Introduction

Thank you very much. Thanks for the opportunity to be out here. I'm the last speaker of the night, so thanks for hanging out.

Background and Experience

Just a very brief thing about myself. I'm an aerospace engineer by education and I was at U of T in the 80s and 90s when Geoffrey Hinton was there and got exposed to machine learning. At that time, went to all the industrial seminars and lectures on machine learning while I did my graduate and undergraduate degrees there.

And then I ran IBM Canada's machine learning practice for five years as one of the first worldwide users of IBM's machine learning technology. So I've been in the space for 30 plus years.

I feel the world is having the moment today that I had by myself in the 90s when I got super excited about supervised learning. But then I realized that supervised learning is not the solution to all problems.

And so I kind of brought back some of my aerospace roots into the solutions that I've built.

Founding Daisy Intelligence

I founded a company called Daisy Intelligence. We ran it for 20 years, automating retail processes and automating merchandise planning, you know, what products to promote, what prices to charge, how much inventory to allocate, no human in the loop automation. and insurance. We did claims automation and fraud detection, again, with no human in the loop. And we did medical diagnosis for brain injury with no human in the loop, systems that ran in real time.

And I'll talk about that experience and what I learned and share some of my philosophies of AI. I'm really, this hype makes me absolutely, every time I hear the word AI, I just want to shoot myself. Makes me crazy.

Observations on AI and Society

So the one thing I'd like to observe is that since the 1970s, disposable annual household income has declined by minus 7%, which is probably a lot of the reason we're in this crazy political environment that we are today, that people are not getting wealthier, right?

Another stat that's depressing is that the reason people take a job is love of the job is the fifth most important thing. So there's four other reasons than love of the job for taking a job. I think these are sad facts.

And I'd like to see, can we use artificial intelligence to help turn some of these facts around?

So the goal today, I want to share a little bit about my personal philosophy, share my experience in implementing this autonomous AI or agentic AI. We invent new terminology every day. The terminology has changed faster than the technology during my career. Started out as data mining, advanced analytics, statistics, machine learning, AI, all happened like in 10 years, right?

Autonomous AI and Decision Making

So I want to talk about, first of all, we don't separate the behavior of systems from decision making, right? So if you're using AI as a system to help with decision making, it's simply closed loop control theory.

This has been done in engineering for decades, since the 1950s and 60s. But we need to separate the dynamics. How does the system operate versus what is the optimal decision given the system operating?

So the system could be anything. It could be a business, a process. It could be a machine like a car, a plane, a spacecraft. It could be a game like Go, chess, or other things.

So you have an input into the system. Then based on that input, you want to make the best decision You run that decision through what will happen in the system if I make this decision.

What will happen if I put Coca-Cola on the front page of a flyer? What's that going to do to sales? I measure the outcome and then I feed that back and I adjust my decisions to get the desired outcome that I want.

And so all human decision-making really is about what decision should I make to get a desired outcome, right? Whether that is I want to fly a spacecraft to Pluto, I pre-plan a trajectory. What's the trajectory for optimum fuel consumption?

You fly that or you say, I want to grow sales 3% this year. What promotional mix should I execute to achieve that? right? So that's what all humanity is about.

Complex Systems and Data Limitations

But complex system dynamics can't be learned from the data. I think we think that we can learn everything from data. This is not true.

Some systems have uncountable inputs or uncountable system states, and you can't learn those systems from data. So you have to specify those by coming up with mathematical laws that describe the system, like in science. That's a scientific method.

You invent a set of differential equations that govern the system. You test it. You collect only the data that you need you don't need all the data you know in retail there's like nine facts you need so you collect nine facts you simulate the the you know the possibilities are infinite but the system dynamics are not necessarily so whoops

Reinforcement Learning

1So we use reinforcement learning to find the optimal control policy. So in some cases, you can learn the dynamics from data, but definitely, you can apply AI, or in this case, reinforcement learning to learn the optimal decision-making policy. If you're familiar with reinforcement learning, it's really where the future, I think, is going, which is closed-loop control.

So for an example, a self-driving car is a finite system. There's probably 10 to the sixth. There's roughly a million state action pairs like steering wheel position, gas pedal position, brake position, gear. There's maybe a million possible control combinations you can do that.

