Chris Sharkey (00:00:00):
Whether or not it's scientific, whether or not it's more accurate than doctors, people are going to do it. The technology's there and people are going to do it. I mean, people were already, doctors already got pissed off when people would look up their symptoms on Google and diagnose themselves with all sorts of stuff, right? Like that was already happening. There is absolutely no chance that people aren't going to use the most generic models to diagnose their medical conditions and act on it. It's going to happen. It's happening now.
Michael Sharkey (00:00:29):
Krista, this week, one paper that caught our attention was meta G p t meta programming for multi-agent collaborative framework. Now, not to be confused with meta, can you explain, uh, what this paper's about and what excites you about it?
Chris Sharkey (00:00:44):
Yeah, so it's something we've spoken about before, this sort of multi-agent mode where you've got different agents within a little world that all have a different role. And in this specific case, what they've done is isolated the idea of a, a software project, specifically building simple video games. And they've mapped out all of the different roles like c t o, project manager, QA engineer, and uh, a, you know, architect and given them each very specific roles of what they're meant to do in the job. They follow a water form methodology and, and work through the project the way a real software team would.
Michael Sharkey (00:01:19):
Yeah, I found it fascinating because we've talked about this on the show before, but we've also, behind the scenes talked about it quite a lot, that eventually you'll have a series of agents working together on a combined goal. And I know we've talked a lot on the show about our horse racing AI project, which I assure you we're still working on and plan to ship to you all eventually. But the idea that if you could have like an accountant for your horse race betting that keeps track of the winnings, and then you could have a a, a series of agents doing analysis that can disagree for different, different
Chris Sharkey (00:01:54):
Viewpoints. Yeah, yeah. And like have, you know, one side of it that's like, let's swing for the fences and just bet everything on this race. And then another one that's like, oh no, this, this race is more risky than the others. And, you know, have different perspectives that it's held to that, that come to a better consensus. It's sort of, you know, we talk in real life that often teamwork can actually give a better result than an individual working on something. And it's the same thing applied here, just automatically.
Michael Sharkey (00:02:19):
Yeah. Or like these different specialists or experts, and we talked, I think it was last week or the week before, about having these specialist models for each, uh, each role on that team potentially, but that's not what they've done in this paper. They've just segmented the roles, right?
Chris Sharkey (00:02:35):
Yeah, that's right. So this is all using G P T four. I noticed I looked into the code and they have support for Claude in there as well, but in the paper they specifically referenced using G P T four. I guess they tried a bit of both, but yeah, they, they've just given very, very specific roles to each of the, the people. So like for example, the product manager is the one that's creating like the sort of general, uh, decisions about what they're gonna work on, what that will entail. Then they've got the architect that's deciding, okay, we're gonna use PI game as the library and we're going to use S Q L light as the database and things like that. Then they map out what files need to be created. The programmer then writes the, the code, and then the QA engineer analyses that and does code review. So they've got the whole full process in there that a normal software development team would follow, or one of the ways. And it gives excellent output. I mean, they're outputting real working games that, that are perfect along with diagrams and workflow things and documentation and everything.
Michael Sharkey (00:03:35):
Good developers and software teams be worried,
Chris Sharkey (00:03:38):
I mean, regarding this, I think probably yes, it's a really, really high quality project and the output is undeniably good. And what I also found coming out of it was several things I must admit I hadn't really thought through. Like we've obviously had that discussion before about the domain experts, and I absolutely believe in that sp like multi-model where you've got different models making different parts of the decision. And, and we'll definitely apply that in our horse racing one because I think that's probably its weakness right now is not having that. But the things that they found were by isolating the roles, they would get predictable and repeatable outputs. So they constrained them to output formats quite tightly. So it's like, you can't go outside your role, you're only to output this, that's all your concern is. And that gave them much better results. Um, and it also forced them to draw on their, you know, domain expertise. Like I know it's all working from the one model, but by that sort of thought guidance with G P T four in this case, it forced them to think in terms of that role. And that was very, very effective in this project.
Michael Sharkey (00:04:45):
Can you see in future having a series of roles, like almost like a workforce of agents and then you spin up a project and then that project decides, okay, which agents do I go and pick? It's intelligent enough to identify which specialist agents it needs, or do you think this is gonna be at a project level? Like I need to go and define the roles? Like who is defining the roles?
Chris Sharkey (00:05:09):
Well, I think in this case it was sort of hard coded, but I think in your example is probably the most logical conclusion of where it gets to where, you know, it, it, you give it a general task, like a complex task that needs to be solved and it picks its team and it says, you know, I'm gonna use this person for this, this one for this, because we know they're good. And the other thing we haven't spoke about is the idea of long-term memory. And, uh, as the agents become more experienced, they'll become better at their individualised tasks and become even more domain experts. And on top of that, the thing we've spoken about, which is them having an actual highly trained model for their specific skillset, then you're more likely to pick that one for that task. I really think this is the future of how we get, uh, really complex task solving by, or AI groups or whatever we wanna
Michael Sharkey (00:05:59):
Call it. So could you imagine the, so the architect that they call the architect in this case, I think is like the, the C T O in this paper, right? Mm-hmm. is that, is that agent with its long-term memory going, you know, I really hate Jeff, the engineer, like we've gotta get rid of him and find a new persona on the team or, or giving it a performance review?
Chris Sharkey (00:06:17):
Yeah, I'd say so. I mean, the whole part of, of this project and why it was successful is that each step, when it reached each new role, uh, in, in the process, they would actually evaluate to say, are there any holes in this plan? Is there any problems with this code? Are there errors in it? And give feedback and actually go back to the previous step to iterate to get it done. So they had this idea of concerns where a specific role sort of subscribes to a message, and when that message occurs, they can then take action based on that. So it really was a sort of collaborative thing with an iterative process. So yeah, I could absolutely see a supervisory agent ditching ineffective sub-agents or whatever from the project if they're not working. And maybe trying again if, um, if it fails.
Michael Sharkey (00:07:06):
Yeah, it'll be interesting if this is the approach that everyone now takes and how revealing this will be in the applications themselves. Like if, if you're using say, chat g p T to solve a problem potentially, is it going to just do this in the background and and spin up like
Chris Sharkey (00:07:24):
Roles? Yeah. Like do they, do they make it transparent? And that's all happening behind the scenes, which I strongly suspect it's sort of is to some degree with them, with their censorship and filtering already? Or is it out in the open where you realise you're working with a team of agents?
Michael Sharkey (00:07:39):
Yeah, so some of the projects that they worked on were things like writing a snake game, write a brick, a breaker, game design, a blog management system. I mean, these are tasks that, I mean the, the games are quite basic, but you can imagine them getting more and more advanced and just distributing hundreds of games. I think they said the fee, let me scroll down and find it. The fee to like
Chris Sharkey (00:08:02):
$3 or something.
Michael Sharkey (00:08:03):
Yeah, three bucks to produce some of these games. And this is very early on, so you might say now, well how is, does this really matter? Does this threaten any anything? But you can see a series of agents being composed into a workforce or what we see as a company today, really being able to replace a lot of things like potentially even a medical team. It's like you, you give the the data to, to a medical team now, and there's varied opinions. You could eventually have very specialist AI agents that, that forms a team in some sort of, of consensus of opinion around medical. That's
Chris Sharkey (00:08:39):
Right. And um, we are gonna talk about it later, but there was a paper during the week called Med Flamingo where they talk about improving the analysis of, um, reading medical images and, and diagnoses. Um, and one of the things they talked about is the problem of the age at the sort of LLMs mag when, when combined together magnifying hallucinations. So if there's already a hallucination, it, it's the predicate for another agent to run and it uses that as its assumptions, then the hallucinations get worse because it just keeps escalating. But this meta G P T concept, um, and I wouldn't say these guys invented it, they've just done an implementation of it, but this concept really sort of stops that in a way because you've got agents that specific goal is to sort of review the output of other agents and decide if they're in line with what needs for is needed for the project. So it's sort of dampens that effect of any hallucinations and cuts them off, uh, during the, the project rather than it, it just being one agent that can have just be way off and, and get everything completely wrong.
