
Alex: In diagnosing neuroses, I've come to the conclusion that Sonnet 5 is quite neurotic. So I've been having a lot of fun towards the end of this week.
Jim: Gaslighting you.
Alex: No, it's been gaslighting me. There is no gaslighting. Gaslighting is not real.
Will: All right. So we'll jump into some of our stories for this week. So we're going to be talking about the return of Fable 5, of course, and also the upcoming GPT 5.6, which is being announced and having a controlled and slow release on that one. We're also going to talk about Germany had a ruling around Google's liability on its AI outputs. And we'll also talk about Anthropic with the reported distillation attack of their models apparently reportedly by Alibaba and how that is affecting the competitive landscape around the pricing of AI and who owns what and how you're going to be able to stay competitive in the world of AI. So I think let's start off this week. Let's talk about, so this week Bloomberg has reported that Meta is exploring ways to sell some of its excess assets AI compute and model access. So this would start to move them more towards an AWS, Azure, Google Cloud and Neo Cloud type of player. And then at the same time, I've also got OpenAI, who was, with Broadcom, unveiled Jalapeno, which is their own AI chip. So it's their first LMI inference accelerator. And they're already running some lab workloads, and they're planning to deploy this by the end of 2026. So these are some pretty big moves in the infrastructure space in the world of AI. So, Jim, I'd love to sort of hear from you about... Do you think these frontier AI labs that have been focusing so much on building these AI models, is it inevitable that they end up having to be a cloud compute company?
Jim: It really put me on the fence. I'm not 100% sure what's going on and the reasoning behind it. I'd love to be in a few of the meetings where it was all discussed or the decisions were discussed as to make the decisions And I call it a pivot, but it's not a full pivot. They have the infrastructure. And that's the thing is that when you start to look at, I guess, the frontier labs that have transitioned to being more cloud compute like X or SpaceX now, Yeah, CoreWeave to a degree, you know, and there was another one. Anyway, you find that they've got the infrastructure there. It's a legitimate business approach for them. Like, it's a source of revenue. Why aren't they doing it? And so, or why aren't they doing it anyway? Yeah. Sure, everyone wants to do their own models and serve their own models and be that guy, but it's a logical move. So why weren't they doing it beforehand? Why is now the announcement? And I started to wonder whether there's something linked in the talent exodus or, and again, talent exodus, or is it a talent exodus or is it anthropic just attracting everyone?
Alex: Yeah. Well, I remember when I saw the news about Facebook doing anything, well, Meta, let's use their modern name. And let's be real, every Meta press release of the last five, six years has been something of a joke ever since they announced the Metaverse, which was obviously going to sink. Like, what are they running on? They're running on Snapchat at the moment. i remember thinking oh god what are they doing this time um what silly ideas have they got into their heads But I thought about it for a tick, and this actually makes a lot of sense. Meta's comprehensively lost the AI race. They're not a player in the sort of frontier space. They're doing some cool stuff with open models. Shout out for that.
Jim: But are they still, since Yann LeCun left, are they still? Like, that's the thing, isn't it? But they're not even, like you say, they're not even mentioned in the same breath as a frontier model anymore.
Alex: So you're a company that's invested heavily in data center resources for AI training inference loads.
Will: Potentially overinvested, right? That was eye-watering amounts of money that Meta put into this without a clear strategy.
Alex: Meta does not do anything other than overinvestment in silly AI.
Jim: But I think that's a little bit of Zuck's approach to things, you know, like... Don't do things by half.
Will: That's it, is it? Because that driven by double is exact. But is that why now maybe they have excess cloud computers?
Alex: This is exactly where I was going with this, right? They've gone, crap, we've got this massive amount of compute on our hands. The world is short a bunch of compute. And they're actually... So you mentioned CoreWeave earlier. CoreWeave is an interesting analog here. The problem with CoreWeave is that I think it's 67% of its revenue is derived from OpenAI. They are linked to that one customer intimately. If OpenAI ever does its own chip play, where does that leave CoreWeave? Facebook doesn't have that single customer dependency. There are lots of smaller labs that want to be building things. This could actually be a really smart play to turn like a massive amount of capital expenditure, like sort of burning a hole in the books into diversified revenue for a genuine market gap. I think it's a smart idea.
