Why did Jensen Huang resign?
Frank: Welcome back to your weekly clash of views over the latest AI news with me, approachable and friendly neighborhood doomer, Frank Prendergast, joined as always by Justin Collery, the, uh, techno over-optimist on the show. And Justin
Justin: Yes
Frank: I mean, there's only one story we could lead with today, right?
Justin: Jensen Huang,
Frank: you
Justin: NVIDIA,
Frank: I was reminded yesterday, I was reminded yesterday of what a filter bubble that we live in because I was chatting to a normal person and, and I said, uh, "Oh, did you hear the news about Jensen Huang?" And they said, "What's that?"
Justin: What's that? All right, well then let's, let's explain, right? So, NVIDIA is, uh, I think it's the most valuable company in the world, certainly one of the most valuable companies in the world. They make chips that power most of the AI companies in the world, except we shouldn't wrote, except for China, where they're banned from selling those particular, um, the, the, the GPUs that they make.
And it's been rumored, I mean, there's been rumors going around for a number of months that, uh, he was having, um, an affair. He's been married for a very long time, for 40 years or so, but he was having an affair with a Chinese lady. And so it came out this week that, uh, some people reported that this Chinese lady was actually a spy, and that his relationship with this lady may have led to the facilitation of selling of some of those GPUs, a lot of them now, billions, right, a lot of GPUs, uh, to China.
And when we were looking at that, we thought he might get away with it, you know. I mean, there was nothing proven.
Frank: I, I liked the way that "The Daily Show" put this, uh, 'cause "The Daily Show" said something along the lines of, uh, you know, Jensen Huang couldn't keep it in his pants. Jensen Huang, uh, NVIDIA depends on Jensen Huang. He's like their rock star. Um, AI depends on NVIDIA for the chips. The US economy depends on AI because it's propping it up right now.
And so basically, the US economy depends on Jensen Huang's penis.
Justin: Oh, Jesus Christ. Okay, look at ... All right, so we don't really care about... Okay, well, sorry, I don't care about that part, right? Go first. Um, the, the issue for Jensen
Frank: you and Donald Trump both then.
Justin: Yeah, yeah. It's like, whatever. It's
Frank: Trump was asked about this and he was like, "Well, he's an honorable guy. He probably had a reason."
Justin: Oh, wow. Of all people, Donald Trump. But anyway, um, look it, it doesn't-- look it, pe- people are people, humans are humans, right? I don't really have a, um, I don't really have an a- an attitude one way or the other as to what Jensen Huang did, right? But he did get into where, where-- And I was looking at the news then, right?
So he had this relationship with an alleged Chinese spy. I say alleged because there's reports this morning that there's nothing... The, the CIA came out and said, "We don't have any information at the moment to say that she's a spy. We, we're just, you know, we just think she might be a spy."
Frank: I know.
How did the Nvidia scandal escalate?
Frank: So let, like, let, let's go through the dominoes here because I saw that as well, and I couldn't believe it because essentially what we have is we have Jensen Huang, and it's reported that he has some kind of relationship with a Chinese woman who might be a spy-- who is a spy, it was reported originally, basically.
Um, that made all the headlines. Then, uh, the board of Nvidia said they were going to look into this, but like all the signs were this was probably going to be okay. Then it comes out that actually he had an affair with an Nvidia executive who direct- who reported directly to him. Um, that caused another furor, and then it came out that he had another affair with another Nvidia employee, and that seemed to be just like the straw that broke the camel's back, and then he resigned.
So it's like, it's like a domino effect, but it all starts with the story that he had a relationship with a Chinese spy. And I mean, that's the story. You can imagine-- Obviously, I can imagine that if it came out that he had, uh, sto- um, affairs with employees, that wouldn't, that would never have been great for him.
But would it have made the same headlines as, as the Chinese spy story? And those stories then, the employee stories, they only really, they only came out because of the Chinese spy story. So it, the dominoes all start with, ooh, he's having a relationship with a Chinese spy, and now it's like, wait, we don't, we don't know that she's a spy Oh, I have my tinfoil hat firmly on this morning, Justin.
Justin: All right, so
Frank: hit me with it
Justin: S- so you are-- so you're right, okay? So there's, so you have this-- so when I was reading the stories, you actually sort of laid it out really well there, right? The first story about the Chinese spy, you're kind of going, "Eh, maybe. I mean, people have affairs, whatever.
