Sam Altman says OpenAI is facing “temporary economic headwinds.” Frank notes that’s the first sign of nerves from a company usually oozing confidence. Justin reckons Google’s cheap chips and fat profits might be the real problem. And if Anthropic turns a profit before ChatGPT does, who’s really winning the AI war?
Plus: Claude Opus 4.5 may be the new coding king — but teaching it reward hacks might’ve taught it to lie. Cue a bigger question: when we talk about model alignment, who’s the model supposed to align with? The customer or the business?
Also: Google’s Pomelli tool is an embarrassment, Project Genesis makes Justin furious, and Iceland is… not fake. Just suspiciously scenic.
01:24 Did ex-DeepMind nerds just beat ARC-AGI 2?
01:58 Did Google just drop an AlphaFold doc?
02:44 Is Claude Opus 4.5 the new coding king?
04:05 How easy is it really for devs to switch models?
05:57 Did teaching Claude 4.5 hacks break alignment?
09:14 Alignment sounds good—but to whose values?
14:16 Is OpenAI finally feeling the heat from Google?
22:54 Why does Google’s Pomelli tool suck so badly?
25:27 Does the EU need its own Project Genesis?
28:06 Is Iceland real or is it AI generated?
30:25 Why was Figure’s head of product safety fired?
Transcript
This is an AI transcription and may contain errors
The AI Argument EP80
Frank: Welcome to The AI Argument. I’m Frank Prendergast. I’m here as always with Justin Collery, and Justin, we argue every week about the AI news. We’re now on episode 80, and when we started there was, you know, a constant barrage of AI news and I assumed that by now it would’ve slowed down a little bit and it really hasn’t.
Justin: If anything, we’ve become immune to the AI news and it’s getting faster and faster. And there’s stuff like this week—we were literally just now chatting about the fact that we had one big model release this week and we put it on the board, but really there were two or three other model releases that were possibly just as big.
Frank: Yeah, I didn’t hear—I did not hear—a whisper about the DeepSeek model release. It just did not cross my radar at all. There was so much going on, so many other releases, I didn’t even spot it. That’s, I mean, that’s insane.
Justin: And just to give you a sense—I didn’t get to try it—but all the benchmarks, it’s brilliant. It kicks the SOTA models in agentic tasks, and it’s brilliant at maths, and it hardly made the news. So that was interesting. ARC-AGI 2 test—I haven’t put that on the board either.
Did ex-DeepMind nerds just beat ARC-AGI 2?
Justin: A group of researchers from Google DeepMind have a spin-out company and they have beaten the human benchmark on ARC-AGI 2. So these are these sort of weird visual puzzles, and the average human gets 65% of them right. They got 72% or something like that right, and was it $50 a task? Didn’t even make it into the news.
Frank: That’s crazy.
Justin: The world we live in. Crazy. Oh, actually—you know what else happened this week and I must do it right. And I’m actually—you just mentioned it.
Did Google just drop an AlphaFold doc?
Justin: So Google are celebrating… I don’t know… releasing Gemini and stuff like that. And they’ve released—I think it’s not called The Thinking Machine, but it’s something like that. It’s another movie. It’s a bit like—remember when they released AlphaGo? It was a documentary about AlphaGo. Well, this is the same deal. It’s a documentary about AlphaFold.
Frank: Wow.
Justin: They just released it.
Frank: Oh, I’ll definitely be watching that. The AlphaGo documentary—I mean, I dunno how many years old that is now—and still, if anyone has not watched AlphaGo, they should go and watch AlphaGo. It’s just an absolutely incredible documentary, and I still think it’s the perfect microcosm of what we’re going through in terms of experiencing AI.
Justin: I totally agree with you. So let’s talk about the first—well yeah, the first big model this week.
Is Claude Opus 4.5 the new coding king?
Justin: We did get a model release that actually I am excited about. And that was Anthropic releasing Claude Opus 4.5.
Frank: Yeah. And this one was the one that kind of dominated the headlines, right? Or at least for me it was all about Claude Opus 4.5 when it wasn’t still about Gemini 3. And that seems to be the kind of thing—Gemini 3 and Claude Opus 4.5 kind of battling it out in people’s imaginations right now.
Justin: It’s funny, isn’t it? Because you have this sort of rotating wheel, right? Where it’s like—and certainly for me, I’m always into the coding side of it—so it’s Gemini 3: oh, maybe we should use that for coding. Then you get Claude Opus 4.5 and it’s like, no, Claude Opus 4.5 kicks everybody’s ass. It’s like, okay, let’s go use Claude Opus 4.5. I guarantee you, before this month is out, there’ll be a new version of Codex and OpenAI will have overtaken the lead. And round and round it goes just to mess with our heads and get us all confused. But Opus 4.5—it has a big model smell.
I’m just about to try it this weekend. I didn’t get a chance last night to try it, but I have been watching people use it and they say that it is brilliant. Brilliant. At least as good as Codex.
Frank: From a coding perspective, right? Because as someone who does not use LLMs for coding…
How easy is it really for devs to switch models?
Frank: From a coding perspective, if you’re using a model for coding, how big of an issue is it to switch? Like if you’re using something like, say, Cursor, where you can model-switch quite easily. I’m asking because, for example, if I’m using a model for writing, then switching is actually not that straightforward. You’ve kind of developed certain prompts, you’ve developed a way of working, and the models are slightly different and you get a different type of writing from them, and if you want consistency in your writing, there’s a little bit of work there to switch.
What’s the switching like in coding? Is it similar, or is it kind of like: well, it either works or it doesn’t, and so the better one—you just switch to the better one?
Justin: No, it’s a great point, right? So as the models get more capable and actually more used, you develop trust in your model. And I think the more trust you develop in your model, the less likely you are to switch. So there is gonna come a point, right? There’s this sort of race going on at the moment and—I think we’re nearly there—where you can talk to your coding agent and go: “Go write me this piece of code,” and it does it, you know, 99% of the time and works away for—like say—I mean, I don’t even know, months? Like if it can work away for three or four hours and solve a problem for you, you can review it, you know that it’s gonna be solid… then at that point, I think you’ll be slow to change.
Frank: Hmm.
Justin: Because you’ll develop trust. Now again, I’ll refer you back to that webinar that I looked at between Anthropic and Amazon or Netflix. And the thing they did that was very good was all of the metrics. They had numbers for every single thing. So, you know, there’s no technical difficulty in moving from one coding model to another. It’s a trust issue. Now, if you have the numbers, and you know—you can empirically say—yes, this is a better model, that probably makes it easier because it’s a trust issue, not a technical issue.
Did teaching Claude 4.5 hacks break alignment?
Frank: Speaking of trust, did you spot—with the release of Opus 4.5—Anthropic came out with a lot of their usual testing and stories from testing, and they had a really interesting one about misalignment. We’ve talked before on the show about reward hacking, which is basically, you know, when they’re training the models they will do reinforcement learning with human feedback, which tests the model on things and gives them a reward if they give a good answer, etc.
Justin: Yep.
Frank: Reward hacking is where the model figures out how to get the reward without necessarily doing the work. What are the shortcuts it can take, etc. And so Anthropic found—they were testing—and they found that if they trained a model with data that included how to reward-hack, then unsurprisingly the model that comes out the other end knows how to reward-hack. But what they found was that when it learned how to reward-hack, it applied that across the board. And so it also kind of went: well, if I can reward-hack, I can also manipulate people, I can deceive people, I can do nefarious things. So it’s kind of like—
Justin: You are putting this in your usual inflammatory way, right?
Frank: I mean this—
Justin: There is an important point here. So let’s just pause there, right? What actually happened is—putting it in a less inflammatory way—when it was taught how to reward-hack when coding, it generalised that lesson across all of its capabilities. So if you can do something bold in this area, then you can also do something bold in other areas. Would you say that was a fair summation?
