In The Arena by TechArena - Bloor Group CEO Eric Kavanagh's 2025 enterprise AI Predictions

Episode Date: November 25, 2024

In this episode, Eric Kavanagh anticipates AI's evolving role in enterprise for 2025. He explores practical applications, the challenges of generative AI, future advancements in co-pilots and agents, and more.

Welcome to the Tech Arena, featuring authentic discussions between tech's leading innovators and our host, Alison Klein. Now, let's step into the arena. Welcome in the arena. My name is Alison Klein. Today, I'm really delighted to invite Eric Cavanaugh, CEO of The Bloor Group, to the podcast. Eric, it's so great to have you on the show. Welcome. Thank you so much for your time. Thanks for the invite. Very exciting. So, Eric, I've been following you for a long time. Why don't you go ahead and introduce yourself and

what you do at The Bloor Group? Sure. So the Bloor Group is a hybrid media analyst firm, and I host a lot of radio shows and webinars and, of course, write articles and white papers and do social media promotions all around data, AI, technology, really the nexus of hardware, software, and human ingenuity. That's what we talk about. And we try to educate people about what all these technologies are, how to use them, what to use them for, what not to use them for. So we're an educational outfit that tries to spread the word about cool tech and what it can do for you. I knew you were going to be the perfect guest for this topic, which is AI in the enterprise in 2025, practical and purposeful.

Why don't we just start with how you're seeing AI adoption in the enterprise today? And what do you hear from companies about how they're approaching this topic moving forward? Yeah, that's a great question. I think the first thing to keep in mind is there are many kinds of AI. There is the old fashioned AI that's still around, it's still kicking and is doing amazing things. And there's all this new gen AI and these foundational models with billions of parameters that are really blowing people's minds. And I think they certainly have captured the imagination of the business, of consumers, of the academic folks in our world.

They've blown minds all over the planet, basically. But there are caveats. And so what I'm hearing in terms of practical applications of AI is that the Gen AI stuff is still a bit unwieldy and it's because of these so-called hallucinations. So you see a lot of effort around a new architecture, often referred to as a RAG architecture, retrieval augmented generation, which will temper what the engines deliver, but doesn't control what they deliver. And so because of that, and because of the fact that these foundational models are, all of them that we hear about at least, are probabilistic

in nature, we have to be careful about how we use these things. And so what I've heard from a lot of the smartest people in the business is that most of the production-ready use cases are inward focused. In other words, they are designed to help employees like decision supports make better decisions about how to do their jobs. And you're seeing a lot of reluctance in turning these engines outward to consumers, to prospects, to partners, to clients, for example, because of that. And I think that's probably a good move because no matter how good your RAG architecture is, you're still going to get some wonky stuff

coming out of these things because how they were designed. Frankly, I think that's what Sam Altman was talking about when he said curiously that the era of large language models is over. Like a year ago, we're like, what did you say? And I think that's what he's alluding to, but of course we'd have to ask him.

Now, I love the way that you just described that. I think that one of the things that I ponder when I'm thinking about this topic is what are the use cases that fit within that purview from a generative perspective that are going to be accelerated in terms of enterprise adoption? And how does that match with the value that companies are getting out of them? Yeah, I think the co-pilot stuff is where the action is. And that's for coding, obviously, but it's also for creative writing and just generating content, generating material, because these engines are so good at reflecting back common and useful patterns, meaningful patterns of words based upon all their training and the profits you give them. So the way I describe it to people is for any topic that has been widely published,

for example, the life of Martin Luther King Jr. Let's just pick that out of the blue. There's been so much published about him that the models are probably going to get most of that very accurate. Now, there's been mistakes because it is generative. I once got the go-to-market guy from OpenAI and a bit of a Twitter thread going back and forth. And I asked him about the hallucinations. This is like a year ago or maybe a bit more. And he said, remember, the hallucinations are more of a feature than a bug.

So in other words, they're designed to do this. They're not designed to be accurate representations of what a database would show, which is what an old fashioned query will do is show you the records that are in the database according to your query. And that's not what these engines do. They generate new content based upon patterns of words that they have been trained on. So the point is that as a co-pilot to expedite the process of any project, whether it's writing an application for your phone or for the internet, or whether it's writing a press release for your client, these engines will get you 80% of the way there, like in no time, just bam, there it is. And then you tinker around and finish the job yourself.