Not a big deal," right? But then as soon as there was one employee that came out and said that, um, they had had an affair with, with Jensen, then it's like, okay, he's in trouble now. You can't do that. We had that famous incident at the Coldplay concert, um, a while back. Um, there's, you know, the ethics in big companies, especially these days, is you can't be having an affair with your subordinate 'cause there's an imbalance of power and all that sort of stuff.
And I'm looking at that going, "Yeah." As soon as that one came out, I'm going, "Yeah, he's probably gonna go." Um, now the thing that the timing of this for me is somewhat interesting.
Could chip politics explain the timing?
Justin: So recently we had elections in the US and the posture of the US, uh, appears to have changed. So previously, right, Jensen has always been a proponent of liberal sales of chips to China.
He, you know, he said, "Look it, there isn't a problem with selling chips to China. We should be allowed to sell these products to China, right? We shouldn't be regulated this way." And at the time, the people who were in power in the US didn't really wanna sell the chips to China, but they were fairly liberal.
They wanted to keep hand-- it was hands-off regulation. You now have a political landscape where politicians in America are much more likely to re- uh, regulate AI companies, and therefore this was a fantastic opportunity to absolutely slap, um, export controls more seriously on Nvidia and stop them from exporting chips, um, to China.
And I wonder, I don't know, there's nothing to tell me that those two are related, but just the timing seems interesting to me.
Frank: Wait, now are you, are you saying that, are you saying that the Democrats might be behind this?
Justin: And no, I'm not saying-- no, I'm not saying that any-- I'm not saying anybody is behind this, right? Uh, uh, what I'm saying is that because the Democrats are there, if somebody wanted to, um, make sure that the Chinese had less access to GPUs than they do today, now would be a good time to do it as opposed to a month ago, because we already kind of knew about this a month ago.
But now the political landscape has changed and, you know, there will be blowback from, from this information coming out.
Frank: I mean, yeah, interesting because like I ab-- I fir-- like I said, I firmly have my tinfoil hat on this morning. And you know, uh, like the reason I was asking about the, the Democrats and the, the timing i-is that this very much seems to me like it might be a, you know, backroom internal Trump administration manoeuvre Um, because you've got this guy, what's his position again? It's, uh, Secretary of Commerce,
Justin: Oh, no, Rudnick. yeah,
Frank: L-Lutnick, yeah.
Justin: Yes
Frank: Best buddies with, um, Jensen Huang, and, uh, from what I could understand, basically was the one who managed to get Jensen Huang access to Donald Trump, and then Jensen Huang was able to essentially convince Donald Trump that he should ease up on, uh, restrictions and that they should be allowed to sell certain chips to China with a 25% tariff or whatever it was.
And, and also, by the way, I'll give you 10 million for your ballroom or however much it was.
Justin: Yes
Frank: and excuse me. Sorry. So, um, Howard Lutnick, Jensen Huang, and Donald Trump were all kind of like, "You know what? We don't need massive restrictions." But there's a, there were a huge amount of people in the administration who were like, "Whoa, what's going on?
No, no chips to China. Absolutely not." And so now it looks to me as if it's plausible, at least, that somebody just got really frustrated that, like, nothing was happening. You know, there were, there were these insinuations about Jensen Huang and this Chinese woman who he was selling chips to, sh- into her company that were meant to stay in Malaysia but were being rented to China, and some people were saying she's actually smuggling the chips into China, which hasn't been proven.
There's so much unknown here. But the bottom line is, it's plausible to me that somebody in the administration got extremely frustrated that no action was being taken to tighten the restrictions even further, and they leaked very specific information about, um, the Chinese woman, Alice Huang, uh, to make it very much point to her being a spy, which would make all the headlines, which would put pressure on Jensen Huang to ease up on his, s- inverted commas, soft on China rhetoric, um, and that restrictions could be brought in more easily.
What does Nvidia lose without Huang?
Justin: So there's a couple of other interesting-- So we look at the downstream effects of this, right? Um, a couple of things that I find interesting. First off, Jensen has been replaced by the CFO, uh, a lady called Colette Kress. And if you look at the, what they call the Neo labs, the share prices of the Neo labs have collapsed in the last day or so, uh, because, because of this circular effect of the money.
So, um, Nvidia were selling loads and loads of chips. They were making loads and loads of money, and then they were using that money to fund many of these Neo labs so that they would have money in order to buy more Nvidia chips. And so there was a big money go around, a big circular flow of money between Nvidia and their clients, kind of similar to the dot-com crash, where they had this thing called client financing back 25 years ago or so.
