Conference Recaps and AI Updates – Techstrong AI Podcast EP11

Hello and welcome to our next AI podcast. I'm Amanda Ani, and with me today is of course, Mike Baard. How are you doing today?

I'm doing great. Good to see you as well. Yes.

We've had a busy week. Lots to share. So to start off, you were at the, the Google conference this week and a lot of news there.

So can you share with us what some of the big, the big things coming outta that conference? All right, well, the answer is generative ai, and doesn't matter what the question was, that's what we got outta the Google thing. So it's being applied to just about everything that they do.

Um, Gemini, formerly known as Bart, is their foundational LLM that they're using to extend into the realms of cybersecurity and any other use case that you can get to, which was all, you know, very interesting and very compelling. What was really kind of the big takeaway in my mind, and, and frankly I don't think they drove the point home hard enough. 'cause I think they're a little worried about the cultural impact still.

So they kind of downplay it. But if you look at what's going on with these LLMs, the reasoning engine inside them is getting smarter by the day. And as they add more parameters and other things that extend the capabilities of these things, they can go deeper and handle more tasks in a way that they can orchestrate.

So, uh, we're not too far off from where I can just say to, and then 'em build me a website and then it will go and create the various steps required to build a website and then link all those steps together and generate the entire website for me. Now, I may go back in and tweak it, or I may customize it a little bit, but in the days when I'm writing these kinda low level scripts to do things, they're coming to a close quickly. And in effect, we're all gonna have some type of AI assistant, it's gonna be a really smart assistant, but you know, all you're gonna need to do is put in the prop.

I want you to build me X and then it will take care of all those steps and organize it accordingly. And you may interact with it along the way to kind of fine tune it, but we are kind entering a new era rapidly, and the cultural implications for this are phenomenal. Absolutely.

I think all business leaders are trying to harness AI in one way or another. And, and, and I think you have it correct that eventually they're all gonna have assistance and it's, it's gonna help us in so many ways, especially increasing, uh, innovation. And each of these AI companies though, is just trying to soft pedal the cultural issues because they're afraid that people are gonna perceive that this is eliminating their jobs.

I think we just need to come flat out and straight say, look, jobs are gonna evolve and you know, what you were doing yesterday is not gonna be what you're gonna be doing tomorrow, and you need to figure out how you're gonna supervise these AI assistants and how we're gonna coordinate those AI assistants to do a more complex set of tasks. But, um, rather than just kind of throwing it over the fence and hoping people figure that out, I think we need to get to the point where we're proactively saying, Hey, these are the functions and things that are about to become, uh, shall we say, handled by your personal digital assistant. And then you need to figure out how to organize the ad and manage it, and that's your value add.

But right now, I feel like we're just kind of running around talking about the great quote unquote productivity benefits without actually discussing what does it mean for the rank and file. Yeah, absolutely. And the fact is, there are gonna be some jobs that are not necessary after a while, and they're gonna have to be able to adapt and pivot and, uh, and new jobs will be created as, uh, you know, more opportunities open up utilizing this ai.

And I'll be honest about it, on any given day, there's probably about 50% of the things that I have to do that are manual in nature, including loading content into things like WordPress that I'm pretty much excited about the notion that maybe some digital assistant will do for me. A lot of the stuff we do, we don't like doing. So why get upset when it goes away?

Exactly. Some of these things are very mundane tasks and, you know, and, and looking at it even from a sense of, you know, manufacturing and, uh, and, and specific tasks that people are doing all day long, the same task repetitively that, um, utilizing AI and machinery is gonna get rid of those tasks and, and who really likes doing those tasks? There you go.

All right. Well you were at a separate conference, uh, hosted by NTTI believe, and I, I imagine there were similar conversations going on there. Of Course, there were.

Yes. So I've been at upgrade 2024 here in San Francisco, uh, and, uh, NTT uh, originated outta Japan, but they have a, a huge, uh, US presence now and, uh, I got to tour some of their demonstrations of the many projects that they either have, um, already accomplished or are working on. And there are some really neat things that I will tell you.

Um, they're really focused on sustainability and one of the neat things was how they, 'cause we were talking about this a few weeks ago in a podcast, how could we use AI to reduce the energy consumption of ai? Um, and so they're working on things like that and then they're working on things like, um, their own, um, harnessing photosynthesis to, to put energy back, all sort of anything. But, um, a few of the things that came to mind, um, for us to talk about today was, um, uh, there, and let me look at my notes so I don't wanna forget.

So, um, I really thought it was cool that they have partnered with Harvard University Center of Brain Science to do some studying and researching in how AI actually learns moving forward. Um, so there are parts of AI that people still don't understand, um, when left to its own devices, how it will learn, how it's, um, um, capturing and taking information. So this partnership, they're hoping to get to the bottom of that and, and then be able to use that, uh, moving forward.

