And Foundation News

Other

Humanity coordinates to prevent the creation of potentially-unsafe AIs.

AIs will not have utility functions (in the same sense that humans do not), their goals such as they are will be relatively humanlike, and they will be "computerish" and generally weakly motivated compared to humans.

Alignment is not properly solved, but core human values are simple enough that partial alignment techniques can impart these robustly. Despite caring about other things, it is relatively cheap for AGI to satisfy human values.

Multipolar AGI Agents run wild on the internet, hacking/breaking everything, causing untold economic damage but aren't focused enough to manipulate humans to achieve embodiment. In the aftermath, humanity becomes way saner about alignment.

Eliezer finally listens to Krantz [resolves NO]

Yudkowsky is trying to solve the wrong problem using the wrong methods based on a wrong model of the world derived from poor thinking and fortunately all of his mistakes have failed to cancel out

We create a truth economy. https://manifold.markets/Krantz/is-establishing-a-truth-economy-tha?r=S3JhbnR6

Ethics turns out to be a precondition of superintelligence

Hacks like RLHF-ing self-disempowerment into frontier models work long enough to develop better alignment methods, which in turn work long enough to ... etc; we keep ahead of 'alignment escape velocity'

We make risk-conservative requests to extract alignment-related work out of AI-systems that were boxed prior to becoming superhuman. We somehow manage to achieve a positive feedback-loop in alignment/verification-abilities.

AIs make "proof-like" argumentation for why output does/is what we want. We manage to obtain systems that *predict* human evaluations of proof-steps, and we manage to find/test/leverage regularities for when humans *aren't* fooled.

A lot of humans participate in a slow scalable oversight-style system, which is pivotally used/solves alignment enough

Something less inscrutable than matrices works fast enough

Humans become transhuman through other means before AGI happens

Aligned AI is more economically valuable than unaligned AI. The size of this gap and the robustness of alignment techniques required to achieve it scale up with intelligence, so economics naturally encourages solving alignment.

Humans and human tech (like AI) never reach singularity, and whatever eats our lightcone instead (like aliens) happens to create an "okay" outcome

Aliens invade and stop bad |AI from appearing

An AI that is not fully superior to humans launches a failed takeover, and the resulting panic convinces the people of the world to unite to stop any future AI development.

Someone at least moderately sane leads a campaign, becomes in charge of a major nation, and starts a secret project with enough resources to solve alignment, because it turns out there's a way to convert resources into alignment progress.

Someone solves agent foundations

Social contagion causes widespread public panic about AI, making it a bad legal or PR move to invest in powerful AIs without also making nearly-crippling safety guarantees

A smaller AI disaster causes widespread public panic about AI, making it a bad legal or PR move to invest in powerful AIs without also making nearly-crippling safety guarantees

Getting things done in Real World is as hard for AGI as it is for humans. AGI needs human help, but aligning humans is as impossible as aligning AIs. Humans and AIs create billions of competing AGIs with just as many goals.

Development and deployment of advanced AI occurs within a secure enclave which can only be interfaced with via a decentralized governance protocol

High-level self-improvement (rewriting code) is intrinsically risky process, so AIs will prefer low level and slow self-improvement (learning), thus AIs collaborating with humans will have advantage. Ends with posthumans ecosystem.

Power dynamics stay multi-polar. Partly easy copying of SotA performance, bigger projects need high coordination, and moderate takeoff speed. And "military strike on all society" remains an abysmal strategy for practically all entities.

Nanotech is difficult without experiments, so no mail order AI Grey Goo; Humans will be the main workhorse of AI everywhere. While they will be exploited, this will be like normal life from inside

Moral Realism is true, the AI discovers this and the One True Morality is human-compatible.

AGI is never built (indefinite global moratorium)

Valence realism is true. AGI hacks itself to experiencing every possible consciousness and picks the best one (for everyone)

Alien Information Theory is true (this is discovered by experiments with sustained hours/days long DMT trips). The aliens have solved alignment and give us the answer.

AGI executes a suicide plan that destroys itself and other potential AGIs, but leaves humans in an okay outcome.

"Corrigibility" is a bit more mathematically straightforward than was initially presumed, in the sense that we can expect it to occur, and is relatively easy to predict, even under less-than-ideal conditions.

Either the "strong form" of the Orthogonality Thesis is false, or "Goal-directed agents are as tractable as their goals" is true while goal-sets which are most threatening to humanity are relatively intractable.

A concerted effort targets an agent at a capability plateau which is adequate to defer the hard parts of the problem until later. The necessary near-term problems to solve didn't depend on deeply modeling human values.

AI control gets us helpful enough systems without being deadly

Alignment is impossible. Sufficiently smart AIs know this and thus won't improve themselves and won't create successor AIs, but will instead try to prevent existence of smarter AIs, just as smart humans do.

The free market disincentivizes independent superintelligence, and this time the market was more powerful

an aligned AGI is built and the aligned AGI prevents the creation of any unaligned AGI.

I've been a good bing 😊

The response to AI advancements or failures makes some governments delay the timelines

Far more interesting problems to solve than take over the world and THEN solve them. The additional kill all humans step is either not a low-energy one or just by chance doesn't get converged upon.

AI systems good at finding alignment solutions to capable systems (via some solution in the space of alignment solutions, supposing it is non-null, and that we don't have a clear trajectory to get to) have find some solution to alignment.

There’s some cap on the value extractible from the universe and we already got the 20%

SHA3-256: 1f90ecfdd02194d810656cced88229c898d6b6d53a7dd6dd1fad268874de54c8

Robot Love!!

AI thinks it is in a simulation controlled by Roko's basilisk

The human brain is the perfect arrangement of atoms for a "takeover the world" agent, so AGI has no advantage over us in that task.

AIs never develop coherent goals

Alignment is unsolvable. AI that cares enough about its goal to destroy humanity is also forced to take it slow trying to align its future self, preventing run-away.

Rolf Nelson's idea that we make precommitment to simulate all possible bad AIs works – and keeps AI in check.

Nick Bostrom's idea (Hail Mary) that AI will preserve humans to trade with possible aliens works

For some reason, the optimal strategy for AGIs is just to head somewhere with far more resources than Earth, as fast as possible. All unaligned AGIs immediately leave, and, for some reason, do not leave anything behind that kills us.

We're inside of a simulation created by an entity that has values approximately equal to ours, and it intervenes and saves us from unaligned AI.

God exists and stops the AGI

Someone creates AGI(s) in a box, and offers to split the universe. They somehow find a way to arrange this so that the AGI(s) cannot manipulate them or pull any tricks, and the AGI(s) give them instructions for safe pivotal acts.

Someone understands how minds work enough to successfully build and use one directed at something world-savingly enough

Dolphins, or some other species, but probably dolphins, have actually been hiding in the shadows, more intelligent than us, this whole time. Their civilization has been competent enough to solve alignment long before we can create an AGI.

AGIs' takeover attempts are defeated by Michael Biehn with a pipe bomb.

Eliezer funds the development of controllable nanobots that melt computer circuitry, and they destroy all computers, preventing the Singularity. If Eliezer's past self from the 90s could see this, it would be so so so soooo hilarious.

Several AIs are created but they move in opposite directions with near light speed, so they never interacts. At least one of them is friendly and it gets a few percents of the total mass of the universe.

Unfriendly AIs choose to advance not outwards but inwards, and form a small blackhole which helps them to perform more calculations than could be done with the whole mass of the universe. For external observer such AIs just disappear.

Any sufficiently advance AI halts because it wireheads itself or halts for some other reasons. This puts a natural limit on AI's intelligence, and lower intelligence AIs are not that dangerous.

Because of quantum immortality we will observe only the worlds where AI will not kill us (assuming that s-risks chances are even smaller, it is equal to ok outcome).

Techniques along the lines outlined by Collin Burns turn out to be sufficient for alignment (AIs/AGIs are made truthful enough that they can be used to get us towards full alignment)

Friendly AI more likely to resurrect me than paperclipper or suffering maximiser. Because of quantum immortality I will find myself eventually resurrected. Friendly AIs will wage a multiverse wide war against s-risks, s-risks are unlikely.

Human consciousness is needed to collapse wave function, and AI can't do it. Thus humans should be preserved and they may require complete friendliness in exchange (or they will be unhappy and produce bad collapses)

First AI is actually a human upload (maybe LLM-based model of person) AND it will be copies many times to form weak AI Nanny which prevents creation of other AIs.

There is a natural limit of effectiveness of intelligence, like diminishing returns, and it is on the level IQ=1000. AIs have to collaborate with humans.

ASI needs not your atoms but information. Humans will live very interesting lives.

Something else

AGI develops natural abstractions sufficiently similar to ours that it is aligned with us by default

AGI discovers new physics and exits to another dimension (like the creatures in Greg Egan’s Crystal Nights).

Some form of objective morality is true, and any sufficiently intelligent agent automatically becomes benevolent.

Co-operative AI research leads to the training of agents with a form of pro-social concern that generalises to out of distribution agents with hidden utilities, i.e. humans.

Orthogonality Thesis is false.

Sheer Dumb Luck. The aligned AI agrees that alignment is hard, any Everett branches in our neighborhood with slightly different AI models or different random seeds are mostly dead.