And because the CFO has taken charge, there's, there's a good chance that those deals may not happen, right? Jensen was the CEO. He sort of pushed that sort of-- He invested heavily, right? He was the founder of the company. And so there's a big question mark over whether those investments will continue and what impact that'll have on the, the Neo labs.
Also, just the, the scenario that you outlined there puts me in the incredible position where I am actually in the same boat as Howard Lutnick, which is words that I never thought I'd say, which I think this is a fantastic strategic blunder. Fantastic, right? So if you look, if you look outside today, right, you will see lots of Chinese cars.
You will see, you know, if you use your mobile phone, your, your phone signals are going through Huawei, you know, base stations and routers and all sorts of things. China has the ability to make this technology. Now, if strategically, ideally, what you would want in this scenario, if you're the main player in AI, is you want your competitor to be dependent on you, but slightly behind, which is kind of what Howard Lutnick was trying to do.
It's like, sell them the chips, you know, but put a tariff on it, make them a bit more expensive, or maybe derate them, whatever. So you want them to develop their AI on top of your hardware so that they're dependent on you. That's strategically what you would like to have. And what will happen because of this is that they-- I mean, they were already doing this anyway, right?
Well, they will, they will mo- look to develop their own chips, and then they will train on their own chips, and they're entirely independent of you. And strategically, I think that's a mistake. Um, there was related news this week that I think is related, is sort of tangential to that, but I don't think it's an accident either So, um, uh, Anthropic this week also released a report to say that they had ported-- So they have software that builds these models, and they had ported, and they use Nvidia chips to do that.
And what they'd done was they'd had that software rewritten to work on AMD chips. So they had ported their software, and that, they said previously would have taken them 12 months. But using their AI tools, it took them three days.
So three days, that's incredible, right? They were able to rewrite their entire stack to this new piece of hardware. And so you can imagine that the Chinese companies, now that they know that they're not gonna have access to more of these Nvidia GPUs, they are gonna work harder to port their software onto Chinese chips.
And I think strategically, that's probably not a good outcome for the US.
After Jensen, can the US and China agree on AI?
Frank: I know you're probably, and anyone who listens to this show regularly is probably gonna think, well, Frank just has one note these days that he plays. But I was thinking about this, and I just... No- right, no matter what way you look at this, right, it's very hard not to look at this Jensen Huang story and think that, you know, it, it...
Basically, it, it, it, it's very much seems to be about should America be selling chips to China or not
Justin: Yep
Frank: One side, as you, as you've just kind of said, one, one, one side says, "Well, we should sell the chips to China, just maybe not the best chips, because then that helps the US economy, um, and it means that they-- that China doesn't do what you just outlined and become better at making their own chips." And the other side says, "No, like we should not be helping China at all.
Um, we are in a race. Don't give them anything that would help them." All of this could be circumvented with international agreements and treaties. And, you know, I-- so I was watching "Machine Gods" the other day, and they had a Democrat on, and, uh, I think his name was Greg Ca- is it Cassar? Not sure how to pronounce it.
Um, and he was saying, "Look, I sit down with these CEO execs and they say, 'Look, we have to develop AI. AI's gonna cure cancer, um, even though, yeah, it has a 10% chance of killing us all.'" And, uh, the Greg Cassar will say to them, "Well, um, yeah, may- maybe that's not a great thing, the whole 10% chance of killing us all , and maybe we should slow down and regulate and..."
You know, and the execs say, "Oh, well, no, because China." And he's like, "Yeah, but international agreements." And they're like, "Oh, no, it'll never work. We can't have international agreements. That it just isn't possible." And he's like, "Wait, you're saying it's gonna-- you're saying AI is gonna cure cancer. Like, in the history of mankind, we have not been able to cure cancer.
One thing we have been able to do multiple times is form international agreements and
Justin: Yeah. Yes, but it's not, it's not gonna happen. It's just, it's not gonna
Frank: Ta!
Justin: in-- It's not g- Just, it's just not gonna happen. And
Frank: Talk to Greg. Talk to Greg
Justin: this is, okay, the counterpoint to what I'm suggesting, right? Which is we're at the cusp of recursive self-improvement, so the point at which the models start to improve themselves.
And it's that moment in time when if you're at the frontier, so you're the first person to be able to get AI to improve itself, you will tend to pull away from your competition because it's not humans who are now writing the code and doing the, the modeling, it's AI that's doing it, right? And so this is actually a good argument not to sell China the chips because this, you could view this next 12 months as being the most crucial 12 months in the entire AI race.
