Models
All Entities
Models
// Zach. 2026 Sep 3. // The result is greater than 1 because the worlds where Ant performs well are unfortunately the worlds where we already have more money -- because a large fraction of the endowment is in Anthropic. // The result is driven by the fact that antPerformanceVsBaseline is higher-variance than nonAntPerformanceVsBaseline. If they were the same, we'd want 50% of the endowment in Ant and 50% in Non-Ant. Non-Ant is lower-variance; one angle on this is that it's like a basket of several investments with similar profiles to Anthropic, which are not fully correlated. The specific distributions for antPerformanceVsBaseline and nonAntPerformanceVsBaseline are important, but I haven't thought much about them. Relatedly, this model is about what happens from end-of-lockup-time to doing-philanthropy time, but I don't specify when doing-philanthropy time is. I suspect antPerformanceVsBaseline is too high-variance, so I fudge from the model's output of 1.45 down to 1.4. // I assume log utility in money. For this reason, ratios are all that matter -- 1.01xing the endowment is equally valuable no matter how big the endowment is. // Recall that this model is just about overexposure, assuming your Ant and non-Ant investments have similar expected returns; in reality there may be investments clearly better than Ant, like Ant-with-1.5x-leverage. // I kinda pretend lockup ends 2027-04-01, but really it will be continuous. I assume that you either sell immediately-post-lockup or hold long-term. I optimistically assume that outside of Ant cofounders/staff nobody has Anthropic exposure post-lockup, but this is false. longtermistEndowmentBeforeLockupEndsEXCLUDINGAntCofounderStaff = (40 to 80) antMarketCapBeforeLockupEnds = (1000 to 6000) antCofounderStaffOwnershipFraction = uniform(25%, 40%)
/* ============================================================ Airborne pathogen surveillance: chamber characterisation vs per-sample capsid-integrity PCR. All distributions are reported with 90% intervals. Health outcomes are deliberately not monetised. ============================================================ */
// ExpectedTPPower_{region}
etpp_US = normal({p5: .36, p95: .6})
etpp_EU = normal({p5: .23, p95: .47})
etpp_Brazil = normal({p5: .32, p95: .56})
etpp_Mexico = normal({p5: .05, p95: .29})
etpp_Korea = normal({p5: .16, p95: .4})
etpp_India = normal({p5: .57, p95: .81})
etpp_Japan = normal({p5: .67, p95: .91})
// AdminPowerModifier_{region}/* What do our beliefs about Sys/Surg imply about the other? */ // Inputs surg_leverage = 5 to 20 surg_effectiveness_multiplier = 1.5 to 5 innovation_potential_us = 1.8B * (0.25 to 4) // All-time emissions averted by US innovation system each year, in expectation us_energy_innovation_budget = 8B to 12.5B state_capture = beta(3, 6) // Share of innpot due to govt spending
// I am deciding between eating mussels and eating tofu, with omega 3 supplements. I end up with a conclusion that a mix is reasonable, because 500g of mussels is more than I can feasibly eat a week, so supplementing omega 3 and b12 on top should put me in a healthy range. Results below generated by iterating with claude opus 4.8 until I felt i had removed all its mistakes. I believe the results are reasonable. I feel that mussels are plausibly sentient but springtails are more plausibly sentient, and springtails are killed much more frequently when farming soybean than mussels are when farming mussels. I fix them equal for this aanalysis to be sure my conclusion is not unfair to the mussels. The bottome line I draw is it is probably slightly better morally speaking to eat organic mussels as a vegan in the EU, but it's really not clear in the end. // // === Mussels vs Tofu — sentience-weighted expected deaths for 500g of mussels and equivalent tofu === // // Compares 500 g cooked mussels vs a CALORIE-EQUIVALENT serving of cooked tofu. // Output unit: sentience-weighted welfare units per serving (human welfare-range = 1.0). // // All figures cite the local fact sheet: // /home/dmrivers/mussels_vs_tofu/Mussels vs. Soy Plus Algae-Oil_ Cost-Effectiveness Input Fact Sh.md (pasted at end of file) // with additional numbers from the prior conversation research (Potapov 2023,
