Seed 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.

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

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

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.

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

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

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

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.

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.

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

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

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.

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.

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

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)

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

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.

"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

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 😊

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

Aliens invade and stop bad |AI from appearing

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).

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

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.

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.

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).

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.

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

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

Stretch-and-fold after mixing, 3x every 30 min

3 iterations of stretch-and-fold, at any time during the 14h waiting period. Minimum wait time between iterations 1 hour

Place small ice cubes between parchment and pot instead of water

Add 1tsp of diastatic malt powder per 3cps of flour

Use tap water instead of fancy RO water

put butter into it

Toast the bread

Put a non-lethal dose of any rat poison.

Donate the bread to a food pantry, homeless person, or someone else in need

Add lots of butter (0.2 ml per gram)

Half yeast

Bake it with your best friend.

Use whole wheat to improve the nutrition of the bread

Bake for 5 more minutes

Bake for an amount of minutes equal to the percent this market answer is at when it comes time to begin baking. (Maintain the ±3 minute tolerances and the 2:1 ratio of time before:after the water spritz.)

Sprinkle 3 grams of flaky sea salt on top of each loaf before the second bake

Increase water by 50 g

it needs more salt

Replace all water spritz steps with a basting of extra virgin olive oil.

Diastatic malt (~1% baker's percentage) = happier yeast

Serve the bread hot

Use Iodized salt

Do a second rise

Create indentation, fill with melted cheese and butter

don't eat anything for at least 2400 hours before eating the bread

Cut into the dough right before baking looks destructive to improve the appearance

Sell your bread at an auction and donate the money to those in immigration detention prisons.

Do all the changes in the top 5 open options by probability, excluding this option

Sprinkle sesame seeds evenly over the top

Do it with a good spirit in your heart, or ask someone with a good spirit to do it for you. But don’t watch while they do it.

Try baking a little more "bien cuit". If the image is indicative, your loaves may be quite "blonde".

Substitute 75 g of your flour with spelt flour

Short advice: Start baking at 260°C for strong rise, then reduce to 230°C and uncover halfway to achieve even browning and a crisp crust. 🍞

put ketchup and cheese on it

Add garlic

Use sourdough instead of yeast

Add slurs to it

Give ChatGPT your current recipe as well your take on what optimal bread tastes like, then take that advice for your next bake

Don't automatically "Heat water to 30±1 °C". Instead, aim for a desired dough temperature (DDT) of 25-26°C. 30°C water is too hot for summer, and potentially too cool for winter.

Ask ChatGPT (GPT-5, with thinking enabled) for suggestions on improving the bread, with this market description, then do all of them.

Get recipe from gemini with experiments done, taste preference, then confirm it with chatgpt for final suggestion.

Make all of the ingredients from scratch.

Make the bread taste good

If your city uses artesian water, replace plastic bottled water with tap water. It will add natural, healthy alkalinity to your bread.

throw your recipe into a black box Bayesian hyperparameter optimization framework and iterate as necessary (see the Google Vizier Cookie Recipe paper for example

Don't eat anything for at least 48 hours before eating the bread

Invest in a "Bakers Steel" for better heat retention and oven spring. It would mean graduating from a dutch oven though.

Watch the video

Ditch current process, do everything the same as the video

Bread flour, 3x yeast, cut rise to ~3h

Use lemonade instead of water.

Eat the bread while punching @realDonaldTrump in the face

Eat the bread while watching your mana balance steadily tick to (M)0

Throw the bread at a telescope

Put a baking rack in the Dutch oven before putting the loaf in, raising the loaf off the floor and lofting it over a layer of air.

Cut bread into loaves before serving

Cut bread into ≤0.4inch slices, toast before serving

Invite your taste-testers to make the bread with you

Tarriff the bread-making process with a 10% reduction of all ingredients where actual physical money is required to purchase them, until it “shrinkflates,” but try to keep the same volume. Do not reduce any free ingredients.

Standardize a separate list of process features to keep track of independently of all other tests and use the cross entropy method to tune them to maximize your bread preference

Taste the bread

put oats on top of the dough just before baking, press them in so they don't fall off

Buy chemical bread improvers

Just freeze the ready bread, then slowly bake it until it’s hot inside. It will give you a crustier crumb, contain less moisture, and taste better.

Have someone sell the bread to you at an expensive price

Add cheese inside it and also sausage

Add 6.25±1.25 g lemon juice when mixing in water to yeast and salt jug

Add melatonin to the bread and eat before you sleep (do safely)

Don't eat anything for at least 24 hours before eating the bread

While the bread is baking, tip every user who voted "Yes" on this option 25 Mana

Use a food-grade, human-approved vitamin D supplement in the correct dosage for testers with vitamin D deficiency

Add an amount of non-smoked paprika equal to the current salt content

Strawberry jelly filling

Use a convection oven/setting

Make banana bread

Buy bread from a michelin star restaurant.

Replace part of the flour in the dough with freshly crushed hemp seeds. It will make the bread a little bit sweeter, especially appealing for Canadians.

Eat it slowly and focus on the taste

Give Gemini your current recipe as well your take on what optimal bread tastes like, then take that advice for your next bake

Keep the bread in a high quality expensive beeswax coated textile bag

Make the bread great again

Bake the bread into a fun shape, like a fish, or an octagon

Decompose it into infinite spheres, then a few parts per sphere, rotate the spheres by arccos(1/3), unite them and you will find 2 chilis (Banach-Tarski)

Bake the Manifold Crane into the Bread

Make the bread edible then throw it in

Drink vodka while eating the bread

Do FOLDED

Quit Manifold into the bread.

Kill the bread into Manifold.