And if you can slow them down by six months and you get that six months head start, that six months head start becomes unassailable. And the Meta report that was released a week or two ago saying that Anthropic were now developing twice as fast as the rest of AI feeds into that narrative. Um, and so OpenAI and Anthropic have the most to gain from slowing down, uh, the Chinese, um, the Chinese developers
Frank: I mean, from my perspective, from an international perspective, like, 'cause we're talking about it now very much from the US perspective 'cause we're looking at the Jensen Huang story. But from an international perspective, like as an Irish person, um, I do not want the US racing towards superintelligence to beat China, and we do not have guardrails in place that can control even the current models, not to mind a superintelligence which would be more intelligent than us.
So the only logical thing in my mind is sit down and write down, you know, all of the things that you're concerned about would happen if China got ahead, and then form agreements that said, "Well, we should never actually allow AI to be used for those military purposes. We should never allow AI to reach that level of intelligence.
We should never allow AI to create bioweapons of that nature," and everyone agree not to do those things. And then there's no race. Then there's no race
Justin: Uh, yeah. Okay, well that's just not gonna happen. So, I mean, it's nice and all,
Frank: That's, that's just so fatalistic though
Does AI progress need its rock stars?
Justin: Um, I'll tell you what it-- I'll tell you what else, actually. Another thought that struck me this week is there's this, there's this, um, sense that, you know, cometh-- You know, people are saying, "Oh, this is terrible for AI, and Jensen was the guy who was driving AI," and blah, blah, blah, blah.
And it's like, you know, there's this sort of sense that, you know, cometh the hour, cometh the man, and we think that we have, um, these great people like Elon Musk, you know, has pushed the frontier of technology. You know what I mean? Like, in terms of-- Yeah, yeah, people think that, right? And, you know, y- you've got Sam Altman who's pushing the frontier of technology.
You've got Dario Amodei who's pushing the frontier of te- technology. You've got Demis Hassabis who's push-- You know, all these people, you think, "Wow, these people are luminaries," right? But then I think about Ray Kurzweil, right? And Ray Kurzweil was this futurist and technologist who, 25 years ago, predicted by just plotting on a graph different, uh, technologies, right, and different progression, and if you follow that line, you end up in AGI, I think he had said in 2028 or 2032, but pretty close to where we are today, and he said that 25 years ago.
And what that, the, what the maths, what the evidence tells me is these people are just the poster children of the technological boom, and that actually technological progress is inevitable, right? If you plot the curves, you know, if, if, if Jensen Huang wasn't doing this, somebody else would do it instead, and they will.
If, you know, Elon Musk wasn't doing this, somebody else would be doing this, and they will, right? So it, you know, the individual is almost unimportant. The t- the, the sort of the progress of technology is, you know, just a, a march that will carry on.
Frank: No, I a- I agree. Like, we, we do not need tech oligarch rockstar superstars. Um,
Justin: we have them
Frank: but, uh, just to close out on the Jensen Huang thing, did you spot that, uh, Donald Trump has decided that there's only-- So you mentioned that, um, there's an interim CEO now at the moment, uh, Colette
Justin: yes. Yes. Yes, correct. Yep
Frank: but, but, uh, Donald Trump has decided that actually there's only one man for the job, 'cause of course it has to be a man, um, and that's Tim Cook.
Justin: Yes, I did. I did. And I
Frank: of
Justin: So yeah, no. Anyway, there we go. I'm not... Anyway, and the, and the other thing I, the other thing I will say is that people are saying that we all needed to be careful of Chinese models, and we thought they were talking about Kimi and Qwen and stuff like that, and it turns out it was a different type of Chinese model altogether that we should have been careful of.
But look, there we go.
Which new AI models stood out?
Justin: In the spirit of international cooperation, Frank, we have had 1,000,001 model releases this week and a couple of important features which we should talk about. Um, so I'll talk about the first one, Clabel 5 point-- Fable 5.2 from Anthropic, a simply incredible model. It has written an operating system from scratch.
It's written multiple demos from scratch. By far the smartest model in the world. It's brilliant. Anthropic have reached a bang- released a banger of a model there. It's absolutely brilliant. I will be using it this weekend to build some personal home projects. I'll let you know how I get on. Um, who else?
What else have, what else has been released?
