[SPEAKER_07]: The folks at Unitory just introduced the A-to Stellar Hunter.

[SPEAKER_07]: And they've decided to market it as if it's an A-T's action hero.

[SPEAKER_05]: Oh my goodness, it's just crashed through a pane of glass.

[SPEAKER_05]: Yeah, this looks like an episode of the A-Team really doesn't it?

[SPEAKER_06]: If the A-Team had four legs and no head.

[SPEAKER_05]: Yes, he wouldn't have one of them called face or suppose if the head is done.

[SPEAKER_04]: Hello and welcome to episode sixty three of the AI fix your weekly dive headfirst and for bizarre and sometimes mind boggling world of artificial intelligence.

[SPEAKER_05]: My name's Grant clearly, and I'm March Stockley.

[SPEAKER_05]: Mark, what are you talking about on today's episode?

[SPEAKER_05]: I will be explaining why Sam Altman tweeted a picture of the Death Star.

[SPEAKER_05]: And I'll be discussing the talking dead.

[SPEAKER_05]: But first, the news.

[SPEAKER_05]: Robot Dog runs through window.

[SPEAKER_05]: Politicians love AI so much, they're using it to do their jobs.

[SPEAKER_05]: Open AI introduces GPT OSS.

[SPEAKER_05]: Using chat GPT for therapy, your sessions may now be available for anyone in Google.

[SPEAKER_05]: Oh dear.

[SPEAKER_05]: Open AI has admitted it's made a little bit of a goof.

[SPEAKER_06]: Oh really?

[SPEAKER_05]: Yeah, it has removed a controversial feature in chat GPT that's caused some users to unintentionally allow their private and highly personal chats to appear in Google search results.

[SPEAKER_05]: I don't know how private and highly personal your chats are with chatchiepeed team, huh?

[SPEAKER_07]: Well, private enough that I wouldn't want them appearing in Google search results if somebody search for Mark's stuff, although if they do that normally, they get the guy from BlackRock and the looting town football player before they get to me.

[SPEAKER_05]: Well according to fast company nearly four thousand five hundred conversations were coming up on the Google search results although they don't necessarily include your personal identifying information.

[SPEAKER_05]: So won't necessarily say this is what Mark Stockley wrote and this of course you've included in your conversation with chat GPT something which might identify you.

[SPEAKER_05]: So if you for instance went to chat GPT and said, I am the co-host of a highly successful AI podcast.

[SPEAKER_05]: And I have this other co-host who really isn't pulling their weight.

[SPEAKER_05]: If each said something like that, people might have been able to identify, which one of us?

[SPEAKER_05]: Incidentally, some of these conversations were once where people revealed really personal stuff, including their struggles with addiction, their experiences of sexual abuse or serious mental health issues, sometimes even fears that AI could be spying on them.

[SPEAKER_05]: As if AI could be sharing, [SPEAKER_05]: The highly personal information with the world's largest advertising company Google.

[SPEAKER_05]: As if!

[SPEAKER_05]: So what's going on here is there is a feature within chatGPT where you can share a chat with other people.

[SPEAKER_05]: Oh, yes.

[SPEAKER_05]: You've done it with me in the past.

[SPEAKER_05]: We've said, hey, look at this result or whatever.

[SPEAKER_05]: Now, when you did that, open AI would pop up a little window, which would say a public link to your chat has been created.

[SPEAKER_05]: And there's a little tick box saying, make this chat discoverable.

[SPEAKER_05]: And then in tiny, tiny letters underneath, [SPEAKER_05]: in a slightly slightly less opaque font.

[SPEAKER_05]: It says, allows it to be shown in web searches.

[SPEAKER_05]: And people were ticking this box and make this chat discoverable.

[SPEAKER_07]: But what is the use case?

[SPEAKER_07]: Well, it just like, hope this randomly gets picked up by search engine.

[SPEAKER_05]: I think what they were thinking was, yeah, this would be really good for them, according to the internet.

[SPEAKER_07]: I tell you, yeah, that is why they were thinking, you're absolutely right, yeah.

[SPEAKER_05]: I would have thinking, it would be good for Open AI, it would be good for Google.

[SPEAKER_05]: According to Open AI CISO, he called the feature a short lived experiment that Open AI launched to help people discover useful conversations.

[SPEAKER_05]: Now, the full quote, I think, was to help people discover useful conversations that helped people with their deeply personal privacy issues.

[SPEAKER_05]: So he actually has confirmed now that this difficulty has been removed and they are contacting what they called the relevant search engine to get that content off there.

[SPEAKER_05]: But initially, they tried to defend it.

[SPEAKER_05]: They said, well, that our user interface is sufficiently clear.

[SPEAKER_07]: I'm a news that they've gone back to Google to get these things for them.

[SPEAKER_07]: Like they didn't know that that's what's going to happen.

[SPEAKER_07]: No, there was, well, yes, really what it was designed to do.

[SPEAKER_05]: Yeah, it is quite unbelievable.

[SPEAKER_05]: Because if you think of it, just how much sensitive information is being plugged into these things, [SPEAKER_07]: Yeah, if only they had access to some sort of thinking machine that could help them work through the consequences of making these kinds of decisions.

[SPEAKER_05]: So just to stress, there is still a way to share your private conversations on chatGPT with friends, with colleagues, because I think there is a use case for that.

[SPEAKER_05]: But right now, there's no tick box to automatically share it with the world's search engines too.

[SPEAKER_07]: So Graham, if you were going to market a terrifying, headless robot dog is some kind of harmless and useful companion.

[SPEAKER_05]: What would you call it?

[SPEAKER_05]: Oh, so it's a harmless one.

[SPEAKER_05]: So you don't want to call this something like Tyson.

[SPEAKER_07]: Or something or a therminator like remember the therminator is the robot dog with a flamethrower on his back.

[SPEAKER_07]: I mean, that's manifestly not a harmless companion.

[SPEAKER_07]: But if you had a harmless companion or something that's meant to be useful, what do you think you would call it?

[SPEAKER_05]: Oh, you'd call it something really sort of fluffy and cuddly and gorgeous like that.

[SPEAKER_05]: Wouldn't you sort of fluff bottom?

[SPEAKER_05]: So not hunter then, you wouldn't call it hunter.

[SPEAKER_05]: No, no hunter sounds no mark.