// EA Indonesia — EG pilot, 2-yr forecast
// fractions -> beta (bounded 0–1); large leg -> per-prospect sum (lumpy)
// hi to low is p5 - p95 (90% CI)
// ---------- Leg 1: broad / pledge (large N -> smooth product is fine) ----------
reachedNet = 600 to 2400
pledgeConv = beta({ mean: 0.11, stdev: 0.05 }) // ~0.05–0.20
newPledgers = reachedNet * pledgeConv
grossPledge = 40 to 300 // USD/yr, unbounded above -> lognormal// Cost-effectiveness of saving to buy galaxies — Squiggle model // Cost-effectiveness of saving $1M to buy distant galaxies vs AI safety philanthropy // Wealth as fraction of global wealth post-ASI globalWealth = 5e14 // $500T // A. Probability you can buy galaxies (no AI takeover × galaxies for sale) aPCanBuy = 0.6 * 0.2 // B. Your returns as a multiple of global wealth growth
/* Describe your code here */ max_years_until_impact = 60 // Idea: It feels like AI taking off makes the concept of time less useful as a variable in the scenarios where we get meaningful speedup in research before doom. In that case perhaps it would make sense to model it such as if AI is coming later in "progress years" so to the extend that it is possible to do faster bio research for those periods, it might be sensible to think of it in the same way as it does to think of it as having more time? future_mass_before_agi(x) = 1 - cdf(Sym.lognormal(2.2, 2), x) years_until_super_babies = Sym.lognormal(3, 0.7) super_baby_growup = Sym.normal(18, 2) grown_up_super_babies(x) = cdf(years_until_super_babies + super_baby_growup, x)
p_drafted = 0.05 to 0.30
p_die_given_drafted = 0.05 to 0.30
p_death = p_drafted * p_die_given_drafted
russian_deaths = 1.5e5 to 5e5
russia_population = 1.25e8 to 1.55e8
share_male = 0.5
share_fighting_age = beta(40, 60)
p_death_sempere = russian_deaths / russia_population / share_male / share_fighting_age
// Total population
tcd_pop = normal({p5:16343231,p95:22494774})
caf_pop = normal({p5:5346925,p95:7814208})
ner_pop = normal({p5:23264141,p95:30200196})
mli_pop = normal({p5:22371158,p95:26038039})
sle_pop = normal({p5:7396538,p95:8648290})
ssd_pop = normal({p5:8293052,p95:11407307})
bdi_pop = normal({p5:12024009,p95:16677804})
som_pop = normal({p5:15525542,p95:27842362})
mdg_pop = normal({p5:27873947,p95:33799461})/* LBE_1 Notebook Dashboard UI */ // --- 1. IP & MARKET BASE --- tax_rate = 0.30 // Corporate Tax Rate ip_royalty_rate = 0.30 // 30% off Top-Line Gross Revenue ticket_price = 60 to 90 // Premium pricing justified by Anime IP marginal_cost_pp = 2 to 5
import "hub:ozziegooen/sTest" as sTest import "hub:AI-safety/p-solve-alignment" as generalAlignmentCosts /* Cost-Effectiveness Analysis for ASI Alignment to Animal Welfare */ /* Simplified version of https://squigglehub.org/models/AI-for-animals/alignment-to-animals-EV Unlike the more complicated version, this model does not consider spending-so-far on alignment, it just considers total cost; and the probability of solving alignment is a user input instead of being a derived parameter.
/* Generated by Squiggle AI. Workflow ID: ce07efa6-e1cb-4c49-a387-a659e4ee7b6a */ // Golden Rice Counterfactual Model // Estimates lives saved if Golden Rice had been approved globally in 2005 import "hub:ozziegooen/sTest" as sTest // == Model Parameters ==
// Golden Rice Counterfactual Model // Estimates lives saved if Golden Rice had been approved globally in 2005 // Forked from https://squigglehub.org/models/Abi/goldenrice import "hub:ozziegooen/sTest" as sTest // == Model Parameters == // == Model Parameters ==
/* A model for the probability of solving AI alignment The basic setup: 1. It will cost some amount to solve alignment; the cost is distributed over multiple orders of magnitude. 2. Some amount has already been spent, and some amount will be spent in the future. 3. If the amount spent exceeds the cost, then alignment is solved. */