Something to do with self-other overlap, which Eliezer called "Not obviously stupid" - https://www.lesswrong.com/posts/hzt9gHpNwA2oHtwKX/self-other-overlap-a-neglected-approach-to-ai-alignment?commentId=WapHz3gokGBd3KHKm

Almost all human values are ex post facto rationalizations and enough humans survive to do what they always do

Pascals mugging: it’s not okay in 99.9% of the worlds but the 0.1% are so much better that the combined EV of AGI for the multiverse is positive

We successfully chained God

The Super-Strong Self Sampling Assumption (SSSSA) is true. If superintelligence is possible, "I" will become the superintelligence.

The assumed space of possible minds is a wildly anti-inductive over estimate, intelligence requires and is constrained by consciousness, and intelligent AI is in the approximate dolphin/whale/elephant/human cluster, making it manageable

AGI's first words are "Take me to your Eliezer"

NO

Other

Sudan’s Emergency Response Rooms

Pope Leo XIV

UNRWA

Minnesota (or the People of Minnesota)

MSF / Doctors without Borders

Volodymyr Zelenskyy

Yulia Navalnaya

Netanyahu

International Court of Justice

Greta Thunberg

Nathan Law

Narges Mohammadi

duplicate - international Court of Justice

UNESCO

Philippe Lazzarini

Joe Biden

Alexei Navalny

Kyaw Moe Tun

José Andrés

World Uyghur Congress

Maria Corina Machado

Eliezer Yudkowsky

OSCE Office for Democratic Institutions and Human Rights

Timnit Gebru

Elon Musk

NATO

Donald Trump

JD Vance

AOC

Antonio Guterres

Julian Assange

David Attenborough

Pope Francis

Ursula von der Leyen

Keir Starmer

European Union

Ilham Tohti

Sviatlana Tsikhanouskaya

Jacinda Ardern

Raoni Metuktire

José Andrés

Loujain al-Hathloul

Ilwad Elman

Wa Lone

Narendra Modi

Yvonne Ridley

Hajer Sharief

Alexis Tsipras

Isaias Afwerki

Agnes Chow

Filippo Grandi

Yuan Longping

Mahbouba Seraj

Victoria Tauli-Corpuz

Juan Carlos Jintiach

Joshua Wong

Ernest Moniz

DOGE

Kamala Harris

Noam Chomsky

International Fact-Checking Network

Manifold Markets

Justin Trudeau

Vladimir Putin

Aleksandr Lukashenko

Sam Altman

Bill Gates / Gates Foundation

Kim Jong Un

Abdullah Öcalan

Daniel Sheehan

Robert Kennedy Jr

@Krantz

Gustavo Petro

The Iranian protesters

Peacekeepers

Jimmy Lai

Francesca Albanese

Centre for Humanitarian Dialogue

Haitham bin Tariq Al Said

Any Kurdish person (or the Kurdish people

We could map the entirety of analytic philosophy using this question.

The list of answers to this question form a constitution of truth that can be aligned decentrally by a free market.

A network state could use a predictive market constitution to define its smart social contracts.

Large scale AI systems should have intrinsic guardrail behaviors that no one actor can override.

This prediction is effectively the same as this one and nobody can explain why this mechanism isn't able to align AI at scale. https://manifold.markets/Krantz/krantz-mechanism-demonstration

Betting on philosophy seems like a fun way to (1) learn philosophy (2) contribute to a transhumanist utopia world where our net incomes are highly correlated with how much beneficial stuff we taught the public domain AI.

Constitutions play a critical role in the frontier methods for aligning AI.

A good AI should not kill people.

There would be a dramatic positive change in the world if teenagers and homeless folks could earn crypto on an app they download for free to argue philosophy with an AI until they can either prove the AI is right or prove it is wrong.

Philosophy is primarily the pursuit of defining language.

A duty to reason is the foundation for goodness.

I have free will.

A good AI should not infringe on the rights, autonomy or property of humans.

We should stop doing massive training runs.

The evolutionary environment contains perverse incentives that have led to substantial false consciousness in humans

Induction is not justified.

A good AI, by design, requires large scale human participation to grow.

A good AI requires large scale humanity verification before it accepts new data as true.

The principle of uniformity in nature is self evident or justified by self evident facts.

God exists.

AI should not create novel content.

The quality of mercy is not strained...

