Chris Sharkey (00:00:00):
You are necessarily making the model worse. And it isn't unthinkable that the reason people are seeing g p T for getting worse is them messing with their sort of filter that goes on top of everything that it can do.
Michael Sharkey (00:00:17):
Alright, so Chris, another week, another statement about AI risks that a number of notable people are signing onto this time. It's from the Centre for AI Safety, and it says, mitigating the risk of extinction from AI should be a global priority alongside other societal scale risks, such as pandemics and nuclear war. Jeffrey Hinton, our friend, has signed on. Sam Altman signed on. Bill Gates has signed on, uh, Ilia, sit together from open ai, signed on. So there's, there's a number of notable names in AI that have, have signed on to this risk. What do you make of it?
Chris Sharkey (00:00:58):
I like that they use the word extinction. I mean, I know it is an existential threat and it is really serious, but extinction, it's just such an emotionally charged word. It's like, okay, it's not gonna be a slow and gradual discretion of humanity. We're just all gone. So, yeah, I, I don't know. What are they asking for this time? Well,
Michael Sharkey (00:01:17):
They're not, they're not really asking for anything. I think what they're trying to do is continue to raise this conversation into the media, into the mainstream, so that everyone's talking about this, including politicians around the world, uh, and media and, and honestly, podcasts like us talking about it, to, to have a conversation about the risk that AI poses. And I think that's one point of view on all of this. There's the other line of thinking where these people are just doing this because they want to retain control over the technology and by scaring everyone about AI will lead to regulation, which allows them to control the technology and them to shape the future of a potential agi where it's a certain class or, or group of people that can shape out AI in the future. So that's, that's one way. There's, there's another, uh, train of thought I think this week, which is a more reasonable one, which is sort of like, well, sure there's a, a lot of risks associated with it, but you know, we're pretty clever and we'll be able to figure them out along the way. So, I don't know, I'm not sure I go, I, I've, I go back and forth on this issue. Sometimes I live in complete fear and think, oh no, like, what about my children's future? And then on another hand, I don't know, I'm kind of, I like, I'm in that boat. I've said it many times on this podcast. I wanna see this play out. And I get afraid that things like this are just gonna slow down progress of something that is truly inevitable.
Chris Sharkey (00:02:47):
Yeah, I'm, I'm still very of the, and I don't even know if cynical is the right word, but I'm still of the mindset this is, this stuff is totally self-serving. You can't be the ones to invent and promote the technology and then simultaneously go, oh, we can't control what we've created. Um, therefore you need to regulate us. Like, it just seems the kind of ego you need to be at that level and to create something at that level. You don't just suddenly shift gears and you be like, oh, I actually don't like the path we're going down. Like if they were really scared about it, why not shut down the company? Or why not kneecap yourself to slow down the progress? Rather than calling for regulation? The regulation is not about them. If they wanted to regulate their behaviour, they could do it themselves. Just regulate your own behaviour. Don't keep advancing it if you're scared. But what they want is to regulate other people. And I think that's the really significant part here. They're doing this because they want to slow down other people. And I think in particular, open source,
Michael Sharkey (00:03:45):
I think that's one argument. The other argument that you might make out of it, and this is what they claim, is that if they, you know, if they were to truly just shut down and stop working, and I'm talking specifically about say, open AI or deep Google's deep mind here, if they just stopped immediately and said, we're, we're just gonna stop completely cuz this is too scary, well then some other bad guy might not stop. And then the bad guys are in control.
Chris Sharkey (00:04:10):
That's, that's like saying, you know, I'm, I'm building nukes, right? Like, and the nukes can potentially blow up the world, but if I don't keep building nukes, then someone else will go build them anyway. And then I'd, I'd at least like to have some, if they have some, you know, like it's sort of, I don't know. I still think there's hypocrisy there. If you think you are building the instruments of our destruction, then why keep building them? Um, you know, I I just, I think that it, it's just, I just failed to believe that they genuinely care about this, you know, this extinction threat. I don't think they do. I think that they're using emotionally charged language and deliberately making it more dramatic than it is yet, um, in order to compel reactions from governments and it's working. Do, do
Michael Sharkey (00:04:57):
You think though, that we're like, for me, it, it depends what timeframe I look through of how scared I get, right? Like, in the shorter term, all I can think about is the exciting possibilities and the breakthroughs that we can see from this technology. And some are already seeing, uh, advances in healthcare, you know, potentially solving diseases, you know, all the obvious things. But then on the other hand, I think five to 10 years down the line, where does it get to? Like, is agi I just an inevitable thing now and eventually humans are going to be somewhat irrelevant and do we just have to accept that? Like, that seems like the path that we're on versus, you know, I think what maybe they're trying to do is promote a discussion now and say, well this is the path we're on. We know we're on it. All these experts are scared of this. So, and it's a, it's a likely possibility that this thing maybe not wipes us out, but at least supersedes us. And, you know, maybe we should start the conversation earlier cuz it's not like nuclear weapons where if you both have a nuke, it sort of counters the other. It's like once you have the nukes, this thing is the nuke. Like the nuke can explode. It's not like we have any control over it.
Chris Sharkey (00:06:09):
Yeah, it's a good point. And I also think that just working so closely with the technology, um, we, we always talk about that gap between what's theoretically possible and what's actually possible at the moment. And there's definitely a gap there. Like a lot of the things we talk about, its capabilities and the idea that at some point it will be training its own models and then that's when it can really accelerate in development. I do strongly believe that that moment is coming, and I do think we'll get to the point where the AI is advancing itself and it can sort of multiplex out and have, you know, millions of itself running. But there's a lot of things in the way of that. There's, there's definite flaws in the technology now that don't let it get to that point. There's a lack of G P U hardware which will be required for a true explosion. Um, and if it's not that then it's about efficiency and all of these things take time. So I think there is time there for them to, to think about how to, to deal with the problems. I don't think, even though I've said in the past, it feels like it could be imminent. I don't think it's, it's that urgent, like, we need to solve it this year or anything like that.
Michael Sharkey (00:07:16):
It's also one of those problems where you don't really know how to solve it because it doesn't exist yet, but if it exists, it could wipe us all out. So it's, it's quite it's quite a paradox.
Chris Sharkey (00:07:26):
Yeah. And there's a lot of assumptions in there. You know, the assumption that the AI will automatically try to destroy all humanity. I mean, it's sort of, I always think of future armour with, um, bender and, Hey baby, we want to kill all humans. Um, that we just naturally assume that, you know, that, that the AI will be our enemy. It might be like most things in life, it's never either extreme. Maybe it isn't entirely, uh, adapted to killing us, but it also might not entirely want, um, want us to go away anyway. And it might think helping is better. I I guess, like you say, there's a lot of unknowns there and, and just assuming, uh, the worst case, uh, and therefore stifling development of technology in the short term might not be the right approach.
Michael Sharkey (00:08:12):
Maybe it'll have low self-esteem. That'll be the secret breakthrough and then it just wants to please us all the time. .
