Avi Goldfarb - The Economic Impact of AI - [Invest Like the Best, EP.321]
What this covers
Patrick talks with Avi Goldfarb, a Rotman professor and co-author of Prediction Machines and Power and Prediction, in a conversation recorded before GPT-4's release. Most of it rests on one historical analogy: the move from 1890s steam factories, where belts tied machines to a central engine, to electric ones. Goldfarb notes that Edison patented the light bulb in 1880, yet half of US workplaces were not electrified until the 1920s. Swapping a motor into an unchanged workflow saved only 5 to 15 percent on energy. From this he builds a ladder of point, application, and system solutions and uses it to place today's AI products.
The range is wide. He defines prediction in the statistical sense as filling in missing information and argues it shifts the world from rules toward decisions. His examples include online advertising's move to real-time auctions, Uber and Lyft combining GPS with demand prediction, and Blockbuster's failure, which he calls organizational rather than a failure of vision. He also covers rapid COVID tests stalled by undecided questions about sick pay and privacy. He contrasts Christensen's demand-side disruption with supply-side disruption.
The more contested turns come late. Goldfarb suggests explicit predictions will expose hiring biases that humans currently bundle invisibly with judgment, and he uses I, Robot to separate prediction from value judgment. He warns that the larger risk is a few companies owning the technology, producing inequality. His practical advice is to sort a company's activities into those that deliver its mission and those that compensate customers for failures, with airports as the illustration. He also says data strategy should begin with the exact prediction needed.
Goldfarb argues AI's economic impact comes from making prediction cheap, which forces system-level reinvention of workflows, business models, and professions rather than simple point-solution automation.
- Historical general-purpose technologies such as electricity only transformed economies after factories and workflows were redesigned around them, a process that took about 40 years.
- Prediction changes decisions by making probabilities explicit and by decoupling prediction from human judgment and bias, which requires coordinated changes across many decisions.
- The biggest opportunities lie in businesses that compensate customers for failures and in complementary assets, while the main risk is concentration of power among owners of AI.
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Although Edison patented the electric light bulb in 1880, it was not until the 1920s that half of US workplaces and half of US homes were electrified, so roughly 40 years passed between recognizing electricity's potential and it affecting most people's lives.
“Edison's patent for the electric light bulb was in 1880... It wasn't until the 1920s that half of the U.S. workplaces and half of U.S. homes were electrified.”
In 1890s factories, a central steam engine powered everything through belts, so machines had to be clustered around the engine in tightly packed multi-story buildings, and workflow was dictated by the location of the power source rather than the production process.
“All the machines in the factory were connected by belts to the steam engine... you want to locate your most power-hungry machines as close to the steam engine as possible.”
Swapping a steam engine for an electric motor without changing the workflow saved factories only about 5 to 15 percent on energy costs, which was not worth the cost of the change for most factories, so they kept their old setups.
“We can save 5%, 10%, 15% on the energy cost... For most factories, saving 5%, 10%, 15% wasn't worth trying to figure out how to remove the steam engine”
Steam-powered machines could rarely be switched off because stopping the engine was costly and risky, so they ran continuously, while electricity let individual machines be switched on and off, enabling new production processes within existing workflows.
“turning on and off the steam engine was really, really hard... With electricity, even if you took out the steam engine... you could start to do different kinds of machines because you can have them flip on and off.”
Prediction makes the world's inherent probabilities explicit, replacing the lazy binary round-up-or-down heuristic with numbers like a 36 percent chance, though most people struggle to act differently on a 36 percent versus a 28 percent probability.
“without a prediction machine, you have a lazy way of thinking, which is we tend to think it's either this or that... And what prediction machines do is they give us an explicit number.”
Blockbuster did foresee digital distribution, but its franchised structure gave franchise owners no incentive to adopt it, so its failure was an organizational inability to change rather than a failure of vision.
“if you actually look at the Blockbuster documents, they saw it coming. They just couldn't manage their organization to move in the right direction. They were largely franchised.”
Online advertising first copied the magazine rate-card model, but prediction enabled one-to-one targeting and real-time auctions, which created an entirely new advertising industry with players like demand-side and supply-side platforms that could not exist without AI.
“The advertising industry is an industry that's experienced system-level change through prediction technology.”
Application solutions sit between point and system solutions, letting companies build new processes that fit into the old system without changing everything, and early successful entrepreneurs usually build these, as with 1990s internet services and electronic data interchange in supply chains.
“Application solutions are at a high level, they're somewhere between the point solution and the system solution. So they allow you to build new processes without having to mess with the entire system.”
Clayton Christensen's demand-side disruption occurs when a startup offers an initially worse product for a niche that incumbents' core customers ignore, while Rebecca Henderson's supply-side disruption occurs when incumbents' standard operating procedures make them unable to adopt a new business model that would change who does what.