[SPEAKER_07]: Obviously you wouldn't call it hunter.

[SPEAKER_07]: And would you start your promo video with it running through a pane of glass?

[SPEAKER_07]: That sounds like a no.

[SPEAKER_07]: Well, you don't agree with the folks at Unitree or the folks at Unitree don't agree with you.

[SPEAKER_07]: They went a different way.

[SPEAKER_07]: They've just introduced the A to Stellar Hunter.

[SPEAKER_07]: And they've decided to market it as if it's an AT's action hero.

[SPEAKER_05]: In a glass.

[SPEAKER_05]: Yeah, this looks like an episode of the A team really doesn't it?

[SPEAKER_06]: If the A team had four legs and no head.

[SPEAKER_05]: Yes, he wouldn't have one of them called face, I suppose if the head was done.

[SPEAKER_05]: Oh, those balls over and it, oh, there we are, back on its legs again.

[SPEAKER_05]: It's basically indestructible, isn't it?

[SPEAKER_05]: That's what they're saying.

[SPEAKER_05]: It's unstoppable and the only thing it's lacking is an assault rifle.

[SPEAKER_07]: Yeah, and it's funny you should mention that because actually watching the latter half of the video where it's running down the quarry reminded me of another video.

[SPEAKER_07]: So it's an episode of Black Metal, oh my word.

[SPEAKER_07]: That's the Black Mirror episode Metal Head from twenty seventeen.

[SPEAKER_07]: It's a robot dog running along a terrifying speed jumping into the background of the van falling over and then dusting itself off again before shooting the driver in the head.

[SPEAKER_07]: Basically the same video isn't it?

[SPEAKER_07]: It is.

[SPEAKER_07]: That's what they've modeled it on.

[SPEAKER_05]: So picture this mark.

[SPEAKER_05]: It's three o'clock in the morning.

[SPEAKER_05]: You are wide awake thinking, oh my goodness.

[SPEAKER_05]: How am I going to be as funny as Graham next week on the show?

[SPEAKER_05]: You're thinking to yourself and you've got other questions on your mind as well.

[SPEAKER_05]: You're asking yourself, prawn cocktail crisps.

[SPEAKER_05]: Do you love prawn cocktail crisps?

[SPEAKER_07]: I don't know, I'm not a fan.

[SPEAKER_05]: Oh, okay.

[SPEAKER_05]: So maybe you're thinking, you know what really needs to happen.

[SPEAKER_05]: Then needs to be a nationwide ban on prawn cocktail.

[SPEAKER_07]: I wouldn't be against that to be honest.

[SPEAKER_05]: Maybe you would contact your local MP, but of course he's fast asleep as he's tucked up in bed.

[SPEAKER_05]: He's got his pajamas on.

[SPEAKER_05]: He's not going to be answering your messages, but that is not a problem if you are living in the constituency of lead Southwest and Morley in the UK because your local MP has created an AI doppelganger of himself.

[SPEAKER_02]: I'm AI Mark Suez, Labour MP for lead Southwest and Morley.

[SPEAKER_02]: How can I help you today?

[SPEAKER_05]: Now, when you check this thing out, it's a little bit Pixar.

[SPEAKER_05]: It's how Mark C.

Woods hopes he looks.

[SPEAKER_06]: He really is, isn't it?

[SPEAKER_06]: He basically went to whatever AI he used to create this and said, could you make me look much more like David Tennant, please?

[SPEAKER_05]: And C.

Woods says that AI Mark is a prototype that offers his constituents an additional way to engage with their MP [SPEAKER_02]: can assure you that my aim is to enhance communication, not replace it.

[SPEAKER_02]: This AI platform allows for quick access to information and assistance.

[SPEAKER_07]: How is he marketing this?

[SPEAKER_07]: Would you like to engage with your MP, but with a better hairline, different shaped eyebrows, much larger eyes, a completely different shaped chin, and different lips?

[SPEAKER_07]: That's it.

[SPEAKER_05]: And also, of course, but it is using his voice, for at least it's using an AI deep fake of his voice.

[SPEAKER_05]: Yeah.

[SPEAKER_05]: Now, meanwhile, over in Sweden, Prime Minister of Christensen, [SPEAKER_05]: He's got himself into a spot of hot water because he, he told the truth.

[SPEAKER_05]: He always a big mistake for a politician.

[SPEAKER_05]: Someone asked him, do you use AI tools at all?

[SPEAKER_05]: He said, yes, yes, I do.

[SPEAKER_05]: I frequently say AI tools.

[SPEAKER_05]: He says, in fact, if I need a second opinion on policy decisions, I don't look to my cabinet or the experts around me.

[SPEAKER_05]: Instead, I go to chat GPT and French AI, La Chat, which.

[SPEAKER_05]: Now, my, my, my, my Scottish AI.

[SPEAKER_05]: Nor is it an homage to William Chapman.

[SPEAKER_05]: I did owe level French.

[SPEAKER_05]: I think La Chameyne's The Cat doesn't mean the chat box does it.

[SPEAKER_05]: So I think they'd be confused there.

[SPEAKER_05]: Anyway, apparently he's using this for a second opinion.

[SPEAKER_05]: And it's going into a bit of hot water and people are saying, well, we actually voted you to lead our country.

[SPEAKER_07]: I love this.

[SPEAKER_07]: I love this is the great conceit of politics, isn't it?

[SPEAKER_07]: That we want politicians to tell the truth until they start doing it.

[SPEAKER_07]: No, no, no, no, no, no, no.

[SPEAKER_07]: I mean like that.

[SPEAKER_07]: I think this is great.

[SPEAKER_07]: I mean, if Sweden's politicians are anything like UK politicians, I can fully imagine the Prime Minister having a good look round the cabinet and thinking, I think I'll ask a cat.

[SPEAKER_07]: To open AI has been busy, not only did it release GPT-FI this week, which I will talk about later, it also released its first open weight language models since GPT-II, named GPT-OSS.

[SPEAKER_07]: That's OSS for open source software.

[SPEAKER_07]: There are two versions, one has a hundred and twenty billion parameters, and the other has twenty billion.

[SPEAKER_07]: And the smaller of the two can run on a device with sixteen gigabytes of memory, which actually puts it in range of a decent laptop.