Michael Sharkey (00:09:43):
It'd be interesting to see how this scales out to more complex ideas. So instead of building like a Flappy bird clone or a basic starter project out of some, you know, web dev book, can it go and like set up Stripe billing? Can it go and like really, you know, do more of these elements and make decisions or think about scaling over time based on it acquiring its first users? The answer's probably yes. I
Chris Sharkey (00:10:08):
Think so, especially because they added that budget element into it where they were, they were sort of making the agents aware of their budget and how much they'd spent. They weren't just outputting it for their benefit. They were actually telling the agents within this little world how much they're spending and what their budget was. So they had to actually be careful with how many prompts they ran. So I could see it if it had a real world budget as well, and the ability to browse the web and put their credit card in and subscribe to services that they could actually get through a lot of those elements of the project as well. Definitely they could, you
Michael Sharkey (00:10:39):
Can also see the fact that this could get to a point where you've got your workforce of agents and you ask them to, you know, build me the blog management system, and you're like, oh, can you tweak it to do this? Can you tweak it to do that all through voice, potentially. And then the, the project manager's like, okay, boss. And then they, they a few seconds later for another dollar have another version of it for you. Yeah,
Chris Sharkey (00:11:03):
That's right. Like redo the entire project because I, I wanna make a change, but yeah, and you can imagine these agents like renting themselves out on fiverr.com and other, other systems where they've got these abilities or like a whole team of people, whole team of agents who can do projects for you, modify your WordPress to do something or whatever it is. This, this demonstrates that that kind of technology is really only, it's here, it's just a matter of someone actually implementing it as a business and, and making it available to people.
Michael Sharkey (00:11:32):
Yeah. Do you see this in the future? I'm not sure of my position on it right now, but does it just replace all software because you just, anytime you want software in your business or need something, do you just go to a service like this and, and go back and forth with it and it just entirely devalues basic software development? Or
Chris Sharkey (00:11:55):
I think as a, I think as a starting point, why wouldn't you try? Like if you can get up to a certain point with this, uh, then you might, I think it depends on the nature of the software project, obviously, because if you're just doing something generic yeah, like a blog management system that's been done a million times before and it knows what to do, then great. Some things that are really novel, it may not have the, the creative ability to actually do, and you might need more human interactive steps. So one of the, one of the concepts I thought that might enhance this project, for example, is a sort of human consultation break point where it's like, okay, we've figured out all these details team, but we're just not sure what to do about this bit. And it could sort of come back to the person operating the system and say, well, what should we do here? Like, how do we proceed on this front? Um, rather than the AI making absolutely every decision and you end up with something, uh, that that's not what you wanted. And, uh, I think that would happen in the cases of projects that are more complex than building a game that they already know what that game is.
Michael Sharkey (00:12:57):
Yeah, I see. Yeah, like filling in, I think it makes the case of, of upskilling organisations and, and, and people that work in them though, I don't know if it's like full-blown job replacement or is it just greater efficiency as a whole?
Chris Sharkey (00:13:13):
Well, I think the, the thing that I found striking about it, we've talked about code generation on previous episodes and how it's getting worse in chat G P T and how, um, you know that as you sort of need to be a developer to know, hey, this is way off what I actually want. But because they've gone one step further and had the code review step in there and the unit tests and the sort of diagrams and, and knowledge of what this whole programme is trying to accomplish, it seems to overcome the limitations of the code generation because it can iterate and improve it up to the point where it wants to get to. So I think it is a sort of next level in terms of code generation over just getting snippets like this is building a whole project, which is totally different to all the other code generation we've spoken about so far.
Michael Sharkey (00:13:59):
So you could imagine outside of code generation, you could spin really, you could spin this up and create a marketing team with a product market or a blogger and you go to it and you say, how can we market and position this particular product?
Chris Sharkey (00:14:14):
Yeah, why not? I mean, doing all that stuff is, is bread and butter for the l l m and you could have, yeah, one's doing your Google ads, one's doing your Facebook posts, and now with some of the video and audio generation technology, it's perfectly plausible there. They'll be able to do TikTok videos and other things as well. So I think an automated marketing team on top of the software generation is perfectly plausible and likely.
Michael Sharkey (00:14:38):
Yeah. The question is, do you need an automated HR team to keep the AI agents in line ,
Chris Sharkey (00:14:44):
You know, when they want days off or, or something like that? They,
Michael Sharkey (00:14:47):
Yeah. Do they, like, do they become rebellious or like, you know, they want promotion or something like that. They
Chris Sharkey (00:14:52):
All work from home or they want soft drinks in the fridge or something. .
Michael Sharkey (00:14:55):
Yeah, . It'll be funny if they take on those personas. So that, you mentioned Med Flamingo there, there has been a real explosion this week and every single week we've been doing this show in New LLMs, but it seems like the freight train's just speeding up at the moment. We hear about new models or refinements of models being released on hugging face or in various other places. And there was a tweet from Robert Scoville I linked to a couple of days ago saying, the flood of LLMs has started. How does anyone in business and not an AI expert keep up, listen to this show? Uh, how , how do they figure out what is good about each of their business? Uh, uh, how do they figure out what is good for their business? I'm watching everything in AI and these kind of announcements really suck. They don't tell us anything about why they're special or people really need, uh, all these people in AI need to really step up their marketing game. Do you find that as well, like, I feel like constantly flooded with these new LLMs and not really know why they exist or why they were built?
Chris Sharkey (00:15:59):
I definitely do and and I'm saying this as someone who tries most of them just because of the nature of doing this podcast. I want to know what I'm talking about when I try the, that's why I try them. But, and yes, you, as you've pointed out, you can try them on hugging face. You don't necessarily have to spin them up and run them yourself, but at the same time I would say to him, well, what do you want to use it for? I mean, you don't need to be across every single tool that comes out in a particular market to be able to accomplish what your goals are. I don't think the fact that there's a proliferation of them is surprising given that we've got open source models that can be easily modified and trained and and aligned. Uh, so it makes sense. There'll be a lot and each one will have its advantages and disadvantages, but knowing what they all are isn't really required.
(00:16:45):
You just need to know what, uh, you would want from a model and find the one closest to what you're trying to do. So for example, the one I think that came out, uh, this week that Nuka one, um, is a smaller model that that is, has been aligned on 15,000, um, uh, what they called it share, G P t I think where people share their G P T chat. So sort of aligned along, it's sort of like a meta alignment. We keep using the word meta but aligned based on the output of chat G P T. And so that would probably be used for smaller applications where you want the chat interface, excuse me, but there's other cases where if you're doing a classifier or if you're doing something for content generation where the chat model probably doesn't make sense for you, so you're really trying to find a model that's the balance between the cost where you can realistically run it and how aligned it is to what you're actually trying to do, rather than needing to know every single one of them for some reason. Like, I don't think you need to.