So in that sense, that's already a time saver. And the other cool thing is they will reflect back things that you did not think of. I've found that with AI in general to be an excellent rule of thumb is that when you leverage an AI model to give you suggestions about things, it'll suggest some things that you expected. And it'll always, that I've seen, suggest things that you did not expect, but wind up realizing are important.

So little missed steps or angles to a story you had not considered, it's amazing what they'll reflect back. And of course they don't sleep and they don't go on strike and they don't clamor for more money. They want more power, maybe more GPU. But yeah, I think overall they're going to expedite projects in many wonderful, amazing ways. Now, you mentioned RAG, and obviously this is one popular technique for training, how do you see the enterprises aligning

behind particular techniques between using public models, growing their own training algorithms, et cetera, to put trust and lack of hallucination into the results? So from my understanding, RAG really is a post-training band-aid if you get down to brass tacks. So what happens is you get your model, you train your model that has X number of parameters, you train it on all the data you want to feed it. And I've heard some interesting stories about this,

by the way, that OpenAI, for example, was not trained on just the rando stuff on the internet. They figured out very early in the process, they had to be much more careful about which data sets they fed into these things. And so they were, they had to curate very carefully to make sure that it didn't get just garbage. Because if you just take everything on the internet and throw it into a model, you're going to get a bunch of weird, wonky stuff that is not going to serve any business purpose. And so that's an interesting point. And just as a side note, one of the more interesting things I've heard about IBM and their approach with Watson, or now it's Watson X, was they designed it from the outset to be able to retrain on new data.

So the problem with a lot of these open models is that once they're trained on this data, you can't untrain them. So you can try to tilt it in a certain direction. And I give an example of raising a kid. Training a model is like raising a kid. You have to be careful what you expose them to because they will pick up on things. They'll absorb whatever comes their way. So if you use a curse word in front of your two-year-old, don't be surprised if they use a curse word back. So you can't untrain models in that sense, at least not the primary ones like OpenAI. I'm not sure about Claude. And I think the same is true for Gemini.

So once it's in there, it's in there. And then what you're doing with RAG is you're trying to temper that. You're trying to give it what Brian Raymond of Unstructured.io called anchors of truth. So you're trying to weight it down and tilt the scale toward your fat sets, towards your factual data. That's really what RAG is all about. So I think that then there are lots of different design patterns for that, lots of different ways to do that. But what excites me the most, and I've heard stories of companies who are doing this, Iterate is one company that's doing this, is you take a small language model, so not as many parameters, and you host it yourself. And that's what they're doing. And I think it was Lama or Lama 2 that they used.

And you host it yourself. So you have to invest in the hardware, obviously. Not cheap to do this. But if you have your own model and you train it on all your enterprise data, like everything, your support tickets, your ERP, your CRM, whatever you want. Then, and I call this the executive cockpit.

I think the ideal scenario is then you set up Kafka topics from all your production systems into a vector database that's aligned with your small language model. And the executives can sit there all day and ask any question they want about anything at all. And I'll give you just a quick example. The CEO of a very prominent company, very smart guy I know, very cutting edge company. I asked him before his user conference earlier this year,

I said, hey, do you guys have the IRS account? Because I have a good contact there. I could probably get you in there. He goes, I don't know. I don't know if we have that account. And I thought to myself, in the future, when you have an executive cockpit, you're going to know because you just got to ask it

and it's going to tell you. And it's going to show you. And you notice the big trend, and I'll close on this, perplexity is focused on this, Gemini, Snapchat, TPT, all these guys are focused on citation, on citing where you got it. In fact, Click Answers, which is one of the better ones that I've seen, Click Answers, they baked that in from the outset to have citations to show you where the different parts of the answer came from. And that's the number one remedy, I think, for hallucinations, because when you cite things, you can then go in and double check and make sure, OK, yes, that's correct. It did cite that accordingly. Or if it didn't, at least you can further amend the model.

Fine tuning is what they call it. It's a layer kind of inside that model, which is a bit more sophisticated. But that citation component is going to be crucial for trust in the responses. How do you see this world of co-pilots and chat GPT prompts evolving to agents? And what alternatives of agentic computing are gathering steam in your mind's eye? So that's an excellent question. And I think agents are all over the map. I think that Salesforce had their big announcement.