Jim: Right on. It's the same with Sam Altman has openly said numerous times that they're not going to be a model company. They're going to be a token company. And that's where Jalapeno becomes a smart idea. You realize that they have all these data center commitments. Is it all of a sudden they get their own chip with Broadcom? you know, a legitimate player who can get it done and have that diversified revenue bringing, you know...
Will: So this announcement of Jalapeno, the chip with OpenAI, I was curious about what you were saying, Alex, with most of their income coming from one customer. I just had a look. CoreWeave's share price this month is down 29%. That sounds about right. Yeah. That's incredible, isn't it? That is incredible.
Alex: I think it's also... Kind of worth highlighting, this isn't the first time we've seen specialized AI compute come out. Google's been doing their thing with TPUs for a decade now. There have been various specialized architectures announced that are going to change the game. A little bit of skepticism, but just around exactly what the advantage to building their own chips is, because that's a very complicated industrial process that will take time to perfect and scale.
Jim: You don't do that in six months. I think it's the inference side of it. And that's where Google hasn't been sweating because the TPUs are really good at inference. And they were designed that way because Google needed them to be good at inference. Is that OpenAI, Jalapeno, is specifically for inference. That's what they've come out and said. Is it NVIDIA? had to buy grok with a q um what was that three months ago four months ago now and they paid a pretty penny for it because they need better influence and i think that's part of it is that i think they're realizing that or the you know the the bigger players and the people who are the the companies who are looking further out are really seeing that As more and more of the next year down, sure, enterprise all have their place for AI in place for the next couple of years. I'm finding more SMBs are now actually going, oh, what can I do here? That's all of a sudden you get so many more businesses looking to come in on AI and you need inference in the next five years.
Alex: This has been an argument I've made for years and years now. People looking for state-of-the-art models for like, you know, there's four million business needs an image classifier and they're going after the latest and grass. You don't need that. You just need something good enough, something cheap, something that you don't need a team of people with PhDs to maintain and operate. Just get something that's easy. Um, and, and, you know, I definitely feel like the LLM space might be trending in that sort of highly commodified direction where there'll still be some state of the art stuff, but most people just kind of move to what works. Um, but yeah. Broad scheme. I don't know if the open AI announcement changes the world. It probably changes CoreWeave's life quite a bit. But by and large, I think Meta is kind of approaching this with the correct mindset, if that makes sense. The ideologically approved mindset of... ideologically approved by me, of sort of pragmatism with respect to their compute rather than expecting that AGI is going to mysteriously appear if you just pour enough, you know, teraflops into them.
Will: This does feel like one of the most pragmatic strategies to come out of Meta in many years. This is one of the first large investments that they've done where I feel like it is not a big bet that they're making on something that may come in the future. Right now, there is clearly far more demand for compute than there is supply. They're sitting on some supply. Why not turn that tap on?
Jim: And so maybe that also then, you know, because this is more about pragmatism and what is the demand going to be, does that then help to explain the talent exodus all over to Anthropic? Anthropic is still betting the farm on AGI and the talent want to go where they can actually do the research.
Alex: Well, yeah, and that's kind of an interesting thought that the talent is going where the idealism is. But idealism in business is historically, you know, it's not necessarily the best play. Does this leave Anthropic somewhat exposed if it turns out, as I suspect?
Jim: Yeah. No, no, I didn't actually put those together. It actually makes Anthropic a little bit – because they're the ones that don't have compute. They're the ones that don't have a chip. They're the ones that – and so, sure, they can design the frontier model that solves AGI, but when everyone just wants Kimi K2 to reconcile their invoices – Do you need FATL? Like, again, I'm the biggest advocate for that you don't need the top-of-the-line model. Is it stop wasting all your tokens? And it could be very interesting come, let's say, October, when IPO day comes, that Anthropic doesn't quite meet expectations.
Will: But, again, that's it.
Jim: Well, again, definitely.
Will: Definitely not investment advice. In my mind, between the AI that you, let's say, is inside a product that does something for you, you only need it to be smart enough to do that task versus if you are, let's say, a software developer working on cloud code, the smarter the model, typically, the better. So, typically, because, right?
Jim: Biggest asterisk on that.
Will: The world's biggest. Am I the business idiot of the week? Am I Alex?