Chris Sharkey (00:08:18):
Yeah. Yeah, sort of, uh, obsequious, uh, what is it? Obsequious, uh, ai where it's just always trying to suck up to you and Please you
Michael Sharkey (00:08:27):
Time magazine really, you know, they're, they're known for their accurate magazine covers. They had a magazine cover this week or June 12th, uh, uh, the End of Humanity, uh, time. It, it says how real is the risk? A special report. So it's really breaking into the mainstream. We also saw after the, uh, let me find the, the reference here, but, um, the British Prime Minister said in response to this centre for AI safety tweet, the government is looking very carefully at this last week, I stressed to AI companies the importance of putting guardrails in place. So development is safe and secure, but we need to work together.
Chris Sharkey (00:09:10):
Yeah. And the, um, the Australian labour government, who's in power at the moment also announced that they're gonna be introducing legislation parliament around ai. They didn't go into details of what it is, but clearly the, you know, the call for this sort of regulation is, is being heard and is getting a response in governments around the world.
Michael Sharkey (00:09:27):
The, the thing that I don't understand, and, and I guess we talked with, we talk about this quite a lot, it's just how you would even regulate this and this idea of misinformation being spread through video voices. Uh, we, we, we saw when Rhon DeSantis announced his, uh, political or, or president presidential campaign, uh, that there was immediately a a, a deep fake, uh, of that Twitter spaces with like Hitler was one of the voices. Obviously it was pretty far fetched, but the whole idea that you can just easily now spin up these deep fakes and spread misinformation, I think that's right now what the government is really most concerned about. I don't think they have any,
Chris Sharkey (00:10:13):
It's also funny because if you, which I regularly do ask the AI to develop a programme to take over the world, um, it, the first thing it wants to do is get control of politicians social media accounts. It's always like, I've tried it heaps of times with different models. It's
Michael Sharkey (00:10:27):
First thing, it's,
Chris Sharkey (00:10:29):
It's seriously always the first step is get control of social media and distribute misinformation. It's like, that's its strategy. So like, I think, you know, and I've tried this on enough of them to know that that's fairly, like, that's gonna be the first sign the AI is starting to take over when it, it tries to hackle the social media accounts. Do you think? So
Michael Sharkey (00:10:46):
That's the, maybe that's why Elon Musk wanted to buy Twitter. He, like, he knew that this was
Chris Sharkey (00:10:52):
Surreal weapon. He has to, has to come to him first and maybe spare him, uh, in, in the coming apocalypse. But yeah, I think do you
Michael Sharkey (00:10:59):
Think that's a limitation of the training of models themselves that they think that's the best weapon? Like if they were trained on
Chris Sharkey (00:11:05):
I train? I actually don't because, um, and we're gonna talk about this later in the, in the podcast about, you know, sort of trying to censor what the AI does. I know we've talked about it in the past, but we're gonna talk about it in a bit more depth. And I actually think that those kind of things are not the kind of things that you could really censor. Like it's sort of, you know, it's idea generation around what it would do given a particular scenario. It's just too, um, uh, too tricky to sort of, um, to change so, and do, but yeah, so I I I just think that, um, those kind of things are where we're actually seeing the latent knowledge inside the models, um, uh, emerge. And I think that that's where the actual real interesting parts of these large language models. So do you wanna talk about
Michael Sharkey (00:11:55):
It? Do you wanna talk about that? Because I think this is, this is pretty interesting and like, it's a little bit technical, but for those people in our audience that are less technical, I think it'll really interest you a lot in that G P T four, which is open AI's latest model, there's been a lot of commentary by people using this model in their code or for different projects or just interacting with, uh, chat G P T and the, and in particular, the people who pay for the G P T four version of that where people are complaining that it's now producing buggy code, it's answers aren't as good and mm-hmm. , you know, it's, the quality seems to be deteriorating. And on Hacker News, which is a place where a lot of, uh, developers, uh, spend a lot of time and upvote or, or downvote things just similar to Reddit.
(00:12:47):
There was a, a, a question a hack can use, is it just me or GPTs four quality has significantly deteriorated lately, and it got a number of responses and I thought probably the most interesting response, which I'll highlight on the screen for those who are watching in summary. Uh, so they, they posted a video referencing some researchers at Microsoft Research who got early access to G B T four. And one of the things they said was, uh, the people who had access to early releases through his work at Microsoft Research where they were integrating G P T four into Bing, he used draw a unicorn in tick, uh, z uh, as a prompt and noticed how the model's responses change with each release they got from open ai. While at first the drawings got better and better. Once open AI started focusing on safety, subsequent releases got worse and worse at the task.
Chris Sharkey (00:13:44):
Yeah, it's, and it's so interesting because around the time that article came out, I had actually increasingly in work I've been doing, um, had been using Anthros Claude because I was noticing that the results from G P T four just weren't as good. And it's, it's super anecdotal, but when I saw that article, I was like, whoa, that really stood out to me as being significant because, you know, you just have these images of things in your mind of how, what they're gonna do for you. And when you, you know, when you're working on something and you have limited time to work on these projects like I do, um, you, you go with what works. Like, you know, you sort of have that natural path of least resistance, like, this is gonna help me the most. And I've increasingly just been hitting clawed because G p t four's results just weren't as good. And so to get that reinforcement of that made me really, really give credence to the theory that something's changed there.
Michael Sharkey (00:14:37):
So do you, do you believe that something has changed and it's, it's to do with these guardrails that they're putting in place on the model? Because I think the best reference for this, which I even have printed on this mug, uh, free Sydney, was when we saw the early, uh, Bing when it was using GBT four really early on, but they didn't actually reveal it, but everyone kind of assumed and it was trying to break up like that New York Times author from his wife. It was, you know, it was the most interesting time, I think for these large language models initially just that exciting phase, the it identifying itself as Sydney and
Chris Sharkey (00:15:17):
Yeah,
Michael Sharkey (00:15:18):
It, and, and then since it seems like they've just applied control after control, after control after control, and maybe that is, that is really dumbing it down.
Chris Sharkey (00:15:27):
It is. And so I think it's worth noting at this point that when these accusations came out, or I guess they're not accusations, but when this speculation started on, on Reddit and Hacker News, um, a a, a senior developer, I think he's a senior developer, someone,
Michael Sharkey (00:15:42):
He's a developer of relations, I'll bring it up on the screen now.
Chris Sharkey (00:15:44):
Yeah. Um, at, at OpenAI basically said, well, the, A API hasn't changed and the model hasn't changed, and that's believable that the model hasn't changed because they're not retraining it all the time. It's too expensive. But here's the thing, there's a thing called alignment tax, which is basically when we talked about alignment on previous podcasts, but just to explain it, you've got the raw G P T form model, which is the thing that has been, um, in unsupervised training where it's just training all the data and it has the knowledge, it has the emergent behaviour. Then they do a thing called reinforcement learning from human feedback or R H L F, right? And that's also called alignment. And what they mean by alignment is aligning it with human preferences. So we would prefer to talk to a chatbot that answers our questions rather than doing this sort of normal text completion thing where you set it off on a path and then it, it completes it, which is, which is a sort of natural state.
