At the Intersection: Insights for Thriving at the Crossroads of Change

How Word Pairing Sparked the AI Revolution

• Craig Francisco • Season 1 • Episode 1

Share episode

In this first episode of At the Intersection, host Craig Francisco sits down with Mike Rojas, a seasoned AI strategist and technology leader who has spent decades helping organizations understand and apply emerging technologies. Mike has been at the forefront of artificial intelligence adoption, guiding executives and teams through the shift from traditional analytics to the new Human + AI era.

Together, Craig and Mike unpack the fascinating journey of artificial intelligence — from Alan Turing’s early breakthroughs to the transformational leap in 2022 with large language models and generative AI.

They explore:

  • The 80-year history of AI and why October 2022 marked a turning point
  • How generative AI tools like ChatGPT are augmenting human intelligence, not replacing it
  • Why AI should be seen as an “Iron Man suit” for leaders and workers
  • The truth about AI “hallucinations” and how models are improving
  • How business leaders can embrace Human + AI to unlock new value

If you’ve been curious — or even cautious — about how AI is reshaping work, this conversation will give you both context and confidence.

At the Intersection is the podcast for leaders navigating the Human + AI era. Subscribe, follow, and share to stay ahead of disruption and thrive at the crossroads of change.

#ArtificialIntelligence #GenerativeAI #Leadership #FutureOfBusiness #HumanPlusAI #AI23

SPEAKER_01:

Mike, thanks for jumping on the podcast of At the Intersection. Very excited to have you on. This is episode number one, just so all the listeners know. And your deep background in AI and your experience, really, Mike, on the tech side of things, has been it's truly remarkable. And you know, I think we'll get into that throughout the story. But what I would love to do is if you could take us back, take the listeners back to the beginning of AI and let this help us kind of walk through the timeline of what we have seen, what has changed, and really what's in front of us today as leaders and humans in this space of AI.

SPEAKER_00:

Yeah, well, thanks for having me on. I mean, this is a uh a great opportunity. Love the show, and uh yeah, let's get into it. I mean, you know, when we hear about AI, it's actually a very, very old science. I mean, it goes all the way back to the 40s when Alan Turing coined the term artificial intelligence, right? So he was thinking about a machine being able to reason. In fact, he even created a test to determine when we would actually build that machine called the Turing test. And we have, as computer scientists, we've pursued AI now for going on 80 years. And you know, a lot of it was fits and and and you know, fits and starts and stops and that sort of thing. But I think what we have finally achieved uh is a machine that can reason. And that is a fundamental change in technology that has wide-ranging ramifications for not just businesses, but you know, society in general. And there's a lot of debate around, well, is the machine really reasoning? Is it just a simulation of reasoning and and all of that? But I think one of the uh one of the things that is fundamentally different about these large language models and their ability to communicate with humans and have a dialogue with humans that sets it apart from all the other attempts at AI is this reasoning capability. And you know, I think so much so that most of the industry leaders that are working in uh the generative AI or the large language model space, they've developed a scorecard. And this scorecard breaks down the intelligence of the machine into five levels, from literally zero to level five. Every AI system that we had built prior to October 2022, they put it level zero, which is no intelligence, no AI. That means all of the AI that we've worked on for like, you know, autocorrection system, decision systems, predictive analytics, all of that AI work is classified as no intelligence. So you can think about how profound that is. Literally every business intelligence system that we have been building in the 2000s, you know, this in essence, over the last 15, 20 years, is considered to have no intelligence. And that's a tough pill for a lot of people to swallow. If it's like, I what are you talking about? I I have devoted my life to building predictive analytics, and now you're telling me it has zero intelligence. But but I think that's a correct classification because if you look at what those systems are designed to do, predictive analytics, they're designed to answer one question. You know, the prediction. What should the stock price be? Is this consumer likely to buy? Right? They're trained with a ton of training data just so they can make an intelligent, intelligent prediction. Language models, on the other hand, are trained for generalities. Like I want to form conceptual understanding of a very complex uh landscape, and then draw conclusions from that conceptual understanding. That means they're very, very good at answering any question, not just what should the stock price be, but what potentially could happen to that price based on economic factors in the future. We've not had that capability before.

SPEAKER_01:

And to me, Mike, the thing that just blows my mind. So, you know, an incredible amount of smart people, tons of money focused on this over the last several decades. And here you're telling me before October of 2022, you know, we've we've been at a level zero. So I I'd love for you to walk us through what happened that made this shift, and then what we're seeing today, and like all of us, like I'm using, you know, I'm a big fan of the chat GPT, and that's what I use every single day. And it's taken it's helped take me to a different level, but this didn't exist. Like this is just this stuff's new, but it is moving at a pace so fast that I think a lot of people are scared to death. And I would love if you could help us understand what happened and made this shift.