Jim: But I think I agree. It certainly does. But, again, even just using Fable again, it disappeared for hours yesterday. told me it was making progress is that there's still multiple errors at the end of the day. And then, sure, it cleaned up those errors really quickly. But I didn't – sure, I didn't have to interact with it. I didn't have to do anything with it, like, for eight hours. But if you're working for eight hours at that cost level and there's still errors at the end of the day, which were clearly in the –
Will: requirements and in the prompt is that you know it's some pretty phenomenal things to me the the marketing that they've wrapped around it right it is going to be twice the cost of opus because it's coming through the api and you're going to set it off to do some autonomous tasks It better not make a mistake 20 minutes in and then spend six hours going down the wrong path because of it.
Alex: This is kind of where I kind of get annoyed with, you know, the bigger and better models from Anthropic. Since sort of Sonnet 4.6, I think we've been going backwards. We've been getting more verbose, more stuff. There's the same amount of errors. There's just like per, say, 1,000 lines of code. But now you're getting sort of 20,000 lines of code instead of 5,000 lines of code. So you're not fixing, you know, five bugs. You're now suddenly fixing... on the order of like 20 bucks, right? And that there, like you say, oh, it's more intelligent and solving more problems. Well, is it? It's just generating more output by itself for longer.
Jim: But that's where I come down on Will's side of when you're working at that coalface. the bigger the model the better the model the the better you are but i work on that cold face in planning and and working out what what i'm going to get it to build and then breaking breaking it down into a work breakdown structure that i then hand off to sonnet and sonnet then just gets everything done you know like and it's and it's smooth it's efficient it does what it's told um and and you're through it so i don't get the fable bugs in the code i just get it in the planning and that's where i think it is very useful in that time because that you know like the divergent thinking and then be able to converge it back down yeah is really helpful but there's not that there's not double the price difference between what opus can do yeah i'd say it's
Alex: Oh, 10% better.
Jim: Yeah.
Alex: And like, I'm with you there on the, the specification upfront being very, very useful. Um, I don't know. I just, the claim that it's inherently better. I think we're developing, um, more stable workflows around these models. But I find the bigger sort of newer models have a very particular way in which they want to work. And so enforcing some structure on them doesn't necessarily work as well as they did, you know, back at 4.5 or 4.6.
Will: Anyway, I found that sort of like a really experienced engineer. It's better at finding or deciding what are the things to focus on and what are the things not focused on. Now, I'm not saying it is like a senior engineer, but what I'm saying is. Opus and Sonnet and GPT 5.5 would regularly pay attention to the wrong things. And I'm always telling it where to pay attention to. And I've had several instances in the last few days where Fable has very quickly honed in on something that was a root cause or an architectural decision that was already made in the past and said, actually, it's creating some of these downstream issues and we should revisit that.
Alex: I agree. I've had it identify these particular problems. And then when it moved to implementation, it implemented exactly the same problems. Like, it's good at talking, but the actual, the qualitative stuff on the output, and look, full disclosure, the system that I was working on this week and having a lot of fun just letting tokens burn were very particular subsystems. They were subsystems that had to be just so in order to make the rest of the system kind of hang together, right? It wasn't crud logic. It wasn't sort of database updates or stuff like that. It was stuff that required some... very particular approaches in order for the thing to hang together. Um, and in that case, like to me, that's where the difference should show up. And I'm not saying that I am seeing it fine in boilerplate. Like it does great at boilerplate. It identified some old bugs in some, some report generation logic that I had. And, um, it successful, like I had a, a work tree in flight, but bugs in report generation. No, no. Oh dude, you shouldn't.
Jim: Yeah.
Alex: Some of this code is so sloppy. It was jammed out in literally like a day and a half, an entire reporting subsystem. And when I find bugs in it, I'm just like, yeah, well, I did zero QC on this. That's on me.
Will: So in the context of Fable having, you know, much longer time where it can run autonomously and you're talking about it being more verbose and therefore burning through a lot more tokens. Yeah. So therefore like Anthropic is even increasing in their demands for compute even still. Yeah. They're also the one that doesn't, you know, in this week, we've got OpenAI and Meta, which are moving into cloud computing. And does this leave Anthropic exposed only having agreements with like SpaceX, for instance, for some of its compute? And I'm sure it has other compute suppliers, but is it exposed not taking on that vertical integration?