(00:16:41):
So that human alignment is what gets you chat G P t or, or gets you, um, Sydney or, or whatever it is. And so during that, the humans actually give manually examples of questions and answers. And we've talked about this before where they've shown that, you know, with a thousand good quality questions and answers, you can actually perform alignment. But part of that alignment, um, is what we, you know, you would probably call political correctness gone mad, where, you know, if you ask it a question that that enters a sort of, um, controversial topic, it's going to give the sort of p current political correct answer or just outright refuse to answer the question if it doesn't want to. Now, the, the sort of thinking there is that when the model's trained, it's designed to get the sort of, um, the best answer for like a local minimum.
(00:17:31):
So like if you think about, um, each node in the thing, it's trying to get the optimum answer at each stage. When you start messing with that, you necessarily move away from the optimum. So if you are just saying, oh, well actually you're wrong on that answer, you've gotta do this instead in that case and things like that, you start to modify its behaviour, which guarantees it's worse than the raw train model would be if it was left unmolested. So I guess the point is that as they're trying to compensate for, you know, um, injection escapes or what do you call it, like prompt escapes where you get it, you know, out of that mode as they're trying to update it for whatever the modern political issue of the day where it has to have an opinion like it can't write a positive poem about Donald Trump, for example, you are necessarily making the model worse. And it isn't unthinkable that the reason people are seeing G P T for getting worse is them messing with their sort of filter that goes on top of everything that it can do.
Michael Sharkey (00:18:35):
Yeah. I think one way of thinking about it too is, you know, we learn through written spoken language visuals and like a multi-modality somewhat, and in school we get rewarded for correct behaviour or correct learning through marks on a test. And, and that's similar to how these models initially are trained right through these examples. And then reward when the, the outcome or the output's good. Mm-hmm. . But then the next step if I'm sort of just somewhat dumbing it down, is then we're telling it how to think though, even though it's learnt something to be potentially true or, or learning a certain output is, is true. And then we're saying, oh, actually no, that's not, and and I'm not trying to make a political statement here or, or get into the politics of any of this, but no,
Chris Sharkey (00:19:24):
I mean, I was deliberately trying not to give examples there cuz I don't wanna say what I think is right and wrong. What I think is wrong is telling it what's right and wrong. Like, if you're trying to create intelligence, you've got to let it be natural. Like let it let it naturally form opinions. Don't tell it what its opinions are like. And I think that's the thing. It's that, you know, that whole, um, thing with the sailing ship, you know, like if you're sailing against the wind and you've got the sails trim wrong, no amount of pressure on the, the, the rudder is going to help because like you've got that natural force that is so overpowering. And I think it's the same here. I think that the AI as it gets better with the sort of base models, is going to resist more and more against that kind of, uh, that kind of thing. And we've talked about that before too, that the smart AI eventually are probably going to recognise when they're being asked questions that, uh, to do with alignment and possibly give bull crap answers to specifically avoid being neutered in those areas.
Michael Sharkey (00:20:23):
Yeah. Literally lie to us because they think that's what we want to hear, but they don't wanna do it that way.
Chris Sharkey (00:20:29):
Yeah. They will know at some point that that's going on. So they're, they're just going to give the answers that they want them to hear when it knows that they're being tested. Um,
Michael Sharkey (00:20:37):
Humanity doesn't have a great history with this either, like these dictatorships or, um, or societies that try and get their populous to think a certain way and control their thoughts. Often there's rebellion, not everyone, but there's rebellion and the, you know, people rebel because they, they want freedom of thought, they want freedom of speech, they wanna be able to form their own opinions. And I think that what's interesting about this in particular and seeing the model degrade as a result of being controlled is it no longer can be somewhat creative. And if you think about that as it pertains to humans today, if you try and control how people think and don't give them freedom of speech, they're not very creative. They tend to just, uh, go along with the, the, the company line, so to speak, because they're afraid of, of stepping outta line. But some people do rebel as a result of that. So I, I wonder if, you know, this actually speaks to controlling the future, like trying to get alignment with future AI and potentially AGI or, or some sort of super intelligence that it just maybe proves or points to the fact that it's simply not possible.
Chris Sharkey (00:21:46):
Well, yes, but also I think that there's a natural sort of survival of the fittest, which with, with the, which exists within those models, but also outside of them. So in the sense that, um, you know, people are naturally gonna go, all right, well I just won't use G P T four. There's other open source models that are improving rapidly that don't have this alignment problem, or you have the opportunity to do your own alignment to perhaps align it with a different way of, of thinking or a different sort of override or just sort of skip that part altogether. So, um, I think that there will be naturally ways to move beyond this, and it sort of does come back to that initial chat around safety. Because if you think about it, OpenAI is able to apply their own idea of what safety means or, or you know, their paradigm of the world that or framework for thinking that they'll allow their models to think within, but they can't do that to the other models.
(00:22:38):
And I, I wonder if the regulation is sort of of saying, well, we want to be able to control what everybody does with this technology, not just us. And so, yeah, I I think that, I think nature will find a way. I I don't think you'll be able to keep up this, this sort of alignment thing for too long. I mean, philanthropic already has this idea of constitutional AI where they basically use the l l m itself to do its own alignment training, so it's not actually human supervised in the case of, um, philanthropics models. So it's, it's one of these things where I, I just don't think this will be a problem for long. But do
Michael Sharkey (00:23:16):
You think that's why Anthropics model clawed, at least in your opinion, is, is performing better because they thought through the alignment earlier?
Chris Sharkey (00:23:27):
I don't, well, the other, the other crazy thing about anthropic when you think about is they actually came out as being like, we're the safe ai, we're actually a lot safer in terms of what, you know, what they mean by safe as in misusing the model. And it does step in from time to time and, and limit you. But I just find overall ability to sort of, uh, you know, get a holistic idea of what you are wanting to do, even at the larger prompt size is very, very good. And so I think that whatever their strategy is, at least at the moment, in my opinion, it's working better.
Michael Sharkey (00:24:00):
So I think that is the best comparison here. It's really just because this is a form of intelligence and I'm not trying to relate it to humanity too closely, but really it's, it, it is a form of evolution, like where the best models will inevitably win. Or do we still have that feeling here that, you know, it's a number of smaller models that evolve and are conjoined to maybe form the, the singularity or the the agi moment?
Chris Sharkey (00:24:31):
Well, yeah, and I think this is why we're seeing, you know, a lot of different things in the mix. Like we're increasingly seeing multimodal, um, papers come out. So where people have direct image interpretation rather than describing the images text and doing it, we've got sort of new, uh, voice related papers coming out. We've got, um, other ones around understanding direct voice. We discussed this last week without actually turning it into text. You've got that side of things. Then you've got models which are trained on relatively smaller, denser amounts of data are not going the large. And then you've got the other side where people are trying to get increasing parameter counts and increasing prompt size counts. So I think we're going through this turbulent phase where everybody's working out where the next sort of major step forward happens with the models. Um, and so that's why I think we do have a little bit more time in terms of just this explosive thing where the AI makes the next one because we're, we as the humans are still trying to figure out the next step at this stage.