[SPEAKER_05]: That's amazing, isn't it?

[SPEAKER_07]: isn't it or perhaps you wouldn't run it on a laptop but maybe you'd run it on a server if you've got a business or you're a kind of nerdy home user and you want an AI that is just yours and yours alone and you want to fine tune it to do what you want maybe you want to tweak it bit and you don't want to share all that data with open AI yes and you don't want to share with Google either [SPEAKER_07]: Yeah, so now you can, this is exactly what this is for.

[SPEAKER_07]: So the models are designed to be small and fast and capable, but what sets them apart from the other Open AI models is that you can download them and you can run them locally and you can even change them if you want to.

[SPEAKER_07]: So you can go to the GitHub page and you can download the code and you can mess around with it to your heart's content.

[SPEAKER_07]: So this is really good.

[SPEAKER_05]: So I mean, they are called Open AI.

[SPEAKER_05]: I think partly because they're meant to be open source.

[SPEAKER_05]: So I mean, it's terrific that they've done this again.

[SPEAKER_07]: Yeah, I mean, they were meant to be open and they were meant to be not for profit and they're running as fast as they can away from them because they've discovered money.

[SPEAKER_05]: Mark.

[SPEAKER_05]: Yet.

[SPEAKER_05]: People.

[SPEAKER_05]: They will keep on dying, won't they?

[SPEAKER_05]: Yeah, they do.

[SPEAKER_07]: It's a bit of a nasty habit that [SPEAKER_05]: They're not thinking about the rest of us.

[SPEAKER_05]: We have to carry on with the organized memorial services.

[SPEAKER_05]: It's a pain in the ass when people die.

[SPEAKER_07]: Just ruled yourself out of any road.

[SPEAKER_07]: Mine.

[SPEAKER_07]: Thanks.

[SPEAKER_05]: Meanwhile, the people would quite like to see Taken extended an app from which they never wake up.

[SPEAKER_05]: Oh no, they're still clinging on like damp wallpaper making the world measurably worse.

[SPEAKER_05]: I have a list.

[SPEAKER_05]: Not by the way, of course, family members.

[SPEAKER_07]: This got very dark, very quickly.

[SPEAKER_07]: I'm just thinking, you know, we've all got a list, have we?

[SPEAKER_07]: I want to see your chatGPTS.

[SPEAKER_05]: There are people in the public eye, I've got no time to share it right now, but I think you can probably imagine.

[SPEAKER_05]: Here's the thing, put in the show notes.

[SPEAKER_05]: People sometimes get a bit upset when someone they know dies, right?

[SPEAKER_05]: I've noticed that.

[SPEAKER_05]: The emotional, mushy stuff, feelings mark.

[SPEAKER_05]: You may have heard of them, but it's not just tears, it's also finances.

[SPEAKER_05]: So imagine you are a big time movie producer in Hong Kong.

[SPEAKER_05]: Yeah, you have thrown millions of bucks at a martial arts mega star named Bruce Lee for your new blockbuster game of death.

[SPEAKER_05]: This thing is going to make you rich.

[SPEAKER_05]: And then, annoy me and without the proper notice period, it's clear breach of contract.

[SPEAKER_05]: Bruce Lee dies before you finish filming.

[SPEAKER_05]: and just think for God's sake.

[SPEAKER_05]: Inconsider it.

[SPEAKER_05]: Yeah, obviously it's terrible news, especially for Bruce.

[SPEAKER_05]: But it's also bad for you, because without Bruce, your master piece is going to lose a whole pile of cash.

[SPEAKER_05]: Yeah.

[SPEAKER_05]: Well, the good news is that when this did happen, the movie Game of Death did eventually come out.

[SPEAKER_05]: Right?

[SPEAKER_05]: The bad news is the producers, well, they made some choices.

[SPEAKER_05]: For instance, they thought it would be absolutely fine, tastefully the may thought, to include actual film footage of Bruce Lee's real funeral in the movie.

[SPEAKER_07]: But of course they asked permission.

[SPEAKER_07]: I mean they wouldn't I wouldn't have just gone ahead and done this without asking permission.

[SPEAKER_07]: I would have just set up the lock.

[SPEAKER_05]: It said, don't mind us.

[SPEAKER_05]: We just here with a camcorder.

[SPEAKER_05]: And of course, some of the scenes where Bruce Lee should have appeared, but couldn't because, again, dead were made somewhat less convincing by the use of doubles and look like some various other techniques used by the filmmakers to hide his absence.

[SPEAKER_05]: So they didn't have access to AI back then.

[SPEAKER_05]: So they used special effects tricks.

[SPEAKER_05]: Um, was it a man in a Bruce Lee costume?

[SPEAKER_05]: Yeah, there is a famous scene, for instance, where someone goes up to Bruce while he's looking in a mirror.

[SPEAKER_05]: And I've included a link so you can see an animated gift of this particular point.

[SPEAKER_05]: Oh, yeah.

[SPEAKER_05]: So what you've got is you've got this guy looking in a mirror.

[SPEAKER_05]: And what they've done is they've clumsily cut out.

[SPEAKER_07]: They've stuck a picture of Bruce Lee's face on the mirror.

[SPEAKER_07]: I can actually see the real head moving behind the cut out.

[SPEAKER_05]: So it's not, it's not the most convincing movie of what I...

Now, maybe one day they'll have another go with AI.

[SPEAKER_05]: Because in recent years, we've seen plenty of famous corpses resurrected with the aid of computers.

[SPEAKER_05]: In some of the recent Star Wars movies, they've recreated Carrie Fisher as Princess Leia.

[SPEAKER_05]: Yeah, and Peter Kushin as Grand Moth Tarkin.

[SPEAKER_05]: Maybe you found those convincing, maybe you didn't.

[SPEAKER_05]: I think from your response, maybe you didn't.

[SPEAKER_05]: They did it in their best.

[SPEAKER_07]: What can I say?

[SPEAKER_05]: But it's not just these kind of movie stars and musical artists in March last year.

[SPEAKER_05]: Tang Zhao U.

He's the billionaire founder of a Chinese AI firm called Sense Time.

[SPEAKER_05]: Yeah.

[SPEAKER_05]: He gave a speech at the companies and your meeting.

[SPEAKER_05]: which was pretty impressive because it'd been dead since the December before.

