Models by mdickens

/*
Expected value of influencing USAID dollars via campaigning for the Presidential candidate who's more pro-USAID, compared to donating to GiveWell top charities.
*/

// actual budget 2025 was $34B pre-cuts
usaid_dollars = normal(34B, 10B)

num_voters = 150M

// relative to GiveWell

// Model for the expected value of donations to a US House of Representatives campaign. Based on Eric Neyman's model of the expected value of donating to Alex Bores' 2026 campaign: https://ericneyman.wordpress.com/2025/10/20/consider-donating-to-alex-bores-author-of-the-raise-act/

@name("Key Inputs")
inputs = {
  @name("Dollars per Vote ($)")
  @doc(
    "Cost per vote in a House race. Eric Neyman assumed a typical campaign costs $100 per vote based mainly on 'numbers thrown around casually by experts', then multiplied by 3 because New York has higher costs.
  
  I spent 15 minutes looking for literature and I found:

/*
Reproduction of Eric Neyman's model for the expected value of electing Alex Bores to US Congress.

source: https://www.lesswrong.com/posts/TbsdA7wG9TvMQYMZj/consider-donating-to-alex-bores-author-of-the-raise-act-1#Comparison_to_non_AI_safety_opportunities
*/

@name("ΔP(Doom) If USA Behaves Wisely")
@format(".0%")
@doc("Absolute change in P(doom) given that the USA behaves wisely rather than unwisely.")
dp_doom_given_usa_wise = -5%

// Model estimating the change in probability of AI doom from electing a pro-AI-safety representative to the U.S. House.

// TODO: AI wrote all the docstrings (based on my suggestions), I have not reviewed for accuracy, all I did was remove the four million "this value is highly uncertain" tics that the AI loves.

@name("Include Weak-Bill Path?")
@doc("If 0, zero out the weak-bill path to impact. If 1, include it.")
include_weak_bill_path = 1

@name("Include Strong-Bill Path?")
@doc("If 0, zero out the strong-bill path to impact. If 1, include it. Removing the strong-bill path still allows strong bills to be passed via weak bills or via warning shots, but not directly via voting/advocacy.")

// Alternative for the expected value of donations to a US House of Representatives campaign. 
// Original model here: https://squigglehub.org/models/mdickens/congressional-campaign-donations

@name("Key Inputs")
inputs = {
  @name("Dollars per Vote ($)")
  @doc(
    "Cost per vote in a House race. Eric Neyman assumed a typical campaign costs $100 per vote based mainly on 'numbers thrown around casually by experts', then multiplied by 3 because New York has higher costs.
  
  I spent 15 minutes looking for literature and I found:

// Cost-effectiveness of messaging campaigns directed at elected representatives. 
// I created the core model and then ran it through Squiggle AI to create the boilerplate. Then I wrote all the docstrings.

// ==== Inputs ====

inputs = {
  @name("Proportion of Votes Within 1% (%)")
  @doc(
    "Proportion of vote outcomes that fall within one percentage point of the pass threshold. Based on historical voting records. The number has varied over time—it was 2.2% from 2000 to 2010, and 5.1% from 2015 to 2025.
  

/*
This model compares the expected impact on future AIs' values from two interventions:

1. Conventional animal activism — persuading people to care more about farm animals / eat less meat
2. Creating a friendliness-to-animals LLM benchmark and advocating for AI companies to use it in post-training.

My numbers are intentionally overly optimistic toward conventional activism and overly pessimistic toward an LLM benchmark because I was pretty sure the latter was going to look better, and indeed it still looks better by >2 orders of magnitude.

I came up with the input parameters by doing approximately three seconds of research, I'm sure they could be much better.

/*
DEPRECATED. See https://squigglehub.org/models/AI-safety/usa-pause

Is it a good idea for US actors to accelerate an AI arms race with China?
*/

// Probability that an aligned AI has good values regardless of what its creator wants, i.e., an aligned Chinese-built AI will refuse to let China take over the world.
p_aligned_ai_enforces_good_values = 0.1

// Probability that China would take over the world if it could.

/*
 * EV of AI safety technical research vs. policy advocacy, in 
 * terms of change in probability of extinction.
*/

@doc(
  "Open Phil granted $69M on AI safety in 2023; a LW post[1] says $43M of that was on technical AI safety; but most of that was not directly funding alignment research (generously, only $28M was alignment research, the rest being forecasting/field-building/etc.). There are some other philanthropic funders, plus for-profit companies do some internal safety research. Internal research is the bulk of the money but I expect most of it is bad ex ante (e.g. fake safetywashing research). It would be feasible to estimate spending on AI 'safety' research, but harder to estimate how much of that spending is real, so I kinda just made up a number range.

[1] https://www.lesswrong.com/posts/adzfKEW98TswZEA6T/brief-analysis-of-op-technical-ai-safety-funding"
)

/*
A cost-effectiveness analysis for starting a communist revolution in the United States.

Critics of EA like to say that it's impossible to create a CEA for a communist revolution. I am making this to prove that it's possible.

This wasn't even hard, it took me like half an hour. I'm sure I could do a much better job if I invested more time.
*/

@doc("Conservative estimate. Surely communism is at least 20% better than capitalism but it's likely much better than that.

/*
 * Cost-effectiveness analysis for Uganda Community Farm's grain cleaning facility
 */

/*
 * Model Inputs
 */

// As of 2024-05-24
gbp_to_usd :: usd/gbp  = 1.27

/*
 * Cost-effectiveness analysis for Uganda Community Farm's program to 
 * buy sorghum seeds and supplies necessary to grow sorghum.
 * 
 * Some parts of this model are copy/pasted from
 * https://squigglehub.org/models/mdickens/Uganda-Community-Farm-facility 
 */

// As of 2024-05-24
ugx_to_usd = 1 / 3828