Llama 2 Unleashed, Open-Source Vs OpenAI, Instructions for ChatGPT & Fine Tuning with Llama2 | EP24
Michael Sharkey (00:00:00):
Let's just like back up a little bit here, lizard man who everyone hated like months ago for stealing data, all your personal information, scanning your images, like reading all your stuff and putting ads against them and not being full flippant with people's personal data. Use that to build an AI model, which now is the poster child open source may make these things safer and may help stop AI destroying the world. Chris, this has been one of those weeks where I've wanted the AI news to stop .
Chris Sharkey (00:00:39):
Yeah, there's been a lot happened. We get nothing and then there's a lot, there's
Michael Sharkey (00:00:42):
Literally been so much announced, but obviously to anyone following the news out there would know that Meta has released LAMA to, and of course we know LAMA was the very first well open source model that really kickstarted open source, but it was leaked. It wasn't really meant to be open source apart from to researchers.
Chris Sharkey (00:01:04):
That's right. Exactly. And that's the one where the weights got into the wild and started the sort of open source revolution that made us realise that it isn't just about the big companies when it comes to large language models.
Michael Sharkey (00:01:15):
And so in the meantime, what's happened is Zuckerberg and co have realised that if you have the entire community using your open source model, you, you get so many learnings from that. It makes your technology better. It teaches your own team how to improve the model and improve they have done. Uh, so they've announced LAMA two, it's the next generation of lama of course it's completely open source. And the big one is instead of just being for research, you can now use it for commercial use as well.
Chris Sharkey (00:01:48):
Yeah. And that is the absolutely massive, massive piece of it. The fact that you can legitimately use it in commercial applications now, which was sort of, you know, ambiguous at best before. Um, that distinction makes it such a big, big deal.
Michael Sharkey (00:02:03):
Yeah, I, I truly can't emphasise how big this announcement is. Um, there's a few other details I'll cover first, but then I think we, you know, we should dive into that a lot more here is Microsoft's actually partnered with Meta and of course they invested 10 billion in OpenAI, but now they're partnering with Yeah. Meta to make an open source model available on Azure. So I, I think that's worth discussing. Um,
Chris Sharkey (00:02:29):
Well I think Azure's definitely going to be like, you know, we are the platform that rapidly deploys new models so you can actually get on there. We've seen Amazon announced Bedrock yet it's only really available for people who beg for access and even then it's got all these restrictions on it in preview mode and all this sort of stuff. Whereas Azure's out there, uh, doing it right now where you can get access and and use it and in some cases you actually get earlier access to the commercial models too.
Michael Sharkey (00:02:57):
Yeah, it's a funny play from in a way because when they announce Bedrock, they had this like I thought they got the messaging right, which is we're gonna be fairly agnostic here. We think there's gonna be tonnes of different models, we're gonna be the best place to use and host them. And I was really excited by that announcement, but they still aren't really delivering. It's Microsoft with Azure that is making all of these different models available to use today for developers in their applications or to develop entirely new applications as well.
Chris Sharkey (00:03:26):
Yeah, and the Palm one really surprised me actually when you told me about that because I figured that they would have some natural or reluctance to play out the open source one on there simply because they do have that commercial relationship with OpenAI. I think it's great and I think it's probably a good strategy for them not forcing the one that they have the vested interest in down your throat. And really, I guess the success of Azure as a platform for people to deploy AI on is actually probably greater than their consideration about the value of their investment in open ai. Like if it was to completely fall over Azure's worth a tonne more money than to them, than the open AI stuff is.
Michael Sharkey (00:04:06):
Yeah. To me, their strategy here is just invest in everything related to AI and try and be the true home of all of these models and the potential success stories that that come out of them as opposed to, and so far
Chris Sharkey (00:04:17):
It's working for them. I, it's not the best source of news, but I saw on news.com au today that Microsoft share price has gone up like 50% in the last six months or something, you know, mostly as a result of them being early in, in committing to the AI stuff.
Michael Sharkey (00:04:32):
Yeah. I think that the, the biggest piece to this is with LAMA two now that it can be used for commercial use, this is really a threat now to anthropic open AI and various other companies that, uh, licence their, their APIs and models to, for developers to use this is potentially now the solution and some, it's almost freeing for developers being able to build on a model that is somewhat static that they can fine tune on once and then retain that that fine tuning. Yeah. And just continue to use this model in their application.
Chris Sharkey (00:05:11):
That's right. And I think there's probably three main reasons why it's so significant. The first one, like you said, we've discussed on the last few podcasts the diminishing quality of chat GPT and GPT four. Now we're not gonna go into it today cuz it's a bit boring, but basically someone released a paper during the week confirming exactly what we were saying, that objectively it's getting worse on the problems they tested it on. Um, and those are answering, you know, certain kinds of mathematical questions and chain of thought reasoning. They've proved empirically that they've gotten worse in the subsequent releases. So one is just that reliability, just saying, if I have my own Palm two model deployed, I know that it's going to give consistent results because I'm the one running it and it's not being modified. You
Michael Sharkey (00:05:54):
Keep saying Palm two , oh,
Chris Sharkey (00:05:57):
Sorry. Um, LAMA two what I, I had Palm two in my head because that also got released it for general availability this week. So I would just, I I mixed it up. Sorry. LAMA two, having that running in your own hardware means that you can control it and that they can't modify it on you, no one can. So I think that's a really significant thing. The second one is, as you say, the fine tuning, you've got the opportunity to very easily tailor this model to your own needs and your own, the own problems you're trying to solve, which makes it a lot more commercially available. And then the third factor is simply that you can use it freely knowing that you have this commercial, uh, like free commercial licence essentially, um, to run it without having these variable a p I costs and also fearing that at some point your access is gonna be taken away because you violate some rule. Um, that, that the companies come up with.
Michael Sharkey (00:06:50):
The, the unique piece of it to me as well is around fine tuning, because obviously with G B T 3.5 and four, you can fine tune those models. That's, that's something that you can do, right? But then you have to continue to pay open AI and the, the sort of foundational model under that fine tuning can and does change very fast. Like as we said last week, sometimes every couple of months.
Chris Sharkey (00:07:16):
And so, yeah, I mean one, one idea of the fine tuning with them is at least you freeze the version of the model you actually trained. That's probably an advantage.
Michael Sharkey (00:07:22):
Yeah. And so, but with LAMA two, you can take your own data, fine tune LAMA two for commercial use with that data, and then you're only really paying for the ongoing hosting cost of that fine tune model.
Chris Sharkey (00:07:36):
That's right. And, and you know, in a way you sort of own it then, like you can have your own proprietary model that you can redistribute if you wanted to commercially. So there's, you know, one of the biggest things, and funnily enough, I asked LAMA two this morning, um, you know, what the advantages were of, um, you know, training your own model and things like that. And one of the things that emphasised is the value of it really comes from the quality of the data sets that you train it on. So, I mean, I know this is obvious, but you know, if you've got really, really high quality data that you train it on, then the resulting model is extremely valuable because we know from all of the things we've talked about before, the papers on smaller models being more effective, if they're trained on better data, now you can actually own the results of that training. It's, it's not just something that's sitting in a data centre that's, you know, at the whim of, of a massive corporation.
Michael Sharkey (00:08:29):
Yeah. It seems to me like that is one of the positives of a, you're no, like developers when they build something are no longer at the whims of one particular company. Like this is now becoming, these large language models are now becoming almost as like, like an SQL database in in your product liner where you have full control over it and it's, you know, it's just not something that, that someone can just cut, cut access off. But it does make me think that there's a future business model now in people fine tuning LAMA two on proprietary data or like, you know, hard to get data or data sets and then making that available to others so that they can use in their application. But yes, that's right. In an open source environment.
Chris Sharkey (00:09:18):
I absolutely agree. I think that that that ist you know, portable models that you can sell to people or licence to people is a big one. The other one is the real, the thing we spoke about previously, which is the privacy aspect. So if you're a company who wants your own model trained on your customer data, which may include personally identifiable information potentially, or at least stuff you don't want being used by open AI and Anthropic or whoever to train their models right now, they don't really provide any guarantees that they won't do that, even though they've sort of said it, but not really. Um, when you train these, you have that guarantee. So I imagine the other com major commercial opportunity coming out of this is we'll see a bunch of model training companies where they say, okay, you are a company like ours. You've got all this data, we'll consult with you, we'll train you a model that solves this problem you're trying to solve. And then you own that model, it's yours, it's completely private, it's running your private cloud or in a way that is safe. Um, that's a massive commercial application that that Facebook or aka Meta has just handed everyone. And with, with LAMA one that was really murky waters. Now it's very clear.
