Capitolo 1
When Prediction Becomes Cheap, Everything Changes
The year 2012 marked a turning point in human history that most people missed. At the ImageNet competition, a neural network suddenly achieved a dramatic breakthrough in image recognition, slashing error rates by an unprecedented margin. This moment sent shockwaves through the tech industry, triggering Google's $600 million acquisition of DeepMind (a company with minimal revenue) and eventually prompting the Chinese government to invest billions after watching AlphaGo defeat the world's best Go players. What happened? According to economists Ajay Agrawal, Joshua Gans, and Avi Goldfarb, we witnessed something profoundly transformative yet deceptively simple: prediction became cheap.
"Prediction Machines" has become required reading in boardrooms across Silicon Valley and beyond. Elon Musk reportedly keeps a copy on his nightstand, while Microsoft CEO Satya Nadella distributed it to his entire executive team. The book's genius lies in stripping away the mystique surrounding artificial intelligence and reframing it through a practical economic lens that business leaders can immediately grasp. When something fundamental becomes drastically cheaper-whether it's artificial light in the 1800s or prediction today-it transforms society in ways both predictable and surprising.
Capitolo 2
The Economics of Artificial Intelligence
Economists view technological breakthroughs differently than most people-not as revolutionary innovations but as simple price reductions. When prices fall, usage increases. The commercial internet wasn't a "New Economy" as many claimed in 1995, but rather a dramatic reduction in the cost of distribution, communication, and search. Similarly, AI represents a collapse in the cost of prediction-filling in missing information using data you have to generate information you don't have.
The impact of these price drops can be deceptively powerful. An improvement in prediction accuracy from 85% to 90% might seem modest, but it reduces mistakes by one-third. Going from 98% to 99.9% reduces errors twentyfold-a transformation that can completely reshape industries. Credit card fraud detection exemplifies this evolution: improving from 98% to 99.9% accuracy means the number of fraudulent transactions that slip through undetected drops by a factor of twenty.
What makes prediction machines seem magical is how they're transforming activities we never previously thought of as prediction problems. Google Translate didn't improve overnight by adding more grammatical rules-it reframed translation as a prediction problem: "Given this French text, what English text would a human translator produce?" Self-driving cars don't follow pre-programmed routes; they continuously predict "what would a human driver do in this situation?"
As prediction becomes cheap, we'll use it more frequently and in novel ways. But prediction is just one component of decision-making. A complete decision requires prediction plus judgment-determining the relative rewards of different outcomes. While machines excel at prediction, humans maintain the advantage in providing judgment. This creates a new division of labor where machines handle prediction while humans focus on specifying what matters.
Capitolo 3
From Data to Intelligence
If prediction is the core of modern AI, then data is its lifeblood. Machine learning systems require three types of data: input data (what you want predictions about), training data (examples pairing inputs with known outcomes), and feedback data (information about prediction accuracy that enables improvement). Each type plays a crucial role in developing accurate and reliable AI systems.
Cardiogram, an app that detects irregular heart rhythms with 97% accuracy, illustrates this perfectly. The app collects heart rate data from users' wearable devices (input data), compares patterns against training samples from 6,000 users including 200 with diagnosed conditions, and continuously improves through feedback when predictions are verified by medical professionals. Similar applications have emerged in other medical fields, like dermatology apps that analyze skin conditions and diabetes management systems that predict blood sugar fluctuations.
Organizations must make strategic decisions about data acquisition's scale and scope. How many types of data should we collect? How many examples do we need for training? How frequently should we collect it? These decisions should be driven by the specific prediction problem and balanced against collection costs and privacy concerns. For instance, a facial recognition system might need millions of images across different angles, lighting conditions, and demographics to be effective, while a specialized industrial defect detection system might work well with just thousands of examples from a specific manufacturing line.
While data technically shows diminishing statistical returns (the first hundred observations provide more insight than the next hundred), it often creates increasing economic returns in competitive markets. Google's advantage with rare search queries comes from having more data on such searches than competitors. While each additional search technically provides less statistical value than previous ones, having the most comprehensive dataset creates disproportionate market advantages. This phenomenon is evident in other domains too - Netflix's recommendation engine benefits from vast viewing history data, and Tesla's self-driving capabilities improve with every mile driven by their vehicle fleet.