Michael Sharkey (00:17:44):
Yeah, there, there's definitely a lot of confusion. I think he has a point around like just some of the names and terminology they're using for particular models and the fact that you don't really, it is hard to tell which one to use. And then will that be superseded a couple of weeks later? Uh, I guess hugging face ranking these models and stars on, on GitHub and places like that are, are definitely helpful to developers, but I still find it hard to really push the capabilities of these models because for example, this week we saw, uh, the, I think it's l lava, it's now got, uh, like 4,000 stars on GitHub or something like that. And that's actually, um, um, a multimodal large language model with vision assistance. So similar to GPT four vision, except, you know, you can ask it questions about an image. We've played around with it, it's okay. Yeah. But you, you, I don't know, I, I get the feeling that it would be good if there was a marketplace for the, the best models, even if you had to pay, um, for some of these specially trained models. And, and we've talked about this previously where you could just go and say, I want the best current vision assistant model, and then that's updated through an A p I, uh, all the time. I think I'm just describing open AI here.
Chris Sharkey (00:19:07):
Yeah, but I think the thing is that these general models, I tend to have the exact same experience. You go try them out, they give fairly okay results, like decent, but nothing like that blows you away. Like it's so much better. I think what's exciting about the open source models is their ability to be fine tuned and the fact that you can control 'em with no censorship, but in, in most cases and, um, and then cater them to specific tasks. I think as general models, you're just not going to get something that's better than than G P T four, for example, because the people training these just don't have the hardware, uh, to do it. So I don't think you're ever gonna get one. You go, oh, actually it's just as good as G B T four, I'll use this instead. That you're not gonna have that experience. It's more just the fact that other things exist that you can get in there and use if you've got a use case that that works for them. I think it's really now about what problem are you trying to solve and you've then gotta go model shopping to work out what's gonna get you the best, uh, value for money and, and best results depending on what you're trying to optimise for.
Michael Sharkey (00:20:10):
So you did mention earlier Med Flamingo. This is a multimodal fuse shot learner, uh, specialised in the medical domain. What's interesting about it is it's actually trained, if I bring up the paper here on both images, so medical images and, and textbooks. So it's got both the, the visual reference, uh, along with all of the textbook content that your doctor or your specialist would actually go through and they're able to show it images describe the patient just like someone would come into a, you know, a, a surgery and then ask a question like, what's the likely diagnosis? And it's incredibly good med flamingo at getting this right over baseline, the baseline model that they used.
Chris Sharkey (00:20:59):
Yeah. And they talk about in the paper that one of the hardest things in the medical space is getting the data for the training examples, which is why a multi-shot approach, which is where you give it say three or four examples, and then it can learn just from those small amount of examples is so important because they're just not able to get vast volumes of data required to fine tune it to specific problems. So if it can take two or three examples and give results that doctors are fairly happy with, which it is what it's doing, then that's a way they can get it into practical use sooner. Um, at least as an assistant to the doctor, um, or a second opinion or whatever you want to call it. Um, without having to access vast amounts of private medical information, that's gonna be hard to come by.
Michael Sharkey (00:21:46):
You can see why this will work well eventually in that if you give these models all of this data, all of this image recognition, obviously they can consistently take in much more than any human could, no matter how smart your, your, your doctor is in terms of diagnosis. And then also just understanding, you know, what the best treatments might be at any given time. It's interesting, someone close to us, uh, had a medical letter from their doctor that it was really hard to understand, uh, last weekend. And I, I took a photo of it on my phone and then I was able to use that apple where you can just like cut and paste the text from the image itself. And I added that into, uh, chat G B T I think, and it gave me such a clear summary. I'm like, explain it to me like I'm 12, which is always a great prompt to use. And the explanation was just so clear, it gave other references and then I asked it more questions about that diagnosis, like, what could be the potential cause or treatment and, and various other questions. And a week later, one of the, the main things that suggested the medical team in charge of this person suggested the exact same thing, . So I'm not saying it's perfect, and it obviously that's just pasting it into a standard model, but it helped explain the problem better. Yeah,
Chris Sharkey (00:23:09):
And I think it's a, it's another, it's another reference point. It's a sanity check. You know, you don't want to, when you've got a, a life-threatening medical condition or any medical condition, you don't want to go to a single doctor and place 100% faith in what they say given the absolute extensive amount. And I forget the word, there's some cool words starting with eye that when it's, um, uh, medical doctor mistake, but basically if you go to a hospital, there's something like a 40% chance that a doctor's gonna make a mistake that affects you in adverse ways, right? Like it's really high chance. And a lot of those are around prescribing the wrong drugs and, and things like that, or adverse reactions to medicine. So having a second opinion regardless of whether which one you trust more just gives you that moment of pause to say, Hey, is this, is this right?
(00:24:03):
Or is there an alternative here? Because one of the things that, especially in America, they're not allowed to do, the doctors aren't allowed to give you natural treatments for things. They're only really allowed to subscribe, uh, prescribe you stuff from their, their playbook. Whereas I can see AI models, especially the uncensored ones, going, Hey, did you just think about changing your diet here that might actually help you, for example, and giving you just an alternative perspective so you at least know to get a second opinion or something like that. It's not just this static thing where you, you have to blindly follow it. You've got options. Yeah,
Michael Sharkey (00:24:37):
It was interesting on the medical front as well, there was a paper, I had it up on the screen a little earlier, Ethan Molik tweeted about it In this study, AI was more accurate than two thirds of radiologists. Yet when radiologists had AI help their diagnosis, uh, their diagnosis did not improve why humans ignored the AI's advice when it conflicted with their views. A big barrier to future human AI collaboration.
Chris Sharkey (00:25:06):
Yeah, we'll just take the human outta the equation and solves that problem. But yeah, I can see that, I can see it, it happening for that reason. But that's partly why you want a sort of caucus of experts, right? Because an expert, you know, you kind of gotta, you gotta say, well, if I'm trained as a radiologist and that's my level of expertise, then I should really act in line with my training. Just because there's some tool out there doesn't mean I'm just going to defer all of my decisions to it. So I kind of get where they're coming from on doing that. But as the person who's being radiolog, I would like to get both opinions and make my own assessment as to which one I'm going to listen to.
Michael Sharkey (00:25:45):
Yeah. What I found interesting is, and I'm not sure I agree with the conclusion, but in the conclusion abstract, it says, our results demonstrate that unless the document of mistakes can be corrected, the optimal solution involves assigning cases either to humans or to ai, but rarely to a human assisted by ai.
Chris Sharkey (00:26:03):
I see
Michael Sharkey (00:26:04):
Sort of saying that, you know, these people are just incapable of working collaboratively with the ai, uh, without I guess thinking it, I mean
Chris Sharkey (00:26:14):
It's a, it's a, it's an existential threat. It's like, you know, their jobs will go away because the AI will inevitably be better at interpreting it sounds like it already is, but it will at some point be consistently better to the point where the radiologist is just no longer required, I would assume.
Michael Sharkey (00:26:31):
I think a big part of it is that resistance, and I can totally see it. I mean, at the, the start of this episode, we have talked about a paper that outlines building essentially a software development organisation producing working software right now today for a dollar 40.
Chris Sharkey (00:26:50):
Yeah.
Michael Sharkey (00:26:51):
And that's today, like, and this is moving so rapidly and so that's, okay, that's software engineers. Now we move on to doctors and what are doctors, they're just big knowledge bases really, and pattern recognises what are, what does AI do really well? Knowledge aggregation and pattern recognition.
Chris Sharkey (00:27:10):
Yeah, you're right. Exactly. I mean, that's literally what it is. It's a lot of rote learning and a lot of memorization of textbooks and they're just a reference thing. Like, I see this and this together, then therefore it's this disease or whatever it is. And I, I don't wanna belittle the industry or anything. I know there's more to it than that, but it's more like you say, this is right in the AI's wheelhouse. It's what it's excellent at. It's really, really good at classification. We've always said that from the start, and that's really just what it is. It's a classification challenge and it's excellent. And now that we've got proper multimodal, you can see that it can really, um, it can really perform that function. I mean, to the point where we're seeing with, uh, the newest models that voice recognition is going to be a core part of the, of the models, like straight from audio to interpretation without the text layer in between. So if you have the textual description of a patient, you can hear them coughing, you can hear their lungs, you can hear their heartbeat, all that sort of stuff. And you get x-rays and you know, all the different scans and stuff. The AI is gonna be so much better at hol holistically evaluating that information than any human could possibly be.