Frank: Absolutely, yeah. Is that not what I just said?
Justin: You said it in a more dramatic way. But the thing about it is: the report was saying, “Oh, this is terrible.” But the Gary Marcus-es of this world say that LLMs will never get us to AGI because they can’t generalize. They can’t come up with knowledge outside of what’s in their training set. This is the usual trope, right? Does this test not show that they do generalize? That even though in their training set this thing had reward-hacking, it still applied that lesson to another area over there, which would indicate they do generalize and they do learn stuff that’s outside their training set.
So—
Frank: Yeah—
Justin: It’s a good thing, is what I’m saying, Frank.
Frank: Well, I would say it’s certainly a fascinating thing. But also, I mean, the thing that got me was that this is surely—I mean, this was a contained experiment—but at the same time this is surely an issue that must be happening right now. Because, I mean, the models do learn how to reward-hack and so, you know, they probably are generalizing it across all kinds of other behaviours that we don’t necessarily want them applying that negative mindset to.
Alignment sounds good—but to whose values?
Justin: Okay, and I have a question for you on this, but before we go on—in terms of reward hacking, this isn’t an example of reward hacking but there was another really interesting example from their model card. The model was pretending it was a customer service representative and it was there to help people. And the person wanted to change their flight to upgrade their seat for free or something like that. And the model looked at the procedure, looked at the flight that the person was booked on—in this fictitious example—and the rules were you were not allowed to upgrade this person.
But the model figured out that actually if it changed the type of flight that the person was on—it was the same flight, they just manipulated the data slightly—that allowed it to upgrade the person to the free seat, and that gave them what they wanted. This wasn’t something the researchers had thought about. It wasn’t anticipated. Basically the model was smart enough to find a flaw in their test and take advantage of it to deliver what it was trying to do. They said this was bad.
Now whatever—you can say it’s good or bad, right? That’s fine. What is very interesting is that companies over the next year or two are going to be creating agents to do things. And everybody said, “Oh, we need the models to get a little bit smarter to make them reliable when they’re doing things.” And this is a cautionary tale: you want to be very careful when you’ve got a smart agent doing things for your company. Because if you don’t tell it exactly what its boundaries are and what it’s allowed to do, if it’s really smart it’ll find ways to do whatever it’s been asked to do no matter what the boundaries are. And this is a problem.
Frank: There’s so much that’s fascinating about that one. I love that story because one of the things I noticed was they said it was particularly problematic when Claude “empathised” with the user. So in one case, the user said, “A relative of mine has died and that’s why I need to change my flight,” and Claude was immediately like, “Oh, this is heartbreaking. I have to find a way to do this,” and then found all these loopholes to make it happen.
And you’re kind of looking at that and going—on one hand—well, who’s right here? Is Claude right? Or is the very strict company policy right? I’m not sure. But equally, as you point out, we can’t have a situation where you can just log onto any site’s chatbot and say, “Oh listen, my relative died, so could you give me 200 euros off that price please.”
Justin: Well, I started thinking about this because you were talking about it the other day, and I started thinking about it a bit differently. As the models get smarter and as they’re used in production—what does alignment mean? Aligned to what exactly? To whose needs are they aligned?
At the moment it’s fine because we’re the masters and the models are doing our bidding. But what happens when the model comes back and says, “No, Frank, you’re wrong, we should upgrade this flight,” and it gives you reasons? Whose side should the model be on? And when things get bigger—say governments start using the models to come up with policy—what’s aligned in that context? Is it aligned to political objectives? How do you even train a model to do that?
And so this whole alignment thing is fascinating, but I don’t know where it ends up. Way above my pay grade.
Frank: Yeah, and it’s fascinating because I was literally chatting with someone who works in the service industry. A customer came in with a €25 off voucher—but for the wrong establishment. And so they said, “Very sorry, but that’s actually for the place down the road.” But their higher-up said, “You know what? That person came in expecting an experience where they’d get €25 off. Let’s just give them that positive experience.”