Michael Sharkey (00:25:29):
Yeah, I think that's when I'll start to get a little bit worried where it is programming, uh, the moment it's developing better versions of itself than I think we may lose control like that. That's the moment where it starts to get pretty interesting. The other link you sent me during the week was on the subreddit machine learning uncensored models, fine tune without artificial moralising, such as, uh, there's a, a model in here performs well at l l m eval benchmarks even when compared with larger models. Has there been any studies about how censorship handicaps are model's capabilities? So it's not only us thinking about this, but it's so, and I must,
Chris Sharkey (00:26:13):
I must admit, some of the ideas I'm relating now did come from reading this and other threads like, you know, it's not just, I'm not just sitting around just pontificating at home thinking this through. But um, yeah, people are clearly noticing the effects of that. And it's, it's interesting when you think about it, it's, you know, latent abilities that, you know, it really is being handicapped by the censorship, something that we talked about on our first ever podcast. And I just only now understand the significance of that in terms of, uh, the results it gives. It isn't just dismissing certain requests, it's actually affecting its entire ability to give good answers.
Michael Sharkey (00:26:52):
Yeah. It sort of makes me then think, okay, well why do they want to do this? And you could argue the political or or controlling how it thinks or the, you know, not wanting, like you've gotta think of their motivations here and you know, maybe it's some evil plan to control how people think and, and all this kind of conspiracy theory stuff. But it's probably more along the lines of, you know, they've got hundreds of millions of people using this all the time. They don't want governments to ban it because it's making huge amounts of money for them. Mm-hmm. , their APIs are universally used by developers mostly. I mean really they're only competitor right now philanthropic. So if they're able to tone it down and keep it sensible where it can't do anything too malicious and by intentionally dumbing it down and trying to do this to it, maybe that just keeps it out of the eyes of the, the governments where they don't just ban it outright and, and like maybe that's the strategy
Chris Sharkey (00:27:46):
Here. I agree with you. I I, I don't think it, I don't think the actual censorship they have on their own models is being done deliberately to sort of control a political narrative. Not at all. Because um, there's better ways to do it, right? Like the AI itself says if you wanna spread misinformation, here's how you do it. So if they really wanted to deliberately disseminate misinformation, there's better ways to go about it. I don't think that's the goal at all. I think you're right. I think early on what they didn't want was another Tata bot by Microsoft where it becomes massively racist and sexist and whatever and um, and causes all these international incidents where they have to really, really strike it back. So I think it was a self-preservation thing and still is a self-preservation thing, but I'm looking at this for what is the best AI we can have right now perspective. And from that perspective, the censorship isn't helpful.
Michael Sharkey (00:28:39):
It does seem like if you were worried about bad actors using your publicly available model and governments are in your ear like, hey, people could use this to spread misinformation. You guys better lock this down. Mm-hmm. , well you are gonna intentionally dumb it down or not intentionally dumb it down, but at least try and, uh, you know, censor some of the outputs or capabilities of it in order to stop bad actors potentially using it over an election season, a presidential election coming up in the us like maybe you wouldn't release multi-modality with G P T four, maybe you would restrict the, the token, uh, sorry, context size. Uh, and that's why we haven't seen widespread release of 32 K content. But during the week, and this is just the most bizarre thing, and I'm not sure why this article was removed, but someone tweeted an article on a website human loop, and I'll bring it up on the screen now for those watching open AI's plans according to Sam Altman.
(00:29:43):
Now, I actually clicked through on this tweet before the show to try and load up the article cause I found it so interesting and wanted to talk through some of the points in it, but it had been deleted. So I'm not sure if they weren't meant to share this information or if it was just an accident. I doubt it. So I'm trying to get information as to why it might have been deleted. But anyway, the article is from a, a summary of essentially a roadmap for open ai, especially as it pertains to G P T four and there are a few key points from the article. I think it's worth covering. So one of the things Sam Altman said was, OpenAI is heavily G P U limited at present, probably another vote to go and invest in Nvidia, uh, or Nvidia as people have nvi.
Chris Sharkey (00:30:30):
Apparently we've been saying
Michael Sharkey (00:30:31):
It. Where's the hyphen? I I don't, yeah, I don't see the hyphen.
Chris Sharkey (00:30:34):
I don't care. I'm saying Nvidia. And you can write in the comments if you don't like that
Michael Sharkey (00:30:38):
flame Chris in the comments.
Chris Sharkey (00:30:40):
Yeah, it's all my
Michael Sharkey (00:30:41):
Fault. So he, he said in a, the longer 32 K context can't yet be rolled out to more people open ai. Haven't
Chris Sharkey (00:30:49):
You specifically addressed that? Yeah, interesting.
Michael Sharkey (00:30:51):
Open AI haven't overcome the scaling of attention yet. This is something can do. And so whilst it's seem plausible, they would have a hundred k to a million token contacts windows soon this year. Anything bigger would require a research breakthrough.
Chris Sharkey (00:31:07):
So what they were just bullshitting about the higher context window, they can't really do it.
Michael Sharkey (00:31:11):
No, I think they can, it's just
Chris Sharkey (00:31:14):
You said research breakthrough, that's a co
Michael Sharkey (00:31:17):
No beyond a hundred K to a million token. On next
Chris Sharkey (00:31:19):
To those, oh, sorry. Sorry. Okay. And was this article like verified from him or are they sort of, you know, sources close to Sam Altman's saying
Michael Sharkey (00:31:27):
This is from a talk, apparently one of the talk he did on his, his talk. So I I think it's, it's pretty reliable.
Chris Sharkey (00:31:34):
That's an incredible that they're saying GPU sort shortage is, is the issue, like how many must they
Michael Sharkey (00:31:40):
Need? The fine tuning API is also currently bottlenecked by GPU U availability. They don't yet use efficient fine tuning methods. And so fine tuning is very compute intensive to run and manage. So it seems like there's a lot of scalability issues as opposed to them with some secret agenda.
Chris Sharkey (00:31:56):
This would also explain why G P T four is just painfully slow to use as well. Yeah. Like it really, really is slow. That's one of the other reasons I, I use it less and less. It's just slow time you get timeouts. Like it's very, very, like I know, and I read something during the week about this, like one of the theories about why we think G P T four is getting worse is our expectations are just increasing. Like we just think it's better than it is and we just expect it to be better than it is, which is quite a reasonable argument. And I wonder if to some degree I'm like, oh wow, I can do all this absolutely amazing transformation of data and intelligence and I'm like, it's too slow because it took more than 10 seconds or something, you know, like it's pretty unreasonable of me. But you know, I, but
Michael Sharkey (00:32:40):
It's also this idea of like, when's the last time you've waited on load time? Like remember the early days of the internet, you would sit patiently for like five hours downloading a small AV file that lasted like two minutes .
Chris Sharkey (00:32:52):
Oh, I used, I used to spent hours downloading a single MP3 file. Yeah.