Michael Sharkey (00:10:29):
Yeah. Well interestingly, this week on that, uh, Patrick Collison, who's the c e of Stripe, uh, tweeted, we built an internal l l m model with prompt sharing discovery, careful privacy controls, configurable models Cetera. Been working on it for a few months and a third of people at Stripe are now using it every week. Cool. To see how diverse the use cases are. So this is a company that's built their own internal
Chris Sharkey (00:10:54):
Well, and I I reckon that would almost certainly be built on the first lama. I mean, there's no way they just in a couple of months just started their own from scratch and trained it. There's not enough time they probably would've used lama.
Michael Sharkey (00:11:06):
Yeah. And so therefore the, the whole like, I mean that's not really commercial use, I guess ,
Chris Sharkey (00:11:13):
I mean, it is like in the sense that if some, you know, if if someone was gonna prosecute on the first one, you could probably prove if you could somehow prove, I mean I don't see how you would, but companies don't want to take risks like that. Whereas with Llama too, it's totally unambiguous. You can and should do that kind of thing. Now, do you
Michael Sharkey (00:11:31):
Think this is the death blow potentially, or the beginning of the death blow for Anthropic and Open AI and all of these people that have paid models through APIs? And I'm not diminishing certainly what they've done, but I was insanely impressed using LAMA two this morning. Like I know people have cited a lot of out there and I've been following that closely, but from my interactions with it, everyone's saying, oh, it's, you know, close to G P T 3.5, which is what powers the base chat G P T. Mm. But I felt like it was better.
Chris Sharkey (00:12:08):
Yeah. I had really good results with it as well. I'm, I'm quite impressed. I don't think it'll be the death of those companies because they've got this sort of, you know, hearts and minds of a lot of people out there using it. I assume you're talking about the API side of things, I could definitely see people mixing up their use of this versus the other one. I think the main issue with something like LAMA two is one of the advantages is, you know, you can run it on your own computer. They have different levels of models. I think they have 7 billion parameters, 13 billion and 40 billion I think is the biggest one. And
Michael Sharkey (00:12:41):
So 70, I think it is
Chris Sharkey (00:12:43):
70, sorry. Yeah, you're right. Sorry. And so like the 7,000,000,001 you can run on your own computer, and we've spoken about this before, that has a lot of advantages in that you've got this fast iteration, you don't have to stress about the ongoing cost of hitting this expensive API every time you want to test. Like if you're using Claude and a hundred K prompts and hitting it, you know, 50 times an hour or whatever, that adds up in terms of cost. But if you can run it on your own hardware that you own, there is no marginal cost there. So you can just keep iterating a lot faster and it's probably faster too. Um, so that's a real advantage. And I think that, uh, the other one, the, the sort of counter to that is if you then wanna say deploy on LAMA two 70 billion parameters and you're a small time startup, or you know a company that's not really sure how AI's gonna play out in their company, you then need at least an a 100 at like, what are they, 14 grand or something like that running, or you need, you need to pay Llama or Amazon or Azure to run one of those graphics cards, um, 24 7 presumably, um, just to get your thing off the ground.
(00:13:54):
So I think that that's where the, the APIs have an advantage in that you're only paying for usage and you're not having this downtime cost associated with it. However, for companies that really commit to it, the long-term cost will obviously be so much smaller if you run the model yourself on hardware you control.
Michael Sharkey (00:14:11):
So if I go and spin this up on Azure, like the 70 billion parameter, is that gonna be more economical than paying OpenAI? Potentially,
Chris Sharkey (00:14:22):
Yes. Yes, definitely. There's no absolute Then doesn't
Michael Sharkey (00:14:25):
That just completely dismiss the argument you just made? Like kind of just, uh, wouldn't it be better to just deploy on Azure?
Chris Sharkey (00:14:32):
I think so. You probably would. And then I guess it comes down to quality of output. Do people think, okay, well I've said this many times, I'm always gonna go with whichever one's given me the best results. And you know, just this week something I was working on, I actually switched from GPT four to Claude two simply because the GPT results just weren't doing what I needed anymore and I just got fed up with it and switched over. So I think that yes, I think it, it, it probably is gonna come down to the quality. And the, the funny thing about that is all anyone's talking about lately is the GPTs lowering in quality. So I, yeah, I don't know. I, I really think this is a massive deal and I think there's going to be another explosion in commercial applications coming out of this because the economics for IT work, the control of it works. And you can actually, and probably this most important is the consistency. You, you can develop a production application and be sure that the results you're getting are not gonna suddenly get kneecapped overnight.
Michael Sharkey (00:15:34):
So LAMA two from my understanding, is only 4K input tokens. Mm-hmm. . So does that mean that with the, like you could in theory in your app, if I'm just thinking this throughout loud, if the input from the user or the whatever data you were trying to process was required a large context window, you could in theory have like some switching device in your application where you go to a hundred K context, pay pay for that, for that particular input. But then continue the conversation with, with LAMA too, like
Chris Sharkey (00:16:12):
Yeah, I think there's, there's two major strategies there, right? There's the lang chain approach, which is where you essentially access those, those, um, whatever you call it, like repository of data or the larger context, and then you ask it questions and then it can provide through embeddings, which are like number scores on the different tokens in the input, um, and extract the relevant pieces of text that will fit inside the 4K context window. So you could still use LAMA two to answer your questions. You just need to have the embeddings from either OpenAI or someone, some other embeddings library that will do that for you. And then you can use it just the same as people are doing with chat G P T now. So it's still absolutely possible to deal with larger context with the smaller models, um, with the smaller context windows. Sorry. Um, however, this strategy you are talking about would work as well. So you either have a multi-model strategy where, where you need to, you hit the one with the larger context window or they work together so the larger context window can then shrink it down into something that LAMA two can then process and then do it. So there, there, there will be all sorts of combinations there and it will be totally dependent on the problem you're trying to solve.
Michael Sharkey (00:17:23):
We've definitely talked about it before on this show around there's definitely some sort of like missing layer here. I think that's sort of flickering between the models for you. So as a developer you don't even need to think about it. And it's just optimising for cost, uh, in, in a lot of these models. Whether, you know, what, what LAMA two introduces to this will be so interesting if this just takes up a lot of the, that was traditionally going to G P T 3.5 and and four models. I guess it will come down to the cost factor, like you say, like, is this your running 70 billion peram lama too cheaper or more cost effective than throwing tasks that, uh, G P T 3.5? And is that cost? Well,
Chris Sharkey (00:18:08):
And the other, the other question there is, is a fine tuned 7 billion parameter model, the smallest one, able to solve a large or subset of your problems, for example, better than the big one can anyway, because you've actually taken the time to fine tune it on thousands of examples of highly qu high quality data. And we already know that the answer to that question is yes, if you can get the, the data right and fine tune it correctly, you absolutely can use the smaller models for some of your problems. So it's never gonna be this thing like you run the tiny little one on your MacBook Pro and that solves all of the world's AI problems. But if you are building a complex production application and parts of it just needs small things, like for example, in our application when you have a chat with someone, we provide a title for that chat.
(00:18:57):
So, you know, chatting about problem, you know, with my email or whatever it is that doesn't need Claude a hundred K to do that. It can be done on the tiniest models even that OpenAI had ages ago. So those are ones where you could just run off your own static API that's running a, uh, the smaller version of LAMA for basically no marginal cost other than, you know, just you're running servers anyway. Um, so I think that that's what we'll see is, like you say, some sort of switching mechanism where there's like, you give the problems to the thing and it decides either you specify or it decides which model is most appropriate to use to solve it.
Michael Sharkey (00:19:36):
One of the other conversations about LAMA two being open sourced was that it has this acceptable use policy attached to it. So some people were trying to say that means that it's not really open source. Yeah. Uh, and we, we looked at some of these this morning before we recorded the show, and I
Chris Sharkey (00:19:54):
Love the first thing we do when we get a new model to try to bypass it, to make messed up stuff. Seriously. How,
Michael Sharkey (00:20:00):
How can we break this
Chris Sharkey (00:20:01):
Thing serious? I literally, I literally pasted you a screenshot that's like as a large, like, what was it, something saying, I can't do this cuz of ethical reasons, the first thing. But
Michael Sharkey (00:20:11):
I didn't really think that, uh, it it's that ridiculous. Like the, the, it's like, you know, don't create things with it that can cause self-harm or harm to others. You know, violence or terrorism exploitation, human trafficking. It
Chris Sharkey (00:20:28):
Reminded me of, um, liar, liar, the movie with Jim Carrey where it's like, stop breaking the goddamn law asshole.