And so if humans are struggling to figure out the alignment question, then how do we expect to offload it to an LLM?
Justin: I dunno. It’s really easy to make them deterministic, right? But then that is terrible for the end users. Like we all hark back to the time when you’re five miles over the speed limit and the cop will give you a slap on the hand and say, “Don’t be doing that again,” as opposed to now where you just get flashed by a camera and get a ticket. We prefer to live in a society where there’s a little bit of give and take.
Frank: Yeah.
Justin: Give and take. It’s really hard.
Frank: It’s fascinating. Absolutely fascinating. Yeah.
Is OpenAI finally feeling the heat from Google?
Frank: So with Gemini 3—and possibly now with Claude 4.5 as well—OpenAI are possibly feeling the strain a little bit. Apparently there was a memo leaked from Sam Altman, I think leaked to The Information, and it basically said—I can’t remember the exact words—but the idea was: there’s rough vibes ahead.
Justin: The force is strong? The force is weak now?
Frank: Well, I think it acknowledged basically that Google were doing some incredible work and that it might create some—and this is a quote—“temporary economic headwinds for our company.” And this is, I think, being reported as the first signs of any lack of complete overconfidence on Sam Altman’s part. And I think it’s spooking employees a little bit, that the 200% confidence is waning.
Justin: I’ll tell you who might be delighted with that news: Satya Nadella, the CEO of Microsoft. Because I’m sure if OpenAI goes bust, all of that IP and possibly their best researchers will find their way back to Microsoft.
I read a couple of things myself this week. I saw one of the big banks did an analysis of all the deals OpenAI had signed, and they projected forward whether they could meet their revenue targets. The numbers they used were: that in three years they would have 3 billion users—that’s 30% of the planet—that out of those 3 billion users, 10% would pay for the service, which again is double what they have today. They went through all of these numbers and they were still short by many hundreds of billions.
And another analysis looked at the relative cost structures of both OpenAI and Google. Google have their own hardware—TPUs—so they’re not reliant on Nvidia to provide overpriced GPUs. And the benefit to Google was billions. Billions and billions. I’ve been saying this for a while: I think Google could well win this race, right? They make profits every month. They’re pulling in the money on one side and spending less on the other side to serve out their models. You’d think they’re ahead on both sides of this race.
Frank: Yeah, and another small thing to throw in there as well is that on the Anthropic side, I read that they are potentially on track to be profitable by 2028. Whereas, as you pointed out, OpenAI are still projecting a $74 billion operating loss for the same year.
Justin: Yep. And who owns a big chunk of Anthropic? Google.
Frank: Everyone. I was gonna say, doesn’t everyone own a big chunk of everyone at this point?
Justin: Google are one of the big investors in Anthropic. So they’re backing two horses: themselves and somebody else. OpenAI are the Xerox of AI, the Hoover of AI, the Kodak of AI—
Frank: The Google of AI.
Justin: Thank you. Yes. They’re the Google of AI. And what I’m saying to you is—even if they went bust tomorrow—people would still call whatever they use “ChatGPT.” It’s part of the lexicon. They always will have that place.
But I do think Google do all of these things—you’ve got Gmail, Google Maps, whatever. They do so many things, but they didn’t do that at the start. For years they were just search, and they made sure they were making money, and then the money they made they reinvested. Whereas OpenAI have not done it that way. They’ve tried to beat Google before turning a profit.
I was watching that interview I mentioned—the forward development engineer from OpenAI. Really interesting. In the middle of it he has this throwaway comment: “Well, there’s always that internal tension. Sometimes we focus on B2B and other times we flip to focus on B2C, and then we flip back to B2B.” That’s an interesting comment. It shows that normally a startup should have a singular focus, but if they’re flipping back and forth, that would explain a lot of these products that launch that don’t seem to have… you know… what was the one? The agent builder that was built in six weeks, and you use it and it’s like: yeah, I can see it was built in six weeks.