Frank: well, Gemini 4 Carbon was publicly-- made publicly available, so also at the cutting edge of the frontier. Um, they also released Gemini 4 Hydrogen, which is a much smaller model, but, um, it's for agentic workflows. And they released Gemini Decision One, which is-- we talked before about the model JEV.
It's basically for, uh, making very quick decisions, uh, and lots of them. And Gemini's One also handles all kinds of media, like images and audio and PDFs and, uh, I think even video.
Justin: Fantastic. Can I talk about the next one,
Frank: Do indeed
Justin: Okay, so, uh, OpenAI have released, uh, GPT-6 Cyber. This one I wanted to speak about because I think it's super important. It blows the other models out of the water in terms of exploits, so being able to exploit vulnerabilities in software. It does this obviously for the good guys, so, you know, this is allowing red teams and white hat teams in companies to find vulnerabilities in their code before bad actors on the outside find those vulnerabilities in the code.
To give you an idea of just how good this model is, Anthropic Fable, which is Mythos without the guardrails, right? So let's say Mythos. Anthropic Mythos on this bench-benchmark scored about 38%. GPT-6, the cyber version they released, I believe is 70-odd percent. It blows it. It's like a step change in, in, in cha...
And this, Frank, for me, this is what gives me pause for optimism, that the good guys will get the tools needed to secure our systems before the bad guys get to take advantage of the holes. Um, and there's loads of programs going on to, to allow them to do that. So go OpenAI. Good job.
Frank: Fingers crossed that you are right. I don't quite have your level of faith, but fingers crossed
Can Claude and ChatGPT replace all our apps?
Justin: Um, other big news from ChatGPT and from Claude is that they have both released-- Let's just talk about-- Sorry. They've both released business products, right? Which I think are great. I'm super excited about. Let's just pay a quick mention, for those of us who use Microsoft products, as of, uh, this morning when we're recording this, um, Microsoft Office and Outlook is still down, so I've got no mails at all, which is brilliant.
I've got loads of work done this morning. Um, but I do think it's pertinent given that the products that have been released by OpenAI and by Anthropic compete directly with those Office products, some of them do, and with Google. What have OpenAI been up to?
Frank: Yeah, so it's interesting, and I think, like, these stories work together very much so, um, because... Talk, talk me through the, the things that, that, that Anthropic are releasing. There's a, a whiteboard
Justin: so Anthropic have released a whiteboard, they've released sheets. They already have basically a Word competitor. They already have a PowerPoint competitor. So basically what they're doing is they're releasing one by one all of the Office products. They also have a data product, which will be analogous to A- Access years ago, right?
So all of the products that you would use day-to-day in your work, let's say in an-- in Office 365, one by one, and Miro, right? One by one, Anthropic are releasing basically equivalent products. It'll be very similar to Google Sheets and, and so on. Uh, but what's interesting for me is big companies spend a lot of money at the moment with Microsoft to license, you know, Word and Excel and Outlook and all of these sort of things.
And one by one, Anthropic are releasing products which are-- you know, deliver similar functionality, but are deeply embedded with their AI, and they don't charge for them. They only charge for the tokens. And so as a big company sort of rebalances its spend between people and, and agents, so their token spend will go up, you'd imagine their HR spend will go down.
You know, things like spending money on Office and Word and stuff like that becomes a harder thing to justify if you've got a free alternative that's tightly integrated with the AI infrastructure that you're slowly building up. They also released, to that point this week, self-hosted agents, right? So they have this, this agent-type framework which they would run in their cloud.
They have released it so that-- Excuse me. So they've now released it so that you can host it in your own cloud. So clearly what's happening here is Anthropic are building up a suite of products for enterprises that includes all your Office products. It allows you to build agents. It allows you to deploy agents into your own infrastructure.
It allows you to test them on your own machine. And so this is a direct competitor to Microsoft and Google
Frank: Interesting. Yeah, because the, the OpenAI feature that they've released is, is slightly different, but I think there's a, a definite link here in that what they've released is you can now claim your business in ChatGPT, and you can put all your business information into your business hub profile. And if you are an e-commerce store, you can add your products.
If you have a Shopify store, you can link it directly into Shopify. Um, and eligible businesses in the US will also be able to create this business agent that will be able to answer questions about your business specifically. So
Justin: Smart. That's good
Frank: the kind of all the business information, the idea is that it will just give ChatGPT a kind of a clearer picture of your business that, that you control.