So you can learn how to control a car very easily with reinforcement learning. You don't need to invent the dynamics. You have the laws of physics. So we already know that you don't need to learn the dynamics

In the case of language, you know, the way we've built LLMs, you know, language, there's roughly 250,000 words in the English language. So if you want to predict a sequence of four words, let's say, let's say, you know, not every single word will follow every single possible word. So maybe you have 10 to the four possible combinations with each word you added to the sequence. If you want to do four word sequences, maybe it's like 10 to the 15 combinations. computations you have to make. And that's doable in today's computing, right?

So with the massive GPUs and the massive amount of computing energy we put into it, we can actually calculate a multi-billion parameter model that governs language. But this is just curve fitting. know to call this ai or something i find that i find the word intelligence when we append intelligence to all these things is a long uh leap a huge exaggeration we're doing very sophisticated curve fitting enabled by massive computation that wasn't feasible when i was in grad school doing aerospace computations in the 1980s

It was like my phone has three orders of magnitude more computing power than the computers back then. So this has all been enabled by brute force computing and some smart algorithms and software development. So human beings are the intelligence in today's intelligence, right?

Types of Control in AI Systems

So there's two types of control. There's open loop control with a human in the decision loop.

So Neil Armstrong didn't land the lunar lander on the moon, as contrary to popular opinion. He set the objective. He said, I want to go over there, and this is how fast I want to go through the joystick and controls.

And then the computer fired the retro rockets on the lunar lander. a thousand times a second firing like 16 retro rockets. That's beyond human ability.

So that system is beyond human ability to execute the gory details, but the human sets the objectives. That's one type of strategy you can automate with AI is let the human set the strategy so the human's not gonna tell you to be a discount retailer.

It's not gonna say you should be the best in meat. It's not gonna say you should have e-commerce and flyer promotions. The human sets those rules of the game and the brand and the strategy.

And the AI then figures out what prices should I execute? I have 100,000 products, 400 stores, 10 million customers. I have four distribution centers. I have to make 100 million decisions a day. The computer does that very well. It's beyond human ability to optimize that.

So that's one analogy. Another analogy, closed loop control. Without a human in the decision loop, might be an autonomous race car.

So if you think of an autonomous race car, you wouldn't train it. You wouldn't put a physical car on the track and say, learn to drive, because you'd break a billion cars. So what do you do? You simulate.

You have the laws of physics. There's no labeled training data for training a racing car. It's not like this is a good millisecond, bad millisecond. You don't do that.

There's the laws of physics. The data configures your simulation. You have the force of gravity, the friction of rubber on the road, the air density, the track configuration. And you say, what's the optimal sequence of steering, brake, gas pedal gear to get me the fastest lap around the racetrack?

Applications in Business and Industry

So the analogy to business is a year of business is like a lap around a racetrack. To optimize retail, it's the optimal sequence of what should I promote, what should I not promote, what regular price, what promo price, how much inventory should I put on every shelf in every store, what product should I have on the shelf. That sequence of decisions every day over the course of the year is what you're trying to optimize. And that's just like that autonomous race car.

And in insurance, it's the sequence of... Should I underwrite this policy or not? What premium should I charge? Should I pay this claim with no human in the loop? Is this claim fraudulent or not fraudulent? So you can take every industry and there's a sequence of decisions that you make that optimize that system.

So we're looking at the whole company, not a piece of the company. We're trying to simulate the entire company. And in the case of retail, It's beyond human ability.

There's 10 to the 3600 possible combinations of products to promote. If you have 50,000 products, which is a typical grocer, and you promote 2000 products in the course of a month, 50,000 choose 2000. The combinatorial math is 10 to the power of 3600. There is no supervised learning model because there's only 10 to the 80 molecules in the universe. you can't make 10 to the 3600 label.

So this is not a system that can be solved with supervised learning. This is why we need to create differential equations that govern how business works.

So AI is just math and computer science. We've been developing specialist tools since the dawn of humanity. We engineered seeds for agriculture. Seed engineering was the software development of 10,000 years ago. We bred animals that could help us hunt, that could sniff, like a bloodhound can follow a trail. We had herding animals like border collies that would help you herd your livestock. Again, so animal genetic breeding, that was a software development of the Middle Ages, right?

So we've always created tools, abacuses to do addition. We built a bow and arrow because we could fling that farther than a human could throw a stone to defend ourselves or to hunt or to go to war. So human beings have always created specialist tools.