Assume the chili is not in the interval [0,1], square it for more chili, if it is in (0,1), take the square root, else (equals 0 or 1) add 1 to it.

Assume the chili is in the interval (0,1), square it for less chili, if it is in (1,infinity) take the square root, if it is in (-infinity,0) take the negative of the square of the of the chile, else (equals 0 or 1) subtract 1 from it.

Bake for 15 more minutes

Add 2 tbsp vanilla cake mix

Replace AP with bread flour and cruelly mix the dough in the mixer until elastic and not sticky

Add a tablespoon of sugar

Bake one more minute

Replace 10% of flour with milled wheat bran

Butter resolved yes twice. Double butter again.

replace 10% of flour with farina bona

Follow a recipe online

Grind/powderize all salt used into a fine powder (with pestle & mortar or similar device)

Add 1/2 scoop whey protein powder

Instead of RO water, use lightly rusty water to improve the nutritional value of the bread with soluble iron.

Add at least 20% chic pea flour or some other type of high protein flour

Only buy ingredients from 7/11.

Knead it only 10 times, keep it round like this number

Ask yourself if bread is healthier than fruits? No need to improve my bread

Resolve at least one thing here yes or no while baking bread

A system view is more appropriate. This is a dynamic, multi-variate, biological and chemical system. For e.g. conditioning salt % AND yeast % AND water temperature based on ingredient and ambient temps.

Use soda instead of water (clear, orange, yellow, etc. soda is ok. Don’t use a purple/brown soda as that would make it not look good)

Wear a suit while baking the cake.

Encourage people to participate in the market in good faith while making the bread

Replace salt with sugar

Replace 600+/-5g water with 600+/-50g water (eyeball rather than carefully measure)

Cook the bread with a rod/puck of aluminum foil (or similar) in the core in an attempt to conduct heat through the center of the bread, cooking it evenly like a doughnut.

Get your friends to help you make a batch ten times the size, but add a Pepper X (2.7M Scoville heat units) to the mixture

Slowly add sawdust to the bread until your friends start to notice.

Mail the bread to 1600 Pennsylvania Ave. Washington D.C.

Pray to your preferred agricultural/food deity before baking and before eating

Soak the finished product in vinegar for 24 hours

Ship a piece of the bread to a random person.

Frost it and put sprinkles on it to make it a birthday cake.

Add caffeine to the bread

Add as many Jack Daniel's whiskey barrel smoking chips as feasible to the Dutch oven before baking, physically separating them from the bread as necessary while baking.

Improve the bread

Quadruple salt

Bake your social security number into it.

Bake one fewer minute

Bake the cake while wearing a onesie.

Give the bread a name in a ritual ceremony and baptise it, with pre-blessed holy water if a priest isn't available

Test/filter the water for heavy metals

Eat the bread in front of the White House.

Implement all options that resolved NO

Make the bread inedible then throw it out.

Throw the bread at @realDonaldTrump

Force Feed it to @realDonaldTrump

Cut the bread into the number of traders in the market slices.

make the bread bounce

Implementing every element listed below.

Just make donuts instead

Bake it in an easy bake kids oven

Use a plastic baking sheet.

Eat the bread while betting yes on Cuomo on Manifold

Double oven temperature

Bake the bread very thin and add food coloring to make it have the US flag. Don’t allow it to touch the ground, illuminate at night, fold 13 times properly, and pledge allegiance before eating.

Don’t use usual water (room temperature) for the dough - that water’s only for toilets. Use electrolyte drinks instead with ice cubes; they make the dough taste better and add extra nutrition.

Light it on fire with birthday candles.

Bake it with a microwave

Halve oven temperature

Eat the bread while betting yes on Mamdani on Manifold

Step on it

Bring it to Yemen and put a bomb in it

Bake America Great Again

Sacrifice a lamb

Add MAGA and a splash of Trump juice

Bake in a cat and a dog

Explode it:

5 parts cyanide/ 1 part water/ 1 part sand

say 6 7 67 times before making the bread

Take a fat dump in the dough

Sit in dough 24 hrs

Replace flour with flowers

Let dough sit 24 hrs

Mix half sodium/potassium chloride

Add 2g? of baking soda

Bake in rectangular tin

Add 50g vital wheat gluten

double yeast

halve salt

Double salt

Add 2tsp olive oil

Refrigerate dough instead of room temp wait

Start at 500F, drop to 450F and uncover half way through

Do not mix salt and yeast in water together

Autolyse 20 minutes

Let dough rise on counter only until double volume or 2h max, any time longer in fridge

Think positive thoughts before tasting

Put fork in microwave

Don't eat anything for at least 12 hours before eating the bread

Add 2tbsp vanilla extract

Play classical music while baking

Make naan bread, an easy-to-make bread

Add a pinch of sugar

Bake on upside-down sheet pan, covered with Dutch oven

Eat the bread with friends

Bake vegimite into it.

Bake vodka into it

Bake it in the country you were born in.

Let dough sit 48 hrs

Resolve this option YES while eating the bread

Ditch all the steps. Just buy the bread from the supermarket

Eat the bread over the course of a week.

Bake for 5 fewer minutes

Use 50% whole grain flour

Bake the bread with love

Use King Arthur Bread Flour instead of All-Purpose

Add sawdust to increase the volume of the bread (but only like 10% sawdust by volume max. maybe 20% if it's good sawdust)

More steam! Either spritz with more water (preferably hot) or actually pour some boiling water in just before closing the lid.

Resolve this option NO while eating the bread

Incorporate a whole grain flour (buckwheat for example)

Add 50g sugar

Add 50g honey

Swap 200ml water for milk

Cover bread with damp paper towel instead of initial water spritz. Rehydrate paper towel during 2nd spritz. