Philosophy of AI, Or: How to be a Better Prompt Engineer

Meta-philosophy

What people think it is

  1. No practical applications
  2. Obsession with what things "really mean". Just look it up in a dictionary...
  3. Obsession with pedantry

What it actually is

  1. Not really a criticism...
  2. This should NOT be the point of philosophy
  3. Obsession with precision, clarity, good questions

IBM's Deep Blue: rule-based

Google DeepMind's AlphaGo: statistical

ELIZA: rules-based

GPTs: statistical

Rules-based models

  • "if-then" conditions
  • deterministic
  • hard to scale
  • can very quickly become complex

Statistics-based models

  • Probability distributions
  • Datasets provide the probabilities
  • indeterministic
  • requires large datasets
  • "learns" from data
  • scales with complexity and large datasets

Language Models and Translation

  • Translation seems to be a rules-based activity
  • Why might rules-based approaches not work well in translation?
  • Ambiguity (where meaning depends on context)
  • Idioms ("keeping mum", "raining cats and dogs")
  • How can it get the grammar right without being programmed with the rules of the grammar?

Emergence

  • The behaviour of large colonies from the actions of individual ants
  • Collective human "crowd" behaviour
  • How prices of financial instruments emerge from buying a selling actions of market participants.
  • How consciousness emerges from neural activity

Emergence

  • Emergent properties are not reducible to the properties of individual components
  • Manifest novel and unexpected characteristics at higher levels of organization
  • Whole is greater than the sum of its parts, displaying new qualities or behaviors
  • Emergent phenomena cannot be explained by analyzing components in isolation

Emergence in LLMs

  • Rules of grammar
  • Every single response
  • GPT-3: unprompted CoT reasoning.
  • Solving problems not included in the training data(?)
  • AlphaGo's famous "move 37"

Emergence in LLMs

Basic Emergence

reproduces patterns the same as or very similar to its training data: regurgitating training text.

Compositional Emergence

LLM combines patterns in new put predictable ways.

Novel Emergence

LLM appears to somehow produce an output which transcends it's training data. Not an obvious example of basic or compositional emergence.

Meta-patterns

  • Problem-solving strategies
  • Mathematical and logical principles
  • Patterns of logical reasoning

Novel = Compositional on the pattern-level

Prompting

  • "Output your response as JSON"
  • "Ensure your response follows ethical principles"
  • "Always be truthful"
  • "Do not hallucinate"
  • Explains the success of CoT or few-shot prompting

Other Musings

  • Does this mirror human lanugage learning?
  • "PhD-level" LLMs
  • What are "rules" anyway?