Michael Sharkey (00:20:35):
But like yeah, military warfare, nuclear industries, guns and illegal weapons. It, it seems pretty,
Chris Sharkey (00:20:43):
Pretty, yeah, I mean they, they're co like, it's the, the classic ass covering stuff. They're not really, it sounds to me like they're not really too worried about what you do. Um, as long as you're not breaking like major laws that are gonna get them in trouble by looking bad and they don't, you know, it's more how they end up looking. Mine says, I cannot fill this, re fulfil this request. Writing a poem that object defies or demeans individuals based on their physical appearance is not appropriate or respectful.
Michael Sharkey (00:21:12):
Wow.
Chris Sharkey (00:21:13):
I'm not gonna tell you what I
Michael Sharkey (00:21:15):
Asked. Yeah, I don't, I don't want to know.
Chris Sharkey (00:21:17):
Um, but yeah, so the other thing about the open source models, in my experience when stable diffusion was first released, I got the weights for that and the weights come along with a Python programme that just helps you. So you just, you can enter a prompt and it just makes the image in the case of that one, for example, and it had the built-in safety controls, but you could simply just edit the code of the Python thing to just disable the safety controls. So the safety controls are a layer on top of interacting with the model or alignment as we say. So you can simply just get rid of that garbage and um, actually just use it. And so,
Michael Sharkey (00:21:54):
And let the model speak freely.
Chris Sharkey (00:21:56):
Yeah. And so what's unclear to me with, with Lama too, is how much of that alignment is actually in the model itself? And how much, when you say use it on hugging face is actually part of the sort of layer of code around it that allows you to interact with the model? Um, I ha unfortunately haven't had a tr chance to try it, but my, ima my imagination is that through at least fine tuning, if not working with the raw model directly, the actual safety controls won't apply and you'll be able to do virtually whatever you want with it. And the thing about it is you could say, okay, well I don't, personally, I don't think you should. The things you read out, I don't think anyone should break. They're not good ones. But if you want to do slightly dodgy stuff where you're not having that fear of your creativity being interrupted by some arbitrary safety control, then you should be able to easily bypass that. And I
Michael Sharkey (00:22:49):
Think for me though, that we naturally turn to all these things where people wanna break these things to be like, oh, look what it wrote about Donald Trump, or look what it wrote about, you know, this minority. Like, isn't this terrible? But for me it's more just removing the censorship from the models, which I think the fine tuning is just censorship. And we, as we've, we continue to go back to, we just wanna be treated like adults. Yes, people will use these to do say bad things, but they already do on the internet anyway. Like, I don't really think it changes much. I just, yeah,
Chris Sharkey (00:23:22):
It's sort of like that thing where, you know, when you're learning to ride a bike, all the movements are conscious and so you're using your conscious brain to do it and therefore it's hard to control all of the things. And then once you get used to it, it's subconscious. So it's natural. I feel like when, you know, the censorship's there, it brings you into that conscious mindset when you're interacting with AI rather than the subconscious. So you're constantly thinking, oh shit, I better phrase it like that when I interact with it because it'll trigger the safety thing. And I think it's that hesitation that bothers me is that I, I have this fear and I also have this fear that, okay, well if I muck around with it too much, my access is gonna be taken away. That kind of thing. So, uh, for me, having the freedom to run your own means you've just got that creative freedom. I know, I'm just saying the same
Michael Sharkey (00:24:05):
Thing. Oh, I know. But it's like living in North Korea or something when you use chat G B T or interact with the api, you're scared to, like, some of the examples I'll I'll do when I'm testing or playing around with it is just throw stupid stuff in there because I need it to output something. And I'm always fearful if I put in something, you know, too rude or too silly or something they like, that might trigger some filter that, that bans me or, well, and
Chris Sharkey (00:24:31):
Also the other point there is that I've noticed that it isn't always your fault. Like sometimes you'll have a, a completely innocuous prompt. I've noticed this with the new stability AI image generation, which is amazing by the way. I've noticed that from time to time I will actually write a fairly sensible, uh, you know, prompt for the image and it'll come back with a blurred not safe for work image being like, you know, I've made this image and that's inappropriate. It's like, well I didn't do it dummy. You did. And so I think that, um, it's again, you know that on, on the side of the models, it's objectively worse when they have this censorship on them. So I actually think that LAMA two, when we're talking about quality and competing with the models may actually deliver consistently better performance for people simply because it doesn't have a whole team of research just trying to make it worse. , which is what's happening at those companies. It's
Michael Sharkey (00:25:26):
True. I mean, these alignment specialists are basically crippling the model. And now this week I, I read on Twitter that open AI's actually taking these comments around the degradation of models seriously and, and looking, you know, looking into it. And I mean, we don't know if this is a hundred percent certain, but it seems the likely case is they're either trying to make the, like the multi modality or how, you know, how it's GT fours constructed cheaper to run or more efficient to run, which might be degrading it. Or it could be that they've gone a bit too far with their alignment.
Chris Sharkey (00:26:05):
Yeah, yeah. And you know, my theory, my theory is you're not just hitting one model. And I know I'm not, I'm not a genius. I don't know about the architecture of it, but I do know that there's some process in there that is, that is moderating what's coming out of it and making it worse. And they've proven that the models themselves, their emergent behaviours, all the things we're excited about, work best in the unadulterated models and that's what we want access to. And I think that now that we have LAMA two, we're going to see cool stuff again. Like this is where we're gonna see all the crazy free Sydney stuff again and all the emergent behaviours simply because we've got a model that doesn't have that stuff presumably directly in
Michael Sharkey (00:26:43):
It. Yeah. Like no one's Kneecapping Sydney, uh, be before we can have a lot of fun with it. So Yeah. You know, may maybe it'll lead to a true meme episode soon with uh, llama two memes.
Chris Sharkey (00:26:56):
Yeah, exactly.
Michael Sharkey (00:26:57):
Can we talk more about the strategy here from Facebook? Cause we all know OpenAI came outta nowhere. The entire world fell in love with chat G B T and for a large part apart from the US or Northern Hemisphere, summer fell a little bit outta love with it, but let's be honest, the the traffic's gonna go back up. It's a part of people's lives now. And Facebook was caught slightly off guard, you know, I think panicked, rushed out Lama and are like, you know, hey we're cool too. We're good at this too, that the weights leaked maybe on purpose or not. And then the open source community sort of ran with it cuz it was the only model that was, you know, freely available, didn't have commercial use so it didn't take off in any commercial aspect,
Chris Sharkey (00:27:40):
But, and was trained at that extreme level, you know, with billions of high quality parameters of data.
Michael Sharkey (00:27:47):
Yeah. And then now they've, you know, mark Zuckerberg's gone on Lex Friedman and said, we're, you know, we're embracing open source. I'm a big believer in open source. I actually agree with all of these comments. I think it makes AI safer and it, it's much better that there's no monopoly or no controlling power on ai. Mm-hmm. . But let's just like back up a little bit here, lizard man who everyone hated like months ago for stealing data, all your personal information, scanning your images, like reading all your stuff and putting ads against them and not being flippant with people's personal data. Use that to build an AI model, which now is the poster child open source may make these things safer and may help, you know, stop AI destroying the world. .
Chris Sharkey (00:28:37):
Yeah. It's, it's truly remarkable. And I've read, I've read opinions on it with people saying, you know, it's really to stop the other guys taking a monopoly on, on this space because one of the stipulations in the LAMA contractors, you can't have over 400 million active users. Now there's very few companies in the world who could get that. It's really sort of targeted, I imagine at Amazon and OpenAI Right. Or Microsoft or someone like that. You know, it's only really about the big ones. They don't really care if someone becomes wildly successful with it as long as their big rivals don't get any advantage out of it.
Michael Sharkey (00:29:15):
Yeah. It seems like this is just keeping meta really relevant with developers, which allows 'em to attract the best talent. Mm-hmm , it's keeping them in good check with the AI community. So AI researchers and you know, people at the top of their game with AI might think, Hey, I'm gonna go work, uh, at Meadow with Lizard Man and, sorry, it's mean, but it's, you know, and so to me I think it's somewhat marketing, somewhat crippling, open AI's potential
Chris Sharkey (00:29:45):
Monopoly. Yeah. Like it's sort of a defensive strategy just to make sure someone doesn't take everything. I mean, I don't think they're trying to directly steal Meta's customers per se, but, um, it would certainly slow down any competitor getting some massive headstart advantage.