Data quality is as crucial as quantity. Clean, well-labeled data often outperforms larger quantities of noisy data. Organizations must invest in data infrastructure, including collection mechanisms, storage solutions, and preprocessing pipelines. They must also consider data governance frameworks to ensure compliance with privacy regulations like GDPR and CCPA, while maintaining the data's utility for AI applications.
The timing of data collection also matters. In dynamic environments, historical data may become less relevant over time, requiring continuous updates to training datasets. For example, consumer preference prediction models need regular updates to capture changing trends, while fundamental physics models might rely more on established historical data.
Capitolo 4
The New Division of Labor
As prediction machines improve, the division of labor between humans and AI is evolving in fascinating ways. Donald Rumsfeld's framework of "known knowns, known unknowns, and unknown unknowns" helps explain when prediction machines excel or falter.
With rich data on familiar scenarios ("known knowns"), machine prediction outperforms humans. This is the sweet spot for current machine intelligence, where the machine provides good predictions and we know they're reliable-as in fraud detection, medical diagnosis, and bail decisions.
Humans still excel with limited data ("known unknowns"). While machines struggle with sparse information, humans can recognize faces after seeing them once or use analogy to understand new situations. Scientists are developing techniques like "one-shot learning," but currently, these scenarios require human intervention.
Both humans and machines fail with truly novel events ("unknown unknowns"). As Nassim Nicholas Taleb explains in "The Black Swan," some events are fundamentally unpredictable from past data. Europeans couldn't predict black swans in Australia because they'd never seen them.
Perhaps the biggest weakness of prediction machines is providing wrong answers they're confident are right ("unknown knowns"). This often stems from misunderstanding causality. When machines don't understand the decision process that generated their training data, they make critical errors. Garry Kasparov described a chess algorithm that sacrificed its queen immediately because it learned from grandmaster games that queen sacrifices preceded victories-reversing the actual causal relationship.
This creates an efficient division of labor called "prediction by exception"-machines handle routine scenarios independently while humans focus only on non-routine exceptions flagged by the system. This approach allows one human to leverage a prediction machine's advantages across many routine predictions while providing critical judgment for unusual cases.
Capitolo 5
Judgment: The Human Advantage
As AI improves prediction, the value of human judgment increases. While machines predict, only humans can express the relative rewards of different actions. With better prediction come more opportunities for judgment-determining the rewards of various actions-giving us more decisions to make.
Consider credit card fraud detection. For a transaction nine times likelier to be fraudulent than legitimate, the company will deny the charge unless customer satisfaction is nine times more important than the potential loss. With concrete monetary values-like $20 recovery cost on a $100 transaction-companies can calculate precise thresholds for action.
While economists traditionally treat rewards as givens, determining accurate payoffs requires significant time and cognitive effort. This process involves deliberation-thinking through what you want to achieve-or experimentation-trying different approaches to learn outcomes. The more specific scenarios you evaluate, the more time-consuming judgment becomes.
Navigation apps like Waze demonstrate how prediction and judgment interact. While these apps excel at predicting the fastest route, users often override suggestions based on personal objectives beyond speed-like needing fuel or avoiding stressful driving conditions. Though apps can incorporate some preferences (like Tesla's navigation accounting for charging stations), they struggle with more subjective factors. Humans maintain an advantage in understanding their multidimensional objectives, which are often idiosyncratic and subjective.
As prediction machines improve, determining how to best use their predictions becomes critical. Reward function engineering-determining the rewards for various actions based on AI predictions-requires understanding both organizational needs and machine capabilities. Sometimes this means hard-coding judgment in advance (as with self-driving vehicles), but engineers must be careful that AI doesn't over-optimize one metric at the expense of broader goals.
