Feedback News

2029-2030

2030-2031

2031-2032

2032-2033

2033-2034

2034-2035

2035-2036

2036-2037

2037-2038

2028-2029

2038-2039

2049-2050

2039-2040

2040-2041

2027-2028

2041-2042

2042-2043

2043-2044

2044-2045

2045-2046

2046-2047

2026-2027

2047-2048

2048-2049

2024-2025

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"

YES

Add tooltips that explain how everything works

Let a user filter to markets in which they have positions

If the user enters a date, it should be treated as a range from the beginning of that day to the end, not just the single millisecond at which that day begins.

Make the choice of which markets to store locally smarter, so it won't accidentally cross the 5MB localStorage limit.

Add proper inputs for dates and drop downs for categories

Multi-buy in multiple-choice markets

Add a quickbet button where the user inputs their expected probability, and it calculates and places the appropriate Kelly bet size based on their portfolio.

Add a tiny graph/sparkline on each row to show how the market has changed recently

Generate RSS feeds for user-specified queries (like https://hnrss.org/newest?points=100)

Add support for arbitraging similar markets/answers, either treating them as summing to 1 or specifying a target relative change percentage.

Make a userscript that overrides Manifold's default search with Isaac's search

Add an option to display each answer on a multiple-choice market as its own row in the search results.

A column to display the market price X amount of time ago, input by the user

Allow users to define more than one dashboard, and share them via URL with other people

Option to show dates in a standardized day/hour/minute format that all lines up

Add a filter to only show markets the user hasn't seen before.

Make the search boxes look nicer so they're not just a giant pile of input boxes. Lay them out in columns or hide ones the user isn't using or something.

Add an OR conjunction for search terms, so people can search for something like "question includes 'altman' OR 'OpenAI'"

Open markets in new tab by default

Add a "concise mode", which removes all extraneous text from the table, like the percent sign, M$ sign, etc. and aligns them all nicely

Add fuzzy searching that can catch typos and synonyms

Let people resize columns, or add intelligent column resizing

Let people see the limit orders of a selected user in each market

Dark Mode

Add support for other platforms. Polymarket, Kalshi, etc.

Add tiny arrows in each column header as a visual indicator that they can click there to sort the column

Let people see the position of a selected user in each market

Add reminders to check back on a market after some time

Add direct links to user profile and/or group pages, maybe accessible by shift-clicking on them.

Check to see if any of the code has changed upon page load, and if so wipe the cached market data and reload it.

Let people see the profit of a selected user in each market

Fix the graphical issues that cause fields to unfocus and dropdown menus to disappear each time new data is loaded

Add some way for clicking on a group or creator to wipe the rest of the search. Maybe shift-click? Or maybe it should be the default?

Add more complex diff detection so that the client can catch up on what it missed if it loses internet for a few minutes, or at least tell the user it's out of date

Add hotkeys to make trading and searching even faster

Add more sensible step values to the numeric input fields

Allow people to place long-term limit orders

Make the "last updated time" field include edits to the market description

Somehow convince Manifold to waive the API fee if bets are placed via Predictionary

Make it look not-terrible (this suggestion is for aesthetics only, not functionally-useful layout/display features)

Allow users to see whether a multiple-choice market is select-one or select-multiple

Display who can add answers on a multiple-choice market

Give it a better name than Predictionary. (Suggestions welcome.)

Add support for Polymarket front-running. (https://medium.com/@eightyhi/blockchain-based-amms-arent-fit-for-prediction-markets-bbe6ad7a33ca)

Make a row flash when its probability changes

Add a bulk bet button that can bet on all markets in the dashboard

Fix the bug that causes some resolved markets to not be marked as closed

Add a way to view/query detailed historical market data

Add a one-click feedback form on the page itself

Fix links to Manifold markets (currently all links go to "/{username}/undefined")

Add a filter/sort by recently subsidized

Add an option to change the height of each row such that it's proportional to the width of the interval between the previous market's close time and the next market's close time, so distance down the page represents time in the future.

Fix the bug that causes numeric markets to be displayed as though they were regular percentage markets.

NO