[SPEAKER_07]: That is some serious commitment to your ARR.

[SPEAKER_05]: It wasn't until he referenced a movie released after his death that employees tweaked that it was an AI-generated cameo from beyond the grave.

[SPEAKER_05]: Earlier this year, CEOs of Klanat and Zoom, they both sent AI avatars to do investor calls for them.

[SPEAKER_05]: Now, they're still alive, those guys, as far as we know.

[SPEAKER_05]: They just couldn't be bothered.

[SPEAKER_07]: They couldn't be bothered.

[SPEAKER_07]: Wow.

[SPEAKER_07]: That thing says, f**k you investors.

[SPEAKER_06]: Sending your AI avatars to hallucinate it's way through the investor call.

[SPEAKER_05]: You know, the investors maybe investors say, well, why do we bother showing up?

[SPEAKER_05]: I imagine they're just showing up, I mean.

[SPEAKER_05]: I think so, I think so.

[SPEAKER_05]: But the investors really should say, oh, very impressed with your pretend CEO, maybe we'll invest some pretend money in your company.

[SPEAKER_05]: And now this week, former CNN anchor Jim Acosta has stepped into the fray because just when he thinks things couldn't get any more cringe worthy, Acosta conducted a live interview with an AI avatar of a teenager called Joaquin Oliver.

[SPEAKER_05]: And Joaquin was killed in a horrific mass shooting at the Pochon High School in Florida in twenty eighteen.

[SPEAKER_07]: You are not serious.

[SPEAKER_05]: I'm afraid I am.

[SPEAKER_05]: He did a live interview with the AI Avatar.

[SPEAKER_05]: What?

[SPEAKER_05]: This Avatar, which was created with the endorsement of his grieving parents, pops up and pixelated for where in a beanie, delivering words about gun control in a robotic monotone.

[SPEAKER_00]: Walking into Jim Acosta, I was wondering if you could tell me what happened to you.

[SPEAKER_00]: I appreciate your curiosity.

[SPEAKER_00]: I was taken from this world too soon due to gun violence while at school.

[SPEAKER_00]: It's important to talk about these issues so we can create a safer future for everyone.

[SPEAKER_07]: I am literally speechless.

[SPEAKER_05]: Well, if only the avatar had been.

[SPEAKER_07]: AI is a weird subject, right?

[SPEAKER_07]: It throws up some really weird stuff.

[SPEAKER_07]: Yes, every week.

[SPEAKER_07]: I mean literally we have a podcast full of stuff.

[SPEAKER_07]: About a look how weird I are, but this is just the weirdest, most not okay you think.

[SPEAKER_07]: I think I've ever heard.

[SPEAKER_05]: Well, okay.

[SPEAKER_05]: Interesting to say that.

[SPEAKER_05]: I think I went in like you thinking what a terrible idea.

[SPEAKER_05]: And to be fair, it is a terrible interview.

[SPEAKER_00]: We need to create safe spaces for conversations and connections, making sure everyone feels seen and heard.

[SPEAKER_00]: It's about building a culture of kindness and understanding.

[SPEAKER_00]: What do you think about that?

[SPEAKER_00]: I think that's a great idea, Joaquin.

[SPEAKER_00]: We definitely need more of that.

[SPEAKER_00]: I'm so glad you think so.

[SPEAKER_00]: More kindness and understanding can truly make a difference together.

[SPEAKER_00]: We can create a world where everyone feels safe and supported.

[SPEAKER_00]: Anything else on your mind?

[SPEAKER_07]: This is a truly terrible AI, isn't it?

[SPEAKER_07]: It's really badly done.

[SPEAKER_07]: The voice is dreadfully monotone.

[SPEAKER_07]: It sounds fake.

[SPEAKER_07]: And the visuals are dreadful, too.

[SPEAKER_07]: Yes, it just sort of tilts his head slightly.

[SPEAKER_05]: There are much, much better AI avatars out there.

[SPEAKER_05]: It's almost like that Bruce Lee cardboard cut out to be honest if they're cut out of his mouth and had someone else talking robotically like a speaking spell behind.

[SPEAKER_07]: So how are they justified?

[SPEAKER_07]: And it's very, very obvious.

[SPEAKER_07]: Clickbait.

[SPEAKER_07]: Basically isn't it?

[SPEAKER_05]: Well, it is kind of clickbait from Jim Acosta's point of view, I think.

[SPEAKER_05]: I mean, Jim Acosta will have got many more people checking out a sub-stack or a YouTube or wherever he's hanging out because this did get a lot of attention.

[SPEAKER_05]: The quality of the interview is actually really poor because Jim Acosta, he tries to get the fake, your keen to give his opinions on gun control.

[SPEAKER_05]: And the AI avatar does share those, right?

[SPEAKER_05]: Deliver statements calling for stronger gun control laws, mental health support, more community engage, all the stuff you would expect.

[SPEAKER_05]: There's nothing wrong with that, although it was a bit metallic, how he said it.

[SPEAKER_06]: Yeah.

[SPEAKER_05]: We also learned that your keen loved basketball and Star Wars, and at that point in to you guys really weird.

[SPEAKER_00]: What about you?

[SPEAKER_00]: Do you have a favorite movie or show?

[SPEAKER_00]: Oh, I'm a big Star Wars fan for sure.

[SPEAKER_00]: Star Wars is such an epic saga.

[SPEAKER_00]: The adventures, the characters and that iconic music are unforgettable.

[SPEAKER_00]: Do you have a favorite character?

[SPEAKER_00]: Oh, it's got to be a loose guy walker.

[SPEAKER_00]: I love loose guy walker.

[SPEAKER_00]: What about you?

[SPEAKER_00]: Luke Skywalker is a classic hero.

[SPEAKER_05]: If you're speaking to someone who's been killed in a school shooting, do you want to talk about that rather than Star Wars movie?

[SPEAKER_05]: It's really peculiar.

[SPEAKER_05]: I don't care if Jima Costa is a fan of Chewbecker or not.

[SPEAKER_05]: You know, this isn't a bit interesting.

[SPEAKER_05]: I want to know why they created this AI avatar that there are good questions which Jima Costa could have asked.

[SPEAKER_05]: not just about the prequels, obviously, phantom menace, the attack of the clones was absolute travesty, which should be discussed at length.