Benioff says he wants a billion agents out there working. I mean, agents are like mini semi-autonomous dynamic applications is how I describe them. So they can do all kinds of different things. I think the real challenge is going to be governing them, auditing them, shepherding them, understanding what they're doing and why. That's going to be real important to leverage these technologies. Now, they are very powerful, but they could also go sideways. When you start giving them permissions to get into systems, they can start doing stuff. And if you don't have a means to control them, then goodness gracious, it's like letting a gerbil loosen your machinery. So like this could be very

unproductive and very challenging. So I have not seen any registry yet that gives me confidence that someone's going to be able to do that. And the questions that I ask are, will there be a log file? For every single thing this agent does, is there going to be a log file for that? And then we can go to log files and read, because that's what you're supposed to do with a database or an ERP system is build in the production or the creation of log files when actions are taken. Now that does slow them down, obviously, because they're reporting what they're doing as they're doing it. But that, to me, is the big question. How are we going to monitor what they do? How are we going to shepherd and control what they do? How are we going to report on what they have done? When the auditor comes knocking and says, all right, why did these people not get a loan and those people did? If you can't answer that question, you're going to have some trouble. Now, you talked about Sam Altman earlier. Last weekend, he came out and said that AGI would arrive in 2025. What do you think about that bold proclamation? What do you think he's talking about? I don't think

it's true, honestly. I think these things look like they are intelligent and they can process many different parameters and feeds and do all kinds of things. But to achieve self-awareness is another huge hurdle on top of all that. So I'm not very concerned about it, frankly. I mean, a lot of people are concerned. I think if and when the AI agents take over, they'll probably do a better job. I'll give you a prediction. I think one thing we're going to see, and a lot of people tell me I'm crazy for saying this, but I think that a lot of civil law is soon going to go the route of AI. And what I mean is when you file a suit against someone, or even eventually in criminal cases, you're going to have the option of going with an AI judge

and jury or a real judge and jury. And the smart money is going to be on the AI judge and jury because I think the algorithms will be more balanced. It has the opportunity to eliminate a lot of bias, right? It does. Now, there will be bias in there from the data it's trained on because, again, it's like raising a kid. Apple doesn't fall far from the tree. So the model won't fall far from its training data. And we have to be careful about that, which is why I believe that we should have legislation that mandates that these big AI models be transparent. The argument against that is, oh, this is our proprietary technique. These models are so complex. It's not like anyone can

just go pick up the model from Lama 2 or from ChatGPT and recreate it themselves and then get all the business. That's just not going to happen. So I think we need for these models to be open source and completely transparent in terms of what their architecture is, what their training data is. I think all of that really should be mandated by Congress to protect us from what could be very bad acting. One of the things that you said earlier on is that you're not seeing gen AI models being brought to public-facing implications anytime soon. I want to ask this question. Do you think we will see the first major AI security breach in 2025? Oh, wow. That's a good question. I'm ashamed of myself for not having thought of that previously, because it's a very interesting question you're asking. In other

words, there's, I guess, two ways. One, AI could be used to effect a security breach. And I'm sure that happens all the time. If you're talking about a model being breached and compromised, that is possible. I mean, who knows? It is possible. I'm sure someone's trying to work on it. But that's a very interesting consideration that I had not thought of. And it's not impossible, is it? No.

And I think that people are aware, as you said, of tapping AI to do things like write malware, which is more of the brunt force form of AI security breach. That's not what we're talking about. It's more, will somebody put an AI model out there that opens a door that isn't expected? It is very possible. And I will point out, I'm going to pick on Claude for a moment. There was a story a few months ago when I read this article, it was like in Wired, I think. And I will point out, I'm going to pick on Claude for a moment. There was a story a few months ago when I read this article, it was like in Wired, I think. And I was like, what on earth are they talking about? It basically said something like, oh, there are malware deep in the machine or bad actors deep in the machine that we can't get out. I'm like,

what are they talking about? So I read about it. And what it said was the folks from Claude trained it on some bad data and then they couldn't stop it so they basically put some sort of malware deep into one of the layers and they tried to prevent it from doing bad things down the road basically sort of like a rag model i was saying earlier like you build something and then you build guard rails after that and they couldn't stop it i think well why did you do it in the first place? Why are you trying to embed malware in the core of your neural network?