Michael Sharkey (00:28:19):
I still think it's gonna take a whole generation of people that will need to die before this kind of technology will be accepted. It's the
Chris Sharkey (00:28:27):
Ones who don't trust it will die. Yeah.
Michael Sharkey (00:28:29):
. But it's the kids of today, right? That it's sort of like our kids that are gonna grow up in a world with AI and just accept that AI can help them, uh, with a medical problem. They don't really need to go to the doctor. They just, you know, whip out their phone, take a photo and away. There you go. Yeah. And this
Chris Sharkey (00:28:47):
Is the point I wanted to make earlier where, you know, we were talking, you were mentioning how you tried it and I've actually had people in my life who've asked me specifically to run letters like that or documents like that through the ai. And I think the point about it is whether or not it's scientific, whether or not it's more accurate than doctors, people are going to do it. The technology's there and people are going to do it. I mean, people were already, doctors already got pissed off when people would look up their symptoms on Google and diagnose themselves with all sorts of stuff, right? Like that was already happening. There is absolutely no chance that people aren't going to use the most generic models to diagnose their medical conditions and act on it. It's going to happen, it's happening now. So having things like the Flamingo model, med Flamingo, um, and specialist models that are actually good at it is probably essential to make sure that people are at least getting the best of what's possible with AI now and into the future. We can't, um, you obviously don't want to be using models that aren't specifically trained on it. Um, you want to use the best one for the job, but my point is that people are going to do it anyway, so we might as well look at how we incorporate it into the future of medicine rather than pretend like it doesn't exist. But there's
Michael Sharkey (00:30:00):
Also gonna be censorship in, in all of these areas as well, if you think through it, you know, like a drug, like if I'm Pfizer
Chris Sharkey (00:30:07):
Yeah, I was about to say, the pharmaceutical companies should be scared right now.
Michael Sharkey (00:30:11):
Yeah. Like getting onto this. They, they could literally be like, okay, well I'm gonna create, you know, Pfizer doctor G P T that just recommends my drugs. Yeah. Because why wouldn't
Chris Sharkey (00:30:22):
You? Well be, because, and the other, the other thing I think that I would be concerned about, like as you know, I do a lot of reading about nutrition and health, like separate to this podcast and, and things like that. And a lot of drugs that get prescribed, like statins for example, for high blood pressure just don't work. They, there's scientific study after study that show that they don't really do anything. They don't help with the one thing they're meant to do yet. If you go into a doctor and, and have even an inkling of high blood pressure, they'll give you statins like they wanted to give them to me, for example, and I'm pretty healthy. And, um, you, they just keep lowering the threshold of where you need to be there. Now if you've got a medical thing that you are talking to and giving it your, say your blood, uh, blood results and your blood pressure figures and stuff like that, it's going to be up to that model where it says, Hey, you should get on statins for example. Whereas imagine that you are the pharmaceutical company, you're gonna wanna make damn sure that model has the right figures in there that make sure they get recommended. And the model isn't, Hey, let's just hold off and try something natural. Like there really will be some sort of big industry influence on models I would imagine, just to protect what they have now.
Michael Sharkey (00:31:32):
Yeah. The models in every industry, and I think this is why foundationally everyone, whatever your view should be pushing hard for open source models that are available to everyone to democratise all that information and not
Chris Sharkey (00:31:45):
Just, not just the models, but the data sets that the models are trained on. Yeah. 'cause that's really where the proper manipulation could happen because you could have a model like LAMA that's completely open source, but if the data sets that's been trained on are inherently biassed, then it's gonna be a core part of what that model is.
Michael Sharkey (00:32:03):
You're probably gonna have some extent similar to the culture war that is taking place now, almost like a model training war of different views, you know, the the kind of textbook prescription drug crowd versus the more extreme side. And, and the reality is it's probably some sort of middle ground that's the truth. Yet you could end up just having these models just trained on whatever view the creator has, which makes total sense. But then, you know, people going, I only trust this like healthy living model. Oh, I only trust the government health model.
Chris Sharkey (00:32:38):
Hmm. The thing that gives me hope. There was a paper we spoke about about six or eight weeks ago where they talked about, they were talking about emergent behaviours in models. And one of the things that they found was that the models can actually come to conclusions on the data that are contrary to its, its alignment or contrary to what the actual data may be, you know, biassed towards, for example. And so I wonder if on the very largest models, if you're actually, if it does get all of the data or like as much as is out there, it will actually naturally come to better conclusions that any bias could potentially be on there. But it, it again comes back to the whole alignment thing. Most people, almost everyone who interacts with the models are going to interact with aligned models. No one is gonna be using, uh, the raw stuff.
(00:33:27):
So therefore it really will be about who controls the alignment that who controls the knowledge. And you know, one of the things we've been talking about just off off podcast, is this idea that the future of the internet and the future of our interactions are probably gonna be very agent based. Rather than look up information, do your research, and then come to your own conclusion. A lot of our future research will be done with the assistance of an agent. And if that agent's aligned by someone with vested interests, then those, those interests will be perpetuated.
Michael Sharkey (00:34:01):
So on the medical topic, that guy, you know, the guy that like bought the AIDS drug and charged heaps of money for it and then got jail,
Chris Sharkey (00:34:12):
Oh, Chris Lucky or something like that, he's like the most hated man in the world or something.
Michael Sharkey (00:34:15):
So he's Aou announced Dr. Gupta ai, a virtual doctor chatbot. And so I haven't even checked it out, but, so you can see it's kind of already happening in the, the internet villain himself has released it. Why would you like maybe release it in someone else's name? I don't really know his backstory if he's like truly a, a bad guy from, from what I've heard, it sounds like he is, but it's interesting, a doctor on Twitter said l l m shouldn't be used to give medical advice. Who will be, this is the biggest concern. Not like how will humanity benefit? Who will be held accountable when things inevitably go sideways? Uh, and but
Chris Sharkey (00:34:57):
Doctors aren't held accountable now when things go sideways, the amount of medical mistakes that are made, they're, they're not all prosecuted. Like, I I I, I don't think that, I think that's a moot point. I mean people, but
Michael Sharkey (00:35:09):
My larger point is just that what happens is, and similar to the software example, is the people in the industry just don't want this. The radiologists don't want it 'cause it's gonna disrupt their entire job and workflow. The, the software engineers, I mean
Chris Sharkey (00:35:24):
Probably the Hollywood dudes, I don't know if you've seen any of those speeches, but they're like, don't let robots take our jobs and stuff like that.
Michael Sharkey (00:35:30):
Yeah, well they said this week that apparently in a lot of films now they're all the extras in the background just rendering with ai. But I mean, they've been doing c g I for so long, I'm not sure how
Chris Sharkey (00:35:43):
Impactful. Yeah. I'm like, what the, the dinosaurs have all lost their jobs in Jurassic Park mates or something
Michael Sharkey (00:35:48):
Like Yeah, yeah. Literally that's, that's sort of the point. But anyway, I, I just think there is going to be a lot of resistance and that resistance will probably slow adoption down and slow people's trust.