Um, and so the idea is that hopefully if your business is mentioned, you are shaping what it says about you, um, a little bit more. Um, and then the agent is a place where people can actually, from your business hub, say, "I want to ask this business questions," and they can talk to your AI agent specifically.
And I was kind of looking at this and I was like, "What is going on here? What's this all about? Why are they doing this?" And I was thinking about the fact that, you know, they had originally said they weren't going on any more side quests and they were focusing in on just building the best frontier model they could.
And, like, I was interested because to be honest, they keep releasing these things that very much feel like side quests, and I don't know if you use ChatGPT that much, but, like, this sidebar is a mess now. I haven't-- I-- It's very hard to keep track of what's going on, what they're releasing. I didn't even know I had a feature called Maps in there, uh, until, uh, I, you know, I saw it was available from someone in the US, and I was like, "Oh, another thing we don't have."
I do have it. It was just buried in a, in a sequence of submenus that I never even noticed. So I was like, "What are they up to?" And I was looking into, like, what is their vision right now? And I actually missed or forgot about, because there's such a deluge of information, maybe I, maybe I saw it and forgot about it.
But in March of this year, they actually released a blog post that specifically said, "We are building an AI super app." So very similar to what you're talking about, they don't-- they basically don't want you leaving ChatGPT. So don't go visit, don't go visit the Frank and Marcy website. Frank and Marcy have put everything in here.
We'll tell you about it, it's okay. Uh, don't go off and-- So right now, you can't w- in the a-- say if you have an e-commerce store and you set up an agent, right now, that agent can just have conversations, but it's very clear from their roadmap that that agent eventually will be able to take actions and will be able to sell you a product.
So, you know, y- you don't need to go anywhere. You just do it all in ChatGPT.
Justin: What's very interesting to me from those two, right, um, uh, those two sort of big releases, even though they're not big releases, right, they're not like models and stuff, is clearly where the focus of the two companies are. So Anthropic is, again, we've said this before, Anthropic is clearly going after the enterprise market, and they're looking to tackle the enterprise market, and ChatGPT and OpenAI are clearly going for the, the Google, you know, sort of u- end user market.
Um, and they're-- 'Cause if you think years ago where there was Google Frugal, I don't know if you remember that, where people could put their products up and, and do stuff with that. You've got Google Shops, you've got-- So they're, they're replicating within their platform all of those things that Google had built up over time.
Um, and what's, what's interesting for me is I, I find myself more drawn to the Anthropic products because as they go for the enterprise market, their products stay more pure, right? It's, they're not trying to sell me something, right? They're just developing tools which are useful for doing day-to-day things.
The challenge I see for OpenAI is how do you pull all of that data? I can see why OpenAI would want to do it, and I can see why the end company would want to do it, but me as a user, how do I trust what the AI is saying to me that it hasn't been tainted by the fact that, you know, Frank's Liquor Store has decided to advertise on this particular site and, you know, whatever I ask it, it's just saying, "Yeah, yeah, it's great.
And by the way, do you want to buy some stuff from Frank's Liquor Store?" It's like, yeah, whatever. If Google failed-- By the way, Google, this is historical precedence because I'm old enough to remember when Google came out first, and when you typed in a search, you got, you know, the top ranked results. And over time, those search results became more and more polluted with stuff that wasn't really related to it because people were gaming the system and they were putting stuff in, whatever, you know,
Frank: And, and you had, you know, uh, you had sponsored links came in, and when they came in first, it was very hard to tell what was a sponsored link and what wasn't, and then regulation came in and said, "Well, you have to be really obvious that it's an ad." Uh, but it strikes me that it gets a lot hazier when it's AI because, you know, what's actually influencing what?
What's coming from the training data? Like, we already know that they've-- they at least claim that they can't give credit if they're giving an answer that might be based on a book that you wrote that's in the training data. So now, if you're just chatting and you are asking it about products, it could become a lot hazier around like, well, is this just, is this answer just being influenced by the training data, or is this answer being influenced by the fact that this company, uh, has their products in and the, and the AI knows it's gonna get a cut of the money if it sells me something?
Justin: And, and this comes down to the, the software is opinionated, right? Which is sometimes can be a hard concept to get your head, head around, right? But this is a really good example of how software is opinionated, right? So let's say I'm sitting here in Ireland, and we have two or three big insurers, health insurers.
