← Lectures·L3

Demystifying AI

Zanca · 1 concepts · 9 questions

Concept 1 / 1

Key Concepts to Memorize

The AI hierarchy (nested, largest to smallest):


Artificial Intelligence (1955) ⊃ Machine Learning (1980) ⊃ Deep Learning (2010) ⊃ Foundation Models (2017)

  • AI (1955): Imitate intelligent behaviour in computers
  • ML (1980): AI by learning and predicting from data
  • DL (2010): ML using deep neural networks
  • Foundation Models (2017): Generative AI based on pretrained transformers

Are LLMs intelligent? → No, not really


  • High performance ≠ human-like behaviour
  • LLMs fail on adversarial inputs trivial for humans (e.g., unusual 3D viewpoints)
  • "Error consistency" measures whether AI and humans fail on the _same_ examples — DNNs and humans fail differently

Tool vs. Model distinction:


  • Tool: Statistical algorithm, no claim about human behaviour (e.g., DNN for image classification)
  • Model: Formal representation of a scientific theory, used to explain/predict a phenomenon

Remaining challenges of current AI:


1. Robustness and Safety — vulnerable to adversarial attacks

2. Non-Generalizable Single-Purpose Design — models trained for one task can't transfer


AI Success Stories (specific examples):

SystemPurpose
AlphaGoMastering the game of Go
AlphaFoldPredicting protein 3D structures
MyShakeEarthquake early warning system (smartphone network)
Apple Watch ECGDetecting atrial fibrillation & left ventricular ejection fraction
Autonomous drivingPerception & path planning (e.g., Waymo)

What AI needs: Data & Compute


  • ChatGPT's success driven by growth in both data and compute
  • Europe is falling behind the US and China in AI capacity

Consciousness in AI:


  • No consensus on what consciousness is — subjective and hard to measure
  • Key distinction: intelligence ≠ consciousness
  • Main theories: Global Workspace Theory (GWT), Integrated Information Theory (IIT), Higher-Order Theories, Panpsychism
  • Expert camps: Skeptical / Optimistic / Pragmatic / Ethical ("better safe than sorry")
  • Efficiency gap: 1,000 artificial neurons needed to simulate a single biological neuron