And then, um, they also, um, we're working on, um, aside from AI for sustainability, um, they're, they're working on some really neat things like the autonomous closed loop intervention system, which creates a cardiovascular bio digital twin to help patients. Um, and that one was really neat. They're doing a lot in the healthcare industry utilizing AI and other technologies.

Um, there's so much, it would be a lot to go over right now, but AI, of course there was a huge AI track. There are issues as always with all of these things. I mean, there's gonna be great progress, but right now there's kind of two caveats that we need to think through a little bit and now will lead us into our third segment.

Um, but before we get to that, the issue is debugging these AI things, right? We don't really understand necessarily how they all work. So if something does go wrong, who's gonna um, go in there and figure out exactly what the issue was?

And it's not like I can readily patch an AI model, so therefore I may have to build an entirely new AI model because something went wrong and that's a rip and replace. Um, so there's gonna be issues and that's gonna require some human level of expertise to go in there and sort out what's going on. I think maybe we'll have AI models that monitor AI models and maybe help us diagnose what's happening with that, but, well, yeah, it's not like AI is gonna be fully autonomous anytime soon.

So, um, there will be a need for somebody to get in there with an engineering mindset and understand, you know, what outcome was created because various parts of the system didn't quite work as expected then said that. Moving on to our next topic here, LLMs are expensive as all hell and it's costly to run these things. And, and I know from an end user perspective, you know, Microsoft is doing that on the backend and charges a fee for, uh, usage of that.

And you know, so people do that inside of their office applications and that's well inside, but the cost to running those things in the backend is really super expensive. So if you as a company wanna use an LLM to do something, well, you know, you're paying for the input and the output and the cost of these things are, you know, the more parameters there are, the more expensive they are. Um, so ultimately, you know, we had a story up on cloud native now talking about how an outfit called cast AI has come up with a way to identify which LLMs are being used and then suggest ones that might be less expensive to run or might be better suited for a particular task.

And that's an interesting outcome 'cause it's essentially applying some finops concepts to LLMs. And I think this is crucial because right now most organizations can't afford to launch a lot of projects. It's just too expensive.

So they're picking one or two, three, maybe maximum four things that they might get a lot of ROI from. Um, but it is, um, hampering adoption and of course we know in GPUs they're hard to find and come by. So even if you have a great AI idea and you may not have the resources to go execute it for a little while, and so that's gonna extend their little timeline here by a couple of years, I think.

Yeah, absolutely. Um, uh, uh, on the topic of LLMs, I just have to rewind a bit too and say, uh, one thing I forgot that was really cool here at the NTT conference is they have a, a, a product called zumi, I think that's how you say it. And what they've done is, uh, they've taken, um, the LLM model and they've made little LLMs that do only very specific small tasks and then each little task that they need, they harness it together in an interconnected web of little LLM called zumi.

So, um, and I think that's, we've talked about this before, that that's, um, something companies are looking at at small, very specific LLMs instead of large ones. Alright, well I think that's ultimately gonna be the way to go. 'cause it reduces hallucinations if I limit the amount of data that I'm showing the LLM more training, the LLM on a, it's faster to build and less expensive and then say what OpenAI is trying to do.

But it's more accurate because I've narrowed the data to a confined subset versus, you know, OpenAI has been trained on everything on the internet and you never know what you're gonna get. It's like a box of chocolates because the answers are different every day depending on what it is. 'cause it's probabilistic rather than deterministic.

So, um, you know, gotta understand what it is actually being surfaced because I think one of the issues, and I think we talked about it in a previous show, is we trust these outcomes too much 'cause it looks great and sounds convincing and the machine meanwhile is, you know, let's be honest making it up, Right? Yeah. So I I think people are gonna go more and more that way.

And, um, well next off, let's talk about another open AI controversy. How many times are we gonna hear about AI being in hot water? So, uh, most recently YouTube is looking into, um, um, uh, uh, apparently, um, let's say accusation, um, that OpenAI used a lot of YouTube videos, um, without proper permission or authority to train, um, their model.

And if so, there could be a big, um, suit issuing over this. Um, this is just one of many issues that OpenAI has been facing with what resources it used to train its models. So what are your thoughts?

My thought is open AI is run by a bunch of Jesuits, right? They'd rather ask for forgiveness than permission. So they basically went ahead and they, they knew that these were gonna be, uh, contentious decisions that they were gonna make.

And I think maybe they're hoping they're gonna make more money than they could possibly imagine. So they can pay off these, uh, suits as they go along if and when necessary. And they're still negotiating with all these different content providers.

And, you know, meanwhile the New York Times is leading the charge for, uh, accusing them of stealing their content as well. And this is a significant problem. It is not like, just because I exposed something on the internet that I gave up the copyright laws for that.