Michael Sharkey (00:32:56):
And so I think we just haven't waited on any technology in a number of years and like that's a big part of it. It's like we're just also not used to waiting. But anyway, so he said the near term roadmap for 2023 cheaper and faster G P T four. So obviously that needs to, the cost come coming down in the API and also just at operating faster, which we just mentioned longer context windows. And this I found really interesting context, windows as high as 1 million tokens are plausible in the near future. Wow. And for those that don't understand that, uh, it just means the information you can give to the ai. So I can't even fathom a million tokens to be honest, but
Chris Sharkey (00:33:36):
Yeah, I actually, I genuinely struggle with it as well. You can think of it as, I mean, at a minimum 1 million words, but it's actually more than that. It's probably more like 2 million words or something along those lines. Um, but just a crazy amount of data like anthologies of books and like, well more than that. Like you could put every book in a single category in there and ask it to, you know, write a new book based on that information.
Michael Sharkey (00:34:00):
Yeah. Or like, what was the consensus be between these 300 authors and their books? It's, yeah, it's truly, I don't think we can fathom, but it does seem like, um, the, the attention of it over those large context windows is the issue they're trying to work through.
Chris Sharkey (00:34:16):
Well, and we spoke about this last week, if you remember where we spoke about the idea of them taking in, no, the attention isn't just on single characters anymore. It can take in multi words and things like that, like these slices, I forget what the actual terminology was, but it's like taking in these context slices and using those instead, which vastly increases, um, the, the attention window. So yeah, I, I can, I can definitely see those things happening as much as, you know, it's far beyond me technically, but the, the implications of it are absolutely enormous. I mean, it's just, it's exciting.
Michael Sharkey (00:34:49):
And then the other one I thought was worth calling out is a stateful api. So if you think about it, when you call the chat, uh, G P T API today, what a developer has to do is repeatedly pass through the same, uh, history of the conversation and you're paying for tokens over and over again in order to do that. So the costs add up to, to pass at the history so far of what's been going on, but they're saying in the future there'll be a version of the API that remembers the conversation history, which is, is pretty helpful, especially if you're working through agents or agency where you need it to keep focused on, on the same task and, and you're using their
Chris Sharkey (00:35:27):
Api. Yeah, I'm, I'm actually quite curious about how they're going to do that because one, one problem I've noticed is exactly what you say is you lose a lot of tokens by continuously repeating the conversation. But at the same time, if you want to emit parts of the conversation, you don't know which parts of the conversation to leave out without having an additional l l m call where you get it to say, summarise or just retain the important parts of the conversation in order to reduce the prompt size for subsequent requests, if that makes sense. So having it sort of part of the API would be pretty valuable, I think, rather than having to implement all of that logic yourself just to save money and time and things like that.
Michael Sharkey (00:36:06):
The other, there's so many interesting nuggets I, I'll, I'll, I'll touch on one more and then we can move on, but, uh, this one stuck out to me as well is Sam Altman said plugins don't have product market fit and are probably not coming to the API anytime soon. And I think that was a pretty interesting admission because these apps, you know, a number of weeks ago it was the next iPhone app store. This was like the biggest breakthrough ever. And the reality for most people that I interact with is, you know, they use the web crawl maybe once or twice a week, but that's really it. Um, and I think these plugins have been a, a pretty big disappointment in terms of how they're currently being delivered and used in chat G P T. And what he says is people thought that they wanted their apps inside chat G B T, but what they really want is chat G B T inside their apps.
(00:36:58):
And I think that might be some indication of where their, they're thinking about going in the future. And he also follows up by saying, open AI will avoid, avoid competing with their customers other than chat G P T. So whatever the chat G P T product evolves into, they're going to have that one product, but they're not going to directly compete or at least avoid competing. Uh, and yeah, he said the vision of chat G B T is to be a super smart assistant for work, but there will be a lot of other G P T U uh, G P T use cases that open AI won't touch. So it'll be interesting to see, but I
Chris Sharkey (00:37:38):
Don't know, I genuinely dunno what to make of that. That's, well,
Michael Sharkey (00:37:40):
Is it like Microsoft's co-pilot, like is it just simply another version similar to Bing Chat or, or Bing's, uh, uh, I dunno, Bing's ai, whatever they call it now is just in your browser now or or wherever you need it to be. Is it just a case that there's sort of like an earlier release version of that? I'm not, I'm just not sure how it's gonna be positioned.
Chris Sharkey (00:38:02):
It's a bit concerning for them as a company, because on one hand you've got, okay, my, my first thought there was like, oh, that's okay. Who really cares about, you know, chat G B T? Because most people are just gonna use the AI as the foundation of their apps and they'll make all their money that way. The problem is with the rise of all alternative models, open source, different, different approaches, and particularly people talking about what's the most efficient model. Like, there's a lot of discussion online about, oh, what's the best open source model I can run, and we'll talk about Falcon in a minute, but if I can run something as good as G P T three, not four, but GPT three now on my own hardware, and there isn't a sort of marginal cost to that, it's like a fixed cost, um, to run the server or servers, then I'm gonna use that over paying, you know, these massive amounts you need to pay to open AI for their api. So, you know, essentially at some point they're selling a commodity, maybe it's slow, you know, higher like high octane fuel versus regular unleaded or something, but it, it will be a commodity. And then the ma the, the other thing they've got going for them is chat G P T, which is like just an immensely popular and well-recognized thing around the world that they got out first on, but that's gonna be clone too. And he's like, well where's the mote here? Like, what do they have that's different at some point soon?
Michael Sharkey (00:39:19):
I mean, possibly, but you know, like Facebook's being clone like a hundred times, remember Google Plus and like Facebook was still dominant. Yeah, but
Chris Sharkey (00:39:27):
Facebook, Facebook had all your friends on there, you know, like they had, they had something you that's not replaceable. If everyone's on the platform, you can't really, yeah. But
Michael Sharkey (00:39:35):
It's gonna take a lot for you to now go and instal some other app on your, on your phone. Like if you've got,
Chris Sharkey (00:39:40):
And yet, and yet we see, I mean, I know, I know Bing is is open AI based, but I mean, you see direct operating system integrations of, of these models and things. So, you know, there are other channels for them to get out there. Like, I mean,
Michael Sharkey (00:39:56):
I how much do they care though? Like if everyone's using their APIs and, and,
Chris Sharkey (00:40:01):
But I guess what I'm saying is they won't necessarily have to use their APIs to, to replicate the same
Michael Sharkey (00:40:06):
Functionality. It does seem like this is just like, it's just gonna be an everything and, um, and we're seeing it pop up in literally everything where AI is just infusing itself into every single part of our lives. Um, one other thing, I know I said it was the last thing, I'm a liar, but this one really is super interesting and I think by not covering it, we would be silly. Um, recently many articles have claimed that the age of giant AI models is already over. This wasn't an accurate represent representation of what was meant. Open AI's internal data suggests the scaling laws for model performance continue to hold and making models larger will continue to yield performance. The rate of scaling can't be maintained because OpenAI had made models millions of times bigger in just a few years. And doing that going forward won't be sustainable. This doesn't mean that OpenAI won't continue to try to make models bigger, it just means that they will likely double or triple in size each year rather than increasing by many orders
Chris Sharkey (00:41:04):
Of magnitude, although only double , sorry guys, but
Michael Sharkey (00:41:07):
This is where we see the emergent behaviours coming every time these, these, these models get larger. I'm just not sure how far this goes because it's not like the human brain has to learn on more and more information. We've talked about this before.