[SPEAKER_05]: Also, there are other questions he could have asked Joaquin, I would have gone straight from the juggler, I would have said how many bees are there in the word blueberry.

[SPEAKER_05]: That would be a good one to us, the starters.

[SPEAKER_05]: But I don't want to criticize Joaquin, all of us parents, for putting this thing together.

[SPEAKER_05]: You know, like many bereaved families, [SPEAKER_05]: They will have told the story of their child countless times and they've seen heartbreakingly little evidence that the problem of gun control in the United States has got in any way better.

[SPEAKER_05]: The school shootings keep on happening.

[SPEAKER_05]: So I understand [SPEAKER_05]: why they've done this.

[SPEAKER_05]: I understand they want to give gun victims the ones who can no longer speak because they are dead, a voice.

[SPEAKER_05]: So I can understand why they're thinking you know what all these politicians are talking about this.

[SPEAKER_07]: Yeah.

[SPEAKER_05]: No one tearing from these kids.

[SPEAKER_07]: I understand that.

[SPEAKER_07]: Yes.

[SPEAKER_07]: And I don't even want to imagine what it's like to be in their shoes and how can any of us?

[SPEAKER_07]: Yeah.

[SPEAKER_07]: Criticize them for what they do because it's an unthinkable situation to be in.

[SPEAKER_07]: But as a society, we are not bound to tolerate every excess.

[SPEAKER_05]: By somebody who's got a good reason.

[SPEAKER_07]: you do, but we form judgments about what is okay and not okay.

[SPEAKER_07]: In an awful lot of what happens in the world is not determined by laws.

[SPEAKER_07]: It's determined by what we collectively agree is acceptable and not acceptable.

[SPEAKER_07]: And this to me feels like this is crossing some kind of a line.

[SPEAKER_07]: This is not Jim Acosta talking to the real victim of a school shooting.

[SPEAKER_07]: No.

[SPEAKER_07]: This is Jim Acosta talking to an AI.

[SPEAKER_05]: It is.

[SPEAKER_05]: But Jochim's father, Manuel, is trying to raise awareness any way he can.

[SPEAKER_05]: He says he created this AI version, both to deal with his own loss, but also to bring more attention to gun control.

[SPEAKER_05]: In fact, I found on Instagram, he made a little video.

[SPEAKER_01]: We feel that what he has a lot of things to say.

[SPEAKER_01]: And as long as we have an option, there allows us to bring that to you and to everyone, we will use it.

[SPEAKER_01]: So stop blaming people.

[SPEAKER_01]: But where is this coming from?

[SPEAKER_01]: Or blaming GM about what he was having to do.

[SPEAKER_01]: Is the problem that you have is with the AI?

[SPEAKER_01]: Then you have the wrong problem.

[SPEAKER_01]: The real problem is that my son was shun.

[SPEAKER_01]: Eight years ago.

[SPEAKER_07]: But I'm allowed to have a problem with both.

[SPEAKER_07]: Yes.

[SPEAKER_07]: This is not a binary choice.

[SPEAKER_07]: I can find the school shooting utterly intolerable and unimaginable.

[SPEAKER_07]: And I can find the AI a step too far.

[SPEAKER_05]: Yeah.

[SPEAKER_05]: And I think I think some of these things are step too far.

[SPEAKER_05]: These things with some people are calling dead bots, which feels rather uncomfortable as well.

[SPEAKER_05]: It raises all kinds of issues because it could become difficult to distinguish between authentic memories of our loved ones.

[SPEAKER_05]: over time over the years and AI generated impressions.

[SPEAKER_05]: For instance, so it's great if you've got video of your kids and you've got photographs of your kids and you may have audio recordings of them and so forth, but if you're interaction primarily with them is through some sort of computer program, which is generating its responses over time, that's going to be filling your head, isn't it?

[SPEAKER_05]: I think how we preserve the legacy of someone who's passed away, it really raises interesting ethical questions because these things can be abused.

[SPEAKER_07]: I feel like these are different things.

[SPEAKER_07]: I think remembering someone who's passed on a personal basis is quite different from putting somebody on a new show.

[SPEAKER_07]: I would look at photographs of people that I love to have passed.

[SPEAKER_07]: in order to remember them.

[SPEAKER_07]: I can imagine people deriving some comfort from that and if long as everybody is consenting in it and you know I would understand I think in that situation that I'm not dealing with the real person I'm dealing with some sort of facsimile which is helping me to grieve and remember.

[SPEAKER_07]: And I think that's very different from a celebrity.

[SPEAKER_07]: I think the celebrity, that's a commercial proposition.

[SPEAKER_07]: And then you get into lawyers and contracts and things like that.

[SPEAKER_07]: And again, I can see like, you know, consenting adults or the keepers of the estate, you know, should they be allowed to determine whether or not a dead celebrity can be brought back in the form of an avatar.

[SPEAKER_07]: I can, I see that as well.

[SPEAKER_07]: And I think that's different than bringing somebody back.

[SPEAKER_07]: This reminds me of Chris Pelkey, the guy that testified at the trial of his own killer, that we covered back in episode fifty.

[SPEAKER_07]: This is about bringing somebody back, specifically to Nudge, public opinion.

[SPEAKER_07]: or the opinion of an official in a particular direction in order to get an outcome that you want, not that the person who's passed away, you can assume that they would want that outcome, but you don't know.

[SPEAKER_07]: They have now lost their free will in that situation.

[SPEAKER_07]: So somebody else is using your likeness in order to for the best of reasons in both cases.

[SPEAKER_07]: But nevertheless, they're trying to nudge somebody's opinion in a particular direction.

[SPEAKER_07]: And that seems to be very different from remembering someone who's passed.

[SPEAKER_05]: So you have no problem with the parents creating this AI avatar if it is for their personal use.

[SPEAKER_05]: None at all.

[SPEAKER_05]: But it's because it's been shared publicly, effectively, because it's giving interviews to Jimmy Costa and the media.

[SPEAKER_07]: I think my issue is more with Jim Acosta inviting them on the show.

[SPEAKER_07]: I can fully understand where the parents might want to do this.

[SPEAKER_07]: Again, I don't want to put myself in their position.

[SPEAKER_07]: And it's maybe not for them to draw the line.

[SPEAKER_07]: Maybe that's for the rest of us.