Capitolo 6
Taming Complexity Through Prediction
Prediction machines enable robots and systems to handle vastly more complex environments by identifying more "ifs" and enabling more "thens" in decision processes. Unlike early autonomous systems like the 1980s Mailmobile that required carefully controlled environments and chemical trails, modern prediction-powered systems can adapt to varied conditions and unexpected obstacles. This evolution represents a fundamental shift from rigid, pre-programmed responses to dynamic, context-aware decision making.
Better prediction identifies more situational variables-more "ifs"-allowing machines to react appropriately to complex combinations of conditions. Modern delivery robots can now assess wet surfaces, lighting conditions, human proximity, and distinguish between animals to determine appropriate actions without pre-planned paths. For example, Starship's delivery robots can navigate crowded college campuses, adjusting their speed and path based on pedestrian density, weather conditions, and surface terrain. These systems can even predict human behavior patterns, slowing down near building exits or busy intersections where people might suddenly appear.
Prediction also expands our action options-the "thens"-by reducing uncertainty. Apps like Waze provide accurate travel predictions that eliminate the need for buffer strategies like arriving excessively early at airports. Rather than following rigid rules (like "leave two hours before flight"), better prediction enables contingent rules that optimize departure times based on real-time traffic and flight conditions. Modern navigation systems can even suggest alternative routes mid-journey based on emerging traffic patterns, accident reports, and weather changes, something unimaginable just a decade ago.
Humans often "satisfice"-taking shortcuts rather than making perfectly rational decisions when facing complexity. Nobel Prize-winner Herbert Simon recognized this limitation in both humans and computers. We've developed workarounds like airport lounges and invasive medical procedures to manage poor prediction. These compensatory mechanisms, while useful, represent inefficient solutions to prediction problems. For instance, doctors often order multiple diagnostic tests because they can't accurately predict which one will be most informative, leading to increased healthcare costs and patient discomfort.
As prediction machines improve, they enable handling more "ifs" and "thens," reducing risk and transforming decision-making by expanding options. In healthcare, AI systems can now analyze thousands of variables in patient data to predict potential complications before they occur. In manufacturing, predictive maintenance systems can forecast equipment failures weeks in advance by analyzing subtle patterns in sensor data, allowing for optimal scheduling of repairs and minimizing costly downtime. These advances are not just improving existing processes but fundamentally changing how we approach complex decision-making across industries.
Capitolo 7
When to Fully Automate Decisions
Full automation happens when machines handle prediction, judgment, and action-removing humans entirely from the decision process. This occurs most readily when: (1) everything except prediction is already automated, (2) speed of response is critical, or (3) waiting time for predictions is costly.
Rio Tinto's iron ore mines in Australia's remote Pilbara region demonstrate successful full automation. The company deployed 73 self-driving trucks that operate continuously without breaks, saving 15% in operating costs. These trucks don't need drivers, air conditioning, or even to turn around. AI predicts hazards and coordinates movement, completing the automation process where everything except prediction was already mechanized.
Automation makes most sense when there's "no time to think" (like emergency braking) or "no need to think" (when prediction leads to an obvious action). When speed is critical or judgment can be easily coded, removing humans from the decision loop becomes beneficial.
Even when full automation is technically possible, legal requirements may mandate keeping humans in the loop. This reflects concerns about ethical dilemmas like the "trolley problem" in self-driving cars, where someone must program how vehicles respond to unavoidable harm scenarios. From an economic perspective, the need for human oversight increases with the potential for "externalities"-costs borne by those outside the decision-making process.
Some actions remain inherently better when performed by humans rather than machines. Research shows people find jokes less funny when they believe machines recommended them, even though machines are actually better at predicting humor preferences. Similarly, artistic achievement and athletic competition derive much of their value from human experience and participation.
Capitolo 8
Strategic Implementation of AI
Large corporations must break their work flows into tasks, estimate ROI for AI implementation in each, and prioritize accordingly. While some AI tools can be dropped into existing work flows for immediate benefit, most require rethinking or "reengineering" entire processes-which explains why mainstream businesses may take time to see productivity gains from AI.
The AI Canvas framework systematically separates decision elements: prediction, input, judgment, training, action, outcome, and feedback. This approach helps decompose tasks to understand a prediction machine's potential role, providing discipline and clarity in identifying each component.