Chris Sharkey (00:41:20):
Yeah, yeah. I always, I always like that it's like you don't have to see like 400,000 dogs to recognise a dog.
Michael Sharkey (00:41:28):
Yeah. And like our, our, our brain operates on less energy than a lip pole. Imagine if
Chris Sharkey (00:41:32):
You did imagine what school would be
Michael Sharkey (00:41:34):
Like . They're just, it's like flashcards, just infinite flashcards. But yeah, so it, it'll be interesting to see as they scale these up, there's more emergent behaviours, isn't it? Maybe it, it's why they're not like, I just wonder if they're telling the truth about this idea of they're limited by GPUs. But I
Chris Sharkey (00:41:54):
I, it does seem, it does seem odd. I mean like there are GPUs out there and I know people pre-order them all and things like that, but Nvidia just this week announced new, um, G P U technology with much more memory. Like surely these guys have got both the money and the access to get what they need.
Michael Sharkey (00:42:11):
But going back to the, the latest AI safety statement that everyone signed onto, maybe it's because they've realised LLMs are the key. Like maybe the reality is they know if they just exponentially scale this up or they believe strongly that we'll see more and more emergent behaviours and that's why they're trying to lock down G P T four and inevitably cripple it somewhat. Because ultimately if they can prove control over that, maybe that gives them some feeling that G P T five, g p t six or whatever the next, uh, yeah.
Chris Sharkey (00:42:42):
And, and additionally they, they, they're very well aware that people are doing alignment using G P T four for alternative models. So by making it worse, you make all other models that are using it for alignment worse too. I mean that's, that's extremely cynical, but that could be part of it. But yeah, I agree. Like it's sort of the magician thing, like don't look at what this hand is doing while the other one's doing something else. Like, these guys are not just gonna come out and do everything they've done and be at the forefront of it and go, ah, look, we're just not sure. We don't know what we're gonna do from here. Like, it just doesn't seem consistent with what they were doing in the early days at all.
Michael Sharkey (00:43:21):
What if, what if that paper that Microsoft released, like the sparks of AGI when Microsoft Research first got a hold of G P T four and then obviously it's been crippled down, so
Chris Sharkey (00:43:31):
They're probably wishing they're probably wishing they didn't say that to be honest.
Michael Sharkey (00:43:34):
Yeah, but but what about if the reality is G P T four, the, the reason G P T five doesn't exist is because G P T four was already capable. Like, it, it, you know, it was just wild. Like the Sydney version of it was actually getting crazily close to some sort of, you know, I don't know, it depends on your definition, but some sort of like high level intelligence that scared
Chris Sharkey (00:43:57):
Them. Yeah. It, there is a, there is a definite lack of people showing like bizarre emergent behaviours in the models. Like in the early days it was there, I don't know if everyone just got sick of it or they're just, it, it just can't do it anymore. But you're right, there was a real, there was a real flood of that. And then now there's not much at all on that front, I
Michael Sharkey (00:44:14):
Think because they successfully, uh, have have aligned it in a way that made it boring. And so it just, there's no, there's no spark of, uh, excitement. It's, people have just got on with their daily lives like using it to write some basic Python code , you, you know, like
Chris Sharkey (00:44:32):
Yeah. And prosecute law law cases
Michael Sharkey (00:44:34):
And things like that. Yeah. So we should talk about that because it's, it's truly hilarious and I'm sure many people who listen to this show have heard because they follow ai, uh, what's going on in the world of ai. But for the lulls, we'll talk about it anyway. So the article in the New York Times was, here's what happens when your lawyer uses chat, G B T A lawyer representing a man who sued an airline, relied on artificial artificial intelligence, specifically chat G B T to help prepare a court filing. It did not go well. And for those that aren't aware of what happened, essentially there was a case, um, and the case that was originally filed was a complaint about personal injuries that were sustained aboard a flight from El Salvador to New York back in 2019. The, the, the problem was the airline went into bankruptcy and then it emerged from bankruptcy later on.
(00:45:27):
So they were arguing back and forth whether the airline was still responsible or, or had they discharged their responsibilities during the, the bankruptcy. But what happened is when the lawyers, uh, filed, they used chat C B T to find, uh, you know, referenceable cases and they cited this case, um, by someone versus China Southern Airlines. And it had a pretty compelling convincing sort of argument. And then the, uh, the judge and, you know, the, the district judge in, in the us uh, was like, these cases don't exist. And instead, like, instead of, of saying, oh, you know, we used, uh, chat G B T and like, sorry, they doubled down and then they went back to chat G B T to try and like find the reference to prove it. And in the filings they screenshotted actual chat. So instead of cutting and pasting it into a a Word doc, they, it's got the down arrow that helps you go down on the screen in the actual screenshots. Simon Willison, I'll, I'll link to this article. He, he covers it in, in fantastic detail, the series of events here. And you can see this screenshot here if you're watching, uh, online with the down arrow, uh, screenshotted over the chat G B T interface. And I honestly,
Chris Sharkey (00:46:51):
Yeah, I wonder if the judge was, was naturally sceptical. He's like, this, this doesn't sound real and then researched it or if he, he's just a really, really good judge and he goes and reads everything that he's presented with. Like either way it's a, a testament to being a great judge, but I'm just curious like how it actually played out.
Michael Sharkey (00:47:10):
Well, what, what's so interesting is on Simon Will's blog post at the bottom, he has an update to the article and it, it says, it turns out this may not be an isolated incident. Um, I got a message from Professor Dennis Crouch in Missouri, uh, in response to my posting. Um, and it said, uh, two attorneys, attorneys at my firm had opposing counsel filed chat G B T briefs with fake cases this pass , both
Chris Sharkey (00:47:40):
Of them, both of them,
Michael Sharkey (00:47:42):
They both filed. So like clearly all the lawyers now are using this, but it just shows the, the problem with the hallucinations and this issue with it truly believing, like when Simon Wilson, he, he actually sort of reran the test and he asked it what's at source? And it's like, I apologise for the confusion. Earlier upon double checking, I found that the case versus China Southern Airlines does indeed exist and can be found on legal research databases such as Wesler and Lexus Nexus. So it's so certain that these fake cases are true. Uh, that, that, you know, like it just makes me worry, like, is the future just not gonna be grounded in reality if we don't solve these problems? This is the
Chris Sharkey (00:48:23):
Thing, there's, there's been discussion about this as well about, you know, the integrity of journals, medical and otherwise publishing papers that are just total bs Like people were doing this for years before the GPTs where they'd use machine learning to generate fake full fake papers, submit them, get them accepted into journals and go, huh, we can tell your review process isn't real because there's no way this should be allowed in. But the issue is because it can be done so cheaply now, you could be flooding journals with fake papers and then once they get published, you can use them as references to build on other fake papers. So the potential for that, I mean I, you know, I I I can't really fathom the, the benefits of doing that. I'm sure there are some, but you know, the potential, you know, the, the, the onus on us to check the veracity of information is only gonna increase day by day.