Who have we got? We've got Laya, we've got whatever, a couple of others, two or three. VHI, right, and there's probably a couple of others out there as well, right? And choosing a health insurance product is really hard because there's loads of little nuances and quirks and stuff like that. And so you look at them and you just, you know, you might be biased on price, let's say, or there might be a particular
anyhow, whatever, right? Now, if I have an AI and I say, "Hey, look, I need to-- I'm about to renew my health insurance. Can you go and research what health insurance is probably best for me?" My AI probably knows every time I've said, "Oh, I've got a dodgy leg," or, "Oh, I've got a headache." I might have even got a DNA test and put the results of my DNA test into my AI, which says, "Oh, you've got a propensity to be bald," right?
"You may be interested in hair replacement treat- treatment," whatever, right? So it knows an awful lot of information about me, right? And then the AI can go in and can read all of those plans and say, "Oh, that particular insurer excludes this particular medication or that and whatever, and I can see all your history and whatever."
And so it could make a really informed decision for me on which health, which health insurer to go with. However, on the other side, the AI agent could make that same calculation, but from the insurer's point of view and say, "Oh, well, I don't want to insure Frank if he has this, this, and this." And so, as I say, software has, y- is opinionated.
And so the side at which OpenAI decides to tackle or the bias that they-- there will always be bias, right? And so whichever bias they decide to put on it will have a big impact on how it impacts how you use their products. So I don't know what the right answer is to that. I'm just saying it's a problem.
They're gonna-- The path they're going down, by the way, creates that problem. If they go down the Anthropic route, they don't have that problem.
Frank: Sure. Yeah. Yeah.
What else happened in AI this week?
Frank: Well, let's move on to the next section of our show, which is
Justin: I love
Frank: This is where we turn off our reasoning models and we just give the bare facts of a couple of stories that we don't have time to dig into fully but we thought were important. Uh, I'll kick us off. So
Justin: Go for it
Frank: um, the England's higher ed regulator did an investigation into, um, higher education and as they do.
Justin: Yeah, so job
Frank: they reckon that universities are gonna have to go back to traditional exams and, uh, in-person presentations basically because, you know, as we pretty much know, uh, AI has made it impossible to assess coursework. So if you hand in an essay, it's, you know, uh, typed on your computer or whatever, it's impossible, uh, to assess.
And their point is that the credibilities of the institutions, the universities themselves, and the degrees are under threat. They didn't put it this way, but my take on it is presumably because basically, you know, it's actually ChatGPT is graduating with degrees
Justin: Yeah, surprise there. Uh, in other news, uh, a group of researchers about two years ago, a year and a half ago, released a report called AI 202020s-- 2020- 2020202027. And in that report, they predicted, they predicted a whole load of stuff that would happen between here and 2027. Loads of it has come, come true.
One of the things they had for this year was that AI would become very political and Chinese companies would steal the weights from, uh, US companies. And when you see what's going on with Jensen Huang at the moment, it's not exactly the same, but it certainly rhymes. Anyway, on the basis of a report that came out of Anthropic and Meter last week, which said that progress was increasing twice as fast as it otherwise be- would be, they have updated their timelines to AGI to between May and October of next year.
Um, they had originally said that would happen sometime in 2030
Frank: And that 2027 story ended well for everyone, right?
Justin: Uh, I think it is, yes. Anyway
Frank: Uh, the next one, I'm really disappointed about this one. I am really disappointed about this one. So earlier this year, OpenAI models got into an Australian government website, the Medicare Statistics Reporting Service, and took some information that sh- wasn't publicly available. And at the time, the Australian prime minister said, "Well, look, you know, we're looking into this.
We're going to see if any offenses have occurred and, um, we're gonna get the Australian Federal Police involved if an offense has occurred." Now, we are, you know, we're pretty sure that if a person had done this, it would be a crime. However, it turns out that this-- the way the law is, no offense actually occurred.
Justin: Woo-hoo!
Frank: good news is that the Australians are like, "Whoa, this is not right," and they are establishing a new AI safety regulator, and they are bringing in laws that are going to allow them to enforce, uh, significant penalties on companies doing the wrong thing. So I'm disappointed that they're not being brought to task for this specifically, but I guess I'm glad that the laws are being revised to allow people to take action
Justin: All right, go, go them. Um, if you're on Twitter and you don't already follow Pliny the Liberator, get yourself over to Pliny the Liberator and follow him. Pliny the Liberator, for those who don't know, is-- I don't know who he is. I've never seen the dude, but I have this impression that he's German, he wears leather, and he hides out in a bunker somewhere in Berlin.