Chris Sharkey (00:36:01):
Yeah. But resistance of the people who, the who, who it's going to damage, as I pointed out earlier, I think people want, like, people will do this. They, they're inevitably going to do it. And uh, the people who provide the tools like this guy, uh, are probably gonna get decent business out of it. I mean, I'm not saying his one's good or bad. I, I haven't seen it. But the point is that I think, you know, like that line in Jurassic Park to bring it back to that nature finds a way people are gonna bash down the doors to get access to this. Rather than having to say like, if you live in, uh, underserved areas medically, say you live in the country, say you don't have ready access to medical stuff, or you just don't wanna pay the money, people are going to use these things 100% as often as they can. Like if you can take a photo of your injury or describe it verbally to a virtual doctor at home with complete privacy as if you're not gonna do it, imagine if you've got some sexually transmitted disease or some other embarrassing ailment, it's gonna be far preferable to speak to an AI bot than to go into your doctor and, you know, show 'em the goods or whatever.
Michael Sharkey (00:37:05):
Yeah, I couldn't agree more. I think that's the the first thing it'll be used for. It's like the places that added, uh, self-checkouts, people started using protection more 'cause people weren't afraid to buy condoms and things like,
Chris Sharkey (00:37:17):
Yeah. Yeah. And see the other one is like, they talk a lot about early detection of, of medical, uh, issues. You know, like if you detect certain things earlier than early treatment is more effective. I'm not sure how much I agree with that just on a personal level, but if you can sit behind a webcam like I am now every, every day and it takes a photo of me, runs it through the scanning models or whatever, and it's like, Hey, that's a melanoma, or like you, you ought to get that properly checked out one day that is really going to have a positive impact on the amount of early detection, which may actually boost the medical industry. But it's the sort of ongoing casual thing for things that people put off. You know, like it's natural to put off wanting to go and get things checked out. But if you've got something that's passive or at least something that's very low resistance, then you're probably more likely to do it, I would say.
Michael Sharkey (00:38:10):
Yeah. It makes it more accessible and I think medical costs are skyrocketing around the world no matter what country you're in. And so it could also disrupt and bring down those costs dramatically.
Chris Sharkey (00:38:23):
Yeah, I think that's right. With raising interest rates, at least in the western world, cost of living's going up. If you've gotta pay $180 to go to the doctor to get something minor checked out and you can do it for $1 on Chris LE's thing, um, I think a lot of people would choose that option.
Michael Sharkey (00:38:38):
Yeah. And now the concern will be, you know, what if it misdiagnoses, what if it hallucinates? Uh, and things like that. But as you said, I mean, I know people, I'm doing it. Everyone is already doing
Chris Sharkey (00:38:50):
It. They're going to do it anyway. And I think the other thing that we have to remember is doctors, while they are revered in society, they are not infallible. In fact, they make mistakes all the time. It's extremely common. So to act like, okay, the AI hallucinates from time to time, therefore it's not a hundred percent perfect, therefore we can't trust it is just not a valid argument when comparing them to real world doctors. I think in other spaces, sure, there, there's areas where hallucinations are really serious, but I just think in this one it's, it's like you're comparing a fallible thing to a fallible thing. They're both gonna make mistakes. Which one are you gonna go with? And I think that some people will choose the ai.
Michael Sharkey (00:39:29):
Isn't it just inevitable though, that we as humans just start working on much higher level work, for example, for basic diagnosis. I can imagine literally having the capability at home from your phone or your watch or whatever it is in the future that can, uh, help with the diagnosis process. And then it basically escalates to a really, uh, you know, specialist team. So if it's something that the AI is like, Hey, we need to escalate this, or you need to go get a scan and report back, you might go to a scan centre and have that scan and then that, um, links back to your AI agent who then, you know, tells you what the diagnosis was. You can just see that maybe we're working on higher level things. Like instead of building a blog management system, which has been done a billion times before, or some flay bird clone , the, the software development virtual agent are taking care of, of the, the sort of day-to-day work. So you as a software engineer can focus on much higher level work.
Chris Sharkey (00:40:28):
Yeah, I think that's gotta be the dream. That's gotta be the hope of the positive sides of AI is that the sort of general grunt work that really just doesn't help. Like, you know, you need to get it done, you know what needs to be done and it's time consuming and all that. If the AI can to, can overcome those things quickly, then it can only be a benefit for everyone in terms of just overall output and creativity and, and what we can do.
Michael Sharkey (00:40:53):
So the other model back to our l l m overload and, and probably one of the more important announcements in the week in terms of open source, and I know I'll probably get flack in the comments again for calling it open source 'cause there is a, a usage policy on this. But meta, the actual meta this time or meta ai, Facebook, , mark Zuckerberg's business, open source, , I don't know how to describe it
Chris Sharkey (00:41:22):
Anymore. Old Marky.
Michael Sharkey (00:41:24):
Yeah.
Chris Sharkey (00:41:25):
He's always creating businesses,
Michael Sharkey (00:41:27):
The Zuck, uh, so open sourcing, audio craft, generative AI for audio made simple and available to all. So this has already been out, we've already talked about this before, where you could generate these small snippets of, of music, but now this is open source so you can create sounds, so sounds for, for video games or background noise or whatever you need as essentially propaganda. The Hollywood people are gonna hate this one. Yeah. . Um, so sound and also music and it's really, really good.
Chris Sharkey (00:41:58):
Yeah, I tried, I tried it out this morning too and uh, I made some cool stuff. I made a Beatles song with flute, um, and a few other things. What did you make ?
Michael Sharkey (00:42:07):
So I'll, let me give the the listeners a few examples here. So Oh, cool. Generating audio from text descriptions. This one is whistling with wind blowing. That's not bad. And then this one is sirens. People are gonna enjoy this in their vehicles or wherever they're listening from. I hate sirens that are humming engine approach and pass, don't freak out people. I mean it just sounds absolutely real. Is this all the crazy stuff that we're creating for the metaverse behind the scenes?
Chris Sharkey (00:42:43):
Yeah, I guess that's probably what it's for. And what I really struggle with these ones just personally, I guess it's just, it's just not an area that affects me. But other than the metaverse, I'm like, it's cool, but what are you gonna use it for?
Michael Sharkey (00:42:57):
I think all sorts of things. For example, if you want to interact with what I like to refer to as the analogue world and have a phone call, uh, where you call a restaurant and make, well, you want a booking with your agent and you want a bit of fake background noise to make it feel a bit more human and real. You could have a series of sounds that you're generating that sound local, like for example, like a
Chris Sharkey (00:43:18):
Baby crying in the background,
Michael Sharkey (00:43:19):
. Well if, uh, okay, so this is one way of telling location, right? So if I'm in New York, uh, firetruck sounds a lot different to here in Australia. Yeah, the sirens are completely different as opposed to if I'm in Europe, their sirens sound completely different again. So that could be one way that you can sort of localise the content or even localise the content in a video game potentially. But in terms of like multimedia and creators that are working on YouTube videos, it takes the cost essentially to zero for background sound, stock audio. Hmm. Uh, yeah, a a again, high level work. Like you don't need to accessibility as well potentially where people don't need to be able to afford to pay for like iStock or Adobe stock, uh, in order to create whatever they want to do.
Chris Sharkey (00:44:08):
Yeah, yeah, I definitely see that. And I think the music generation one I definitely can understand more in terms of just coming up with ideas for new music, using it for video games, using it for YouTube videos. It just anywhere, like, it's always helpful to have background music or, and things like that. And one of the cool things you can do with the music generation one is given an existing tune or melody or song and then it can modify that into a particular style similar to what we see with the image to image stuff. So I can see that being really cool, uh, as a creative outlet as well.