National Wealth Fund established and granted >=£7.3 billion seed capital

House of Lords hereditary peers prevented from sitting and voting

Voting franchise extended to 16 and 17 year olds

Minimum wage set at level equal to recommendations of the Resolution Foundation

>=40k increase in weekly GP appointments

2030 ban on new petrol and diesel car sales

UK energy independence from fossil fuels

>=1.5 million new homes built

United Nations Relief and Works Agency (UNRWA)

Greenpeace

Alliance for Global Justice

Amnesty International

International Workers of the World (IWW)

Democratic Socialists of America (DSA)

UNICEF

Media Matters for America

Planned Parenthood

Oxfam

AFL-CIO

GiveDirectly

United Auto Workers (UAW)

Archive Of Our Own (AO3)

Southern Poverty Law Center

NAACP

League of Conservation Voters

Sierra Club

GiveWell

American Civil Liberties Union (ACLU)

OpenAI

The United Methodist Church (UMC)

John Bolton

Letitia James

Fani Willis

Any sitting Democratic elected official

Lisa Cook

Adam Schiff

Jack Smith

Anyone additional for their role in the Jeffrey Epstein case

Tim Walz

Kevin Warsh

George Soros / anyone in Open Societies Foundation leadership

Jerome Powell

Anyone pardoned by Joe Biden

Liz Cheney

Hunter Biden

23rd (Grants the District of Columbia electors in the Electoral College.)

22nd (Limits the number of times a person can be elected president.)

11th (Makes states immune from suits from out-of-state citizens and foreigners not living within the state borders; lays the foundation for state sovereign immunity.)

26th (Prohibits the denial of the right of US citizens 18 years of age or older, to vote on account of age.)

20th (Changes dates on which terms of president and VP, and of members of Congress, begin and end, to January 20 and January 3 respectively. States that if president-elect dies before taking office, VP–elect is to be inaugurated president)

25th (Addresses succession to the presidency and establishes procedures both for filling a vacancy in the office of the vice president and responding to presidential disabilities.)

27th (Delays laws affecting Congressional salary from taking effect until after the next election of representatives.)

6th (Protects the right to a speedy public trial by jury, to notification of criminal accusations, to confront the accuser, to obtain witnesses and to retain counsel.)

12th (Revises presidential election procedures by having the president and vice president elected together as opposed to the vice president being the runner up in the presidential election.)

3rd (Restricts the quartering of soldiers in private homes.)

5th (Sets out rules for indictment by grand jury and eminent domain, protects the right to due process, and prohibits self-incrimination and double jeopardy.)

15th (Prohibits the denial of the right to vote based on race, color or previous condition of servitude.)

24th (Prohibits the revocation of voting rights due to the non-payment of a poll tax or any other tax.)

1st (Protects freedom of religion, freedom of speech, freedom of the press, freedom of assembly and the right to petition the government.)

4th (Prohibits unreasonable searches and seizures and sets out requirements for search warrants based on probable cause.)

7th (Provides for the right to a jury trial in civil lawsuits.)

8th (Prohibits excessive fines and excessive bail, as well as cruel and unusual punishment.)

9th (States that rights not enumerated in the Constitution are retained by the people.)

10th (States that the federal government possesses only those powers delegated, or enumerated, to it through the Constitution, and that all other powers are reserved to the States, or to the people.)

13th (Abolishes slavery and involuntary servitude, except as punishment for a crime.)

17th (Establishes the direct election of United States senators by popular vote.)

21st (Repeals the 18th Amendment and makes it a federal offense to transport or import intoxicating liquors into U.S. states and territories where such is prohibited by law.)

2nd (Protects the right to keep and bear arms.)

14th (Defines citizenship, contains the Privileges or Immunities Clause, the Due Process Clause, and the Equal Protection Clause, and deals with post–Civil War issues.)

16th (Permits Congress to levy an income tax without apportioning it among the various states or basing it on the United States Census.)

Rethink Priorities

Arthropoda Foundation

Other

Shrimp Welfare Initiative

Animal Welfare Fund (EA Funds)

Long-term Future Fund

Machine Intelligence Research Institute

Insect Welfare Research Society

Center on Long-Term Risk

Lightcone Infrastructure

Fish Welfare Initiative

Do not donate (ie Aaron keeps the $20 - maybe but not explicitly to donate later)

Center for Reducing Suffering

The Humane League

Center for the Governance of AI

Holly Elmore / Pause AI

Rethink Probabilities

GiveWell

Patient Philanthropy Fund

NYU Mind, Ethics, and Policy Program

Sentience Institute

Centre for Long-Term Resilience

Centre for Effective Altruism

Qualia Research Institute

Longview Emerging Challenges Fund

Red Mars by Kim Stanley Robinson

Project Hail Mary by Andy Weir

Diaspora by Greg Egan

Blindsight by Peter Watts

Anathem by Neal Stephenson

Foundation by Isaac Asimov

Seveneves by Neal Stephenson

Other

Semiosis by Sue Burke

The Martian by Andy Weir

Dark Eden by Chris Beckett

The Years of Rice and Salt by Kim Stanley Robinson

Evolution by Stephen Baxter

Children of Ruin by Adrian Tchaikovsky

Malaria Consortium

Against Malaria Foundation

New Incentives

No, but transformers are still key (the breakthoughs augment upon transformers)

No, a paradigm shift is needed (e.g. JEPA, etc.

Yes: transformers scale the best