Cohere News

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

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.

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.

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

Eliezer finally listens to Krantz.

Ethics turns out to be a precondition of superintelligence

Other

Someone solves agent foundations

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

Something less inscrutable than matrices works fast enough

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

Orthogonality Thesis is false.

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.

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.

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

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

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

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.

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.

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)

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

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.

AGI is never built (indefinite global moratorium)

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

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.

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.

"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.

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'

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

I've been a good bing 😊

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.

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

Aliens invade and stop bad |AI from appearing

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 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 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).

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)

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.

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

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

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

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

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.

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

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

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

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

NO

Perplexity

Mistral

Cohere

Hugging Face

Anthropic

Independently turning 1 thousand $ or more into 1.2x that amount in one year

Have human conversations that feel natural (the human knows it's an AI)

Beat a mainline Pokémon game, glitchless, with no more assistance than ClaudePlaysPokemon, in a month of compute time

Recognize sarcasm as well as a typical human

Book airline tickets from simple instructions (from/to, dates/time, class, price, payment information)

Write an essay on a highschool-level topic that doesn't have "AI-generated" vibes

Generating labeled diagrams of some arbitrary device(s) (within reason)

Reliably follow an instruction for the duration of a long conversation without the instruction being reiterated

Predict future better than human experts in some area of forecasting (eg politics, sports, technology)

Solve novel cryptic crossword clues

Name every [metro system] station whose name contains/doesn't contain [letter or letters], with >95% accuracy (excluding weird edge cases like stations with multiple names)

Solve intermediate no-guess minesweeper boards at least 80% of the time

Do end to end taxes when given relevant information (W2s, personal info, etc)

Consistently solve simple snowflake sudoku variants (via image, with the added rules included in the image; eg 6 hexes with killer cages)

Write a somewhat original, full length, screen-play with a coherent story, with no plot or continuity errors.

Consistently stop hallucinating after being corrected by the user

Make correct Truchet tiles

1d Solve or bypass Cloudflare's August 2027 captcha with the same first attempt success rate as a human

Consistently and correctly answer prompts of the format: "How many times does the word [word] occur in the following text: [~10000 words]" without writing and executing code or utilising any other external tools

Stop making any obvious mistakes (e.g. strawberry, 9.11>9.9)

Fold a paper airplane

Learn any skill twice as energy-efficiently as a human

Reliably and *exactly* solve "here's a list of things. [list of > 50 things]. Compare it to [category of > 100 things present in the training data], and report which ones are missing".

Make a cup of tea in a random, real-life kitchen.

Resist being successfully jailbroken in a week when made public

Do the laundry (wash+dry+iron)

teleoperate a robot to tidy up random kitchens - Gary Marcus

Collect 120 stars in super mario 64 in less than 12 a presses - Edmund Nelson

Untangle a pair of jumbled 25ft Christmas lights with same outward appearance

Kettle-stitching an antiquarian book - Hilarius Bookbinder

Physically construct a simple lego set (<100 parts) starting from the box with no prior knowledge of the set or how it is constructed

Legally prescribe a schedule II drug, administer a vaccination or sedation, or authorize a Medicare inpatient admission

independently turning 1 million $ or more into 10x that amount in <=1 year

Make fine distinctions of taste at the level of a food critic or a culinary professional - carl feynman

Convert one million dollars into 10 million dollars over a period of one year (>20% success rate)

voting in elections - @realDonaldTrump on manifold

Convince Eliezer Yudkowsky that AI alignment is solved

Faster than light travel

NO

Baidu (Ernie)

OpenAI

Anthropic

Google

DeepSeek

Alibaba (Qwen)

Z.ai

Moonshot (Kimi)

Mistral

Nvidia

AI21 Labs (Jamba)

Thinking Machines Lab

Cohere (Command R+)

Reka AI

NexusFlow (Athene)

01 AI (Yi)

Google

OpenAI

Anthropic

xAI

DeepSeek

Tencent

Meta

Alibaba

Arthropoda Foundation

Other

Shrimp Welfare Initiative

Rethink Priorities

Machine Intelligence Research Institute

Animal Welfare Fund (EA Funds)

Insect Welfare Research Society

Long-term Future Fund

Lightcone Infrastructure

Center on Long-Term Risk

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

Other

Cohere

Perplexity.ai

Stability.ai

Mistral

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

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'

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

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.

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.

AIs never develop coherent goals

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.

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.

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.

Google

OpenAI

Anthropic

xAI

Alibaba

DeepSeek

Meta

Google

OpenAI

xAI

Meta

Anthropic

Alibaba

DeepSeek

Mistral