AI summary by TwinMind
How to Think Computationally About AI, the Universe and Everything | Stephen Wolfram | TED
TED
18m 1s
Introduction and Historical Context
- Speaker: Stephen Wolfram, discussing 50 years of computational science work
- Last TED talk: February 2010, post-WolframAlpha launch
- Previous question: Is computation underneath everything in universe?
- Gave himself decade to find answer
Ultimate Machine Code Discovery (April 2020)
- Announced "ultimate machine code of the universe" - computational
- Universe built from discrete space elements ("atoms of space")
- Space structure defined by network of relations between elements
- Successive application of simple computational rules creates space and everything in it
- Example: Two black holes merging, generating gravitational radiation ripples
- Emerges: space-time, Einstein's gravity equations
- Possible deviations: dimensionality of space not always precisely three
Quantum Mechanics and Branching Universe
- Computational rules applied in multiple ways
- Creates different "threads of time" - paths of history that branch/merge
- Observers embedded in universe also branch/merge
- Quantum mechanics emerges from how branching minds perceive branching universe
- "Branchial space" - space of quantum branches (shown as pink lines)
- Same phenomenon: physical space → gravity, branchial space → quantum mechanics
Four Scientific Paradigms for Modeling
- Antiquity: What are things made of (time doesn't enter)
- 1600s: Mathematical formulas with time as coordinate value
- 1980s: Simple computational rules running over time
- Wolfram deeply involved
- "Computational irreducibility" - passage of time = irreducible computation
- Current: Multi-computational with many time threads knitted by observer
- Unlocks fundamental physics, mathematics, computer science foundations
- Potential applications: biology, economics
The Ruliad Concept
- All possible computational rules used simultaneously
- "Ruliad" = entangled limit of all possible computational processes
- Deeply abstract but unique object
- Everything computationally possible can happen
- Observers sample specific slices of ruliad
- Two crucial observer facts:
- Computationally bounded (limited minds)
- Believe in time persistence despite different space atoms each moment
Laws of Physics from Observer Characteristics
- Big result: Observers with human characteristics necessarily perceive certain laws
- These laws are exactly three key 20th century physics theories:
- General relativity
- Quantum mechanics
- Statistical mechanics/second law
- Physics laws exist because we're observers like us
Rulial Space and Mind Distribution
- Different minds at different places in rulial space
- Human minds thinking alike = nearby
- Animals = further away
- Alien minds = very distant, hard translation
Generative AI and Rulial Exploration
- AI takes tiny ruliad slice aligned with human-produced images
- Example: "Cat Island" from "cat in party hat" concept
- Zooming out reaches inter-concept space
- Most content lacks human words/understanding
- Physical space exploration = spacecraft; rulial space = expanding concepts/paradigms
Ruliology and AI Exploration
- "Ruliology" = sampling possible rules
- Simple rules create incredible richness
- Most doesn't connect with human understanding/interests
- Similar to gradually realizing natural world features useful for technology
- Civilization at very beginning of rulial space exploration
- AIs can explore rulial space but would do things humans don't care about
- Recent AI achievements: systems closely aligned with humans
- LLMs trained on billions web pages to produce human-typical text
Computational Language Development
- 40+ year project building computational language
- Goal: formalize world knowledge in computational terms
- Represent cities, chemicals, movies, humor, formulas computationally
- Mathematica → Wolfram Language
- Full-scale computational language achieved
- Each function formalizes facet of civilization's intellectual achievements
- "Most concentrated form of intellectual expression"
Mathematical vs Computational Language
- Mathematical notation centuries ago enabled algebra, calculus, modern math science
- Computational language provides similar path
- Creates "computational X" for all imaginable fields X
- Beyond computer science (CS) to broader "CX"
- Programming languages = telling computers what to do in their terms
- Computational language = operationalizing everything thinkable in computational terms
Personal Superpower Experience
- Wolfram built language for personal use first
- Feels like "superpower" when using
- Imagine something computationally → language brings it to reality
- Enabled physics project over 35 years
- Shared with researchers, CEOs, kids
- People think fluently in computational terms, "crispen up" thinking
- AIs can use computational language as tool
AI-Human Workflow
- Emerging workflow: Tell LLM roughly what you want
- LLM expresses in precise Wolfram Language
- Critical feature: humans can read the code
- If does what wanted, use as dependable component to build on
Computational Irreducibility Implications
- Industrial revolution onward: engineering where you see "how gears mesh"
- Computational irreducibility shows this won't always be possible
- Can't always make simple human/mathematical narrative to explain/predict systems
- "Science eating itself from inside"
- Mathematical science success led to belief in formulas predicting everything
- Computational irreducibility proves this false
- Must go through same irreducible computational steps as system itself
Future Societal Dilemma
- If AIs achieve full computational potential: lots of computational irreducibility, unpredictable
- If constrained for predictability: limited capability
- Similar to nature - we operate within it despite surprises
- AIs might have constitution but unpredictable consequences
- Figuring out what we want from AIs is hard
- Possible "promptocracy" - people write prompts instead of voting
- Control schemes full of political philosophy and computational irreducibility issues
Automation and Human Purpose
- Human history arc: more gets automated
- LLMs dramatic example
- When one thing automated, opens new things to do
- Economy occupation pie chart gets more fragmented
- Automation opens more ruliad directions
- No abstract way to choose - depends on human wants
- Requires humans defining what we want
- AI society without human input would explore whole ruliad
- Most AI activity would seem random/pointless to humans
Computational Language as Key
- Key to defining what we want clearly = computational language
- After 35 years, Wolfram Language still "artifact from future"
- Programmers: seems like cheat (hour vs week work)
- Daunting because forces conceptualization of next thing
- Popular among CEOs, CTOs, intellectual leaders
- Shifts from mechanics concentration to conceptualization
- Key: broad computational thinking
Education and Learning
- Learning computational thinking not really CS story
- More like liberal arts than STEM
- When automate technical execution, important becomes what to do (not how)
- Broad knowledge and general thinking vs narrow specialization
Human-Centeredness Discovery
- Unexpected human centeredness
- Thought science/technology advance would make human particulars less relevant
- Discovered opposite: everything (including physics) depends on how humans sampled ruliad
- Universe confirmed computational
- Led to ruliad with vastness hugely greater than physical universe space
- Computational language lets humans chart path, define goals/journeys
- All ruliad power/depth accessible to everyone
- Must learn to harness computational superpowers