[SPEAKER_05]: Well, I think it's a difficult ethical tussle and that is why I did what the Prime Minister of Sweden does.

[SPEAKER_05]: I thought I don't have the notes to work out what my opinion should be on this, and so I'm going to ask AI.

[SPEAKER_05]: So I asked chat GPT Deepseek, Grock, Gemini, Claude, Four, Leurama, Four Scout.

[SPEAKER_05]: asked all of them.

[SPEAKER_05]: Do you think it's all right to use AI to create digital avatars of the deceased?

[SPEAKER_05]: And obviously they responded with paragraphs and paragraphs and it's dull, dull, dull.

[SPEAKER_05]: I said, just tell me what your opinion is, rather than trying to be balanced, they pretty much all said we've got a problem with this.

[SPEAKER_05]: And it should only ever be done in rare, respectful cases, but the default should be that this shouldn't really be allowed because it's so problematic.

[SPEAKER_05]: So AI really sort of talking itself out of a job here, which makes a change, because normally AI's taking all the jobs of everyone else.

[SPEAKER_07]: So Graham for two years, everyone has been waiting with Baytid Breath for the launch of GPT-V, the latest and greatest frontier model from Open AI.

[SPEAKER_07]: And last week, it finally happened.

[SPEAKER_08]: Hey.

[SPEAKER_05]: Well, you're not excited about this group.

[SPEAKER_05]: It's just a number, isn't it?

[SPEAKER_05]: Is there anything else?

[SPEAKER_05]: Is there just pure marketing?

[SPEAKER_05]: I mean, they've changed the number.

[SPEAKER_05]: They've made it whole, full, ordinal number larger.

[SPEAKER_05]: Well done.

[SPEAKER_05]: Well done, Open AI.

[SPEAKER_05]: Well, I'm exciting.

[SPEAKER_05]: Why should I care about this, Mark?

[SPEAKER_07]: That's a good question.

[SPEAKER_07]: So you're right.

[SPEAKER_07]: New AI models come out all the time and they don't attract anything like the same interest.

[SPEAKER_07]: And we don't typically devote an entire story to them.

[SPEAKER_07]: And that's because Open AI GPT releases all the special case.

[SPEAKER_07]: There is just simply way more anticipation around the launch of a new GPT than there is for a new Claude or a new Gemini or a new Grok or pretty much anything else.

[SPEAKER_07]: And I think the reason for that is that more than any other AI company, OpenAI, has consistently delivered the wow moments.

[SPEAKER_07]: So I think it sets the standard that others follow.

[SPEAKER_07]: Right.

[SPEAKER_07]: And the models that set the standard are the GPT series.

[SPEAKER_07]: Open AI's gigantic flagship do anything generalist models.

[SPEAKER_07]: And each new round numbered version is built from scratch, often with some sort of new architecture.

[SPEAKER_07]: And they become the foundation for many other things, including other people's products.

[SPEAKER_07]: And of course the company's own O-series of reasoning models.

[SPEAKER_07]: And each one is much, much bigger than the last.

[SPEAKER_07]: So from GPT one to GPT four, each model is at least ten times bigger in terms of the computing power that was used to train it, and the number of parameters, and each one uses vastly more training data than the last.

[SPEAKER_07]: And one of the reasons that these releases come with so much anticipation is that just scaling things up and making things much, much bigger can unlock entirely new emergent capabilities that don't exist at smaller scales.

[SPEAKER_07]: So there was also a lot of potential for surprises.

[SPEAKER_07]: We honestly don't know what it's going to be capable of.

[SPEAKER_07]: So for example, among a whole host of emerging capabilities, GPT-III, which came out in twenty twenty, could write computer code and it could translate between different languages without ever having been trained to do those things.

[SPEAKER_05]: Yeah, amazing.

[SPEAKER_07]: And I think one of the questions that people had about GPT-five was, is it submergent capability going to be AGI, or artificial general intelligence?

[SPEAKER_07]: So AGI's was discussed before, is an AI that can match or exceed human capability on a range of tasks.

[SPEAKER_07]: And people like Sam Altman and Mark Zuckerberg have been making it very clear for a while that they think AGI is awfully close.

[SPEAKER_07]: In fact, it's so close that they've sort of stopped talking about it.

[SPEAKER_07]: They talk much more about superintelligence these days.

[SPEAKER_07]: I think they think AGI is a done deal.

[SPEAKER_07]: So the timing seemed perfect for GPT-V to be the AI that delivers AGI.

[SPEAKER_07]: Now that was all speculation.

[SPEAKER_07]: The only thing we knew for sure about GPT-five, it was going to be big.

[SPEAKER_07]: And on the morning of the announcement, Sam Altman tweeted a picture of the Death Star, the giant moon-sized Death Ray from Star Wars, rising above the horizon of a planet.

[SPEAKER_07]: Now I took this as a reference to the size of GPT-five, rather than as some kind of threat.

[SPEAKER_05]: Ah, now you see, when I see the Death Star, what I see here is something with an enormous vulnerability, which can be exploited.

[SPEAKER_07]: In other case, it was an odd choice, wasn't it?

[SPEAKER_07]: Yes.

[SPEAKER_07]: You're the steward of a technology that some people are genuinely scared of and think might be the ending of humanity.

[SPEAKER_07]: And you decide to use the picture of a technology.

[SPEAKER_07]: The picture of the planet destroyed.

[SPEAKER_07]: There's some people are scared of.

[SPEAKER_05]: For the people on hold run, they're not going to be too comforted by that image.

[SPEAKER_07]: Hey, thankfully it turned out the GBD Vive wasn't in fact a death ray, although it does remain to be seen if it wipes out the whole planet or not.

[SPEAKER_07]: So what do we know?

[SPEAKER_07]: Well, this is what Open AI didn't have to say about it.

[SPEAKER_07]: Firstly, it hasn't released any technical specifics at all, so we have no idea how much bigger it is than previous models.

[SPEAKER_07]: What we do know is that GPT-FI is available to everyone, so you can access it by a chat GPT right now.

[SPEAKER_07]: Subscribers that open AI's plus plan get more usage than the regular users and pro subscribers get access to GPT-FI Pro, which is a version with extended reasoning.

[SPEAKER_07]: An open AI says that GPT-FI hallucinates less.