SPEAKER_00:

Yeah, well, I mean, you've got to go back uh almost eight years now, uh where you had basically people that were operating in data science that were trying to solve a pretty mundane problem. And that was language translation. And they were struggling with how do I build a system where I can convert English to French, for example. And that doesn't sound like business intelligence, right? That just sounds like I want to take a document and it's in English, and I want to put it in in French. So they're working in this space, and they're struggling trying to translate sentences like, I'm sitting on a river bank thinking about my bank account. And they're fixated on this word bank, and they're thinking, okay, wait a second, river bank means something completely different than bank account. And this is the crux of the problem when you're trying to translate a language where context matters. You gotta understand the context of the sentence in order to have a high-quality translation. So using math, they started to play word association. So they would say, What if we weighted an association pair together so that bank and river would pair up, and that would have a strong bond between those two. And then later in the sentence, you would have bank and account, and they would have a different strong association with them. So they started training a model to pair up words, except they didn't just pair up river and bank and bank account, they paired up literally every word in the sentence to every other word in that same sentence. So they would have weak associations and strong associations, and what they realized was this contextual pairing of words really unlocked high-quality language translations. So they just started doing more of that. They built what was called a language model and they trained it with these word pairings, and the first models had you know 60 million pairings, then they built a hundred million, 150 million, 200 million, and so on. Somewhere around the summer of 2022, they had built a pairing model that had almost a half a billion pairings, and it was enormous, and they just kept building it bigger and bigger. That's where the large language model comes from. The large is coming from just a massive amount of pairings. Well, in October of 2022, something very, very unexpected happened. It was no longer just providing high-quality English to French translations. You could give it an English sentence in the form of a question, and what would come out of it is an answer in English. Google published a paper called Emergent Abilities of Large Language Models, a pretty boring paper, but basically what they proved is that it could take standardized tests that had never seen the answers to, and you know, pass the bar exam, pass the LSAT, pass all with with like 90-95% success rate. So this was surprising. And suddenly people are like, wait a minute, contextual understanding through word pairing unlocks intelligence, it appears so.

SPEAKER_01:

So this was let me pause you because this blows my mind, Mike, and I don't and it's hard for me to wrap my brain around it. So this was an accident, this was not something these this group of individuals, obviously, by the time you're at you know 500 billion word pairings, there's a lot of people involved. But did they see this coming?

SPEAKER_00:

Did they did they know that yes, they saw they saw glimmers of it? So there were prior models to uh GPT-4. If you remember when GPT launched, it was GPT 3.5. What came before that was GPT 3 and GPT 2 and so on. So, but what they realized was wait a second, we're getting some interesting output of this beyond just language translation. Like it's starting to form sentences that have rational answers to an input prompt, to a question we give it. So they were like, let's just keep going, let's build it bigger and bigger and see if it gets smarter and smarter. And it turned out to be the case.

SPEAKER_01:

And so they they see this, right? So October 22, the papers released. What happened then? Like it is gone, you know, from zero to 100 miles an hour overnight, it feels like. I mean, everybody is flooded with AI this, AI that there's products coming out. I mean, it is to the point where if you know, as a business owner, I wouldn't know what in the world to do, what to believe. It it's it is absolutely mind-blowing. And it feels like every week that I work with it, with Chat GPT or you know, Claude or whatever, you know, whatever people are using, it gets better and it's getting smarter. And it so help me understand all of this if you can.

SPEAKER_00:

Yeah, so you know, 30 days after that paper was published, chat GPT dropped, right? Which was kind of the the moment the consumers, the world woke up to the fact of just how good uh these GPT style models had become. And people started using it for all kinds of things, including I can have it code software, um, answer questions I normally would have to get a lawyer for, and you know, help me do my job, help me, you know, plan a trip, things like that. So the consumer experience um really kind of woke everybody up to these models have incredible uh potential. Where I think it quickly got confusing is you had a very, very large contingent of the AI community that was working on predictive analytics and decision systems and BI tools and things like that that were you know preaching the religion of data lakes and data warehousing and data science and data analysis. I think they were taken, you know, they were caught off guard by this consumer wake-up. Like they were like, well, it's a hype, it's a bubble, it's gonna burst, it's not, you know, it's not really reasoning, right? So at first they're like, wait a second, this is make us obsolete. And you can imagine if you were a business analyst and your job was to produce dashboards that would provide information to a CEO and to answer his questions. If now that CEO could just simply go to a chat GPT and ask that question and get his answer right away, you're gonna feel threatened, right? Wait a second. I'm maintaining a database of you know trillions of data points, but you're telling me now a GPT can come in and just answer that question directly, and the the the truth is yes, it probably can. So, you know, I think the confusion it can be broken down into several parts. There's a large contingent of the old AI guard, and that's what I would call them. You know, it's like oh predictive analytics. Not that predictive analytics and data science and data warehousing doesn't have its place. It does, it has purpose. Um, but if you look at it at its whole, it's just knowledge, it's just a list of facts with no contextual understanding. And that's why humans get involved to try to make sense of those facts. What do these signals actually mean? You know, where's the economy going? Is the should the price go up? Should the price go down? They're they're trying to provide context. And because that's their craft, I think they're reluctant to employ new ways to apply that context. And certainly large language models um you know threatened that uh you know contextual understanding skill set that they had developed.

