Economics of Innovation: Data and AI - Introduction

Comprehensive practice flashcards covering the introductory lecture on the Economics of Innovation (Data and AI), including economic frameworks, AI definitions, historical developments, machine learning fundamentals, and macroeconomic impacts.

According to the report cited in the lecture, by how much is AI projected to boost the economy by 2035?

By €250bn.

According to research from Trinity Business School, what did AI adoption in Ireland surge to, compared to its 2024 rate?

It surged to 91%, nearly doubling from 49% in 2024.

What three broad job categories are highlighted as being at risk or targeted for AI automation?

Low-skill work (e.g., cleaners, laborers, construction and manufacturing workers), areas of focus for AI development (e.g., transportation, warehousing, clerical work), and data-intensive roles (e.g., mortgage originators, paralegals, accountants).

What are the four key elements that define any system, including an economic system?

Inputs, Processes, Outputs, and Outcomes.

Within an economic system, what three primary capital stocks serve as inputs?

Physical and financial capital ($K$), labour / human capital ($L$), and natural resources or natural capital ($N$, specifically land and energy).

How is 'technology' defined when analyzing the processes of an economic system?

Processes that convert inputs into outputs, conceived broadly by economists to encompass scientific technology, policy technology, and trade.

What mathematical function is used to represent the economy as a system of inputs, technology, and output?

$Y = f(A, K, L)$, where $Y$ is income/output, $A$ is technology or Total Factor Productivity (TFP), $K$ is physical capital, and $L$ is labour.

What is the distinction between outputs and outcomes in an economic system?

Outputs (such as goods, services, and income) are the means to an end, whereas outcomes represent the ultimate ends (such as happiness, health, freedom, life satisfaction, and fulfillment of potential).

What phenomenon is referred to as 'the curse of AI'?

"As soon as a problem in AI is solved, it is no longer considered AI because we know how it works."

What occurred during the 'AI winter' of the 1980s?

Widespread public disappointment over unfulfilled promises of AI led to severe cuts in AI research funding.

What three major factors explain why modern AI is succeeding today after past failures?

1. Availability of cheap, cumulative data (via digital technologies and lower storage costs); 2. Exponential increases in processing power (Moore's law, GPUs, and ASICs); 3. Research breakthroughs in machine learning, specifically Deep Learning (circa 2010).

How does Taddy (2018) define Machine Learning (ML)?

A set of statistical and computational tools to automate predictions based on data; a tool for pattern recognition.

What is the core idea of Machine Learning regarding feature engineering?

Programmers do not need to explicitly program which features to use or model interaction terms; instead, the machine learns the best prediction directly from data.

What standard labeled dataset of handwritten digits, shown here, is commonly used to demonstrate machine learning classification?

The MNIST dataset.

How does Taddy (2018) define Artificial Intelligence in terms of problem decomposition?

AI automates decisions in tasks traditionally associated with human cognition by breaking complex problems into simple prediction tasks.

In an autonomous vehicle, how is a decision (such as stopping or lane changing) derived from inputs?

Raw inputs (Camera, LiDAR, GPS) feed multiple simple prediction tasks (traffic sign recognition, lane detection, vehicle trajectory prediction), which are aggregated into a final decision.

What is the key functional difference between Traditional AI and Generative AI?

Traditional AI is designed for specific tasks based on rules or ML on labeled data (one tool = one task), whereas Generative AI generates entirely new content from user prompts across many domains (one tool = many tasks).

How do Large Language Models (LLMs) learn during the training process?

They guess missing words on text examples sourced from the internet; mistakes are penalized to update the model until it learns what words co-occur and in what order.

Why is AI categorized as a potential General Purpose Technology (GPT)?

Because it automates cognitive tasks previously considered impossible to automate, allowing them to be executed reliably and scaled across the entire economy at almost zero marginal cost.

From an economic perspective, how can technological progress in AI be fundamentally characterized in decision-making under uncertainty?

As a sharp reduction in prediction costs ($\downarrow \text{prediction costs}$), allowing machines to provide certainty cheaply.

In decision-making under uncertainty, what distinction remains between AI systems and human judgment?

Machines provide cheap and accurate predictions, but only humans can assign subjective payoffs and evaluate ethical values for specific actions.

What medical diagnostic example was presented to illustrate AI's superior prediction accuracy over human professionals?

Google's AI successfully detected 92% of breast cancer metastasis compared to a 73% detection rate by human pathologists.

What key ruling did Judge Amit Mehta deliver in the August 2024 US antitrust case against Google?

Judge Mehta ruled in favor of the Department of Justice, determining that Google is an illegal monopolist in search and advertising markets.

In Microsoft's research study 'Working with AI', which occupations exhibited some of the lowest Generative AI coverage scores?

Physical and manual roles, including phlebotomists (0.06), nursing assistants (0.07), embalmers (0.07), plant/system operators (0.05), hazardous materials removal workers (0.04), and tire repairers (0.04).

In the historical demonstration using the 1641 depositions, how did manual event extraction compare to extraction using OpenAI's API?

Manual extraction took 1–2 days per 10 depositions (equivalent to 2–3 years), whereas using the OpenAI API took 0.5–1 minute per deposition (approximately 6 days total).

According to Acemoglu's 2025 framework ('The Simple Macroeconomics of AI'), how does a task-level AI effect translate into an economy-wide impact?

Effect per task maps via the task-per-occupation structure to the 'Effect per occupation', which then maps via the occupational structure of the economy to produce the 'Effect on the Economy'.