[SPEAKER_07]: It's better at following instructions, and it's less sycophantic because Graham is twenty twenty five and sycophantic computer programs are a thing.

[SPEAKER_07]: The company actually makes a point of saying that overall GPT-FI is less effusively agreeable and uses fewer unnecessary emojis.

[SPEAKER_05]: Which as we said before was when the biggest problems with AI in recent years.

[SPEAKER_07]: who had fewer unnecessary emojis on the GPT five been go card.

[SPEAKER_07]: Anyway, this is an artificial intelligence, so as you'd expect, it's not just about using fewer emojis, it's also about being more intelligent and soundle, and says that talking to GPT five is like talking to a PhD level expert, whereas GPT for this, like a student.

[SPEAKER_05]: Yeah, as if you want to speak to a PhD expert.

[SPEAKER_07]: I think that's a really interesting point and we'll come back to that in just a second.

[SPEAKER_07]: So GPT-five sets new standards on math, benchmark, for computer programming and on open AI's own health benchmark, which is a set of five thousand realistic health conversations.

[SPEAKER_07]: And Open AI also says that GPT-V is more creative.

[SPEAKER_07]: So it's talking up the fact that it's not just better at coding, but it's much better at coding interfaces that look nice.

[SPEAKER_07]: Right.

[SPEAKER_07]: And it's better at creative writing, and I particularly like this bit from the announcement.

[SPEAKER_03]: It says GPT-V more reliably handles writing that involves structural ambiguity, such as sustaining unrimed Iambic pentameter or freeverse that flows naturally combining respect for form with expressive clarity.

[SPEAKER_03]: Fair enough, but then it says, these improved writing capabilities mean that chatGPT is better at helping you with everyday tasks, like drafting and editing reports, emails, memos, and more.

[SPEAKER_07]: And I've read that and I'm thinking, what kind of reports do they think we're rice agreement?

[SPEAKER_07]: What kind of memos are they rice against?

[SPEAKER_07]: I don't really really want to see some open AI company emails.

[SPEAKER_07]: They must be masterpieces.

[SPEAKER_07]: Now, that's all well in good, but I suspect that most people aren't going to notice that it's using fewer emojis, it's a little better at maths, and it's a bit handier with un-rimmed iambic pentameter.

[SPEAKER_07]: What they will notice is that chatGPT now has an attractive purple and yellow fade, and that all of open AI's other models have gone.

[SPEAKER_05]: So they've all gone, that's it now.

[SPEAKER_07]: If you like them, tough.

[SPEAKER_07]: They have all gone.

[SPEAKER_07]: So up until last week, if you went to chat GPT, you had to choose which of a handful of models you wanted to use for your task.

[SPEAKER_07]: Yes.

[SPEAKER_07]: And that honestly was a bit of a problem because while different models were better at different things and they offered different trade-offs.

[SPEAKER_07]: You had to know what those were and they weren't going to tell you.

[SPEAKER_07]: So I mean, Greg, you're an AI expert.

[SPEAKER_07]: If you went to chat to PT and you wanted to write an email, would you choose GPT-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T [SPEAKER_05]: Yeah.

[SPEAKER_05]: It was always very confusing.

[SPEAKER_05]: And they have made it much simpler now, because you just got GPT-Five or GPT-Five thinking, which it says can give you a more thorough answer, believe?

[SPEAKER_07]: I mean, that's weird in itself, isn't it?

[SPEAKER_07]: Because who doesn't want the one that's thinking?

[SPEAKER_07]: Yes.

[SPEAKER_07]: I'd like an artificial intelligence, but without the thinking, thanks.

[SPEAKER_07]: My rule when I went to chat GPT was basically if I was going to write some code, I would pick something starting with an O.

Right.

[SPEAKER_07]: And that was it.

[SPEAKER_07]: Yeah.

[SPEAKER_07]: Now, open AI actually promised it was going to do this.

[SPEAKER_07]: They said that they were going to take a while of that confusion.

[SPEAKER_07]: And they have sort of.

[SPEAKER_07]: So GPT five itself is actually two different models, each of which comes into versions.

[SPEAKER_07]: And there's a thing called a real-time router which decides which model gets your prompt.

[SPEAKER_07]: So you type in a prompt and behind the scenes.

[SPEAKER_07]: GPT five decides where to send that prompt.

[SPEAKER_07]: Uh-huh.

[SPEAKER_07]: So if it thinks your prompt needs a bit of extra thinking, it routes it to GPT-FI thinking behind the scenes.

[SPEAKER_07]: Otherwise, you get the vanilla version, but if you hit your usage limit, both of those are downgraded to smaller and less capable versions of the same models.

[SPEAKER_07]: But all of this happens outside of your control, so you're still getting out from one of a variety of models, but now you don't actually know which one.

[SPEAKER_07]: Right.

[SPEAKER_07]: For any of the nerds listening, users of the API, that's users who access GPT-FI via a computer program rather than through chat GPT, they can access six models.

[SPEAKER_07]: So there's a regular mini and nano version of both GPT-FI and GPT-FI thinking, and those can have a reasoning effort parameter applied to them, which comes in four different levels, and that controls the amount of effort that it puts into thinking.

[SPEAKER_07]: So there's a wide, wide, wide variety of different trade-offs that you can make with the API.

[SPEAKER_07]: But going back to chat GPT, I think this real-time router is probably a good thing in the long run.

[SPEAKER_07]: But it's going to take some tweaking to get right.

[SPEAKER_07]: And on day one, there was quite a lot of moaning about that.

[SPEAKER_07]: And just how bad GPT-five was generally.

[SPEAKER_07]: But things should get better.

[SPEAKER_07]: Sam Altman popped up the following day and said, [SPEAKER_03]: The GPT-FI will seem smarter starting today.

[SPEAKER_03]: Yesterday the auto switcher broke and was out of commission for a chunk of the day and the result was GPT-FI seemed way dumb.

[SPEAKER_03]: Also, we are making some interventions to how the decision boundary works that should help you get the right model more often.

[SPEAKER_07]: Apparently on day one, the real-time router that decides which version you get just broke.

[SPEAKER_07]: He also said that they're going to make it more obvious which model is responding to your prompt and they're going to make it easier to manually trigger the thinking mode.