“Rebecca Henderson developed, what you might think of as a supply-side disruption, which is companies have a way of doing things. This is around SOPs, standard operating procedures.”
Joshua Gans recognized that COVID was not a health problem for most people, since more than 99 percent of the population did not have it at any time in summer 2020, but an information problem about who was infectious, which made it an opportunity for prediction.
“At any given time, more than 99% of the population, this was the summer of 2020 did not have COVID. ... Once you understand that COVID is an information problem”
When a situation is well understood, firms can build a digital twin simulation and use reinforcement learning to test thousands or millions of scenarios, generating strategic data for predictions, as Singapore does with a simulation of the entire city-state.
“Singapore has a digital simulation of the entire state... you can simulate it and then try hundreds, thousands, or even millions of different possibilities.”
Uber and Lyft built an entirely new transportation system by combining GPS navigation, digital dispatch, and demand prediction, which upskilled anyone who could drive into a professional-equivalent driver.
“Uber and Lyft and the other ride-hailing services created an entirely different system... they upskilled anybody who could drive could now effectively be a professional taxi driver.”
Hiring decisions bundle prediction of success with human bias, which keeps biases hidden; once AI makes predictions explicit, such as 90 percent versus 85 percent success likelihood, people must confront their biases, which the speaker sees as a hopeful effect of AI.
“once you have a prediction saying, okay, this person is 90% likely to succeed, and that person is 85% likely to succeed, which one are you going to hire? And then we have to face our biases”
In the statistical sense, prediction means taking the information you have and filling in the information you don't have, which is what AI does in diagnosis, lending, and writing.
“By prediction, I mean in the statistical sense, taking the information you have and filling in the information you don't have.”
General-purpose technologies such as steam, electricity, and computing all required businesses to reinvent how they operate before gaining value, and the same reinvention is expected from AI over the next decade.
“That kind of reinvention process where in order to get value out of the new technology, you have to do things differently, occurs over and over again.”
The bigger risk from AI is that the underlying technology is owned by a small number of companies, which could make the owners of capital and machines much better off while skills decline, leading to massive inequality and exploitation of labor.
“I worry about the concentration of power through AI.”
The key insight of electricity was that it was distributed power, which decoupled the location of the power source from the location of the machines, enabling a new factory built around production flow, with single-story modular layouts and the mass-market automobile.
“electricity wasn't just cheap power, but electricity was distributed power. It allowed you to decouple the location of the power source for the location of the machines.”
Many of the rules we follow exist because we lack good information, and they work because everyone following the same instruction coordinates well, but when prediction gives information about individual states, people can make decisions instead of applying uniform rules.
“A lot of the way we do things now is because we don't have good information, so we just follow the instructions because if we follow the instructions, and everyone else follows the instructions, we know things will coordinate well.”
To find AI disruption opportunities, a company should separate its activities into those that actually deliver on its mission to serve customers and those that compensate customers for its failures, since a large share of compensation activities signals a real opportunity for entrepreneurs to skip them with better prediction.
“Category number one is actually delivering on your mission to serve customers. And Category number two is the things you do to compensate your customers for your failures.”
Rapid tests were the best available tool for predicting COVID infectiousness, but they were not useful for a long time because adopting them required coordinated decisions about sick pay, health data privacy, and biohazard waste disposal that organizations had not made.
“the best tool we had for particularly whether somebody had COVID was a rapid test... Unfortunately, that tool for a long time wasn't useful.”
In I, Robot, the robot's rescue choice was based on a prediction of 45 percent versus 11 percent survival, but the real decision depended on a value judgment about the relative worth of the man's life versus the girl's, which humans normally bundle invisibly with prediction.
“the prediction for the rest of the decision is hard... The judge is the relative value of an adult man's life versus a little girl's.”
When arithmetic becomes cheap enough, problems that were not previously solved with arithmetic get reframed as arithmetic engineering problems, as mail, music, and pictures were; pictures moved from Kodak's chemical engineering to arithmetic about 20 years ago.
“Kodak solves pictures with chemistry, chemical engineering... by 20 years ago, we started to really think about pictures as an arithmetic problem, no longer a chemical engineering.”
Major airports like Seoul Incheon are spectacular with shopping and restaurants, but these amenities compensate for hours of waiting, while super-rich travelers use private terminals that avoid that failure, showing that much of an airport's effort addresses failure to deliver smooth travel.
“But how do the super-rich fly? ... They're sheds. Private terminals don't have great shopping and great restaurants”
Moving from rules to decisions usually requires coordinating many complementary decisions across people inside and outside the firm, which makes the process far more complicated unless a single actor controls all of them.