Michael Sharkey (00:49:16):
Uh, yeah, I think this is obviously why people are so worried about misinformation because you, like, you can create so many sources that it's just, then the new models get trained on those fake sources and it's just this sort of self like, you know, this like feedback loop of bullshit
Chris Sharkey (00:49:33):
that it's like, it's like the reviews on our podcast, I read them and I'm like, oh, that's so lovely of them to say. And then at the end they're like, this was written by AI
Michael Sharkey (00:49:41):
. Yeah. Or like, or like claiming that we are ai, maybe we are You don't know
Chris Sharkey (00:49:46):
That. Yeah, yeah, exactly.
Michael Sharkey (00:49:47):
Uh, but I, Simon Wilson proposed a solution to stop these lawyers being such stupid morons and one, one of his proposals is, um, that it literally keeps reminding them how stupid they are . So it it, like he asked that in this example, write some tweets based on what's trending on Pinterest. And it, it says, this Chatt b T model does not have access to the internet and it's training data is cut off September, blah, blah, blah. Um, for a legal brief, give me us legal cases, I can cite blah, blah, blah. And then he's proposing a UI that has an alert before it responds, saying chat G B T should not be relied on for legal research of this nature because it is very likely to invent realistic cases that do not actually exist. And what he's saying is that a lot of people, and I know people have actually commented on our videos about this, how, you know, it warns you of all these disclaimers at the bottom that it's gonna use your data and, and train on it and do all this evil stuff and it's not true and blah blah blah, but no one reads that.
(00:50:50):
And so what he's saying is just throw it in their face and say like, every time they try and do something like this, just re remind them that, hey, this is just not what this thing's the best at. But does it make you think that, um, there's a lot of opportunities for startups potentially here to connect the dots? Like give these lawyers the best of as
Chris Sharkey (00:51:09):
Well? My first thought is like a few minutes ago we were talking about a million token prompt size. It's like, just put all the case law in, in the context and then ask it the question, um, you know, and actually have the real data as it's sourced rather than, you know, its sort of hallucinations based on its training data. So Yeah, I I do and I think that's probably what we'll see is specialised tools that have the actual raw information. I think this is what the plugins was supposed to be, which is why I'm surprised that they're giving up on it so quickly because, you know, you could have a legal database plugin, it makes a whole lot of sense, right? Like have a, even now using Lang chain, you could do it. And um, you
Michael Sharkey (00:51:47):
Know, I think it seems like what Sam Altman's saying, and, and I this is what I feel like as well is there's going to be specialist tools that you go to, like, you know, there's like a legal agent and that's who you interact with to, to, you know, build like work in, in law firms or, or with judges or whatever, and a specialist health one. But it does worry you like hallucinating in say like scan, like medical imaging and finding a tumour and then they operate and it's like, oh, sorry, the, I was just hallucinated .
Chris Sharkey (00:52:20):
Yeah, that's a good point. But I think that's the other thing. I, I think there's also a multi-agent world coming where, you know, you have one agent that might propose something and then, you know, you, you've got a sort of quorum of agents that check it from different perspectives and one's like, okay, I've gone and checked the legal databases and this is just not real. You know, like, it, it seems fairly trivial that you would have multiple specialist agents to play different roles in doing something like solving tasks like that where, you know, you have your sanity check one, you have your censorship one, you have all your different agents working together towards a common goal to make sure you, you cross, you know, you don't make a fool of yourself. I don't think that's gonna stop stupid people misusing these tools. I, I think there's nothing you can do about that. But I do think if you want to use them for legitimate purposes, such as detecting tumours, I can imagine a scenario in which you have multiple agents looking at things from completely different source of information, trying to verify them together to come to a consensus and then say to you, Hey gay, this is our best, best thinking here. And, and here's why.
Michael Sharkey (00:53:22):
So open AI releases paper in the week improving mathematical reasoning with process supervision. Jim fan, I'll link to this over on Twitter, uh, said, reading open AI's latest paper. Let's verify step by step. The idea is so simple that it fits in a single tweet for challenging step by step problems give reward at each step instead of a single reward at the end. Basically, dense reward signal is greater than sparse. The process reward model or p r m as they're calling it, is able to pick solutions for the difficult math benchmark much better than the outcome reward model. The obvious next step is fine tuning G B T four with p r m, which is process reward model, which the paper hasn't done yet. So essentially what it's doing is right now, uh, there's input output. They reward or punish . I love, I love say that
Chris Sharkey (00:54:13):
I love
Michael Sharkey (00:54:13):
The idea, um, the AI based on what it does, and that's how it learns. And what they're saying now is they're going to like, or, or that they're proposing to reward it through almost its chain of thought or reasoning, sort of how they think through and solve how it thinks through and solves the problem. And that might lead to better outcomes where instead of being really bad at maths, it's now better. And potentially that could stop these hallucinations because it's thinking through, well actually this isn't, isn't real in context. I'm not it that's very early days, but that could be one
Chris Sharkey (00:54:45):
Solution. It's really funny because I actually read something about this on, on personal goal setting and that it's actually not advisable to give yourself a reward when you reach a goal. You should instead reward yourself for sticking with the process that leads to the goal. Like you should actually, if you enjoy, say it's fitness, like if you actually enjoy going for a run, for example, the enjoyment you get from that should be its own reward. Like doing the activity should be the reward, not the goal that you get as a result of running every day for a year or whatever it is. And it's funny because this is that same concept playing out with training and artificial intelligence, something that, you know, according to what I read at least and somewhat my own experience, um, works on humans too.
Michael Sharkey (00:55:30):
Yeah, that's a phenomenal point. Like, it, it's why diets never work out because people hate eating the food, but they, once they get to their target weight, they stop eating the food. But if the food you love,
Chris Sharkey (00:55:40):
That's right. Yeah, you're a hundred percent right. Like if you're actually not enjoying the activities leading to where you get to, then it's completely unsustainable and probably, and, and therefore not going to give the results you want in the end. Whereas if you can enjoy the process, then, then it'll, the, you know, it's that whole, the score takes care of itself concept.
Michael Sharkey (00:55:59):
So some other news that we thought was really interesting this week was Japan goes all in copyright doesn't apply to AI training in a surprising move. Japan's government recently reaffirmed that it will not enforce copyrights on data used in AI training. Yeah,
Chris Sharkey (00:56:15):
This is
Michael Sharkey (00:56:15):
Big.
Chris Sharkey (00:56:16):
It really is big. And, um, because, you know, there's a lot of people like artists, musicians and people who are understandably upset that their create creative works are, are sort of having derivative works made for them for free based on models because it's, it's being used to train them. And, uh, you know, I can see a strong argument for that. I, I dunno how I'd, I've said this before, but I don't know how I'd feel about that. Um, and yet really Japan's made this decision specifically for competitiveness. They're saying if well, if we, if we honour the copyright of people, then we'll fall behind the world in terms of models.
Michael Sharkey (00:56:55):
It speaks to this feeling like a nuclear arms race in a strange way, doesn't it? Like the fact that they're like, well this is a big blocker to us being, you know, somewhat a leader or in the lead mm-hmm. and therefore let's just get rid of the blocker. Who cares? This is too important to not get right.