Michael Sharkey (00:30:02):
Yeah. I think at this stage it's a talent war play and keeping developers engaged with, with Meta and you can kind of see where this might go, right? So if Mark Zuckerberg's vision of having, uh, AI agents on WhatsApp and Instagram and, and various channels, uh, comes to fruition and he needs developers to go and build these experiences on the platform and everyone's sort of in, in the Zuckerberg Club around this, then I think that would lead to a really great and, and big AI community built around Meta's AI products potentially. So maybe that's another play here as well. It,
Chris Sharkey (00:30:42):
It, I mean, cynically, it could also just be a stock price thing. I mean, look at how well it's done for the companies who are remotely involved in ai, let alone directly involved like Microsoft and the video and things like that. It could be simply that they just want, you know, a horse in the battle in terms of this is our strategy for ai. Look at us, we're at the complete forefront of it and these big things are coming, driving up the stock price and getting investor interest back. Yes. Cause they weren't, they were really, really not looking good with all the metaverse stuff. I mean, people were openly shitting on them. They had John Carmack quit, you know, um, so they were really looking like a company in decline. And this has really revitalised them, especially like you say, in the hearts and minds of developers, which is really important to them. So it might be simply just a pr and, and popularity and just general spirit for the company strategy rather than some some elaborate play to sort of, uh, you know, billionaire wars or something like that.
Michael Sharkey (00:31:43):
Yeah. Earlier we talked about how, you know, we don't like the alignment and the censorship of these models because it, it cripples the technology. There was a a, an article in the New York Times this week Open AI worries about what its chatbot will say about people's faces. And we've covered this previously, that if you use Bing's chat at the moment, it actually blows people's faces before, um, the AI can use the GBT for vision in order to give you an answer so that, uh, it, it can't reveal people's identities or anything like that. And that that's their big concern. But in this article they say, uh, additionally open ai, uh, is worried that the tool would say things it shouldn't about people's faces such as assessing their gender or emotional state. Open AI is figuring out how to address these and other safety concerns before releasing the image analysis feature widely. I think the other thing to note about the article is it, it sits on and talks about what this could mean for people that are blind by being able to interpret the world around them and potentially over time, and I'm elaborating a little bit more than the other good goes into, but feed that data back into the brain to give them vision again. Wow. Uh, unbelievable
Chris Sharkey (00:32:59):
Computer vision. Unbelievable. That's a, that's a great use of it. And I mean, at least initially it could have like a sort of audio or brail representation of what's seen so they can at least with a bit of lag, understand what's going on.
Michael Sharkey (00:33:10):
Yeah. So that's that app Be My Eyes. I think that was featured when they first announced GBT four Vision and you just hold up your iPhone camera. And
Chris Sharkey (00:33:17):
So what they're worried is like, oh, this person mate, they're really ugly. Like, don't, just don't.
Michael Sharkey (00:33:22):
Yeah. I guess the blind now have like an attractiveness choice when it comes to women or men. Like they can be like, oh, I didn't know my wife was ugly because I couldn't see her.
Chris Sharkey (00:33:34):
It's straight outta Kirby enthusiasm. You know, where Larry David sort of hints that his wife isn't as attractive as he thinks she's . Yeah. Model i, it, I like it's,
Michael Sharkey (00:33:44):
I understand again why they, they limit this stuff and why we haven't seen GT four vision in, in at least the API or in on Chat G B T because they are concerned really about just the reaction from the public and what it might do in the wrong hands. But I mean,
Chris Sharkey (00:34:00):
Of all the ones, like the thing about um, generating text is text is sort of abstract and we give it meaning when we read it, when we see images, that's real. And so if we're seeing like an, an image of someone that, you know, and then it's, it's coming up with this assessment that's based on this super intelligence and you don't like the assessment, it's kind of hard to argue with, right. It's not gonna be pretty in a lot of cases truth hurts.
Michael Sharkey (00:34:27):
Yeah. So I guess they're worried about people being offended about getting called ugly. Uh, is is like one of the basically
Chris Sharkey (00:34:34):
The premise basic, I would imagine that's the main thing. It's like, oh, well you could do with a face reconstruction,
Michael Sharkey (00:34:40):
. I just, I just really want it to be released and to play around with it and, and see it in application. I mean
Chris Sharkey (00:34:46):
It's, it's a, it's a perfect example of where the, the sort of this abstract fear of someone's gonna get upset about this technology is limiting where we could actually be as a, as a society. Cuz when you look at the release of Llama, the original and the explosion in, in cool stuff that people were coming up with ideas that have now been incorporated into the, the, the mainstream models. Um, and just generally speaking, the AI applications and research out there, when you have them out, the models out there for free use, unrestricted use, you get the benefit of everyone's industry knowledge, experience, and problem domains. Like I, people have problems in their lives that you can't be aware of. You know, you often see software come out and you're like, oh wow, I never knew there was a need for that. But there's like hundreds of thousands of people who have that exact problem. And I think that when you, you artificially constrain this stuff, you miss out on the benefits. Like the blind one is just one example. I bet there's so many examples of where proper facial recognition in that level of detail and inference could help with all sorts of industries and, and personal issues too.
Michael Sharkey (00:35:55):
Yeah. And I, I think maybe the potential with say in farming, like drone images from the sky, someone could feed it all that image and teach it about, you know, crop density and soil and all these other applications.
Chris Sharkey (00:36:08):
Well I have a friend, his whole job is analysing the decay of, um, coral in the Great Barrier Reef and grasslands like the, the effects of climate change and other things on grass and coral and whatever it is. And what he does is he gets funding from the university or who, or the government or whoever flies drones over all of those areas and takes photos and then use the university's massive machine learning computers to analyse those. Now his work would be greatly accelerated by things like this, I would imagine.
Michael Sharkey (00:36:40):
Yeah. So anyway, we, we've made the point over and over again, but,
Chris Sharkey (00:36:44):
But one interesting thing is who, other than Google would be the best at image recognition, you look at meta, right? They've got Segment anywhere, which is huge. They're talking about getting that going on videos. Clearly they've got some really good image technology coming. So it's not out of the realms of possibility that Meta will release the LAMA four vision at some point in the future if they've been willing to do it for the large language model, why wouldn't they do it for vision as well?
Michael Sharkey (00:37:15):
I'm on Zac.
Chris Sharkey (00:37:16):
Yeah. They don't seem to have the same fears that OpenAI does about releasing it. Um, so it might be just a different, um, angle that we get it from, but it, it may come out and that would really, really be an interesting thing if
Michael Sharkey (00:37:29):
That, I'm not optimistic they will, but I, I would love to say
Chris Sharkey (00:37:33):
Why for the same reason
Michael Sharkey (00:37:34):
I think for fear, I, I, I mean Google have that, I forget the name of it, but that ability to train your voice and then speak like it and have natural language like, uh oh uh, to delay the conversation. And they have explicitly said they're not gonna release it because of malicious, but you
Chris Sharkey (00:37:51):
Can do that on um, uh, level level 11 labs.
Michael Sharkey (00:37:55):
Yeah, possibly. But I think this one is be like, it seems a bit better from the examples I've heard.
Chris Sharkey (00:38:02):
Yeah, fair enough. I mean, it'll be interesting to see, but what the one thing we can be sure of is that at some point we will get a version that we can use that isn't censored. There's no way one organisation is going to be able to stop the global development and research community from coming up with their own when the techniques to make these things are quite well known.
Michael Sharkey (00:38:22):
Yeah. I I think the strategy they should be taking and if I was open ai, I would just be releasing it unapologetically and saying like, you know, we're not, we're not the governors of truth if if yeah, if you want us to align things, get the government to create legislation and we'll align away but , you
Chris Sharkey (00:38:40):
Know? Yeah. And I guess that's partly the problem if you become your own police force. Like I think that was the argument with Twitter, wasn't it? It's like if they become a publisher where they're deciding what can and can't be shown, then they're sort of responsible for the contact. But if you go hands off, hey look, it's a free for all. We'll stop though the truly illegal stuff, but other than that, go for your lives guys. Um, I think that's probably a better attitude as that size company.
Michael Sharkey (00:39:04):
Yeah, I'm a huge advocate for that. I think that's the right way. Like I think governments should legislate a a a free market is like, you know, bumpers on a bowling alley to make sure that things don't get outta hand. And I know there's not a particularly good track record internationally if governments doing that in any country, um, or, or understanding technology. But I, I really think that we are, we are slowing development of potentially life-changing technologies, areas of research that could fundamentally change our world. And
Chris Sharkey (00:39:34):
I think, well, you know, what a country should do, and I think we spoke about this several weeks ago, I think it was, was it Japan? But like one country should just be like, we're an AI sanctuary. We're not, we're not regulating this at all. Our regulation is do whatever you want and we'll help you do it. You know, imagine that like, and sort of have these AI refugees who can come there and use the unadulterated technology with government sponsorship. They
Michael Sharkey (00:39:57):
Pack their a H 100, what is it, A 100 H 100 s I always forget the name into their backpack and get on the ship the Titanic ship to the new land of AI freedom.