“In most companies, in most contexts where you're trying to really move from rules to decisions, you need to coordinate between lots of people within the company”
Companies that collect data without a clear purpose waste millions, so the data strategy should start by specifying the system-level change and the exact prediction needed, then identifying what data is already available and what must be collected.
“lots of companies say, we're going to try to organize all of our data. They spend millions and millions and millions of dollars... What does that ultimate prediction look like? What are we trying to predict?”
Companies can create proprietary data by launching point or application products that collect data at scale, as with cars equipped with sensors, and the first company to deploy autonomous vehicles would gain a feedback loop that competitors struggle to match.
“you can often create it. And you can often create it by launching products into the market that don't represent the system-level change”
Contrary to the expectation that easy-to-replicate manual jobs like truck driving would be automated first, generative AI has first affected creative and white-collar outputs, because these models are prediction systems over pixels, words, and letters.
“we would replace a bunch of easy-to-replicate human jobs first... it seems to have happened almost exactly the opposite.”
System solutions are harder to find than point or application solutions because they require changing how the whole workflow operates, and the knowledge-based taxi driver example shows how a prediction tool like GPS can disrupt a skill that took three years of training to acquire.
“System solutions for AI are harder to find... a taxi driver... could take a long time to learn your way around the city... that takes about three years of school”
Better AI diagnosis of a disease that is not very rare but hard to identify would let pharmaceutical companies sell far more of an existing drug, increasing profits and helping more patients, because the bottleneck is identifying who needs the drug rather than producing it.
“There's often a challenge to identify people who need their drugs... The pharma company can now sell a lot more increasing their profit and also help a lot more consumers.”
Generative models are most useful when the user already knows what good output looks like, because they make the process more efficient by producing many candidates to choose from and editing toward the right answer, even though outputs are often wrong.
“they make the process much more efficient as long as you know what good output looks like.”
Writing essays turned out to be a prediction problem: ChatGPT takes a query and fills in what the five paragraphs should look like by drawing on similar essays in its data, something the authors did not anticipate even when writing Prediction Machines.
“We never imagined, frankly, even when we're writing prediction machines that writing essays would be a prediction problem.”
AI's future impact cannot be predicted because prediction technology depends on data, and for genuinely new capabilities there is no historical data, so no one knows what will happen.
“We absolutely can't predict it. It's prediction technology, so it relies on data. It's something totally new, we don't have data.”
ChatGPT is a major opportunity for people who have strong skills but are limited by weak writing, because better writing expands their work and income opportunities, and organizations that employ them first will gain the most.
“there's lots of people who have amazing skills that are not getting ready. And so writing limits their opportunities for work, their opportunities for income.”
Automating diagnosis, which is the special domain of the hundreds of thousands of US doctors, will remove their special skill advantage, but it will let millions of nurses, pharmacists, and other medical professionals do their jobs much better and more productively.
“There's hundreds of thousands of doctors in the U.S. and their special skill in diagnosis is going to go away. They'll have to retool... But there's millions of other medical professionals who are now going to be able to do their jobs much better”
Many professions such as law and medicine could become like farming, with far fewer practitioners and much greater efficiency that removes costly intermediaries, so a world with fewer doctors and ten times as many medical professionals would be desirable.
“a world where we have fewer doctors, but 10x as many medical professionals sounds like wonderful.”
The most exciting aspect of AI is that millions of people can now do what only thousands could do, which will be stressful and disruptive for the thousands of current specialists but a major benefit for everyone else, requiring policy attention to the transition.
“millions of people can now do the things that historically only thousands could do. That's going to be stressful and disruptive for the thousands”
Rules-based education that moves children cohort by cohort could be replaced by prediction-driven personalized learning, letting a teacher adapt lessons to each student while preserving social-emotional benefits of staying with a peer cohort.
“there's rules everywhere in education... we can now think through personalized education in terms of what they're learning in the classroom.”
The first company to launch autonomous vehicles at scale would gain a compounding data advantage that is hard to compete with, but regulators require enough prior data to demonstrate safety, which has been a meaningful barrier.
“they're going to collect more and more and more data to be hard for anybody to compete.”
Disruption opportunities arise when prediction enables an entirely new way of operating, but that requires new people and new assets, which can be too heavy a lift for incumbents to undertake.
“a totally new way of operating requires new humans -- a different set of humans who are doing different things, a different set of assets, and that could just be too heavy of a lift.”
Someone still has to invent the interface that lets millions of people use generative AI for writing, whether OpenAI, Microsoft, or a startup, and whoever builds it could upskill millions of people and drive large economic gains while leaving out some who will be worse off.