What is AI? Is a thermostat AI? A thermostat is maintaining the temperature in a room way better than a human being can. Imagine trying to keep a room at 72 degrees Fahrenheit in the winter, we're stoking a fire. You know, you fall asleep, the fire went out, it got really cold, you put too much wood on, it got boiling hot, you had to open the window, like it's so energy inefficient. So a computer can do that very easily.

Add heat to a box, it gets warm, add cold to a box, It gets cold. We can control that. So is that AI?

Is an autopilot on an airplane, is that AI? These things have existed for decades. Is an unmanned spacecraft, is that AI? Like a satellite Voyager flying out to the outer reaches of space. So this definition, this terminology is horrific. It's massively overhyped. It is human invented computer programming and software.

Supervised learning was invented in 1805 by Legendre. If supervised learning was the panacea, you would think in 200 years it would have run its course already, right? Probability was invented in 1754 by Gauss. We understand what a random number generator does in a piece of software. It's not intelligence. It is optimization and curve fitting.

It's important to understand these distinctions because when you're buying and you have expectations of how the system performs, you need to understand what you're actually getting. That doesn't mean that you can't do amazing things with this technology, and that's what we should strive to do. Do things that people can't do that are beyond human ability that help humans thrive. That's what we should be building. This hype about AGI is ridiculous. Let's just build tools like we always have that do things better than we can, that help us live better lives. That's what the focus should be.

1So if a computer can play more games than all of humanity1 has ever played in the game of Go, for example, is it surprising that it wins? Like AlphaGo Zero in three days played more games than every human has ever played by factors of 1,000. So that it won and beat a human, it's not surprising. It's an amazing accomplishment not to diminish the feat, but it's not surprising given the fact that you can play 100 trillion games in three days because you have a GPU farm the size of a football field. Similarly, if a computer can drive a car, you can drive 100 million hours of driving simulation in one hour. It's not going to be surprising that someday autonomous cars will drive better than human beings, right?

I did a hackathon at U of T and we gave students a GPU card and we said, play ultimate tic-tac-toe. I don't know if you're familiar with that game, look it up. There's basically a tic-tac-toe board in every square of an outer game and you gotta play the inner game. And if you win the inner game, you win the outer game. And so the 100 engineering students, we built a random bot that just did 500 milliseconds of random search, like totally random with no memory, and it beat the best that engineering students could do, 100 games to nothing.

So we played the winner of the tournament. The winner of the tournament won every single game. And random bot, which we built just for testing, we said, let's just try it out. 500 milliseconds of random search beats a human being. It's just a statement that brute force computing can win many problems.

So how should we divide the world? We should let computers do what computers do best. They're called computers because they compute. Since the day a computer was invented, it was superhuman. It can add numbers faster than people can. That's superhuman, right? And so let's let computers do computing tasks.

It does that better, that use resources more efficiently. I have a problem using LLMs to build shopping lists, okay? Bill, using 100 megawatts power plant to power your shopping list creator, what a horrific waste of energy contributing to climate decline, right? Data centers are creating all kinds of heat issues for the climate. We should use this responsibly.

Using LLMs to build new drugs that'll help cure cancer, fantastic use case, okay? Building my vacation itinerary, like if you can't build your own vacation itinerary, you know, Really, like, come on, let's get real here, right? And let human beings do things that we are uniquely suited to. I have a problem automating order taking in Tim Hortons and McDonald's, not to slag those companies, but what are high school students gonna do for summer jobs if we automate all these mundane tasks, right?

Let the human beings do what we're good at. Why use computers to write poetry? What's the value to humanity of having a computer write a movie script? Is there value to humanity? There is no value to that.

Human and Computer Roles

So why are we wasting our energy doing things like that? Use it to solve problems that contribute to the benefit of humanity that don't replace humanity. We need jobs. We need to do things. Human beings are good at human interaction. A human customer service agent will always be better than an AI doing that same thing. So let's divide the world in two that way.

So to come up with a theory of business, as I said, retail is 10 to the 3,600. You can't use supervised learning.

What's the theory of business? It's very, very simple.

To improve your performance over time, so d pi over dt, pi is performance. You need to improve the quality of decisions that you make. So if you make better decisions, you will have better performance. Common sense, and it actually works.

I patented this equation in retail and insurance. And you can look up my last name on the USPTO. And I helped grow Walmart Canada sales by a billion dollars a year using this simple differential equation.

And all we have to do is saying, when I make this decision, what happens? And in retail, it's a highly interactive business.