SPEAKER_01:

So you said I mean you said something that really stuck out to me, Mike, about you know, the consumer side, as we all started to embrace this and started to kind of use it and challenge it and using our prompts. Have we as a society, as a world, improved this AI? Have we made it what it is today because of you know everything all that we're doing, our inputs, the the relationship that we're establishing with this technology? 100%.

SPEAKER_00:

So the uh this is a technology that thrives on dialogue, it thrives on interaction. So you basically have the world interacting, having conversations with the machine, and that information is gold to improve the machine, to level the machine up. The machine is learning from the human as much as the human is extracting knowledge and understanding from the machine. And this is a very interesting dynamic. In fact, the way OpenAI and other large AI companies have leveled up their AI is through a process called reinforcement learning with human feedback. So reinforcement learning is a self-improvement process. Human feedback puts the human in the loop to say this is better, that is not. And by guiding the AI to saying this is better, that is not, the AI does better and better things. So that improvement loop um creates a kind of a um a feedback loop where the human is improving the AI and the AI is improving the human.

SPEAKER_01:

Yeah, and it's just at that point, that is, I believe, why we're just seeing the massive change in and really the fear that is being created. I mean, so many people, and you read it, you could read, like you and I talked about before, you can go online and find all these positive things about AI and the same amount of negative. It's it's just it's back and forth. People are are scared. That's what fear brings forward. What is your when you you hear the statements from others that you know AI is going to replace jobs, like it's the negative side of what's happening? What would you say to you know the folks that are tuned in here, which you know are leaders in different industries? Like, how would what would you tell them?

SPEAKER_00:

Well, I would stop treating the AI as it's a a full-blown autonomous human. I don't think that's where we are with AI. I think all the AI companies are are are going for that, but I think we're a long ways off for that. A better way to think of AI, and I'm talking specifically about generative AI, right? Is you can think of it more like a vehicle, right? So would you be afraid of your car if you walked up to it and you're holding the keys, right? It's going to do nothing until you climb into the driver's seat and you put the key in and then you drive it, right? And it will take you places. And you should treat our current state of AI like that. It's a vehicle to get you places. It is not, you know, an autonomous robot that's coming to destroy you. So, you know, and I think a lot of people, I like the vehicle uh analogy because it also means stop trying to go for an autonomous self-driving vehicle. We're not there yet. The vehicle will crash. If you start it up, get out of the driver's seat and set it on its way, it's probably going to do some damage. So, but keeping the human in control, driving the vehicle, steering the vehicle, I think is going to yield a lot more value than job automation. So, and I and it's not just me saying this, many of the industry leaders, Andre Carpathy in particular, is saying we got to stop going for Iron Man robots and understand what we have today is Iron Man suits. Let's augment the human, not automate the human. And once you start to understand that, you realize it's not as scary. It actually opens a lot more empowerment for your workforce than replacement. And it can take you to places that you couldn't get there with humans alone.

SPEAKER_01:

Yeah, that I love the the Iron Man suit, the visual for me, because you're right. It's not today, you're not waking up and you're the AI is not taking control of your life, right? You're you're prompt, you're still prompting, you're still asking the questions, you're still trying to figure out how to leverage it. For me, it has it has leveled up when I'm able to deliver, you know, 10x, 20x, because we all have parts of our jobs that take time, we may not enjoy them, and some of it's data-driven, and the information's available, but you have to put it together. And in what I've seen, for me specifically, just knowing how to use it, and again, knowing that it's not perfect, you have to give it feedback. You I talk to it, I say thank you, please, just like I would any other human. And the relationship that I have been able to establish with how I use it is very, very positive. And it's and so I I want others out there to know that you shouldn't be, first of all, technology is undefeated. I mean, that's one of my favorite phrases. This isn't going away, right? No way. Get on the bus, learn how to leverage this technology, or else it could be, it won't take long before somebody to get really too far behind. And then what happens, right? You're you're in the situation now where the competition has surpassed you, and it's it's going to be very, very difficult to get back into the right position. So, what would you say when people are sitting here starting to make decisions? Going all in is dangerous. All in meaning that all these products are coming out, everyone starts buying everything, and then ignoring it, putting your head in the sand, where should people play? Like, where is the sweet spot?

