AI and Economic Dynamics (AIED) Course

Artificial Intelligence and Economic Dynamics

• Simone Vannuccini is associated with the Chair of Economics of Artificial Intelligence and Innovation at Université Côte d'Azur.
• IPEAI5 is running from 2024 to 2025, under the EUR-ELMI program.

Artificial Intelligence and Economics Dynamics (AIED) Course

• Provides an economic perspective on Artificial Intelligence (AI).
• Focuses on understanding the dynamics of AI adoption and its economic impact.
• Aims to assess how well existing economic theories capture AI's influence on the economy.
• Emphasizes the evolving nature of AI as an industry with interdependent components.

AIED Learning Outcomes

• Grasp key technological aspects of AI.
• Develop critical awareness of current issues in AI development.
• Learn how economic models can explain dynamics within the AI industry.

AIED Course Programme

• Lecture 1 & 2 covers "what is AI? And how new it really is?".
• Lecture 3 discusses AI 'in the wild' and the emergence of the AI industry.
• Lecture 4 & 5 offers basic economics of digital technology and AI.
• Lecture 6 & 7 focuses on model #1: replicator dynamics & AI.
• Lecture 8 & 9 discusses model #2: industry life cycles & AI.
• Lecture 10 addresses the competition and industrial policy of AI, and a recap of the course.

Introduction to AI

• The course begins with an inquiry into whether AI is truly new, and what general concepts define AI.

Defining AI

• AI is described as a 'suitcase word' (Mitchell 2019).
• AI Technology involves software executing tasks typically associated with human intelligence, such as vision, speech recognition, and pattern identification.
• AI algorithms rely on inductive data processing and specific software/hardware architectures like artificial neural networks.
• AI is a diverse field spanning CS, maths, and philosophy.
• Narrow or weak AI aims to solve specific tasks and has the following concepts to consider:
  • "in silico brain": embedding mechanisms of natural intelligence into computers
  • emulation: Performing tasks that humans can perform, as well or better as they do, without focusing on the mechanism

AI Approaches

• Symbolic systems use rule-based execution of instructions (GOFAI).
• Subsymbolic/connectionism learns through artificial neural networks.
• Generative AI is an instance of connectionism.
• Machine Learning uses techniques that allow AI to learn and operate, drawing from statistics.
• Robots are not necessarily AI, but there's an emerging trend of vision-language-action models.

Intelligence and Consciousness

• Humans tend to attribute consciousness to humans, living creatures, supernatural beings, and inanimate objects.
• Developing machines that perform activities associated with consciousness can lead people to imagine that they are conscious.
• Cognition requires wants and/or needs, alongside social embeddedness.

AI as a System Technology

• Current AI is the weak+connectionist version, which are specialized systems that emulate human intelligence without true cognition or understanding.
• Current AI is a system technology (Vannuccini and Prytkova 2023)
• The formula for AI is AI = LA + HDD + CP + DS + talent:
  • LA (learning algorithm) == models such as Transformers/LLMs
  • HDD (high-dimensional data) == Big Data
  • CP (computing power) == Hardware
  • DS (domain structure) == examples being pictures, games, voice, prompt; multi-mode

How AI Works in Practice

• Connectionist/statistical AI uses Artificial Neural Networks (ANNs) with algorithms organized as a network structure.
• Learning in ANNs occurs through training with data, with 'memory' encoded in the structure of connections.
• Once trained, ANNs use new data to make predictions, classify information, and generate content.
• Deep Learning automates structure updating through back-propagation for 'optimal circuit search'.
• Convolutional neural networks and feature extraction, for example, categorize images and extract features.
• Other AI models include generative adversarial networks (GANs), Transformers, reinforcement learning, and Diffusion models.

Generative AI and Large Language Models

• Generative AI (GenAI) broadly refers to AI systems that produce content.
• Large Language Models (LLMs), also called Foundation models, are becoming the industry's dominant design.
• A Language Model is a probability distribution over a natural language.
• LLMs consists of these three criteria:
  • Models text that can be used to generate it
  • Receives large-scale pretraining (large data, not parameters) – at least 1B tokens
  • Makes inferences based on transfer learning (fine-tuning, prompting)
• LLMs can be viewed as:
  • Technology (engine)
  • Products (engine+interface)
  • commoditisation of AI