It is based on the halo. You know, I buy ground beef to make a pasta dinner. I buy pasta, tomato sauce, cheese, bread, wine, salad fixing. So there's a halo effect, right?

There's cannibalization because I bought Coca-Cola. I didn't buy Pepsi because it was on sale. I stocked up at Christmas. I buy different things than I buy in the summer.

Price elasticity, I lower the price. People buy more. There's seven or eight concepts you can learn from the data.

And so when you execute a combination of products, we can measure the dynamical impact and improve the results.

So what does an AI system need? It needs a decision quality metric that measures the difference between every single decision you make at the most atomic level. So in retail, it's the item that you promote, the price on that item, and how many units of that item in every store. That's the most granular decision you make.

And if I can tell the difference between what's a good decision and what's a bad decision, I have a leading indicator that I can then use to drive the system, right? Whether the human beings listen to the AI or not, I can measure the quality of their decisions.

And we've shown our clients that the AI can make better decisions than them, and these decision indicators are leading indicators of the outcome. So I'll show you an example where these leading indicators are 100% correlated to the outcome.

Key Components of Effective AI Systems

You should have guardrails to protect against making mistakes. A baby learns to take one step up a staircase. That baby does not go to the top of the stairs and thinks it can fly. Learning is incremental, so you

You can't believe your BS. Your AI model is only as good as the data it's ever seen. You should confine it to operate within that domain, put guardrails on it so it doesn't do stupid things.

And you should be able to explain why the system works. Explainability is not absolute. In the case of retail, pick a reference point.

Those human decisions you made last year, why is today's decision better? Well, bananas is better than oranges because banana has a bigger halo effect. It is more elastic. It's the right time of year to do bananas. I had a better price on bananas than my competitors did. That's why it's better than an orange. And so this is a very simple explanation.

It's comparing the reference point to what you used to do, measured in terms of the metrics that are part of your theory. So you want to explain it so that you can build trust in the AI and have human beings trust it.

You want to do statistical process control to control it. You have a billion parameter model. You need to measure every parameter every day, every time you run the system so that you measure when a parameter goes out of bounds, the system will likely behave weirdly.

So then you need to say, what's the last safest thing I did? Well, let's just use yesterday's prices because they didn't put us out of business until I figure out what's going on with my system. So you need to have a billion statistical process control measures, like any manufacturing plant has this. When the temperature in the oil pipeline gets 100 degrees above normal, shut the plant down quickly, it's going to explode.

So we do the same thing. My input data is 10 times larger than I've ever seen it. The system is not going to work properly. I have half the data I usually have. It's not going to work properly. Stop. Let's do the last safe thing that we've ever done, that we did yesterday, so that I'm not going to blow my brains out.

If I run my system 10 times in a row, will it give the same answer? If your system gives 10 different answers, it's insane, useless, don't use it, right? It should agree on the majority of the cases. If it agrees eight out of 10 times, you're pretty good, right?

Lunar lander going to Mars, screaming through the atmosphere. One computer says, pull the chute. Two say, don't pull the chute. What should you do, right? You need to have this kind of majority voting thing to protect from making mistakes and making sure the system is not insane.

As I said, in retail, it's about halo effect. Consumers buy meals, not items. And so it's optimizing the combinatorial combination of products, prices, inventory allocations.

In risk management, which would be banking, insurance, disease, are you different than the norm and different than the norm is a very complicated thing so if you go to the doctor they take your measurements your medical measurements are different than the norm that means you're sick you know if you're different than the norm your credit card application is very different than every other credit card application you're probably lying it's fraudulent right And finding different than the norm is not easy.

So there's the obvious orange rock in the middle. But if you group them into peers, round rocks, big rocks, small rocks, rocks with damage, and then find the outliers, it gets very difficult to do that. So AI outlier detection is very complicated. But that's the theory of outliers.

So many human decisions are either interaction dominated, like the halo effect, or they're outlier dominated, like medicine, risk, Insurance, banking, security, the three-letter agencies, border security are all outlier dominated.

So this is an example for a retail client that, so there's the average rank, so where we rank the decision quality of every single item that you're promoting, the average of that when the rank is high, so number one is a good, so a low rank is a good rank. So the lower the rank, the higher the total store sales. That correlation is real data with one of our retail clients in the past.

And with a correlation coefficient of like 0.6 to 0.8, this decision quality indicator correlates to good outcomes. And then weekly total store impact is like 0.9 correlated to year over year same store sales. So you have a leading indicator that is created three months in advance of the decision. And if it's this correlated to the outcome, you should just follow the system blindly.