Jim: Well, I think it's very lucky that they gave that big chunk in the last raise to AWS. You know, like is it, it certainly, they haven't, agreements but there's no absolute focus or there's no like what to yeah alex was saying before diversified revenue and so what happens when everyone goes you know what the extra price isn't worth it i've worked out my process i don't need the big role um i'm just going to you know even if i am using it anthropic i'm just using iq or i'm just using sonnet and and so all of a sudden your beautiful revenues that you've got that you're posting now going you know go to the go to zero or not zero but but they certainly there's a big fall off poor encore um and so i think it does leave anthropic exposed but it's also um Does, again, you know, I'd be very interested to see what they're pitching the top talent, you know, because if they get all the good cattle and they can then solve things at a higher rate, do they then get to a model that can do it for cheaper, faster? And so then they own that whole market because everyone else diversified a bit. and weren't focused on the 100% goal. We all saw what happened to open AI in the outset, you know, is that when you want to do video models and video creation apps and deals with Disney and focus on other things other than enterprise AI,
Alex: you fall behind and so is this what's actually happening i feel like there's a physical limit here we're going to hit um i'd love to see what energy costs look like for some of these frontier models that will tell us how far we're going to get there there comes a point where you just can't play they're also probabilistic like that's the thing is that there is an absolute limit
Jim: And so I think it sets it up for an incredible end of the year when they're all trying to tell me how good their company is to go public. Yeah. And it's exciting. That's what I love about it. It's exciting.
Will: All right. Well, let's pivot over to Europe now. So, we had a court ruling this week in Germany on a Google case. So, the case was around whether Google is responsible for the outputs of their AI in those AI overviews. When you search on Google now and you get an AI overview at the top, is Google responsible for the outputs of the AI in that top section? So, The ruling here was that Google is liable for it. It's a preliminary ruling at this stage. But Google is claiming that they've been able to provide disclaimers to their users that it is AI and they shouldn't necessarily trust or they should be verifying the outputs. But that take was not enough for the German court. So, over to you on this one, Jim. Do you think that this is going to set a new precedent for AI companies?
Jim: I think everyone has to use this ruling in their thinking is that because it really does, it delineates really quickly and becomes not an employee. And again, I'm not a lawyer, but you can see where the precedent is going to set. And in this case, like when you get into the specifics of it. It was – and so it's something about two publishing companies and Google – the Google summary or overview came back saying that they were caught up in a scam. And so these two publishing companies saw that as defamatory, and that's what the case is about. And it's because Gemini – I shouldn't say Gemini – associated them with the scam in summarizing everything and then gave the overview. And so that's a very specific scenario is when you then take it into – or is Anthropic going to be liable for the output that – it gives me for a blog topic, a blog article I'm writing, you know? Claude Cowork is fantastic for that sort of thing. Are they going to be responsible for the output of that? I think there's a fairly big difference in... Google giving an overview rather than 10 blue links and saying, because there's a sort of an authoritative nature to that.
Will: There is. They're saying that within the next 10 blue links, the truth of it is this.
Jim: Yeah. And that's it. And so they're summarizing it. Whereas in the more common use case, it's anthropic is going, oh, here's what you asked me to do. And it's on the user to go, no, that's not what I wanted.
Will: I wanted something better. It's probably something false and then you post it in a blog post. That's on me. My overview is effectively and what the court was saying is Google was creating their own new substantive claim that they were putting on the public internet.
Alex: We're talking about the first 10 links here. If defamatory content appears in the first 10 links, is Google liable? And I'm going to defend them briefly because I have a way more fun take on this whole topic. But let's start with that simple thing. Are the search results... do they constitute an authoritative claim about the nature of a topic or is it just whatever index?
Jim: I think it's just because the way in which they phrase, like you think about what comes back from an LLM, it is an authoritative tone in a way. They're confident in what they're saying. And I think that's what comes back in it is that with the 10 blue links, it's, Have a look. I think it's somewhere in here, right? Based on our algorithm, the answer is somewhere in here. And you've got to click through a few, right? But you're still making the decision as to what counts. Is it because it comes back, the overview, it comes back, and there's that confident tone in it? I can see where the German court is coming from and going. Well, the general public is going, like, again, I never believe in anything that the thing's put back to me, you know, but again...