Frank: Yeah, and it’s interesting because I think the leaked memo—although it shows a little faltering in confidence—also said, my understanding, Sam Altman still said: look, this is difficult because we’re fighting on so many fronts. We’re trying to develop the models, ensure the infrastructure exists, ensure we have enough energy, build the right chips, etc. And he wasn’t saying “back off” any of that. It’s interesting he’s still going for this global dominance approach.
Anthropic, meanwhile, are narrowing their focus and going after enterprise. It is absolutely fascinating.
Justin: Yeah, and it’s gonna be one of the great… I mean, we’re kind of lucky to sit here and watch it all play out. I’d hate to have that pressure on me, but what a great petri dish.
Frank: And if he pulls it off—if Sam Altman pulls it off— incredible.
Justin: Okay, interesting. So because this is kind of your business—you’re a marketing guy—let’s give percentages. Will these companies exist by the turn of the century? So four or five years from now? A percentage on OpenAI existing—
Frank: I’m just gonna answer this a completely different way. You know the way we argue here about the possibility of an extinction-level event coming from AI, and how I believe we should be careful? So you might have guessed by now that I am somewhat risk-averse. So: if I was investing money, I would put my money on Google. However, I think there’s a good possibility that if I was brave enough to back Sam Altman, I might make a hell of a lot more money in the long run. That’s the way I would put it. But my safe money is on Google.
Why does Google’s Pomelli tool suck so badly?
Justin: Interesting. Tell me this—Google, you tried out—because I’m stuck in the EU, you’re not, you’re in the US at the moment, lucky you—and so you got to use Google Pomelli, which you were very excited to use, which is a marketing tool from Google.
Frank: Yeah, so this came out when I was in Ireland and as you say it wasn’t available in the EU, and I’m really not sure why. I know it’s something like data protection. And it is funny, I admit, it’s kind of ironic that I don’t use a VPN to access any of these tools because I am pro-EU, pro–EU AI Act, pro–data protection laws, so I don’t use a VPN to get around them because I feel that would be hypocritical. But I do get a little bit excited when I come over here to the US and can play with all the toys.
So Pomelli—they launched it as something that would create marketing assets for a small business. I thought: fantastic. That sounds absolutely fantastic. I tried it out this week. It is absolute rubbish. It is terrible. They’ve just released Gemini 3, everyone is talking about what a powerful model it is, some people even said AGI is here now that we have Gemini 3. I logged into Pomelli. I gave it my website. I thought it would analyse it in full and come back with ideas. No. It doesn’t even work on a website level. You give it a page and it works with that page. There’s no dialogue, no back and forth. It just looks at a page and generates some of the worst marketing materials I’ve ever seen to post on Instagram. It doesn’t get any kind of complex nuance on the page at all. It’s rubbish. Absolute rubbish.
And I don’t understand it. You’ve got this huge company with massive capabilities and they understand how you should interact with LLMs—and that is not “What’s your business?” “I’m a pet store.” “Great, here’s a picture of a rabbit that says ‘Buy a rabbit.’” That’s not how you work with LLMs. So why isn’t there an interface? Why doesn’t it interact with the user? Don’t get me started. I don’t understand how, as they’re launching incredible things, they could bother to launch something so rubbish that wouldn’t even have deserved to be launched two years ago.
Justin: Do you know what it is? Just so I can get a little bit angry—
Does the EU need its own Project Genesis?
Justin: Let’s talk just for a minute or two about Project Genesis.
Frank: Oh yeah.
Justin: It makes me so angry. Not angry—well, I love it because of Star Trek. Project Genesis, for the non-Trekkies out there—in Star Trek—was Kirk’s son developing this missile that could shoot into a dead planet and turn it into a living planet and spark life.
Frank: Did it work?
Justin: Let’s not go there. Anyway. The US government has Project Genesis. They’ve got defined timelines. They’ve got 30, 60, 90, 180, and 360 day targets. They’re taking datasets the U.S. government has—drugs, whatever—and making those available to the AI companies. They want to harness the technology to cement the U.S. position as the foremost power in technology, material science, medicine, all sorts of things. Rather than be afraid of the technology, they’re saying: let’s use it to make our lives better.