Impact of AI on Business Performance

We can put human beings on the moon. We can build nuclear power plants. We can build accelerators to accelerate protons to 99.999% the speed of light. We can automate business processes today to the benefit of humanity.

If I help Walmart make a billion dollars, they put that back in our pockets. They lower prices and they increase customer service levels. And if we do that in all industries, we can lower the cost of living for humanity.

This is an example of a US grocer that took a year of historical promotions, and then we took that decision quality indicator where we rank every item from best item, second best item, down to 100,000th best item. And then we colored green and then red. The top 10% of items are green. And then the next 15% are yellow. And then the bottom 2 3rds is red. So I just combined the yellow and reds in this graph. And this just shows that every single week, green decisions outperform yellow decisions outperform red decisions with 100% accuracy.

And if you eliminated all the red and yellow decisions and did just the green, that's worth 3% of total company sales. This was an $80 billion grocer. That's worth $2.4 billion in increased sales with no increase in cost overhead. So it goes straight to the bottom line. Grocery is a 1% net margin business. That means we tripled an $80 billion company's net margin if they're willing to listen to the machine and act on it.

1Technology is not the challenge, human change management is. Getting people to play along and not be afraid of losing their jobs, that's what the real challenge is.

Which one of these flyers was created by AI and which one wasn't? Better decisions are invisible. You can't tell. If it's doing it right, you can't tell. The one on the left was the AI one. That one on the left drove 30% more transactions and $13 million more in company revenue between the two.

just by optimizing that halo and the and those effects that i talked about this is an insurance this is the false positive rate based on the quality metric of is this claim suspicious or not so the higher the suspicion the lower the false positive rate right so one of the challenges in fraud detection is you have a huge false positive rate If you're predicting a 1% rare event with a 90% accurate model, 90% of the time you will be doing wild goose chases. The false positive rate is 90%. That's why doctors tell you to get a second opinion. If your AIDS test come back and says you have AIDS and it's a 90% accurate test, it's only a 3% chance that you actually have AIDS because of the false positive rate, right? can help with that.

So this showing that when the system has a high degree of suspicion, the false positive rate is very low. So this is why using an AI non-supervised learning system generates better results than predictive modeling. And this is an insurance.

I told you that for retail, we increased total company sales by 3% to 5%. I did this for 20 retail clients around the world. This is insurance, about a $400 million business.

We showed that we took the claim automation from 5% of claims with no human touch to 30% of claims with no human touch, saving $50 per claim in processing cost. And we found us, you know, $3 million in fraud, finding one to 2% of claims as suspicious, which they validated only a third of them, but we found that that's worth about 2.6 million. So in a $100 million revenue during COVID, revenue went down quite a lot because travel was a way, there's a travel medical client. So we found 6 million in incremental savings across $100 million business using automation.

Characteristics of Ideal AI Problems

So what characteristics of problems are ideal for AI or reinforcement learning? These are problems that are beyond human ability, that are highly mathematical, have huge amounts of data, are highly repetitive, and have long-term objectives, right?

Because, you know, human beings are, we're not systematic. We don't do the same thing twice given the same situation. We get tired, we get bored, we can't,

Phone numbers are only 10 digits because that's all we can keep in our head at one time.

The fact that we say that we don't understand what an LLM does, yeah, no kidding, it's got a billion parameters. There's no way we'll ever figure out what that does.

AI never gets tired. It's consistent. It takes into account all the data and it can achieve a long-term goal.

Conclusion

So I said at the beginning, annual disposable household income is declining at 7% a year, horrific, and creating all kinds of issues, which is why I would say one of the reasons Trump won, because he talks about the economy. People care about the economy. I care about the economy.

The fact that job satisfaction, liking the job is the fifth reason, this is horrible. If we divide the world like this, using AI for computing tasks that help humanity and let us do what we are the best at, what people are good at, what we enjoy, then I hope that we can turn this around, that we can make household disposable income back on the rise and we can make, I love my job is the number one reason you take a job, right?

If we make all companies efficient, they will invest those savings in price and innovation and customer service level. Smart companies do not dividend that money out.

And this is how we should use AI to solve narrow, specific problems. This is where the conversation needs to go. And it's 100% doable today. I spent 20 years doing this, delivering these results.

Thank you so much.