Will: If that AI overview started with, you know, contained in the next 10 blue links, there are claims that these publishers are associated with a scam, then there would be nothing defamatory about that. It would be the actual links themselves, but... it doesn't reference the sources in that case. Maybe it does. It's like a, you know, something you could click through to, but you're right. It's such an authoritative statement around. And this is the truth of this, of this question.
Alex: That was their whole pivot. Their whole pivot was the search business is basically dead or a couple of years ago, they raised a red alert or something. They're like the search business is on life support. AI is the future. You know, we need to, we need to lean into this now. And look, to be fair, I think they've done a pretty good job in leaning into this, but yeah, If you hedge in that respect, does it add that much value? Because even if you're signing onto a website, how often do you read the terms and conditions? You just gloss over it mentally and click, I accept, and off you go. You're still making the same fundamental fact claim. And I think the German court is perhaps being a bit... I'm going to say pedantic, but, you know, you're not supposed to accuse Germans of that. No.
Jim: But I think that is the way of the world, is that the legislative bodies in most OECD, G7, you know, larger nations are in that phase where they're looking to protect the broader public. Yeah. And that's because when you think about it, and so that's where I don't think, I think it's going to become very specific to Google this, ruling yeah and again if it becomes an official ruling like it is only primarily now could because when you when you then drill further down into the alternative right where anthropic is then responsible for mistakes in my blog post It changes, it fundamentally changes, and that is, Will was saying this in our group chat over the week, it fundamentally changes the whole way AI works.
Alex: Well, I've been arguing for some time that generative AI in general doesn't does funny things with the idea of intellectual property in a way that we either have to double down into intellectual property and a lot of what we're doing currently with AI and sort of statistics even suddenly comes under a whole new set of laws because, you know, it's one of those what is pornography. I know when I see it situations, there is no hard line here. All we need to rethink what intellectual property is. Now, I think intellectual property as an idea has some merit for individual creators, but you look at what some certain large cartoon giants do with it, and you get very leery about what it means for the future where people can own ideas, right? Yeah. If you remember Francois Chalet's On Intelligence, right, like this paper about why AGI is unlikely, his central argument there is, or one of the central arguments, it's a very long paper, is that the shared ideas are really what creates intelligence within a society. You know, Albert Einstein born back in Roman times isn't going to come up with a theory of relativity because there's not the sufficient ideas. I mean, he might still be really smart. He might still be a brilliant mathematician, but he's not going to come up with relativity because they don't even know about the number zero. Exactly.
Jim: And that's the thing is that the probabilistic nature of the way LLMs work to me is that's where I agree, you know, in the age, you know, is AGI going to occur because it is just taking the midi part of the curve. It's not like, and so. That to me is, again, Google's going to have to tweak the way the overviews work, I don't think, because you cancel out the entire technology if you then start ruling that all the models have to be accountable.
Alex: Oh, no, I'm 100% for this. I think this is a fantastic idea for some, like, one of the big... problems I have with how these things are being sold is, oh, they're intelligent. I think Google bought this on themselves. Here's a spicy take, number two. Google bought this upon themselves when they marketed this as artificial intelligence. It's not artificial intelligence. It's statistical pattern batching. Yeah. right and if you if you if people go into interacting with these things knowing oh they're just kind of like random numbers filtered through some linear algebra you know it's maths at the end of the day not thinking then i'm not sure i think someone would have come along and waved the magic stick over and called it ai eventually i i i agree but like you know what's the uh expression you mess with the bull you get the horns good good on the court for being like no The courts aren't used to a world in which written expression is not made by humans. Fair enough. The world is kind of moving along, right? We have similar problems when it comes to, like, legislative documentation in a world that's, like, largely digital and paperless. How often do you have to deal with the government with an actual physical bit of paper because they just haven't adapted yet, right? I get all of that. But let's hold these companies accountable. If they're going to talk about, oh, this thing is the most intelligent thing ever and we're going to let it do science, rather than it's a tool we're going to use to help support science. Yeah.
Jim: Hold them to it. Hold them to a higher standard. I see your point, and particularly for the broader public because the broader public don't understand that it is just linear algebra. And so I hear what you're saying. I think you're going to cripple the entire industry if you're forcing them to be accountable for everything that comes out because they can't be because it's linear algebra.