Atomwise exemplifies this approach with their pharmaceutical drug discovery tool that predicts binding affinity between molecules and proteins, dramatically accelerating the identification of promising drug candidates. Their AI tool processes millions of possibilities, creating ranked lists of molecules most likely to bind to specific disease proteins, while pharmaceutical companies retain the judgment role in evaluating disease targets and potential side effects based on their strategic priorities.
Three critical insights emerge from decomposing tasks for AI implementation: First, task decomposition reveals where prediction machines add value, enabling cost-benefit analysis and prioritization of AI tools by ROI. Second, the AI canvas provides structure for this process, forcing clarity about data requirements, prediction objectives, judgment criteria, possible actions, and outcomes. Third, identifying the core prediction often triggers existential discussions among leadership about true organizational objectives.
Capitolo 9
Reshaping Business Strategy Through AI
C-suite leadership must directly engage with AI strategy rather than delegating it to IT departments, as AI can fundamentally change business models when three factors align: (1) a core trade-off exists in the business model, (2) uncertainty influences this trade-off, and (3) AI reduces uncertainty enough to tip the scales toward a different strategic approach. This strategic involvement requires executives to understand both the technical capabilities and business implications of AI implementation.
Consider how improved prediction could transform Amazon's business model from shopping-then-shipping to shipping-then-shopping. If Amazon's AI could accurately predict what customers want to buy, it might become more profitable to ship products before customers order them. While today's 5% prediction accuracy would make this impractical (requiring too many returns), a sufficiently accurate prediction machine could fundamentally change Amazon's strategy. This concept, known as anticipatory shipping, could revolutionize e-commerce by reducing delivery times from days to minutes, creating a significant competitive advantage.
Otto, the German e-commerce company, demonstrates this perfectly-they used AI to predict with 90% accuracy what products would sell within a month, allowing them to reduce inventory by 20% while cutting returns by 2 million items. This prediction capability resolved their delivery-time dilemma without expensive inventory holdings, fundamentally changing their logistics strategy. Their system analyzes 3 billion past transactions and 200 variables including weather data, website clicks, and economic indicators to make its predictions. The success has led to Otto applying this AI approach across their entire product range, demonstrating how predictive capabilities can scale across operations.
AI implementation in one area can affect other parts of the business, potentially redesigning workflows and company boundaries. For example, when AI handles routine predictions, employees can focus on higher-value activities requiring judgment and creativity. Prediction machines increase the value of complements like judgment, actions, and data-potentially changing organizational hierarchies, enabling higher-level optimization, and creating competitive advantages for those who own the actions affected by predictions. Companies like Stitch Fix demonstrate this by combining AI predictions with human stylists' judgment to provide personalized fashion recommendations, creating a hybrid service model that neither humans nor AI could achieve alone.
The ripple effects of AI implementation extend to organizational structure, requiring companies to rethink traditional departmental boundaries. For instance, when AI predictions span multiple functions - from supply chain to marketing to customer service - organizations often need to create cross-functional teams and new coordination mechanisms. Companies like Ping An Insurance have restructured their entire organization around AI capabilities, creating integrated platforms that combine multiple services based on predictive insights.
Capitolo 10
Societal Implications and Trade-offs
The rise of AI presents society with three fundamental trade-offs that will shape its development and implementation. First, the productivity versus distribution dilemma: While AI will undoubtedly enhance economic productivity through automation, process optimization, and improved decision-making, it may simultaneously exacerbate income inequality. This occurs through increased competition for remaining human tasks, displacement of middle-skill workers, and disproportionate benefits flowing to highly skilled workers who can leverage AI tools. For example, AI-powered legal research tools make skilled lawyers more productive while potentially reducing demand for paralegals and junior attorneys.
Second, the innovation versus competition trade-off highlights how AI's unique economics could reshape market structures. The technology exhibits strong scale economies and increasing returns - as AI systems process more data, they become more capable, creating powerful feedback loops. This dynamic, combined with the massive computational resources required for training advanced AI models, may lead to market concentration and monopolization. Companies like Google, Microsoft, and OpenAI already dominate certain AI domains, raising concerns about innovation stagnation and market power.