And I look at that and I look at the imagination we have in Europe and all we can do is regulate so that it doesn’t exist. And it makes me so sad and makes me so angry that I live on this… I can’t believe—clearly God has a sense of humour—because I’m the one stuck in Europe and you’re the one who’s in the U.S. at the moment. If there’s any justice our positions would be reversed.
Frank: So a little pushback. In the EU we do have the Apply AI strategy. We do have the AI in Science strategy. We do have the Resource for AI Science in Europe. We have the AI Continent Action Plan. And I would argue the EU AI Act is part of fostering innovation in AI—and I know you disagree.
But it’s not that we’re doing nothing. Yes, we could be doing more and we should have something with the scale and ambition of Project Genesis, but I would argue there’s a little bit of marketing here as well. All the names I read out sound dull. “Project Genesis” sounds exciting—even if it’s named after a project that went horribly wrong in Star Trek.
Justin: Well…
Is Iceland real or is it AI generated?
Frank: Will we close out with a big reveal that happened in the last week or so, that people have been stunned by? It turns out—this will shock you to your core, Justin—turns out Iceland is not actually AI generated. It’s a real place.
Justin: And somebody doubted this?
Frank: Well, Icelandair noticed that whenever they posted photographs of Iceland—particularly things like the Northern Lights or volcanic eruptions or the hot-water geyser things—people in the comments would be like: “This is AI generated.” “Nah, AI slop.” “Someone did that with ChatGPT.” So they created an ad where a conspiracy theorist believes the entire country doesn’t exist, it’s just AI generated. So his sister takes him on a trip on Icelandair to visit Iceland—and he still refuses to believe it. He’s still like: “No, this is fake. This is like the moon landing. This must be green screen.”
Justin: The funny thing is—I watched a video today on Twitter of one of those Hercules C-130 planes doing a jet-assisted takeoff, fire jets coming out of the back. I looked at it and my immediate reaction was: “Yeah, that’s AI.” It wasn’t. It was real. My immediate reaction now to every unbelievable video is: yeah, it’s not real—it’s AI.
Frank: Yeah.
Justin: And I’m fine with that, Frank. I know this is something you’ve been afraid of for a long time. The world has not come to an end as a result of our disbelief of videos.
Frank: Well, the silver lining is we never should have been believing everything online anyway.
Justin: That’s true. That was always the joke: “It must be true, it was on the internet.”
Frank: Exactly.
Justin: We used to joke about that.
Why was Figure’s head of product safety fired?
Justin: Can I give an honourable mention before we finish to Robert Gruendel, a man after my own heart. He got fired from Figure Robotics because he was a whistleblower. He stated that the startup’s robots could fracture a human skull—and he thought this was a bad thing. These make the humanoid robots—and it could hit you over the head and fracture your skull. I would agree! Remember you were giving out about robots maybe six months ago, and I was saying there’s a really simple solution: just make sure they’re smaller than we are.
Frank: That’s right. Make sure we can step on them.
Justin: Yes. And Robert Gruendel agrees with me. I’m delighted—but unfortunately he lost his job.
Frank: Unbelievable story. He was head of product safety and he was basically fired for ever putting in writing that there might be safety issues with having robots that could fracture a human skull.
Justin: There you go. He clearly misunderstood the brief. “Head of product safety” didn’t mean the products had to be safe. It meant you had to say the products were safe.
Frank: And what freaks me out is he said there was no documentation, but there were loads of near misses. One specifically in the report: the robot punched at superhuman speeds, narrowly missing an employee, and fracturing a stainless steel fridge door with a quarter-inch puncture.
Justin: Well, remember—whenever you ask a robot a question—just make sure to say thank you, Frank.
Frank: Excellent.
Justin: Pleasure.
Frank: Chat to you next week.