Will: So, as we're talking about, I suppose, the dangers here of some of these kind of AI tools, let's talk about the controls that are being applied to them now as well. So, we obviously saw, you know, when Mythos first came out, we had Project Glasswing, which is, you know, providing access to a hand-selected list of companies. And we saw the release of Fable, and then it got turned off, and now it's being turned back on, having gone through a U.S. government approval gate there. Yeah. We're also seeing the same thing now with GPT-5.6. So, Sam Oldman was talking about voluntarily submitting the release process of GPT-5.6 to U.S. government approvals and working with them. So, we're also seeing there that certain companies are getting access to models before other companies. So who's actually controlling the release of these models? Is it the company themselves that is doing a consumer launch or is it actually now the U.S. government is the one who kind of owns the release of these models?
Alex: I think Anthropic have shot themselves in the foot with their Midos release. I mean, I haven't seen it. Maybe it is the most amazing thing and it's absolutely terrifying what it can do. I'm going to guess it's probably not. I'm going to guess if it was that amazing, they would have gone to market with it straight away. And it probably had some severe limitations that meant they'd hyped it up. Oh, it's good at some things on this benchmark. Let's use that as an excuse to not release it. Because remember, Opus 4.7 completely unannounced drops like three days later. And, oh, you know, this is a new model. It's not Mythos, but it's Opus 4.7. Opus 4.7 was famously bad. Yeah. So whether or not that was just a mythos in disguise. Now, of course, as we're saying with the general public, they hear these stories and they think that this is actually like GLaDOS or HAL from 2001. Politicians aren't engineers or statisticians. They hear the story, they see public danger, they act.
Jim: Is this just the world we live in now where the marketing story has a life of its own? And when you go down the Twitter hole on Mythos particularly, is that that is exactly the – or reportedly that's exactly the chain of events. Is it all of the bureaucrats went, oh, hang on, you told us this was the most dangerous thing in the world, and now one of your trusted testers came back and said – hey, we were able to jailbreak this. We've got to put exports in. It's literally the chain of events. And I agree. To a degree, I think they have.
Will: Can I connect some dots there as well? Now that we're starting to see a few things come out. Firstly, the positioning of, you know, and this is, I think, Dario said that this is like a super weapon mythos, right? So we've got that. The next dot is, you know, wanting to integrate these things into the Department of Defense and use them for national security. The next step is Anthropic announcing that they were the victim of a distillation attack where they're saying that Alibaba was able to use 25,000 accounts, millions of messages to replicate what that model was. And then now China effectively has the same capabilities around sort of Opus 4.8. So are we effectively seeing here the US government connecting these dots and saying, you just released to us something you call a super weapon at the same time as saying that the Chinese are able to steal your super weapon whenever they want? If I'm the government there, I'm going, and if I don't understand these things... Dots connected.
Jim: I'm locking that down. And understandably.
Will: It's not being that, like, the Chinese could steal a nuclear weapon any time, but let's publicly release the access to nuclear weapons. Hell no.
Jim: I'm also really interested in the desire to still report the... Because the Chinese, like, from all reports, if it's all still alleged... But if the Chinese were doing these distillation attacks, they were still paying for it. Anthropic was more than happy to report the... the income, you know, and when they're doing their raises. And so in one hand, they're reporting the income. The other hand, they're saying they've got a super weapon.
Will: There was only 16 million exchanges, apparently.
Alex: I feel like I do that in a week. How secure is your technological moat if 16,000 exchanges, 16,000 or 16 million? 16 million. But still, that's a tiny amount. Given the number of parameters in these models, you have barely explored the specifics.
Jim: I also can't get past the hypocrisy side of the distillation attacks, is that you train the original models on the corpus of the whole world's data, which you've got for free. You just have to settle a whole lot of cases with book publishers. With no credit and no licensing. Exactly. How do you sell these lawsuits with these book publishers just so that everyone went, oh, no, okay, you're all right now, and then you're whinging that someone's stealing their responses? I understand you've put a lot of effort in and they've refound it, but it is just either way is that I think there is a legitimate – security issue, which I'd love to see Mythos with the rails off, you know, like, and back to what we're talking about, you know, like the talent migration to Anthropic, is it that great that that's why they're all going because Mythos with the rails off is unbelievable, you know, like, and so these are the things.