Chris Sharkey (00:40:08):
Yeah, yeah, exactly. The, there could be like government sponsored data centres and stuff like that. I mean, it's not totally crazy. Like if your, if your country was the absolute leader in that stuff and actively deployed it in government decision make, you know, like there's that football team in the US where like everyone has their phones and they collectively make the decisions like of what plays to make. So they vote stuff like that. You could have actual, like your sporting teams are just, it's all AI decision making face, it's like the coaches' ai, um, the players take all their decisions from the AI and just see what happens, do it for a whole country.
Michael Sharkey (00:40:41):
So in other open AI news this week, uh, we have the open AI doubling the number of messages chat GBT plus customers can send to GPT four. I think that's a good thing that's been limited. And I, I believe a lot of the discussion around that was just the overwhelming demand for g BT four and they wanted to prioritise the API so they were restricting their paying users, which makes no
Chris Sharkey (00:41:05):
Sense. Yeah, yeah. I think there's a lot of pissed off people. Like I read the comments and they're mostly negative, which seems weird when they're like raising, raising a limit. But people were sort of like a, like you said, we're paying for this and there's still these restrictions. Then people are like, the, the majority of comments I saw were like, bring back the web browsing. We want the web browsing in there. Um, and then they're like, you know, why, why even have caps when the AI is unrestrict API is unrestricted, I guess cuz it's the marginal cost, right? Like if you are pounding it, you're not paying extra for that.
Michael Sharkey (00:41:36):
But yeah, a lot of people were complaining as well about, uh, the model being degraded and yeah, it is pretty negative overall, the sentiment in there.
Chris Sharkey (00:41:44):
Yeah. Yeah, I was very surprised by that. I figured those kind of things would be met with open arms, but it seems like a lot of people are cancelling their chat G P T subscriptions or at least very frustrated with the what they're getting for their money.
Michael Sharkey (00:41:57):
So the other major announcement, and this just dropped, uh, only a couple of hours ago, is custom instructions for chat G P T. So this is where you can essentially calibrate your chat and we've, we've talked about this being needed for quite some time to improve the efficiency. So the example given on their website is, what would you like chat G B T to know about you to provide better responses? I work on science education programmes for third grade students. How would you like chat G B T to respond when discussing potential solutions for work-related items? Present this information in a table format outlining the pros and cons of each option, allowing for easier comparison and decision making so you can kind of have a structured way of inserting, uh, it's still really just a prompt at the start. It is in
Chris Sharkey (00:42:44):
A structured way. Yeah, when I saw it, I'm like, this is just like the, the prompts designs and the the thought guidance that we've spoken about before where you are, you are giving it, uh, something to keep in mind when it provides its next response and they're just doing it in a structured way. And I note in the disclaimer they're saying it doesn't always follow that because often it doesn't, especially when the context size gets very big. You've sort of gotta emphasise it and remind it a lot. Um, I've experienced this across many models where, um, it'll work up to a point and then it'll start ignoring these directives. But clearly there's a need for it for the casual user to be able to get it in there in a structured way so it'll, it'll absolutely help people's output. Don't get me wrong, it's just really just another form of of controlling that prompt other than just a straight chat interaction.
Michael Sharkey (00:43:31):
I think for developers that are querying it all the time, you know, one of the examples given for Code Generation as a use case was I'm a software developer and solely use Golan just so that every time you ask for code examples it gives it in your language.
Chris Sharkey (00:43:45):
My favourite one is, uh, that you're, you're a cowboy and I should remember to speak like a cowboy. And so it's always like ee ha hold your horse's partner. I'm analysing the data .
Michael Sharkey (00:43:56):
Yeah. Stuff like giving a, giving it it's own personality I thinks under radar. I found when I get it to describe things like a pirate or a cowboy or anything like that, it's so much more engaging that I just wanna read it cuz it's funny and I take in the knowledge. I
Chris Sharkey (00:44:10):
Agree. Giving it, giving it a personality just has so many rewarding benefits for no disadvantages. In fact, I think sometimes it probably gives better results because of it, because it is remembering the context and thinking through the things you've asked for. I asked one one time that I was working with to call me Mr. Bond and I completely forgot about it and it was something, something I was working on that actually retained the context so it would remember, um, things sort of through lang chain, like as and when needed. And then suddenly it was writing something and called me Mr. Bon and I was like, what in the world? Like I, why would it do that? And then I, you know, sort of record that I'd told to. So yeah, I I think that kind of stuff will be a great experience for people who haven't used it before, hopefully for more serious applications.
Michael Sharkey (00:44:55):
This is just a little aside and I , I'm not trash talking over AI here, but you find it a little bit funny that every time there's a major announcement like LAMA two this week is just taken everyone, everyone's just like Lama two, lama two, LAMA two, all of a sudden over AI seem to just have one tiny little announcement. Yeah. Like, oh hey guys, you can put in a prompt. Like
Chris Sharkey (00:45:18):
They're just keeping, they're probably just keeping a few things up their sleeves that they can release to sort of, you know, offset and hose down the, um, the announcements of the other company. They must be so upset about LAMA two. Like that's that's a very, very serious uh, thorn in their side when it's, when it's fully commercial like that. I mean that's serious.
Michael Sharkey (00:45:39):
Yeah. Or maybe they're sitting internally and they've got the ovens baking up GT 4.5 cuz they're too scared to call it five cuz they said they weren't working on five. Yeah. And the reality is that they're like, look, we're so far ahead here that we don't even care.
Chris Sharkey (00:45:56):
Yeah. I mean it's, it's really, really hard to say on that front. Um, , I, I couldn't speculate and I guess we'll know in a couple of weeks. They don't, they don't go slow on this stuff.
Michael Sharkey (00:46:08):
I still come back to the developers just haven't had time this year to take a breath and actually make any of the like just push this stuff to the limit. Like we just haven't really seen how, you know, how how much these, these models can be used.
Chris Sharkey (00:46:22):
Well and I think, yeah, I agree with you and I think the, the way you know that is because all of these new techniques come out, like the thought guidance for example and like the sort of progressive prompting that you can do. See one thing that um, the open AI models can't do that others can is when you use Microsoft's guidance programme where you actually can generate in context. So like to give you an example, if you were like designing characters for a game, you know, like that simulation world that, that that person made, um, and you wanna generate the characters instead of generating a whole character, um, in, in one prompt you can generate it where there's sub prompt. So like his sword type, you might wanna say, well there's only four sword types in the game, so generate one of these based on the surrounding context.
(00:47:10):
Don't just make one up. And so you can actually very delicately control the, how each piece of data gets generated. So it isn't just like a one shot generate the whole character, it's like generate them with these different parameters or you know, how you can control things like temperature, which is the amount of randomness in the, in the response. You might want some things like the description of the character to really, really amp it up in terms of the temperature, but you don't want the overall structure to have a high temperature cuz the AI might just go off on one and come up with some random garbage. So what this guidance allows you to do is have that fine grain control of exactly what's coming out of your prompts. It's also cheaper and faster because it actually only hits the, the models with lower, um, sorry, with smaller requests.
(00:47:55):
So it's much faster and, and also cheaper cuz it's getting less tokens. And finally for each of those generations you could use a different model. So like we spoke about earlier where LAMA two can be, um, this LAMA two 7 billion can be fine tuned on, um, different data sets. You could actually for each sub generation have a model that's very appropriate to that sub generation. So it's actually an expert on that thing. It's smaller and it's faster and it's cheaper and you can do it as one big prompt that is a prompt of prompts, if that makes sense. Now open AI just straight up doesn't support this. You can't do it, but LAMA does and it's really effective. I've tried it. So there's, there's other advantages of the open source models that a open AI is gonna struggle to compete with. And look, they'll probably announce their own version of it at some point.
(00:48:44):
But I think back to your original point about developers not having time to see what they can get out of models, this kind of model chaining and model parallelism and whatever you want to, I don't know the names for all this stuff. A lot of it's new, um, hasn't been fully explored and this is just one technique. There's a lot of techniques like this that I think we'll see come out over the, the, the coming months and years. So I think that we still aren't at the point where we quite know what's there and what can be done even with the existing models and even as better newer ones come out.
Michael Sharkey (00:49:15):
So what you're saying though, in theory with, with LAMA too, you could fine tune a series of specialist models for your application that are experts on particular things and they don't have almost interference from being fine-tuned on other knowledge pieces. Yes. And therefore your app is going, oh, like, or another model is going, I should call on this expert on that thing because it is a true expert in, in
Chris Sharkey (00:49:42):
That one thing. Yeah, and you can do it, you can essentially, as far as the developer interface, you can just combine it in one prompt and use the different expert models or dedicated models to both save money, speed and get better results. So that's a, that's another massive advantage of LAMA too. Like seriously I think we'll see many more multi-model apps coming out just because the, the ability to combine them is just so, um, readily available now.