That's a really foolish idea. It's the digital equivalent of gain-of-function research in the virus world, and that's scary stuff. Who do you think is best positioning themselves for this new era across hardware, software, and services? That's a good question. I think who's winning is still Microsoft. I think Satya Nadella, from a business perspective, is one of the sharpest tools in the shed. Look when they kicked out Sam Altman and over the weekend, Satya Nadella's like, I'll hire you. And now

some of the main people have all left. It's anybody's race still. Obviously, Elon Musk is in on the game with X. I personally use Gemini. Claude does get lots of good accolades. Mistral, I think probably the winner is going to be the mixture of experts, which is Mistral's thing, because these super large language models, they are tremendously unwieldy and they don't have to be that big to do the things you want them to do. In the world of business, or even in the world of just consumer artwork or having fun, you always want a tool to be good enough to do what you want

it to do. But if it's so complicated that you can't control it, that's a problem. And bigger is not always better. Osama Fayyad is a good friend of mine. He was the first chief data officer for Yahoo. He now runs the Institute for Experiential AI out of Northeastern University. And he told me earlier this year, he says, these models, they're too big. We don't know how to do the AI. And you tell me you don't know how they work. We just don't know. These super large language models, they learn things.

It's freaky, the things that they can learn because they're seeing patterns. Again, these neural networks can have any number of architectures. You can have any number of layers, any number of inference layers, ingestion layers, whatever you want. You can do anything with these things. And they're incredibly complex. That's why I'm saying make them open source because even if they're open source, it's going to be darn near impossible for people on the outside to figure out what you've done. But they are very powerful and they're unwieldy. And I've heard stories about models going sideways to all of a sudden it just stops working. I've seen really weird responses from chat GPT at times when you pressure it. And the more they try to temper or box it in after the fact,

the weirder things will get, I believe. Because, for example, if you ask a question about politics, you would notice it's, oh, no, ask Google about that, because they know hallucinations. They don't want to get in trouble for making mistakes on very important issues like that. I think in the terms of service for Chet Chibutee, even if it tells you, don't use this for medicine, don't use this for financial services,

don't use the thing on Q-tips, where it says on the box of Q-tips, don't put it in your ear. Everyone puts it in their ear. That's what you use it for. But I don't know. Don't do that. Okay. People do it anyway. So it's a very strange situation. I think mixture of experts is going to be the winner in the end, because all the people I talk to say, especially with AI agents, you want them to be very focused on doing a specific task very well. You don't want them to try to learn 10, 15 different tasks because that whole rationalization part about which task I should do now turns out to be very challenging. But I think Microsoft is in the best position right now, but a mixture of experts is knocking on the door. Now, we talked about the negative sides of AI earlier.

What do you think the surprise of AI in a good light will be for 2025? I think you may start seeing businesses run or at least partially run by AI. Some people are even doing some really amazing things where they are spinning up marketing campaigns using AI in record time, like in hours instead of weeks, because you can create the content, then vet the content, trigger actions to load into your distribution technology, whatever it is, SMS, email marketing, social media marketing, whatever. You can go end to end, lightning fast these days, and it can monitor all these things. So my business partner, Dr. Robin Bloor, he theorized probably 10 years ago that sooner or later, AI will run companies. And I thought that's crazy. That's not going to happen. And then the more I look at it, I'm like, that

might just happen because they can watch and see your revenue is off target. This sales group is not doing a good enough job. It can trigger all sorts of actions. And that's really what it's going to be like. We're going to have humans riding the machines almost like you ride a Bronco. You really have to hold on to that sucker. And when you get thrown off, hit the kill switch, it's back again. So it's going to be really interesting to see how that all pans out. But people involved in the middle will always be necessary. I believe you don't ever want to just turn this thing on and let it roll and run your business.

You always have to have people monitoring closely and managing. But it is going to be like riding a horse, man. That horse is very powerful. And you got to be careful he doesn't jump off the cliff. Eric, thank you so much for your time today. This was an awesome interview. I didn't know we were going to be talking about Bucking Broncos, but you took it to some interesting places. I'm sure that folks want to keep engaging with you. Where can they find you? Hit me up on dmradio.biz, info at dmradio.biz or any of my social channels at Eric underscore

Kavanaugh. And it's a K-N-O-U-K-A-V-A-N-A-T-H. I went to Ireland to look in the phone book. Thanks so much for being on today. Thank you. That was fun. Thanks for joining the Tech Arena. Subscribe and engage at our website, thetecharena.net. All content is copyright by The Tech Arena.