And you know, I think, um, open AI is trying to argue that, you know, there's a fair use doctrine that is true. When I'm not sure though it applies or extends to the training of AI models because basically you're screen from sites and taking their content, you're not just kinda like referring to them as you might in a fair use kind of case. So, um, we'll see how this all plays out, but it's gonna be years before anybody gets anywhere near a settlement.

Stay tuned. 'cause there's gonna be a lot more people waking up one morning going, Hey, how come my content's in this AI machine? Right?

And, and my question is, how do they even know what content was used? I mean, how do they go back, uh, filter through millions and millions amounts of content to figure out where all this data came from? Well, what seems to be happening is the output, and it's random, so it is difficult to track, but sometimes the output is word for word from something that somebody recognizes and somebody looks at it and says, Hey, you know, like some of the authors have noticed that, you know, there are entire tracks where somebody might say, write me something that, you know, sounds like X.

Well then it turns out it just takes what X wrote, slaps it in there and says we're done. And it's like, but it's word for word what X has already created. So that's kind of where the issues are.

I think there's gonna be better tools to look for that kind of content. And I think a lot of work is being done in that particular space and uh, so it's it's gonna be harder for them to hide. Yeah.

And this may open up a whole new industry of companies creating tools just for that to, to ensure that the copyrights are protected and, and check the data and all these models that are being created. Stay tuned as they say more. Okay.

Well I think, um, next up we have AI as it applies to the elections this year. Um, there's been a lot of, uh, things in the news about AI in the election cycle from AI being used to, uh, try to convince people and persuade people one way or the other, deep fakes being used. Um, but on the other side, AI being used to try to analyze, um, the outcome and the thoughts of, um, the general public.

So, um, what are you seeing? There's two things happening here in my mind. One is, yeah, we're creating a lot of disinformation, no doubt about it, but we're also gonna use AI to kind of maybe fine tune polling so that, you know, we're not surprised by the outcomes of election as much as we used to be.

I don't know if we'll ever get it perfect, but today, I mean, if you look back in time and look at, you know, the election cycles and the predictions there were made, they're deeply flawed. And I think, you know, pollsters don't like to admit they, you know, they'll tell you historically we have a plus or minus swing of 5%. Well, you know, when the election swing is on 2%, that 5% is a big deal.

And if I actually went back and did the analysis of it, I would suspect that it's probably a wider swing than 5%. Um, a lot of the things that pollsters count on is that no one will ever go and look back at what it is that they actually predicted. But you know, it turns out that this is not an exact science by any stretch of the imagination.

I also think the other interesting thing to watch here, I mean, you know, even before my time, there was a moment when, you know, Jack Kennedy and Richard Nixon were running for president, and it, it was the first kind of campaign that was, it used a lot of television to, to actually, uh, put the candidates in front of people. And I think we're seeing kind of the same thing here. This is the first AI campaign is roughly well into the first television campaign and it's gonna change the nature of the game.

I don't know what to do about disinformation because I'll be honest, I mean, I'm a student of history and, uh, they used to create pamphlets back in the day and lie about each other all the time and distribute those things all across, uh, new England and the middle Atlantic states. And one set was funded by Thomas Jefferson and the other one was funded by John Adams. And boy, if you ever read 'em, they were both a pack of lies.

So I'm not sure that this is anything new in that regard. As much as it is just a a the scale is different and I, and I'm hoping that people will come to recognize what it is they're seeing. I mean, a lot of the stuff just kinda reinforces, uh, beliefs, true or not.

Uh, but pamphlets were the same way and didn't stop people from publishing the fact that they were known lies back in the day because they just hid behind the first amendment. Yeah. And I think one thing we're gonna see is that AI is reducing the, the campaign cost, uh, on the marketing side, you know, using generative ai.

Um, it's gonna reduce the, the level of time that it takes to create these, um, TV ads and, um, campaigns across all of social media because of how quickly it can generate, um, you know, similar content for all the platforms. So we may see a lot more advertising because it's more cost effective and cheaper to do so. Is that a good thing?

I don't know, but that's, and I'm just saying that from what I've seen, that's one way that it's gonna be used in the elections. Whether it's a good thing, I don't know. So I I, I might have to pay more fees to my content providers to have them not run election ads.

I wonder if that's a business model. I think, you know, we will pay you not to do certain things, or at least not to distribute that content to me while I'm watching something because well either A, I've already decided, or b, it's just flat out annoying. Yeah.

All right, well that's all we have time for to share today, but there was a lot to cover and um, stay tuned. 'cause we always have more AI news as this technology unfolds quite rest And remember to vote early and vote often, as they say in Chicago. Yep.

Well, thanks Mike.