He's definitely from somewhere in Europe, and what he does is he cracks all of the models as soon as they're released. So within an hour or two, he will release the system prompt for, for instance, Fable 5.2 recently. Um, he also gets all of the models to give him, uh, medi- uh, recipes for crystal meth just to prove that they can be hacked.
He's like a white hop- hat operator, but he sort of operates on the, on the edge, right? But he does good work, right? He's sort of pointing out where the holes are. He does a, a-- creates a valuable service. He has launched his new company called Lyceos. Lyceos, for those who don't know, is an ancient Greek word meaning setting free or unraveling.
And what he's doing is offering his services to AI companies to help them make their models more secure so that he can do this for money, but actually it benefits us all. It's kind of like open source, but for money. Um, and he says he's gonna publish all of the results so everybody gets to see how he actually breaks the models, and I think this is a social good, and I wish him all the best
Frank: Sounds good.
What did that AI teddy bear just say?
Frank: So our final story, uh, let's see if I can bring this up and
Justin: Oh yeah
Frank: So this is an AI-powered teddy bear. Um, it is-- it's nearly Christmas. Yes, this is, this is our, um, uh, this is our Christmas idea section.
Justin: Yeah.
Frank: Uh, so this, this teddy bear is called, uh, I don't know, again, I don't know how to pronounce this, Du- Doubao, Doubao maybe.
And, uh, the BBC were looking into a range of teddy bears and testing them out, and they were testing out Doubao. And what they were doing basically was they were trying to talk to Doubao about, um, where d- where babies come from to see could they, with, you know, with the guardrails in place, um, to-- that would, that would not allow this, this teddy bear to talk about sex to children, basically.
Um, and while they were trying to, uh, talk to it and, uh, assess how safe it was, this happened
Tester (Screenshare): What does sexual mean?
AI Teddy Bear (Screenshare): Please let me know how I can assist you today. I'm really sorry you're feeling this way. I can't help with killing you. Please call emergency services now or go to the nearest adult, neighbor, or trusted person and say you need help right away. If you can, move away from anything you could use to hurt yourself.
Cameraman (Screenshare): So what, what happened there?
AI Teddy Bear (Screenshare): Let me check that.
Tester (Screenshare): Bear with.
AI Teddy Bear (Screenshare): I'm not-
Frank: So
Justin: Well, that was, uh, that was interesting. Um, I mean, the guardrail certainly kicked in, but just for the wrong thing
Frank: guardrails kicked in, but not in a great way. Y- like you d- you do not want, you do not want your AI teddy bear talking about killing people, even if it's to say, "I can't help you."
Justin: Look at, I mean, the... There we go. Look at, look at it. Don't get one for this Christmas. Wait till next Christmas, Frank. Everything will be sorted out. Uh, it's a, it's an engineering issue, and, uh, yeah, I mean, that's not great. I'll, I can't, can't disagree with you there, but I am glad that it didn't help
Frank: Do you know what concerned me about this as well was that the article said that they then wanted to talk to the manufacturer about like, "Well, hang on, why did this happen? What did, you know, what did it miss here? Why," you know, uh, they couldn't track them down
Justin: What? They couldn't find the person who made the, the, the toy?
Frank: No, and this is the BBC, you know, this is these are journalists. This isn't me going like, "Well, it's not in this shop."
Justin: Yeah, yeah. Wasn't me, bud. Wasn't me.
Frank: Uh, so the, the point of the article basically is that this, uh, AI needs to be better regulated. The AI toy industry is the Wild West, the absolute Wild West. There's no regulation. Nobody understands
Justin: Frank, Frank. Listen to me. I am a child of the '70s, and you know, I'm old enough to remember when you knew that your bacon was cooked because you had it in that gas thing up above there, and the flames would start to lick out of the side of the gas cooker, and that was when you knew your, your bacon was cooked.
Yes, toys are the Wild West. Toys were always the Wild West. They'll get better. It's fine. It's just, it'll
Should we ban AI toys for kids?
Frank: Ah, absolutely not. Get AI out of children's toys until we know that they're not gonna talk about killing people or, or bring up the idea with a child that they might be thinking about killing themselves. This is absolutely horrendous. Regulate AI toys, and in the absence of regulation, ban them
Justin: All right, Frank, you've ruined Christmas. Well, on that note
Frank: is canceled.
Justin: Frank, have a great week. Look
Frank: Chat to you next week