AI summary by TwinMind
Nick Bostrom on the Joe Rogan Podcast Conversation About the Simulation | AI Podcast Clips
Lex Fridman
21m 41s
Simulation Argument Part 3: Probability Logic
- Core question: If somebody eventually creates simulation, why does that mean we're likely in one?
- Alternative 3 leads to more simulated people with our experiences than non-simulated ones
- By end of time, counting up total would show more simulated than non-simulated beings
- Extra step needed: moving from "more simulated exist" to "we are probably simulated"
Bland Principle of Indifference
- Indexical statement problem: determining probability that "this person right now" is simulated
- Sets of observers: one much larger than other
- When can't tell from internal evidence which set you belong to
- Should assign probability proportional to size of sets
- Example: If 10x more simulated people with your experiences, 10x more likely to be simulated
- Rationale: Without enough information, rationally assign probabilities based on set size
Time Requirements and Civilizational Development
- Don't need infinite time for simulation argument
- Need time for universe to produce intelligent civilization with technology for ancestry simulations
- Once first simulation created, others start creating simulations
- Different planets/civilizations may start at different cosmological epochs
- Some planets take billion years longer for monkeys/bacteria development
- Staggered timeline across universe
Connection to Doomsday Argument
- Both involve anthropic reasoning about indexical propositions
- Simulation argument requires much weaker assumptions than Doomsday argument
- Doomsday argument discovered by Brandon Carter (theoretical physicist, 1970s-80s)
- Developed by philosopher John Leslie (book ~1996)
- Other versions by Richard Gott (physicist)
Doomsday Argument Mechanics
- Argument: We systematically underestimate probability humanity goes extinct soon
- Most people think something wrong with argument, but hard to identify exact flaw
- Urn analogy explanation:
- Two urns: one with 10 balls (numbered 1-10), one with million balls (numbered 1-million)
- 50/50 prior probability which urn
- Draw ball #7 randomly
- Strong evidence for 10-ball urn (10% chance vs tiny chance from million-ball urn)
- Bayesian update makes you virtually certain it's 10-ball urn
Doomsday Argument Applied to Humanity
- Two hypotheses: 200 billion total humans vs 200 trillion total humans
- Start with 5% prior chance extinction at 200 billion (optimistic view)
- Your birth rank: ~100 billion (roughly how many humans existed before you)
- Each person has numerical position in human sequence
- If only 200 billion total: unremarkable middle position
- If 200 trillion total: remarkably early (what are chances you'd be #100 billion out of 200 trillion?)
- Should update toward lower total human population (doom soon)
Self-Sampling Assumption
- Core assumption: reason as if you're random sample from all humans that will ever exist
- You know you're human #100 billion, alive in 2020
- Literature exists on supporting this assumption
- Intuitive appeal: assume you're average/typical case
- Many find conclusion counterintuitive - weak premises leading to striking implications
Anthropic Reasoning in Cosmology
- Need similar assumptions for scientific inference in contemporary cosmology
- Multiverse theories: all possible human observations made in sufficiently large universe
- Different observers observe different values due to local fluctuations/statistical anomalies
- Larger proportion observe "true" value, few observe wrong value
- Self-sampling assumption allows concluding our evidence supports correct theories
- Enables inferences about cosmic background temperature, fine-structure constant, etc.
- Without anthropic reasoning, these scientific inferences don't work
Simulation vs Doomsday Argument Comparison
- Simulation argument's "bland principle of indifference" much weaker than Doomsday's "self-sampling assumption"
- Doomsday: reason as random sample from ALL humans ever, despite knowing you're #100 billion in 2020
- Simulation: if you can't tell which group you're in, assign uniform probability
- Observer classes in simulation: original history people vs simulated people
- Question: which group contains you?
Supporting the Bland Principle
- Sequence of cases where fraction simulated approaches 1
- Limiting case: if everyone simulated, deduce with certainty you're simulated
- As percentage increases (90%, 99%, 99.9%), probability should approach certainty
- Shouldn't be discrete jump from 50/50 to certainty
- Fraction of simulated worlds approaches (but doesn't reach) 1 over time
Physical Constraints on Simulation
- Limits to information processing any civilization can perform
- Finite matter available due to positive cosmological constant
- Accelerating universe creates finite sphere of reachable matter
- Lower limits on computational losses (erasure, matter decay, proton decay, gravitational waves)
- Eventually leads to heat death/cold death of universe
- Finite but large computational capacity
Stacked Simulations
- Simulations can spawn other simulations
- Each level has fewer resources than parent level
- Uncertainty about which level we're at (level 1, 42, 100?)
- All computations in nested simulations must be powered by "basement reality"
- Finite compute power in basement reality limits tower height
- Large overhead per level suggests most people would be lower in tower
- Tower height naturally limited by computational constraints