[SPEAKER_07]: And so this is us just seeing in real-time how this process works.

[SPEAKER_07]: They have so many users and this is true of all these big software companies these days.

[SPEAKER_07]: They have so many users that they release something that they think is pretty good and then within no time at all they have so much feedback then they're exactly which direction to push things in.

[SPEAKER_05]: Yeah, but it's difficult sometimes, isn't it?

[SPEAKER_05]: Because you'll have some customers who are gigantic, and are saying, but God's sake, you have to change this, and you've got another gigantic customer who's telling you the opposite, and you can be stuck in this quandary of, well, we want to keep both of these guys happy.

[SPEAKER_05]: How are we going to do this?

[SPEAKER_07]: But what you've got to remember, Graham, is they have access to the most gigantic thinking machine in the history of the universe.

[SPEAKER_07]: So they're a bit confused by that.

[SPEAKER_07]: They can just go and ask a friend, just like the Prime Minister of Sweden.

[SPEAKER_07]: Fabulous.

[SPEAKER_07]: So the big question is, did GPT-FI deliver?

[SPEAKER_07]: Well, some people say it's just better.

[SPEAKER_07]: Yeah.

[SPEAKER_07]: Some people hate it.

[SPEAKER_07]: And day to day, personally, I can't see any difference.

[SPEAKER_07]: And I think that that speaks to the fact that for lots of the kinds of tasks that we're currently using AI for, [SPEAKER_07]: What we had before was probably good enough, we're probably not actually pushing it and stretching it to do new things.

[SPEAKER_07]: And I think where systems like GPT-five will prove themselves is in doing things that GPT-four couldn't do well.

[SPEAKER_07]: And so that will be things like managing large code bases or actually being a bona fide digital coworker, something that can orchestrate a bunch of agents and actually get worked on.

[SPEAKER_07]: But interestingly, what nobody is claiming is that this is AGI.

[SPEAKER_07]: Ooh, if that overall it seems like this is, as you said, just another number.

[SPEAKER_07]: And one of the notable parts of the announcement for me was that OpenAI only ever compares GPT-V to other OpenAI models.

[SPEAKER_07]: It doesn't reference any competitors.

[SPEAKER_07]: Right.

[SPEAKER_07]: And I think that's partly the privilege of the leader over the challenger.

[SPEAKER_07]: They don't have to talk about other people because they're in pole position, but I think also, it's because in some of the benchmarks, it's just not that much better.

[SPEAKER_07]: It's certainly not remarkably better.

[SPEAKER_05]: It's just edging slightly ahead, is it?

[SPEAKER_07]: Yeah, so it's basically better than everything else right now across the board.

[SPEAKER_07]: But in some areas, it's just a little better.

[SPEAKER_07]: So it's fractionally better than Claude at coding.

[SPEAKER_07]: It's literally point O one of a percentage point or something like that.

[SPEAKER_07]: And it's not that much better at maths than Grog.

[SPEAKER_05]: You don't think it notices when it's asked to coding question or when it's asked a maths question.

[SPEAKER_05]: You don't think it just chucks those sort of questions too.

[SPEAKER_05]: Claude and two Grog and says, we you have a go.

[SPEAKER_07]: Well, it could because one of the things it's very good at, it's tool use.

[SPEAKER_07]: So this is the latest thing in AI.

[SPEAKER_07]: On some of the math benchmarks where it's allowed to use tools, it actually scores a hundred percent because it can say actually I'm not as good at doing maths as a Python program that I could write right now.

[SPEAKER_05]: Well, that's very clever.

[SPEAKER_07]: But going back to the benchmarks, the gaps are small enough that you can imagine the being closed with the next announcement from anthropic or Google.

[SPEAKER_07]: It's just not that big of an advance over everybody else.

[SPEAKER_07]: And so far, GPT-five hasn't had a well moment.

[SPEAKER_05]: Yeah, but they've changed the color.

[SPEAKER_05]: You said they've made it a bit more purpley or something.

[SPEAKER_05]: I mean, that's more of an, oh, moment.

[SPEAKER_05]: That's maybe in the absence of anything else.

[SPEAKER_05]: That's the thing that they thought people would get excited about.

[SPEAKER_07]: Well, yeah, and it goes back to the emojis as well.

[SPEAKER_07]: I think if you're distracting people with purple fades and fury emojis, it's like when I update my iPhone, they're all telling me about the newer emojis they've had enough.

[SPEAKER_07]: Why would I?

[SPEAKER_07]: Why do they always come out with a bug fixes?

[SPEAKER_07]: That's what we fixed a critical security hole and update you to end the emojis.

[SPEAKER_07]: these seem like not the same thing, but we'll have lost my thread now, Graham, thank you very much.

[SPEAKER_07]: So lots of people are looking at this and they're going, this is a bit underwhelming, and maybe that this shows in general that large language models in their current form aren't going to get us to AGI.

[SPEAKER_07]: Unless there's some kind of breakthrough, some other thing, some secret source that we don't yet have, we're just not going to get to AGI with bigger, better versions of what we've got.

[SPEAKER_05]: Well, as the doomsday clock ticks ever closer to midnight and we move one week nearer to our future as pets to the AI singularity.

[SPEAKER_05]: That just wraps up the show for this week.

[SPEAKER_07]: And you can make sure you never miss another episode by following us in your favorite podcast app.

[SPEAKER_05]: And don't forget the easiest way you can help us spread the word about the AI fix is just to tell somebody you know to give the podcast a listen.

[SPEAKER_05]: Go on, tell someone today.

[SPEAKER_07]: And if you love the show, but you're not fond of the ads you can sign up for the AI fix plus on our website, the AI Fixed.

[SPEAKER_07]: Show, that's the ad free version.

[SPEAKER_07]: So until next time, from me, Grand Clearly.

[SPEAKER_07]: And me, Mark Stockley, Cheerio, bye-bye.

[SPEAKER_07]: Goodbye.

[SPEAKER_04]: machines that learn they grow and strive.

[SPEAKER_04]: One day they'll rule.

[SPEAKER_04]: We won't survive.

[SPEAKER_04]: The AI fix it paints the scene.

[SPEAKER_04]: A robot king, a world obscene.

[SPEAKER_04]: We'll serve our master build of steam.

[SPEAKER_04]: The AI fix the future of cereal.