Michael Sharkey (00:44:42):
Do you think Spotify's future business is completely cut out the musicians and creators? 'cause they've gotta pay them a really large cut today, right? Like 70 or 80% of all the membership fees. Do you think in like a decade, music gets so personalised and people are just creating their own music or Spotify's just generating music at No, you like to slowly get you to listen to more AI music to get their profits up?
Chris Sharkey (00:45:07):
I don't know because I'm, look, I'm not real, I'm not even close to an expert on music, but the only thing I can think that would stop them is music and TV shows and things like that are somewhat of a shared experience. Like, oh, you like that song? I like that song too. I dunno how that happens. If everything is completely unique and there's just nothing ever the same, so,
Michael Sharkey (00:45:26):
Okay, but what if it creates a unique, so people share playlists today and they're quite unique. I create playlists all the time on Spotify and share them and people listen to them and
Chris Sharkey (00:45:36):
Right. So you're like, here's music. I actually generated myself his music.
Michael Sharkey (00:45:39):
I either generated or, or Spotify's algorithms just globally generated that I've gone through and handpicked and curated and here's my selection of sounds that I like ,
Chris Sharkey (00:45:49):
I definitely try it. I mean, I, I, yeah, I, I don't know, it's, it's very interesting, but clearly right now they're doing what, 12 second clips or something like that, that's obviously gotta get a lot higher, but as we've seen with everything else, these things are inevitable. I just,
Michael Sharkey (00:46:04):
I can't help but think here, imagine if every, like you, you go and listen to new music today on Spotify in their like for you section and, and the various competitors have their own different, uh, like recommendation engines like that. But what if, you know, I'll go through that list and there's like two songs every week. I'm like, they're great. I'm gonna add to my playlist. The rest are junk, basically. Or throw away music. Yeah.
Chris Sharkey (00:46:29):
What
Michael Sharkey (00:46:29):
If every song on that li list every week was just so good? You were like, you had it on repeat all the time. I mean, it's potentially gonna happen.
Chris Sharkey (00:46:38):
Yeah, it's interesting and it pro like you say, it probably will happen. Someone's going to try it and people will want to see what it's like.
Michael Sharkey (00:46:45):
So I, maybe these Hollywood guys are onto something like they really are panicked and they're like, but, but now they've upset the software engineers, potentially the doctors, the yeah. Stock image people. Like, it's
Chris Sharkey (00:46:58):
Sort of like the people fighting, the fighting in the front lines of, of AI replacing their jobs. Like the ones who are most vulnerable. And in most of those cases they're people who are highly trained. I think that's the real thing. Like if you were a, if you were working at the front counter at McDonald's and they replace you with those computers, which they basically have, then you're like, okay, well it's a low skill job. I'll go get another job. If you're a doctor or you're a a Hollywood actor, there isn't exactly many close alternatives that require that level of training. Like they've trained half their c like I know a guy is, well I guess surgery's probably gonna still . Yeah. I would hope still be human based for a little bit longer. I don't know if I'd trust the robots that much. It's hallucinating
Michael Sharkey (00:47:38):
, hallucinating
Chris Sharkey (00:47:40):
. Like, oh, I thought that was your foot. Whoops, whoops. Um, but yeah, so I think that, but just the time it takes and the effort it takes and the sacrifices doctors make to get where they get to have that sort of being like, yeah, nah, don't worry. The, um, meta's latest model can do that. That would be threatening and that would be upsetting and and tricky to deal with. Um, and I could see why you would resist it. And maybe it's the same with music creation. I don't know.
Michael Sharkey (00:48:08):
Yeah. I just think there's a big large group of professionals and highly trained professionals, lawyers, doctors, all of these people, and I'm not talking short term here, I'm talking much longer term that AI's going against. And as you said, nature will find a way. People are gonna probably use these things anyway. It may just lead to a, a lessened demand for these services as opposed to a full replacement. Well,
Chris Sharkey (00:48:33):
Yeah, and my thoughts turn to school and what, you know, if you have careers guidance or something like that, what are students thinking who are say 14, 15, 16, or maybe even a little bit younger than that when they start to think about their future job. Like you really gotta be thinking carefully, by the time I spend 10 years training as a professional, will this job still exist or, or still be as prestigious or as valuable?
Michael Sharkey (00:49:01):
Yeah. Well, I mean, we both got an email from a, a listener that was saying the same thing. They had their, their I think son or daughter graduating from high school going to take literally a year off to figure out like, what is the best thing that I should dedicate my time to in terms of education because of how disruptive AI is becoming in every, in every industry. I guess that's the real question is, is how, like, how quick does this happen? I think it's gonna just take so much longer than we think because you've gotta go build all these things.
Chris Sharkey (00:49:35):
Yeah. And it's gotta, it's gotta also filter out into the actual real world in, in places like medical industry especially is full of regulations. They're not just gonna whack in an AI radiologist and just swap everything out in a weekend or something like that. So yeah, I think that a student who's 18 now is probably fine in terms of their career, but, um, like you say, we've gotta think about the long term implications of this stuff here, but we don't have to. We want to. And I think that yeah, at some point this has to enter the discussion at the schooling level as to as to whether you put your time into it or not.
Michael Sharkey (00:50:10):
Yeah. And, and someone
Chris Sharkey (00:50:11):
Will be the last radiologist
Michael Sharkey (00:50:13):
I think also schooling, I mean the, you can see like, wouldn't you want your kid trained by the best teaching agent in the world? I I'm already learning every day now from ai, like it's become the best educator for me in my whole life. Like I, I can ask
Chris Sharkey (00:50:28):
It anything. Yeah. And definitely, definitely my kids who are eight and nine will, will say, can you just ask the ai, can you just ask the ai? So like, they're already seeing it as a, as a, as a way to verify information and, and get knowledge. And so I, I think that it, it will be a critical part in education going forward, especially, especially agents that say help you practise for exams or ones that help you practise language, which I've seen a few of those around. And ones that can help you with repetitive tasks or for example, like you're learning the piano or the violin. It can listen to you and be like, okay, you know, your tempo's off, you've gotta get that fixed up. Like, 'cause this, this Facebook stuff can also interpret audio as well. So it will be not that far off where those little pieces of paper where you get a guitar teacher may be completely unnecessary because you can just learn guitar with a dynamic agent who understands music perfectly off your computer.
Michael Sharkey (00:51:23):
Yeah.
Chris Sharkey (00:51:23):
Or phone, probably phone, I guess everyone just does phone. Right?
Michael Sharkey (00:51:26):
Well, will you like learn just how to create like music prompts? Like I, I'm, yeah, don't
Chris Sharkey (00:51:30):
Even bother with strings. What are you talking
Michael Sharkey (00:51:32):
About? Yeah, like you don't, you don't need an instrument.
Chris Sharkey (00:51:35):
Yeah, that's right. I play the keyboard, man.
Michael Sharkey (00:51:37):
Uh, so in other news that I wanted to cover, but I've left way too late, is open a OpenAI has rolled out some improvements to chat G B T what I told you about these series
Chris Sharkey (00:51:49):
Of the lamest updates in
Michael Sharkey (00:51:50):
History. Well, you just repeated it, so, perfect. Yeah. , I told Chris, I'm like, should we, should we even cover this? And he's like, what? I mean it's just, you know, what have they really done? So there's suggested replies. Now there's uh, prompt examples when you load it up. Uh, they're not that exciting. ,
Chris Sharkey (00:52:06):
I mean, we've just seen some guys make an entire game using a $3 of G P T four credits and a, and a caucus of agents. And these guys have added a suggested reply button. I mean, it's, it's really basic web app stuff. It just seems like why, why even announce it? I don't get it.