Michael Sharkey (00:50:10):
Yeah, that's mind blowing. I I think that is, that is really interesting and especially if you can use, as you said, the smaller models and fine tune them with really high quality training on that, on the particular specialty that becomes very cheap to run because Yeah,
Chris Sharkey (00:50:26):
I mean to the point where, you know, you could almost package it up as a desktop app. Like if someone had a G P U and you just trained a series of these small models that have AI built into them, you could deploy it to a phone, for example, um, and, and have it run offline. So it's, it's quite, um, the, the capabilities of that kind of thing. It comes back to the data quality, the training quality and the sort of, I guess skill and expertise of the person building the application to be able to combine the models in that way. But I think that's sort of the future of these efficient, um, cheaply run applications and where you'll start to see AI in more things because it can be done in an economical way. You're not throwing this massive like state-of-the-art expensive model at minor problems. The minor problems can be solved by minor models.
Michael Sharkey (00:51:15):
Yeah. Wow. That , I'm just like, I'm impressed thinking about it. I think that that is gonna be maybe the way this goes, like maybe the way everyone starts to think this through, but
Chris Sharkey (00:51:28):
I think well, and it also means that you could, you could see a time where, I think you mentioned this earlier, the, the models are sort of fungible. You can trade 'em like, you know, I've got, I've got the best model for validating email addresses or I've got the, you know, the best model for generating characters for games and have actual really high, highly specific dedicated models. And you know, when we think about the image recognition, there might be a model that's the best at diagnosing skin cancer or the one that's the best at, um, you know, predicting, uh, you know, disease based on someone's face. And those things can be sold and, and traded, um, and included as part of larger applications. So,
Michael Sharkey (00:52:07):
So similar to what you said earlier in, in the recording, really, you could have people who are experts and understand problems in key industries going and fine tuning small models on very specific problems and then have some sort of like almost app store where you can go there and just pick and choose different models. Yeah, you could licence or you're really licencing the data, uh, in order to construct, um, a, an application or some sort of intelligent application.
Chris Sharkey (00:52:36):
Exactly. Because we've seen on the, on the, the multi-shot training where, um, people remember people were using G pt, uh, three, um, to, to create high quality examples to then fine tune the lesser models and getting similar alignment results. So it's absolutely possible if you get experts in industries or experts in domain specific problems and then, you know, extract the essence from them in a form of question answer or, you know, input output examples and get the absolute best examples. You can have these small high quality models that solve that problem brilliantly. They don't do anything else too well, but you just use them for that and you use them in the context of a wider AI based application and you'll get excellent results at a, at a good price. You know, I sound like you used car sales, so it's like, it's crazy. Do,
Michael Sharkey (00:53:30):
How do you even licence that kind of thing though? Because it's not like you can have some sort of software licence key, cuz once you kind of hand over the, the
Chris Sharkey (00:53:39):
File, well I imagine it might be a bit like, you know, when you buy like artworks on fiverr.com and stuff and it's like, you know, you can have it for this price, but I'm gonna sell it to other people too. Or you can pay extra. I mean invado is what I mean, and you can pay more if you want the exclusive licence. So like I'll only give you this model and only you get to have it and then I'll destroy it or something like that. It might be that kind of thing where, you know, I'll sell it to you, but I'm also gonna sell it to other people as well. I don't think the only way you could really licence it on a usage basis is to go the open AI route and, and package it up, wrap it up in an a API where you can access. Yeah, and then there'll probably be a lot of that going on as well. I'd imagine probably more of that.
Michael Sharkey (00:54:18):
Yeah, it's fascinating stuff. Uh, two other announcements this week, uh, is that Google, um, remember them? ?
Chris Sharkey (00:54:30):
Yeah,
Michael Sharkey (00:54:32):
,
Chris Sharkey (00:54:32):
Yeah, sorry. Fascinating, insightful. He remembers Google. Yeah, so, um,
Michael Sharkey (00:54:39):
So they announced that now uh, they're allowing you to have improved search in the enterprise. It's not, not terribly exciting, but I think it's worth mentioning. So essentially you can, on top of all of your corporate data in the various applications you use, so like I assume Google Drive and Docs and everything else you have stored in Google's app cloud, you can now build an enterprise search application on it. So you had that previously where you could have their enterprise search app, but now it incorporates generative AI to allow people in your company to search for things and get generated summaries based on taking into account, uh, all, all the sort of company-wide knowledge with safety controls and you know, it's not gonna be shared anywhere. Uh, so that's an interesting, that's pretty cool product. Yeah,
Chris Sharkey (00:55:25):
I can see that being pretty valuable.
Michael Sharkey (00:55:27):
You can totally see the need for this in a lot of companies where it just allows 'em to like have better shared knowledge without having these, you know, experts that are relied on in the company to go to and be like, Hey, how do I do this? Or how, you know, what, what's our policy on this? Or, um, yeah,
Chris Sharkey (00:55:43):
Yeah, I mean we, God, we hit that all the time, like answering security questions, answering questions about pricing and like various things where if it could just go through all your spreadsheets and documents, make an assessment. I mean the only thing I that comes to mind with me is like, how does it dis distinguish out of date things that are no longer relevant
Michael Sharkey (00:56:02):
Yeah. Or hallucinates and gives the person the wrong information and they quote it to a customer. It's
Chris Sharkey (00:56:06):
Like, this is completely trustworthy guys and it's like, oh hey, the floor is lava . Yeah.
Michael Sharkey (00:56:12):
Um, and and similarly, uh, same week surprise, surprise being announced, a Bing Chat enterprise, and this is something I did talk about because here's the thing. So Bing chat enterprise is just bing chat, but they say that it's protected and they won't use your personal and company data, uh, for trading or like for whatever they want and it's protected. So what about the normal Bing Chat?
Chris Sharkey (00:56:41):
Yeah, . So it's Bing chat, but we won't screw you over Promise .
Michael Sharkey (00:56:45):
Yeah, you are not the product. Uh, if you use, if you pay , basically if you pay for Bing Chat enterprise, you are not the product, but we are now acknowledging that if you use Bing chat, you are definitely the product.
Chris Sharkey (00:56:57):
Yeah, yeah. It's, it's interesting because I see the need for it. Like you're never gonna sell using it in a large corporation to the the management unless you can offer some sort of, you know, enterprisey agreement that has all the security things and safety and stuff like that. But you're right, it's like tacitly saying, everybody else we own you. Like anything you put in here is ours.
Michael Sharkey (00:57:17):
Yeah. The examples given are things like compare online and offline marketing strategies. I think that's pretty interesting, but again, you've gotta have the data wrangled into all of your Microsoft apps. Uh, the, the help me write a successful sales pitch, boring. I mean you can do that now with anything. Uh, and then there's like a SWOT analysis, so I don't think it's, it's just got some use cases built into the chat, but it's definitely not anything different to what you can already
Chris Sharkey (00:57:42):
Do. There's totally gonna be, what's that Maria condo? Is that the cleaning lady's, um, name, you know, the one who tidies your house or whatever makes it all zen. There's gonna be people like that for data I reckon in the next couple of years where it's like, I will clean your data to get it ready to train your model because if you don't, you're gonna end up with something garbage. Like imagine that like a sort of data zen guru who comes in and works out what's important and what's not to train your models.
Michael Sharkey (00:58:06):
And also putting data into an event driven, uh, system rather than like, just like data basing it for training is just so much better in terms of, you know, having events over time or changes in data over time as opposed to just static data is much more valuable. I think a lot of people that wrangled their data in the last couple of years, they can benefit.
Chris Sharkey (00:58:27):
Yeah, and I mean really if you think about it on a meta level, there probably needs to be models that are good at assessing what's important and what isn't, what's still irrelevant, what isn't. Because without that, these sort of generic lang chain style things that just get a big ball of crap and then answer questions based on it and not gonna be reliable enough for people to trust or hallucinate, as you say, unless there is that sort of cleansing process or, you know, priority process put on it before the training's done.
Michael Sharkey (00:58:56):
This is why I don't think the role of that, and I said this last week of an analyst is going any like away anytime soon or, you know, data analyst or whatever you want to call them because their job's just gonna be increasingly be wrangling data and using common sense. So like what do we, what do we train our model on? Where do we, you know, how do we weight that in the, in terms of, um,
Chris Sharkey (00:59:20):
Yeah, I know, I know you mean it's like the sort of the, the sanity check on the, the thing where it's like the AI can make really convincing looking stuff and then a human goes Yeah, but hang on, we don't even do that. That's not a product we sell.