SPEAKER_00:

Well, we're we're in an interesting uh phase of this technology. I mean, to your point, this is very, very new technology, even though it's been out you know for almost two years now, it is very much still the wild, wild west of generative AI and where we're going. I mean, it's getting better all the time. And we're discovering techniques. And I think that's one of the key words here. We're all on a discovery on what is possible with technology. And this is this to me is the most exciting part of this. It is anybody's game, anybody you know can just access the technology, play with it, and figure out what works. I mean, there's some guideposts that you can use where it will definitely level up your interaction uh with generative AI. And that is, you've got to understand it's a human simulator. So therefore, it is trained on all the interactions that us humans have had on this planet. So treat it as if you would treat a human. Don't give it agency, it's not human, but the interaction that you have with it, like telling it thank you, and great work, and let's do more of that. And I'm confused here. I can you explain that in more detail? Is how you would interact with another human, right? Scolding it, talking down to it, that sort of thing, you're gonna have it shut down on you. Right. Right, just as a regular human would, because that's core to its training. So, you know, there is a lot to discover here. Like we're discovering new ways to communicate with AI all the time. We're discovering new ways to make sure that it can be grounded, it can not hallucinate. Like hallucinations, which is one of the most cited um flaws with generative AI out there, it's usually the go-to for a lot of the naysayers saying, look, it hallucinates, it makes stuff up, it lies, it you know, it misdirects you. Well, that's because it's trained on a lot of human interaction, and in that human interaction are lies and misdirection and that sort of thing. So naturally, it's going to adopt some of that behavior. But one of the ways that we're realizing that uh AI will hallucinate is how it was trained from the very beginning. I mean, initially it was trained by data scientists that basically said, Hey, I want to train it with as much information as I can find, like all the internet data it can find. And we're just going to train it in any order we want. Well, as we study this technology, we find out the order of the training matters. So, for example, if you if it is trained with eighth grade algebra and you want to train it for calculus, you don't give it calculus, you level it up through the various stages of math, and you give it trig and you give it geometry and you give it pre-calc, and then you train it on calculus. If you do it in any other order, that's where hallucinations come from. So as we're on this journey, the large AI companies have realized this. The training sequence matters, the the process of training matters, and they can build models that will definitely have fewer hallucinations.

SPEAKER_01:

Yeah, just just it's really fascinating, Mike. And it, you know, to have you on the show and help really help our listeners understand, you know, what has happened in really in a short period of time. I think history is also important to know that you know, this was the goal, you know, a long time ago, 80 plus years ago, but we're here. And I just I firmly believe that anytime you know the internet, this is bigger than the internet. Like this to me is this is so big that you know, I've heard people say it's like electricity or it's like discovering fire. Like, I mean, it's it is going to change how we live, how we perform, it it's going to change everything. And I I want to just encourage people to embrace it for what it is. Learn, you know, take time every single day, learn and share it with your kids, talk about it. It don't ignore that it's here. And for companies that are putting rules out there where they're saying nobody's allowed to use AI, a lot of this is happening, guess what? Your employees are still going to go home and they're going to be using AI because it's they're fascinated with it, and they're hearing it from their kids and their their relatives. And so, you know, I guess if you had a parting message of your wisdom to share with our listeners to be able to go forward confidently with their business and how to maybe handle what they're seeing and experiencing today, what would it be?

SPEAKER_00:

Yeah, I mean, just keep a healthy uh level of curiosity, like experiment for yourself, try it, discuss it, compare notes with others, and you got to remember we're all in the same place right now at the same time, right? So we're all coming from the same level of understanding, uh, which means there is a lot to discover, there's a lot to learn uh from each other. Lean into that, be curious, use it every day, see how you can improve your life with it, and just have a conversation with it, have a dialogue with it.

SPEAKER_01:

Yeah, no, that's great. And I really appreciate again, Mike, your time. It's been very helpful. We'll have you back on the show for sure. I mean, it this is evolving again, like you and I talked about. I feel like every few weeks it's there's more and there's changes. And we're gonna the goal with with at the intersection of this podcast is to provide a platform for leaders across the globe to be able to plug in on a regular basis, you know, hear some discussions with industry experts or just insights of what's happening in the space and to try to really feel like they've got a place to go to hear the truth, not someone trying to sell them something, because I think a lot of that is happening out there too. So it's it's really just taking the time, um, healthy interaction, like you said, and to to really question what you're seeing, share notes. I love it. I think it was it was a magnificent message and episode. And again, can't thank you enough for giving us your time. Absolutely. Glad to do it. All right. Thanks, Mike.