The most underrated area of innovation in artificial intelligence isn't the computing, or even the development of algorithms and techniques for data collection, but the human ability to reframe problems in terms of prediction.
Speaking at the Fiduciary Investors Symposium in Toronto, Ajay Agrawal, a prominent economist and academic, said it's helpful to think of AI and machine learning as “simply a lowering of the cost of prediction.”
Agrawal is the Jeffrey Tabor Chair in Entrepreneurship and Innovation at the University of Toronto's Rotman School of Management, where he is also a professor of strategic management.
“AI is computational statistics that makes predictions,” Agrawal says.
“That's it. So on the one hand, it seems very limiting. On the other hand, what's so remarkable about this is that we've discovered all that can be done with high-fidelity prediction.”
Agrawal said that, simply put, prediction is “using information you have to generate information you don't have.” He said that “the creativity of people who take a problem that nobody in this room thinks of as a prediction problem and turn it into a prediction” is the foundation of AI's growth and potential.
“Five years ago, no one in this room would have said that driving is a matter of prediction.”
“Few people in this room would say that translation is a prediction problem. Few would say that replying to emails is a prediction problem. And yet that's exactly how we solve all these problems today.”
Whether it's predictive text when replying to emails or improving investment performance, the AI systems that power it are “all implementations of statistics and predictions,” Agrawal says.
These predictive models culminated in large language models (LLMs), where machines are trained how to predict the next most likely word in a sequence of words that make up an entire sentence, paragraph, or response.
“If you think about a language, say English, every book you've ever read, every poem, every Bible is the same rearrangement of letters — 26 letters and a few punctuation marks rearranged over and over again to create every book. What if we could do that with behavior?” Agrawal said.
From LLM to LBM
The LLM (next most likely word) principle is currently being applied to large-scale behavioral models (robots) by training them to predict the next most likely verb or action.
“Then you could take all the tasks. Think about everyone you know, all the jobs they do. Every job has probably 30 to 40 different tasks, so hundreds of thousands of tasks. But what if all those tasks were actually just a sequence of a handful of verbs?
“What they do is train a robot to teach 50, 80, 120 verbs. And then, just like with chat GPT, you give the robot prompts. You say to the robot, 'Can you open that box and put the tools on the shelf?' The robot listens to the prompts and predicts the best verb sequence to complete the task.”
Agrawal said this is “another application of forecasting.”
Agrawal said businesses and industries are now facing a “wave of problems being reframed as prediction problems.”
“So we’re now turning to machine intelligence for many of these things.
“The problem is that things have gotten so sudden and difficult that people seem to be struggling with where to start and how to actually channel this into something useful.”
Agrawal said it's helpful to be very specific about the metrics and performance indicators that need improvement.[point] “It's AI.”
“AI is a mathematical optimizer, and you need to know what you're optimizing for,” he said.
“If the problem is that there's a wave of new solutions available and we just don't know how to leverage them, then the way to think about solutions is short-term and long-term strategies.”
Agrawal said the near-term strategies are primarily productivity improvements, with implementation within a year that aims to improve productivity by 20 percent, with a payback period of less than two years.
“And here's the key point: There's no change in workflow,” he said.
“In other words, this is really a technology project that you just put in place and the rest of the system stays the same.”
A true game changer
Long-term strategies take time to implement, but they are true game changers that will bring 10 times more benefits than short-term implementations. But the key is that they will require a redesign of workflows. Agrawal said that AI is a general-purpose technology, just like electricity, and
A useful analogy is when factories were first electrified and began to transition away from water-powered engines.
For the first 20 years after electricity was invented, its penetration was very low, with less than 3% of factories using it, and even when they did, “its primary value proposition was reducing input costs by, for example, replacing gas lamps.”
“Nobody wanted to destroy the existing infrastructure to make a small profit,” Agrawal said.
“The only people experimenting with electricity were entrepreneurs building new factories, and yet when it came to designing the factories, most people said, 'No, I want to stick with what I know.'”
But a few entrepreneurs realized there was an opportunity to completely reimagine and redesign the factory to run on electricity, since it no longer needed to transmit power through long steel shafts from an engine outside the factory to run the factory's machines.
When the shafts were eliminated, the large columns within the factory that supported them were also eliminated, allowing for lighter, less expensive construction and allowing the factory design and layout to be all on one level.
“They redesigned the entire workflow,” Agrawal said.
“The machines, the materials, the material processing, the flow of people, everything [was] It has been redesigned. In some factories, it has increased productivity by up to 600 percent.”
Agrawal said initially, the difference in productivity between electrified and non-electrified factories was very small.
“For people who have non-electrified plants and want modern electricity, they might find it more hassle than it's worth,” he said.
“But electricity is only just starting to improve productivity.
“Now we're seeing the same thing with machine intelligence. [and] “AI adoption rate”
This is what we learn from
But Agrawal said, “What makes AI different from any other tool in human history is that it is the only tool that learns from humans.”
This, he said, explains the sudden rush to develop the technology and the large amount of capital being poured into it.
“The way AI works is that whoever gets an early lead has better AI. And the better the AI, the more users you get, the more data you get, the more data you get, the better the AI predictions become,” he said.
“And once the flywheel starts turning, it's very hard to catch up with them.”
Agrawal said AI and machine learning are developing so rapidly that it is virtually impossible for businesses and entities to keep up, let alone implement and adapt.
“My focus isn't on technology capabilities, because obviously that's important and it's advancing rapidly,” he said.
“But what I'm looking at are the unit economics of companies that experiment with it first and then put it into production,” he said.
“Costs continue to come down as AI learns and improves, so my sense is you have to pay very close attention to the unit economics of how much it costs to do things.”
“And then you can really take a deep look at any product or service stack and start applying these machine intelligence solutions to that and see how that changes the unit economics.”