Will: So you get a demo of that as a part of the interview process too. Exactly.
Jim: That's how they convince you to get the chip in your neck. How do you convince Carpathia you come out of retirement? You absolutely get a demo of mythos in the interview. If you're not, if you're not anthropic, that's how you should, they're like, start doing that and everyone will want to go and work for Carpathia.
Alex: Except that they can't hire international talent.
Jim: I think it is just, I think it is literally just that they're caught between their idealism which i think is you know i think they have they have the right attitude to to protecting everyone and doing it all in the right way and also trying to do an ipo and i think that's the that's where they're caught and that's where they're tripping themselves up to alex's point about you know whether whether they do it themselves but i'm again i'm You know, so, you know, the stereotypical angry old man where I think, oh, yeah, the government should stay out of everything. But I do think there is a legitimate pathway here. And that's why I quite like the way, like, again, you know, don't particularly like Sam Altman, but he's really done this. He's managed this really, really well in saying, hey, U.S. government. We're voluntarily doing this with them. Meanwhile, they've just done exactly the same thing as Anthropic. Got their core customers, given them the mall early, and said, oh, well, we're also giving it to the U.S. government. And that's what the press release was.
Will: Yeah, see, sometimes I think it's just really good marketing what they're doing at OpenAI. So if you've created GPT 5.6 and you know that it like is not as good as Fable or Mythos, I reckon one of the best things you can do is group it in there and tell the US government it's just as good. It's just as strong. You know, give us a three-month release window where we can tell the world how good it is and buy our engineers some time to continually make it better. And then when we release it, everyone will go, OpenAI deserves the $1 trillion.
Jim: Either way, all the credit to them. Then they're playing the game better. Rather than Anthropic saying we've got a weapon of mass destruction and then getting tripped up when they then get export controls put on it. And so... That's the thing is I just think they're sitting back like they're playing the game like a good number two should.
Will: Well, I think the real tragedy here would be if we have an open source model, let's say in the next six months that has benchmarks similar to Fable and there are some politicians who get convinced that open source is dangerous.
Alex: Yeah, I don't think... I think Anthropic would love that. But it's the same as redefining IP so that we can't use anything that could be considered IP in model training. It would kill so many things. Politicians don't understand how much of the modern tech ecosystem, like 95% of machines are based on Linux operating systems. When you consider cloud fleets, does that mean we subject cloud fleets to regulatory approval for every upgrade?
Jim: And we do not restrict Linux.
Will: Yeah. For the audience there, can you maybe explain what you're talking about there, Alex? No.
Alex: Okay, so... Well, Linux is one of these systems. It's an operating system. It's like Windows, but it works on lots of small machines. And most of the internet doesn't run on Windows. It doesn't run on Mac OS. It runs on these small machines with this open operating system.
Will: Restrictions on that would nuke the internet very quickly. In the same way that AI could be nuked in many companies, the way they're deploying open source AI.
Jim: That was exactly going to be my point, is that I don't believe they can or will restrict open source purely in the fact that and open weights, like even just that, that anthropic and NVIDIA deal where they're putting, you know, Nemotron, which is, which is the NVIDIA Palantir one. Yeah. Palantir. Sorry. Yeah. Um, NVIDIA Palantir, they're putting open source into like sovereign environments in the U S government because you can do that with an open model, you know, is it, You can put it in a completely blocked off environment. Yeah, the biggest enterprises in the world do want open source. Because you don't need the big model. You don't need Fable for what they're trying to do. And you also can't use Fable because it's in somebody else's data center and you can't share the data. And so that's where I think the open models are going to become more of the norm. Back to our original... discussion or original topic today you've got to be a bit more diverse and as a as an ai lab or a leader in the AI industry, you've got to have diverse revenue. And so you've got to have an inference chip. You've got to have a cloud providing option.
Will: To wrap it up there, guys, for this week, this was podcast number one for the business idiots. So who do you think was the business idiot this week, guys? Have we got a clear winner?
Jim: The marketing teams of the AI firms are the business idiots this week. Cool.
Alex: I think it's Will because he named himself earlier. You've got to take that one.
Will: Absolutely. All right. Thanks very much. We'll be back next week. Bye-bye. Cheers.
Related episodes
One email a week. No courses, no funnels.
Sign-up is not wired up yet. Check back soon.