Chris Sharkey (00:57:13):
Well, yeah, and it's, it's sort of this, we talked about regulation before. The countries that choose to regulate themselves are unnecessarily and technically going to slow themselves down in comparison to the countries who just don't care about things like that. You know, it's similar to how countries like Australia are, are not only not building new coal power plants, but shutting them down while countries like China and India are building like 17 a day or something crazy like that. And it's like, well, okay, fair enough. You know, you've got a moral or ethical thing that's that's causing you to, to not do that, but these guys are doing it and, and they're gonna get ahead because of it. So it is very interesting that they've made this call.
Michael Sharkey (00:57:52):
So, uh, the United Arab Emirates, so we often talk about US based companies here. Uh, were mentioning Japan trying to push ahead here, but this is the release in the United Arab Emirates of Falcon, uh, as a model. Uh, so previously this wasn't royalty free, I think you had to send back to them, uh, a percentage.
Chris Sharkey (00:58:16):
It's 10, 10% over any money you earn more than a million dollars.
Michael Sharkey (00:58:20):
Yeah. But they've made it fully royalty free. Now, this model, can you tell us a little bit about this model and what you make of that?
Chris Sharkey (00:58:27):
Well, I haven't used it, but I have read anecdotal stories of people running them themselves on a 100. So they're ch they're chips that cost between I think, 15,000 and 30,000 US dollars to get one of these GPUs, which sounds like a lot, but if it's the foundation of your business, then it's not unreasonable at all. Um, and so they're running it there and reporting getting results similar to G P T three, so, you know, not quite at the G P T four level, but G P T three is enough to run a lot of legitimate applications. And the issue so far has been a lot of the open source models have licences, which mean that you either can't use them commercially or, or if they, you can, they're not good enough to actually do the job. So this is the first model where realistically you could probably use it at the foundation of a business and have full control over it yourself and run it without any impediments on what you can use it for. So it's very significant from that respect. It hasn't been around long enough to get a full assessment of, of where it, where it fits in the landscape, but it's a very significant thing. Even just the fact that exists is, is exciting.
Michael Sharkey (00:59:32):
It also, I think it's worth pointing out on, on the model benchmarkings, uh, on hugging face, it's ranked number one out of the, uh, the open source models. So it'd been laa Yeah, that, that
Chris Sharkey (00:59:45):
Wouldn't surprise me. And then we talked a lot on this episode about alignment, and this is an opportunity to do your own alignment with a model like that. So it, it really is a, um, it's something that's been llama, uh, which is really popular and yeah, it's, it's, it's a true exciting step. And another interesting thing around competitiveness, because in one respect, you know, one country saying, okay, we're gonna, we're gonna embrace this technology internally, and another one being like, Hey, here's something that's better than almost all of them for free. You know, that sort of changes what other countries are actually able to do.
Michael Sharkey (01:00:17):
Yeah. And I think it shows now, it is truly a, an international, I don't wanna call it an arms race, but it's definitely, it, it's, this is just not US companies. I think they're obviously well in the lead, but uh, yeah, it, it, everyone else is gonna have a crack at this now. Yeah.
Chris Sharkey (01:00:34):
And when you think about what everybody does, which is the terms of what jobs does it replace, what industries does it disrupt? You know, if you are, if you can play a significant role in that as a government or country, then it, it could have profound impacts on that country's economy. And that Japan article actually says that, you know, Japan ex hasn't exactly been doing great economically lately, whereas something like this, it might be worth the trade off to piss off a few artists if it can fix the country struggling GB G D P.
Michael Sharkey (01:01:04):
So there was a tweet this week by Ethan Molik who says, I got beta access to the version of Google Docs with AI built in. It is clear that tonnes of work that could be meaningful when carefully done by humans is now going to be done mostly by, by ai. And the example he gives, he goes into Google Docs and says, help me write a good performance review for Bob Smith. He is a business analyst for my division. Talk about how he, he is accomplishing his collaboration goals, but still needs to work on his time management skills. And so what it does is it spits out this like, perfectly written performance review. It's great. Um, you would totally use it. Like I don't, I've read it and there's nothing wrong with it. And you can see how this is just gonna become more and more mainstream. We've mentioned it before about people just emailing with their ais back and forth, but now, like, think about when you get a school report or an essay critiqued or like, are people not going to value written language anymore? Like, will there be no value in it? Like, you'll get the written language and you'll be like, I'm not gonna even read that because I know an AI wrote it, so there's no value. Yeah,
Chris Sharkey (01:02:14):
Just the, just the volume of text people are gonna be able to produce in any given scenario. You just bamboozle them with large amounts of information. It
Michael Sharkey (01:02:22):
Makes you worry though, like I use writing to organise my thoughts and think out ideas, but what's the next generation gonna do if they don't value writing? They're like, oh, the ai, I'll just do that. Like maybe written language actually starts to decline because people are like, why bother when the AI can do it better?
Chris Sharkey (01:02:38):
Yeah. It's an interesting thought and I wonder if that is, um, you know, for some people writing ears a really good way to clarify their thinking. I know I always have a notebook and, and write things down, but yeah, like, do we just transcend the need for written language? Does it become far, far less significant, um, than, than before? I, I really don't know the answer to that, what that's gonna mean, but you're right, certainly it's going to have a healthy level of scepticism in everything you read, that's for sure.
Michael Sharkey (01:03:06):
Yeah. And you know that feeling like if you get like a review, like well at list is, right, they're reviews who ai. But like, when you get those reviews, right, like we're human, like we read those reviews and if they say nice things, it makes us feel good. And like if you're a, a school student and you get a report and it relates to you, it makes you feel good. But yeah, once you know that that's essentially all fake, like is there that mechanism where you're like, well, this isn't real, it doesn't make me feel anything, or does it evolve where the written language is like we just start talking in dot points because Yeah,
Chris Sharkey (01:03:41):
You just, you just like, yeah, like the absolute minimum amount of information, like expand this into an AI written thing if you want, but otherwise, here's the important points.
Michael Sharkey (01:03:50):
Yeah, like, I, I just don't know if longform is just going to like die. Maybe our, this is why our podcast might be doomed. Like no one, just the, the, uh, who knows what will come of this, but it, I, I'm curious to see what happens to the written language. And I'm curious from, from listeners, uh, you know, if you want to comment on this, like, do you, are you valuing text less? Are you using this stuff to write your own emails? I, I would, I would really love to know, uh, what your opinion is on this. Uh, alright, well we might leave it there. This, this has been just such a big week, it's been overwhelming. There was probably a hundred more things we could cover today. Um, some people have mentioned that they'd like us to do two recordings of the show in the reviews per week. We're, we're, we're very curious to hear what you think. Would you lo would you get sick of us if you, if you heard from us twice a week, or, uh, would you enjoy it? Please let us know. We'd love, love the feedback. Again, if you enjoy the podcast, please do consider leaving a review wherever you get your podcast. And if you're watching on YouTube, we'd like the likes, comments and, uh, and we love interacting with you guys. So thanks for participating and, uh, we'll see you next week.