Third, the performance versus privacy tension becomes increasingly acute as AI systems hunger for data. Modern AI algorithms demonstrate significantly improved performance with larger training datasets, especially those containing personal information. This creates an inherent conflict between maximizing AI capabilities through comprehensive data collection and protecting individual privacy rights. For instance, medical AI systems could provide better diagnoses with access to complete health records, but this raises serious privacy concerns.
China's approach to AI development exemplifies these trade-offs while highlighting its three distinct advantages. First, massive government investment: China is committing unprecedented resources to AI development, with individual cities like Tianjin allocating $5 billion - exceeding entire national AI budgets like Canada's. Second, population scale provides China with an unmatched data advantage: its 1.4 billion citizens generate vast quantities of training data across diverse applications, from facial recognition to consumer behavior. Third, China's relaxed privacy protection policies give domestic companies like Face++, Baidu, and SenseTime significant advantages in developing sophisticated AI systems, particularly in areas like facial recognition and personalized services.
This situation creates potential for a "race to the bottom" on privacy standards as countries compete for AI leadership. Nations may feel pressure to relax data protection regulations to remain competitive, even as their citizens express concerns about privacy. Different jurisdictions are already taking divergent approaches: while the EU emphasizes individual privacy rights through GDPR, China prioritizes data access for AI development, and the US maintains a relatively fragmented approach. These varying strategies reflect different cultural values, economic priorities, and governance models, suggesting that a unified global approach to AI development may prove elusive.
Capitolo 11
The Economics of Our AI Future
Prediction machines are valuable because they produce better, faster, and cheaper predictions than humans. These machines excel at pattern recognition and can process vast amounts of data to make forecasts about everything from consumer behavior to medical diagnoses to weather patterns. As prediction becomes cheaper and more accessible, human prediction skills naturally decline in value, while complementary skills like data collection, judgment, and action become increasingly valuable. For instance, while AI might predict a patient's likelihood of developing a disease, doctors' judgment in determining treatment plans becomes more critical.
The AI revolution parallels trading with a fictional "Robotlandia" - economists view this as beneficial free trade that creates overall prosperity despite disrupting some jobs. Just as international trade reshapes labor markets, AI is transforming employment patterns. As prediction becomes cheaper, jobs focused on providing judgment will grow in value, particularly in reward function engineering - the complex task of defining what success looks like for AI systems. A key example is autonomous vehicle development, where engineers must carefully define what constitutes "good" driving behavior.
The labor market impact is nuanced: while high-skilled prediction jobs (like radiologists and financial analysts) may decline, new opportunities emerge in judgment-focused roles. For instance, while AI might handle routine medical image analysis, healthcare workers increasingly focus on patient communication and treatment decisions. Similarly, while algorithms handle stock trading, financial professionals shift toward relationship management and complex strategy development.
Current AI remains narrow in scope - AlphaGo Zero mastered Go but can't make a sandwich or adapt to board size changes. This limitation is evident across AI applications: image recognition systems excel at specific tasks but lack general understanding, and language models can generate text without truly comprehending meaning. Several well-funded companies, including DeepMind and OpenAI, are explicitly pursuing human-like intelligence, investing billions in advancing AI capabilities beyond narrow applications.
The National Science and Technology Council believes general AI remains "decades" away, but the eventual move beyond prediction machines to general AI or superintelligence will fundamentally alter economic principles. This transition could reshape labor markets, productivity, and economic structures in ways current models struggle to predict. Some experts warn of potential disruption far beyond historical technological revolutions.
For now, understanding AI as cheaper prediction provides a powerful framework for navigating this technological revolution. Organizations can leverage this perspective to identify where AI can add value - from improving inventory management to enhancing customer service. By focusing on economic principles rather than technological mystique, companies can make strategic decisions about AI deployment, transforming not just operations but entire business models. For example, retailers use predictive analytics to optimize supply chains, while insurance companies employ AI to improve risk assessment and pricing models.