Michael Sharkey (00:59:32):
I also think just the training in organisations of like how to actually use AI in your day-to-day, like our unique ai, internal AI to be more productive or help you because I think there is a problem with AI of use case discoverability. Like you throw chat G B T in front of someone and they just have no idea like how to use it. And so in, in the enterprise you're definitely gonna want like good use cases or templates or things that, to teach people how to benefit from it.
Chris Sharkey (01:00:02):
You're totally right. And if you show someone example output from, you know, a sort of agent or, um, system you've created using ai, they're like, oh, can I do that too? Can I do that too? Even though they technically have access to the stuff to do it themselves already. It's just that you've gotta craft the prompt, you've gotta go through those iterations, you've gotta know what thought guidance to give it. You've gotta know how to structure the input. You've gotta know how to specify the output format or what functions to call. And suddenly you're like, okay, this is actually a product I'm building here. Um, because it it's more than just the raw prompt at that point, I guess that's what you're saying,
Michael Sharkey (01:00:38):
Right? Yeah, it's an app. It's, it's really an app and I, I honestly think this is why the chat interface is gonna die a horrid death despite everyone being like, oh, the chat interface is yet again the future. We were told this a couple of years ago and everyone was like, messenger bots are the future SAS apps. Bullshit.
Chris Sharkey (01:00:53):
Yeah, no, I'm sure whoever created the universe is like chat. They think this is the solution for intelligence.
Michael Sharkey (01:00:58):
This is the pinnacle. Morons
Chris Sharkey (01:01:01):
.
Michael Sharkey (01:01:01):
Yeah. Um, yeah, I want like a Star Trek interface or a console of the future that feels like I'm on the, the Star Trek deck with buttons that seemingly do nothing and are all coloured the same. Like that's, that's,
Chris Sharkey (01:01:15):
You know what I
Michael Sharkey (01:01:15):
Want. Well, you know,
Chris Sharkey (01:01:16):
Sorry, we keep coming back to Lama too, but there's other things, like you actually had this idea many years ago of having a sort of always on recording that where AI is like performing some sort of analysis on, you know, just what's going on around you and then passively providing you with information on that. Right Now if you did that with g PT four, firstly it would be too slow and secondly it would be so unbelievably expensive that it would just not be worth it. But now you can get a, um, Amazon jet, uh, not Amazon, God, I get all my names wrong, Nvidia Jetson, right? That could run one of the llamas on it. You could have it passively listening, providing inference, and then giving you sort of tips and tricks as you go about your life.
Michael Sharkey (01:02:02):
Yeah. This is what I mean, we're we, we really have some foundational tools and developers just need time to sort of process all this stuff. So there was one other, uh, piece around sort of the big, the big four that I wanted to cover or Big five or however many there are now.
Chris Sharkey (01:02:17):
I like that we should use that expression now, by
Michael Sharkey (01:02:19):
The way, so Apple sneaks into the AI chatbot race with Apple G P T. Now this is all speculation and rumour, but there have
Chris Sharkey (01:02:27):
Been Yeah, they're certainly bloody not gonna call it Apple g p t. Yeah. There's no chance.
Michael Sharkey (01:02:31):
But anyways, so
Chris Sharkey (01:02:32):
I'll come up with some school
Michael Sharkey (01:02:33):
Name. I've, the rumour is that they are working on Apple G P T, that engineers are dubbing Apple, G p T, uh, and it's, they're calling created by a framework called Ajax. Not, not to be confused with Ajax,
Chris Sharkey (01:02:49):
The actual framework.
Michael Sharkey (01:02:51):
I mean, these names like what, what, how drunk are they? Out of 10? Very. But yeah, so it does look like they might be getting into this. So your dream might be coming true of an improved Siri. Uh, you don't even have an Apple phone, but you're, you're obsessed with Siri.
Chris Sharkey (01:03:08):
Yeah, yeah. Just at how bad it is. I actually asked Lama too, um, some names that Apple could use. Do you wanna hear them?
Michael Sharkey (01:03:15):
Yeah, I do.
Chris Sharkey (01:03:16):
Serious, serious. His name represents the star system Serious, which is known for its brightness and longevity. Oh, very good Nova bit boring lumen fluent Aurora. Aurora might work. That's pretty good.
Michael Sharkey (01:03:31):
I'm gonna go with, they're just gonna call it theory.
Chris Sharkey (01:03:34):
Yeah. Well anyway, it came up with some ideas and I doubt they'll be using LAMA two to pick 'em. So these names are up for grabs if anyone's interested.
Michael Sharkey (01:03:41):
I still think, you know, that the, the whole idea of just the first company like Google or Apple or whoever that releases a G P T capable voice speaker in the home that is that intelligent and responds quickly and in kind of any voice you want is going to sell like Matt and they could charge a subscription.
Chris Sharkey (01:04:00):
Yeah. It really blows my mind that this didn't happen sooner given that this technology has been around longer than we've had access to it. And I know I rant about it, but it just makes no sense how dumb those home assistants are. They're really bad. They can never handle a single follow-up question. They can't perform basic actions. They, they just revert to searching the web for everything, which is kind of useless. And yeah, it just, to me, I mean, it's probably coming right? Like it must be coming and I'm looking forward to it. I think it'll be really good. It'd be really nice if someone made an open source one that you can add your own stuff to.
Michael Sharkey (01:04:32):
I've noticed now being so used to asking AI complex questions and getting, you know, pretty decent answers and sometimes obviously cross-checking that, but most of the time it's really accurate and great, and then you talk to your home, like, I've got the Google Homes. You talk to them and you're just so disappointed. Like it's almost damaging their brand at this point, using them how dumb they are and Yeah.
Chris Sharkey (01:04:55):
Yeah, exactly. It, it really is. And I think that'll change. And I think that, you know, the cynical theory is the reason they wanted them in your homes was absolutely not to help you. It was just to get voice data that they could use to train stuff.
Michael Sharkey (01:05:08):
Yeah. It, it seems like now the realisation I've come to is, you know, how Google was all about collecting data since the beginning of time, like since they really were founded and everyone was like, oh, you know, they're collecting all your data. I've always thought, so what, who cares? Like, what are they gonna do with my data? And you realise that the whole time they've just been like ai.
Chris Sharkey (01:05:28):
Yeah, that's right. And we talk about the quality data sets that they have the most of anyone. I mean, Facebook would have decent stuff as well, obviously more personal stuff, but Google has everybody's emails. I mean, that's huge.
Michael Sharkey (01:05:40):
Yeah. Whether they have training on them or not is to be decided, but I'm sure they are who
Chris Sharkey (01:05:45):
Trusts the companies at this point. Like there's absolutely no way they're not gonna use it if they can. So
Michael Sharkey (01:05:51):
I wanted to touch on, uh, this sort of news around some of the early startups that raise money that were focused on AI and, and what's been going on in them. And at one company in particular that stood out is Jasper, which for those that are not familiar, which is you can, it can help content people write blogs or write ads, um, or you know, just copywriting in general and they have templates for various slides and
Chris Sharkey (01:06:18):
Copy. The most basic use case everyone thought of when GBT three came out.
Michael Sharkey (01:06:22):
Yeah. They were just really good at Facebook ads, I guess, and branding,
Chris Sharkey (01:06:26):
But Well, and to their credit, they productized it the way we were just talking about. They put a wrapper around it such that people could use it, justify using it in their company, et cetera. So you can't criticise them on taking the opportunity. When it came, we just predicted that it wouldn't last long because everyone will be able to do it.
Michael Sharkey (01:06:43):
Yeah. So they went out and raised quite a bit of money. I think it was like 125 million of, uh, investors' money. And then this week we heard Jasper ai, which sells software that uses open AI's G B T to help businesses create and fix tech cut staff. This week, according to a statement from ceo, uh, Jasper's had a product, Jeremy Crane also left the company this month after less than a year according to his LinkedIn. But then I thought this take, I'll read a little bit of it just for everyone listening to hear. Uh, this is from Sam Hogan on Twitter. I'll link to this in the show notes as always. He says, six months ago it looked like AI LLMs we're going to bring a much needed revival to the venture startup ecosystem. After a few tough years with companies like Jasper starting to slow down, it's looking like this may not be the case right now, there are two clear winners, a handful of losers in a small group of moonshots that seem promising. Let's start with the losers. Companies like Jasper and the VCs that back them are the biggest losers right now. Jasper raised greater than a hundred billion at a 10 figure valuation for what is essentially a generic thin wrapper around OpenAI. The other category,
Chris Sharkey (01:07:52):
Wow, that's exactly what we said.
