Economics of AI Lecture Notes

General Introduction

Good morning, good evening, everyone; lecture starts after a brief wait.
Today's topic: Economics of AI with four main ideas:
- Microeconomics: Budget planning and costs of adopting AI in organizations.
- Macroeconomics: Global impact of AI.
- Labor Economics: Role of humans in the AI era.
Objective: Provide an understanding of where costs are incurred in AI projects and broader economic implications.

Computational Power:
- Measured in FLOPS (Floating Point Operations Per Second).
- Necessary for training large data sets and models.
- Increasing significantly since around 2008 with the advent of deep learning.
Data:
- Size and diversity of data contribute to AI costs.
- Longtail Distribution of Data: Many rare data points create additional challenges and costs.
Machine Learning Models:
- Different algorithms result in varying costs.

Most expenses are incurred during training models, which involves significant computation.
Historical data shows exponential growth in computational resources used for AI models since 2008.
Famous models trained using increasing FLOPS depict the rising cost of computation over time (e.g., GPT-4 is highly computationally intensive).
Increasing computational power impacts AI performance positively but at significant costs.

Economies of Scale: Marginal cost of data collection decreases with scale, but the value of additional data diminishes.
Diseconomies of Scale: Value of further data points declines faster than the cost per unit of data falls.

Longtail phenomena in various domains like search engines and e-commerce.
Graph Explanation: High search volume in a few keywords but a long tail of rare keywords creating high costs.
Building AI models for long-tail data is significantly more expensive.
Continuous learning and retraining are essential in handling longtail data distribution.

Shift from academic to industry-led AI research and development starting around 2014.
Capital and Computational Resources: Industry has more resources to invest in AI.
Focus on computing power and the role of key players like Nvidia in the GPU market.

AI models’ dependence on specific countries for computational resources (e.g., chip manufacturing dominated by Taiwan, US).
Geopolitical Control: Concentration of power in tech giants can lead to strict regulations and sovereign AI discussions.

AI models require continuous updating and retraining, especially with heterogeneous data sets and with considerations for the longtail distribution.

Keep It Simple (KISS Principle):
- Start with simple models to avoid unnecessary complexity and costs.
- Simpler models are more interpretable and often suffice.
Consider Continuous Learning:
- Retraining to handle data diversities and new data points.
Optimizing User Experience:
- Design-centric tweaks (e.g., auto-complete) can reduce longtail data complexities and simplify model requirements.
Multiple Models Approach:
- For heterogeneous data sets, splitting into cohorts and training multiple models can be more efficient.
- Example: Different bots for different types of attacks.

Study by Harvard Business School and BCG on 758 consultants to measure impact of AI.
Groups split: one with access to GPT-4, one without.
Key Findings:
- AI improves productivity and quality of work output.
- AI acts as a skill leveler, significantly aiding less proficient workers.
- The scope for improvement is higher for lesser skilled workers compared to more skilled ones.

Different industries will be impacted variably (e.g., legal, finance heavily impacted, some like arts less so).
Insight that high median wage industries will see more AI automation.
Example: Graphic design, AI is cheaper and faster than humans.

Predicted prevalent human-AI hybrids where AI aids humans in complex and creative tasks.

Covered computational, data, algorithmic, and human capital aspects of economics in AI.
Key emphasis on balancing costs, model simplicity, continuous learning, and human capital in AI economics.

Upcoming presentations will finalize the course with assignments focused on pitching AI-driven startup ideas.