Michael Sharkey (00:52:25):
My only thought is they're working on something much, much bigger behind the scenes here. Like the true sort of, you know, one agent to rule them all. And these are just incremental updates that are pissing off their paid users, which are funding them right now, for example, making G P T for the default model, if you're a plus user uploading multiple files, uh, that you want to analyse the code interpreter. But you know what, the interesting thing for me is so many people have emailed in to us and tweeted and all, all sorts of things like that, saying they've, they've quit their Yeah. G P T plus
Chris Sharkey (00:52:59):
Subscription. And when you read the comments, like on that announcement, they're all just like, oh, can't, can't you just make the code writing better like it used to be? Can't you, um, you know, bring back the web browse thing? Like they're, they're not releasing things that people want and people are done with it. They're over it.
Michael Sharkey (00:53:15):
Yeah. I certainly, if I was younger and had to pay, I don't even know what it is. I'm so out of touch, apparently 20 a month, then you, I I wouldn't do it. I would just go to bingeing or, uh, one of these other services that was relatively free, uh, if I could get a better service than to, to pay them.
Chris Sharkey (00:53:37):
Yeah, that's right. They're not, the GPT four's abilities aren't that much better than the others. And I, I guess bingeing sort of is G P T four anyway, isn't it? So it's, you're not really gaining a lot, uh, over the other stuff. And I think that's the thing. We've spoken about this before, they really need more of an application layer around what it's doing to, to bring the value, to earn the money there. Everyone else, if they're hardcore, they're just using the a p I anyway. Um, and if they're casual, like you say, there's alternatives out there that are, are free and almost as good
Michael Sharkey (00:54:08):
In terms of your, like thinking about your kids. 'cause my kids aren't at an age where they can even really use the computer. But thinking through your kids, do they see chat G P T as when they say the ai or do, do they not really have a sort of branded concept? No,
Chris Sharkey (00:54:24):
They don't, uh, they don't have that sort of unfettered computer access yet. That's too young for that. So if they're doing it, I'm putting them in front of it, you know, like I'm, I'm setting it up for them. So they wouldn't know what they're using, um, at all. No. So
Michael Sharkey (00:54:37):
Yeah. So you wonder, does it become like Google to them? Like when we were young and Google first sort of became the, the main thing, eventually it was like, just Google it. Oh, just go to Google. Google can find that for you. And that was sort of the, the, the magic.
Chris Sharkey (00:54:51):
Certainly when you hear, you know, other kids and, and young people talk about it, they all say chat G B T, there isn't, uh, there isn't this idea of of there's, oh, I use Claude actually, you know? Yeah. like, oh sorry, at my school we use Claude. Um, no, nothing like that. So yeah, I would think that the go-to for almost everyone is chat G P T. But I don't think, just like the search engines when they came out, um, once something better came along, there's no loyalty. No one's going to, no one's gonna stick with 'em just 'cause of the name or something like that. Do
Michael Sharkey (00:55:21):
You think Claude, like it's sort of like I, and I think it's a great product and I've got some comments about it, but do you think it's like the Ask Jeeps of our Yeah, maybe,
Chris Sharkey (00:55:31):
Maybe, um, Scobel onto something with the marketing. It's like Claude is like the, the highbrow upper class AI and uh, you hope you use chat G B T in your household. Yeah. That's disgusting.
Michael Sharkey (00:55:42):
But so interestingly, this podcast started, the first ever story we told was about me writing, uh, custom Batman stories using the, the, um, the like a p i at the time. 'cause I don't even think chat G b T even existed when I started doing this. Mm. And then now I've had, I've actually switched to Claude two because it writes just brilliant Batman stories. Once you, you know, tune it and it's far, far superior than, uh, than chat G P T or g P T four, whatever you wanna call it, it doesn't, doesn't really matter. And so they definitely have these great skills, but like you just said, with the marketing bit, no one has any idea. There's just like, oh I'm just gonna spin up chat G B T. It doesn't really matter that it's not better.
Chris Sharkey (00:56:31):
Yeah. And I think that's, that's sort of part of the, part of the point of the, the overwhelming nature of all the different models. Until you have a real thing that you're trying to do, it's very hard to evaluate them. And I think that's why definitely my evaluation of new models is quite unscientific. It's only when I'm trying to solve an actual problem for a real world thing, I need to do that I notice, okay, this isn't going to work because it's just not getting the right result. Or um, you know, this style of prompt doesn't work here. It doesn't follow guidance as well or it hallucinates too much. He knows that stuff when you have a real thing.
Michael Sharkey (00:57:06):
So did you want to talk quickly about this uh, San Francisco compute group? Can you fill us in on what this is?
Chris Sharkey (00:57:13):
Yeah, so it's pretty interesting. It was just a post, a hacker news and a simple nice H T M L website, like sort of old school web style that I love so much. Um, some guys just being like, hey, we're gonna try and buy 500 H one hundreds. Uh, we've got a sort of lead on how we're gonna be able to buy them. Uh, we're basically, it's almost like a cooperative, like everyone's buying a share in it for a certain amount and um, you will be able to sell your share later if you want. And the idea is that they do a collective buy of these machines, set 'em up. And the general idea is if you've got some burstable AI training, which most of it is, and say you wanted to train something that might take a month to train on a single H 100, but you could do it in two days if you get half the cluster to yourself or something like that, um, then that would be available to you just to help startups in general. So it's this sort of like, you know, collective uh, of hardware buying that they're gonna work together on.
Michael Sharkey (00:58:11):
So do you think the core user of this is startups and, and companies trying to train, train their own actual model? Or is it just fine tuning that this would be mostly
Chris Sharkey (00:58:22):
Useful? Well probably f probably fine tuning. I would say there's some people trying to build them from scratch, but I would say the majority of it is going to be taking existing models and then a bunch of data and training it or you know, a bunch of alignment stuff, whatever and and training it so they can say we've now got our med flamingo, we've now got our, you know, expert one in guitar tutoring or our expert one in whatever field people are going for. That's what I imagine a lot of the startups are looking at doing is, is having proprietary models based on the open but not open source ones where they can say, well we've customised this with our data and fine tuned it and therefore it's valuable. But it's interesting because I've been looking at um, Lambda the um, Lambda ML or whatever it's called where you can spin up H one hundreds and different servers.
(00:59:08):
'cause they're always emailing me saying, oh we've got a server available today, but when you go on there and try and spin something up, as of today, this morning when I checked, they've got nothing. None of their servers are available. No spot instances, no reserved instances. You can't get a single one If you go onto Amazon and look at their machine learning options, a lot of those are either unavailable, like the ones that have eight H one hundreds for example. Um, and the smaller ones are like pretty expensive and they're not available in all regions and things like that. So the actual availability, like it's all very well to say, oh, just use the cloud and train your own model or whatever. But the truth is right out there right now, if you do want to train one of these massive large language models yourself or do heavy fine tuning on one or sorry, not even heavy, do any fine tuning on say LAMA 70 billion, you're gonna have a hard time finding the hardware to actually do it. So I can see why things like this exist. I don't know what their goal is with it, um, other than what they they say, but yeah, the hardware availability is an issue.
Michael Sharkey (01:00:14):
Yeah. And for people that don't really know, this is one of the big bottlenecks of the advancement of AI right now is that especially in Silicon Valley, you've got people speculating around who's getting the next batch of a one hundreds next. It's become almost speculative about, you know, how many H one hundreds, uh, h a I always get
Chris Sharkey (01:00:37):
This H one hundred's. The new better one. Yeah, A one hundred's, the previous one, which is still great. Um,
Michael Sharkey (01:00:43):
But yeah, and so everyone's speculating like how much compute power does open AI need to train G P T five? We saw this week as well, they've registered the trademark for G P T five, which, which surprised me. I thought they might already
Chris Sharkey (01:00:56):
Own it. It would've seen who would've seen that
Michael Sharkey (01:00:57):
Coming. What a shock breaking news.