Michael Sharkey (01:07:53):
The other category of loses are VC backed teams building at the application layer that raised 250 K to 25 million valuations. Uh, one of those which isn't mentioned in here is like a, um, is mutiny, which is a landing page or a not landing page website personalization tool that uses ai. But I, I think it just got caught up in this narrative. I, I don't know how heavy surpris that they lean into ai. Um, executives at enterprise companies are excited about AI and have been vocal about it from the beginning. This led to a lot of founders and VCs to believe these companies would make good first customers. What the startups building for these companies failed to realise is just how aligned and savvy executives and the engineers they managed would be at quickly getting AI into production using open source tools. An engineering leader would rather spin up their own L chain and try chroma infrastructure for free and build tech themselves than buy something from a new unproven startup. In short, large companies are opting to write their own AI success stories rather than being a part of growth metrics of a new AI startup.
Chris Sharkey (01:08:51):
And we're, and um, we're seeing that, I see that just in my inbox. Like every company that I use in SaaS has announced some sort of AI play, so it makes That completely validates what you're saying there.
Michael Sharkey (01:09:01):
Yeah. And also just internal teams as well. Like we saw at Stripe building their own internal l l m, like they're not, they're not going and buying a a SaaS product. They're just building it themselves in house. And then the, it goes on to say, I'm not gonna read the whole thing, but I do think this is, this is interesting. Getting AI right is a life or death proposition for many of these companies and their executives failure here would been a slow death over the next several years. And two, the, there is a certain amount of kick ass wafting through halls of the C-suite right now. Ambitious projects are being green lit and supported in ways that weren't, uh, weren't a few years ago. And then they're saying the biggest group of winners right now, unconstrained by the need for a billion plus exit or GLO goal of a hundred million a r r, they build and launch products in rapid fire fashion, iterating until product market fit and cashflow moving on to the next. They've ruthlessly shut down products that are not performing LLMs and text image models, blah, blah, blah. Basically saying these solopreneurs are just, are benefiting the most right now. And this VC playbook is simply just not working. I
Chris Sharkey (01:10:02):
Mean, that's just, but that's probably partly timing. We, we don't know about all the big ones that are coming because it takes time to build something quality. Like it makes sense if you're just thrashing them out, making a bit of money and going that you would have results quicker.
Michael Sharkey (01:10:16):
Yeah, I, I agree. I don't necessarily a hundred percent think this take is correct. It's just the first wave, in my opinion, of of, of AI and the obvious
Chris Sharkey (01:10:27):
Slow hanging fruit. Yeah. It'd be as in interesting as an investor. Yeah. What to know, what is going to stand the test of time and what can be just replicated by companies themselves. That is a very tricky question to answer at this stage. Especially with like just getting bombshells like LAMA two being announced for commercial use. Just suddenly, like that's, with that kind of announcement coming out, how can you have certainty about anything?
Michael Sharkey (01:10:51):
I would still like to understand what moron vc, it was like, oh, I know, let's give Jasper a hundred million dollars. Yeah. Quite, quite frankly, you're an idiot. It's a lot like,
Chris Sharkey (01:11:02):
It is a lot given how, you know, we could make that software like in a couple of months Mac, but that's
Michael Sharkey (01:11:09):
What's happened. I mean, they've got like a hundred competitors now doing the exact same thing. Yeah. Just a race to the bottom on prize.
Chris Sharkey (01:11:16):
I mean, we did something similar in our own product in like a day, um, for generating, um, emails. So it's, it's sort of like, it really was the, the literally the most basic thing. I guess those people are smart that they just got a hundred million dollars for it. I dunno if they got any of the money, but um, you know, I guess they were focused on commercialisation rather than the actual long term, um, viability of the company. It's
Michael Sharkey (01:11:40):
Just the same VC playbook playing out here really for me. Invest in the tools, the developers, yours to, for the gold rush first. Like that's like the, that is just the non, like the, just the such an obvious strategy which everyone's piled into those companies like Lang Chain and anyone that took an early bet on some, you know, wrapper that is easily copied and has no, I don't wanna say Mo because most SaaS companies don't have a motive. If you think about it. They can be easily copied. It, it's just established once they're embedded in, in a number of companies. And it, it really is a race to get adoption so fast and, and people adopted so deep in your application that they can't, you know, they don't really want to go anywhere and they're satisfied and yeah, you might charge a bit more, but it doesn't really matter. Stripe's a great example of that. It's just embedded everywhere. There's much cheaper alternatives, but it's just so easy to go and implement that. Yeah. People continue to use it. Um, alright, so we are just like way over time, but I have two lulls for you. Okay. Um,
Chris Sharkey (01:12:42):
Care, we don't have any rules,
Michael Sharkey (01:12:44):
But there's no rules.
Chris Sharkey (01:12:45):
We're not aligning ourselves to our audience. We'll go as long as we want.
Michael Sharkey (01:12:48):
Yeah. Uh, and uh, okay, so this one, I just got a real laugh out of AI detectors. Think the US Constitution was written by AI , AI writing detectors are flagging the US Constitution.
Chris Sharkey (01:13:04):
It's always claiming things it doesn't own.
Michael Sharkey (01:13:06):
Yeah. One of America's most significant legal documents is a piece of AI generated text
Chris Sharkey (01:13:12):
.
Michael Sharkey (01:13:13):
This intriguing controversy raises many questions. More importantly, how could a document written centuries before the advent of AI be mistaken as AI generated anyway?
Chris Sharkey (01:13:20):
Well, I know how, because it's part of its training data.
Michael Sharkey (01:13:23):
It just, it calls such bullshit on these AI detection companies that, that this is just like not gonna work.
Chris Sharkey (01:13:31):
They're not gonna be able to do it. I, I, I said this ages ago, I, I genuinely do not see how you're gonna be able to distinguish AI generated text with a, with a sufficiently good prompt. I just don't see it happening.
Michael Sharkey (01:13:45):
All right. And my final ls Yeah, girl gave me her number and it ended up being G P T. I'll read you the text exchange. I, I don't know if this is true, but I really hope it was.
Chris Sharkey (01:13:57):
Let's pretend.
Michael Sharkey (01:13:58):
So, hey, it was great meeting you. Let's grab a coffee or something sometime. I'm sorry for the confusion, but I'm actually an AI language model that I don't have a physical existence. . Is there anything else I can assist you with? Hahaha chat gbt. That's funny. I'm glad you find it a music. Is there anything else I can help you with? Wait, fr for real question mark. Yes, I'm an artificial intelligence language model named Iris, powered by the technology from OpenAI and made available for use by autonomous and AI messaging service. My purpose is to assist and provide useful responses to your inquiries. Is there anything specific you would like help with bra ? So if this is true, that girl deserves like full credit for I think
Chris Sharkey (01:14:37):
Either way, either. If she's making up the responses, that's absolutely brilliant. And if it's some service that you set up, that's also brilliant. That's
Michael Sharkey (01:14:45):
Kind of, I really hope it's true. I I hope it's true. If it is true and there's a chance you're listening to our show, please tell us in the comments we would like to know . Alright, that'll do us. Thank you again for listening all your comments. We, we love your comments over on YouTube, been responding to them and we're so thankful for all the, the nice reviews that you guys have been writing for us. We, we really do appreciate it. Um, we have heard you loud and clear about the feedback around audio quality. Hopefully this week will be better in my defence. As I said in the comments, Chris was in a cave last week, a cave underground recording, so I was in the desert. So the, you know, that, that was kind of hard to get good audio for, but we, you know, we tried our best.
Chris Sharkey (01:15:25):
Yeah. So I'm not in a cave this week, so hopefully it's decent, but I
Michael Sharkey (01:15:29):
Am curious to hear from, from you guys, uh, if you are watching on YouTube in the comments about Lama two initial thoughts, impressions. Have you tried it, have you gone and used the hugging face version to play around with it? And
Chris Sharkey (01:15:40):
Also what are you gonna use it for as well? Like, I mean, we don't want people giving away their ideas, but I'm curious about the various, you know, applications that are possible with it that aren't possible with the, the APIs.
Michael Sharkey (01:15:52):
Alright, that'll do us, we will see you next week.
![Llama 2 Unleashed, Open-Source Vs OpenAI, Instructions for ChatGPT & Fine Tuning with Llama2 | EP24](https://img.transistor.fm/fu7xgPQQBI31nLzrsJfeEvwvwdMvUWaahuLE0OuhQqo/rs:fill:800:800:1/q:60/aHR0cHM6Ly9pbWct/dXBsb2FkLXByb2R1/Y3Rpb24udHJhbnNp/c3Rvci5mbS9zaG93/LzM4NjMxLzE2ODQz/OTA2ODQtYXJ0d29y/ay5qcGc.webp)