“somebody has to invent the interface that allows millions of people to take advantage of what this technology can do to write more easily”
Ada Support automated parts of Zoom's customer service by drafting standardized responses for review, and when pandemic demand rose by well over a factor of 10 in March 2020, Ada helped Zoom scale without proportionally adding representatives, an example of an application solution.
“March 2020 came and Zoom's demand for these customer service queries went up by well over a factor of 10. And they could keep up”
A landscaper with poor English built an interface on GPT-3 that turned his terse notes into coherent customer emails, and his business thrived, showing that the interface layer unlocks the value of the underlying model for ordinary users.
“a whole bunch of... a interface with GPT-3 to help the guy to turn his quick note, hello, arrive, Wednesday, nine to a nice cohere e-mail to a potential customer”
London's taxi drivers needed about three years of schooling to learn the city, but GPS and real-time traffic maps let anyone drive point to point nearly as efficiently, so navigation prediction substitutes for accumulated knowledge.
“that takes about three years of school to know enough about the city to be a licensed taxi driver.”
Verafin, a fraud-prediction company in St. John's, Newfoundland, already predicted financial fraud and switched to machine learning with the same workflow, which made it cheaper and better, and it became Canada's first AI unicorn.
“Verafin actually was in that business already in predicting financial fraud, but not using machine learning... It was just cheaper and better, but no one had to change anything.”
Those who will be hurt by AI-driven skill changes need mitigation through income support or training, and designing these responses is a serious policy question.
“how do we mitigate the pain that they're going to experience, whether that's through income or training or otherwise”
Generative AI tools like ChatGPT and DALL·E are fundamentally prediction systems over pixels, words, and letters, which is why they produce hallucinations and wrong outputs while still making the best prediction they can.
“that's all these things are, pixel prediction or word prediction or letter prediction, that's all these things do, which is why they hallucinate”
When the price of something falls, people use more of it, and with technological change the fall can be extraordinary, as with computing, which is fundamentally arithmetic performed very cheaply.
“when the price of something falls, you do it more. So if coffee gets cheap, you're going to buy more coffee. ... the fall in the price can be extraordinary.”
Over the nineteenth and twentieth centuries farming required far fewer workers, yet food shortages are no longer a serious issue in North America, showing that dramatic labor reduction can coincide with abundance and that similar efficiency gains are possible across industries.
“yes, we have a lot less farmers, but food shortages are not really an issue in North America anymore.”
Loan officers once judged borrowers by looking them up and down, and over the twentieth century lending became more rigorous, with banks now using machine learning to predict repayment.
“Over the course of the 20th century, that became more and more rigorous.”
Patrick cites the idea of skeuomorphism, that new technologies are first applied to replicate prior-era forms, such as the iPhone Notes app resembling a notepad, as an example of point solutions.
“They call it skeuomorphism. You tend to take the tech and apply it to something we did in the prior era.”
A writer who spent a day drafting an article can now produce a usable draft in minutes by querying, reviewing, and editing AI output, allowing far more content and new opportunities for creative entrepreneurs.
“maybe you have someone who spends all day writing an article and now that's going to take them five minutes.”
Had a good AI or testing tool predicted COVID infection status accurately, the blanket stay-home rule and the resulting crisis could have been largely avoided, because people could make individualized decisions.
“If we had a good AI for predicting whether somebody had COVID, then there would have been no crisis.”
Doctoral students already receive personalized education because each professor has about two of them, and AI-based production tools could extend similar personalization to the rest of the university, including MBA students.
“Doctoral students largely get personalized because two doctoral students per professor. But the rest of the university, they don't get that personalized education”
In the AI gold rush, the companies that benefit most are the ones selling picks and shovels, namely providers of compute power such as chipmakers and cloud services, and traditional big tech firms currently occupy that position.
“There's going to be huge opportunities for companies that are providing the compute power underlying AI. Right now, that looks like it's the traditional big tech companies”
Investors' comfort with probabilistic thinking will transfer to everyday life, because thinking about relative likelihoods of outcomes rather than repeating the same action is useful across decisions.
“skill you have in investing and picking probabilistically will permeate all aspects of life”
Goldfarb's vice dean, who believed in the economics of technology, structured Goldfarb's early career to set him up for success despite roughly ten initial journal rejections, and Goldfarb only learned of this after the dean retired.
“The first maybe 10 papers I submitted to journals, got rejected. ... I only found out when he retired.”
Goldfarb, who is poor at drawing, uses DALL-E to create images at scale, generating many options of which one is usually useful, showing that generative tools expand what non-specialists can produce.
“I can now create images at scale in a way I couldn't before.”
ChatGPT showed that large language models advanced faster than the speaker had expected, even though he had already seen earlier GPT versions doing impressive work.
“ChatGPT showed me they advanced faster than I'd expected”