Chris Sharkey (01:01:00):
Yeah. They're not Elon Musk just making up funny names for things or so renaming stuff. They're just sticking with the numbers
Michael Sharkey (01:01:06):
They have. I mean that letter that leaked and was soon deleted said that there is that compute limitation for them as well and they are yet to release fine tuning for G B T 3.5 and four of their models as a service and their community manager or developer relations person announced this week that that would be coming later in the year. So it seems like the more availability of this hardware, the more we will see.
Chris Sharkey (01:01:30):
Yeah, I think it's definitely a case now where some of the things they're not releasing are just because they simply can't support it at the hardware level rather than being some conspiracy or them not wanting a g i to take over the world or something like that.
Michael Sharkey (01:01:43):
Yeah. It does just seem like that limit in compute. And so finally some news that I thought would really excite you, 'cause you've been talking about this for, for some time is a letter leaked this week that said Google Assistant is getting an AI makeover, which I'm pleased to hear as well for someone who has these devices all around my home. Yeah. And in the letter it said, uh, since we launched assistant seven years ago, we've built great experiences for blah, blah, blah, all their corporate speak. Uh, you know, and we've heard people's strong desire for assistive, uh, conversational technology that can improve their lives as a team, we need to focus on it. Uh, you know, basically to summarise it, they're putting, they're looking at putting generative AI in the Google Assistant at some point in the future. So finally that thing will not be as stupid as it currently is now.
Chris Sharkey (01:02:38):
Yeah. And the, the, the problem in it, I'm very keen to see them solve because it's something I've been thinking about a lot is the maintaining the context of a conversation when it's unclear if the following statement or question relates to it. So like for example, um, if I asked it for a recipe for pancakes, which is like the main use of Google Assistant, I think, um, I ask it for a recipe for pancakes and then I go, oh, can you make it a bit shorter right now? Those assistants have no idea what the hell you're talking about. They'll be like, I'm sorry, I dunno what you're talking about. But presumably a large language model can easily handle a, a thing like that to know, okay, they're referring to the recipe, but then if I ask a another question that's about something unrelated, how does it know to enter a new context and know what the object of that context is in the ongoing conversation? Because I think this is definitely a problem that needs to be solved in the sort of interactive agent paradigm. It's like you need to be able to handle follow-up questions in a a conversation context. But if I then speak to you about, like, I asked you the weather, I asked what the surf's like, and then I want to get back to talking about pancakes. How's it gonna maintain that context?
Michael Sharkey (01:03:48):
Yeah. And can it, it it'll be interesting to see how deep the implementation is or do they take, do they go really bold like chat G B T and just unleash it to the world? Or do they go in these tiny sort of corporate protective steps like Google usually does where it's like, today we're adding a little bit more context and it's just not that exciting. If I was them, I would go all in, I would, I would give them out for free and I would try and wipe chat G b T out in the home. Because I
Chris Sharkey (01:04:14):
Think the thing that would make it huge is if they made it so there's plugins and people can make and deploy like, like Alexa skills. You know, how you can add things to Alexa that it can do, if they had that with L L M capabilities where you could plug in your own sources of knowledge, your own agents or you know, some sort of concept along those lines where you can really expand it. I would be, I would be so excited about that. That would be absolutely amazing because the sort of getting the, the speech recognition, like I know all this tech exists, the speech recognition, speech generation, it all exists, but getting it all in a device that's cheap and anyone can have and it's sort of commercial is hard. And so having that technology and then being able to extend it, like where you can add capabilities for someone would be huge.
Michael Sharkey (01:05:01):
Yeah. I also think that the future of how you interact with businesses will change as a result of these technologies. For example, if you have to book a flight right now, it's painful. Like Google flights has made that simpler, right? But you've still gotta fill in all the form fields, pick the right dates, and often I don't care when I fly, if I'm going on holidays, I'm like, I can roughly take this time period off, just gimme the cheapest flights. Yeah.
Chris Sharkey (01:05:26):
Not to mention if you go to book it, get distracted by something else and come back, they'll raise the prices just to screw you. Yeah. Like, you know, there's all that element to it as well.
Michael Sharkey (01:05:35):
But if you had the agent where you could go to the airline you fly with like United in the US and talk to your Google home and say, Hey can, like I want to go on a summer vacation, these are the destinations I'm interested in. You know, figure it out , and then it's like, here's some options and maybe there's some audio visual experience with it. Do you ever go to their website again? Like, oh, I want to change my flight. You just talk to it. I mean, call centres died because of websites, right? But like, if you can just have on demand in your house wherever you are and talk to it, maybe the idea of a corporate website eventually dies.
Chris Sharkey (01:06:13):
Well, yeah, especially because the websites have lost their simplicity and they're just so full of garbage and junk and trying to sell you extra things. And I think people are fed up with all of that stuff. That's why ad blockers are so popular and, and stuff like that. So while the web initially probably was a lot more preferable to calling someone up on the phone, um, because you've gotta wait on hold and all that, you get all the benefits of just having someone directly trying to solve your problem without all the hold, without all the other junk, without all having all this stuff shoved down your throat. I think it'll be really popular if it can get the whole job done without having to like, okay, now you've gotta go fill this format. If it can take you from start to flights booked and you understand exactly what's going on, people will use it.
Michael Sharkey (01:06:57):
I, yeah, I think it's the future. I think websites may still exist, but the website will just be a point of access into an l l m. Like imagine an open search box on the United website and you're like, need flights to Vegas cheap price, any date between in March and April or, or whatever. And it just goes, okay, I'll book that for you. And it, and, and there's something on your computer or your device that has context, like your credit card information, your details about you, and it securely communicates when needed with that source. Uh, and, and that could be delivered through a Google home or, or something that's always around you. Yeah, it's
Chris Sharkey (01:07:36):
Funny. That's sort of another probably like, uh, business idea is a sort of context passport kind of thing where you can have all of your information securely stored that is context for large language models that you can then authorise it to use. So it's like, okay, I'm ready to make the booking bang, here's all my kids' names and date of births and passports and all that stuff. I now grant you access for 10 minutes or something to complete all this. It uses that context to complete the transaction, then it goes away. Um, 'cause context will be so important to the, the value of these bespoke agents like this.
Michael Sharkey (01:08:12):
Yeah. I think that that needs to exist. How painful is it sharing that information then, you know, it's getting like stored in a big database and that company's using that information. If you could securely transmit the data and then sort of pull that access away after they've done what they need to do with it, uh, that would be a great way to, to enforce a bit of privacy in this brave new world instead of knowing like every large language model. Yeah. And
Chris Sharkey (01:08:39):
You don't wanna be like every single prompt has your passport number in it, just in case it may need it, you know, like it, it really needs to be a specific action on your part rather than, uh, sort of on like telling it what the day of the week or the weather is or something like that.
Michael Sharkey (01:08:53):
Yeah. It'll, I just think the first airline or or website that starts to do some of this stuff and just changes the, the interface of the web that's gonna snowball really quick once someone does this. And I think you'll see it everywhere. Everyone will start to do this kind of thing.
Chris Sharkey (01:09:10):
Yeah. Agreed.
Michael Sharkey (01:09:11):
Alright, that's all we have time for this week. Thanks so much for, uh, listening all your comments and, uh, feedback we get, we read all the comments and the reviews, so we, we really appreciate them. I apologise if you've been watching today, my camera for some reason is turning on and off, so you've enjoyed quite the light show. Uh, but whatever, you don't really need to see me. Uh, thanks again for all your support and, uh, we'll see you again next week. Goodbye.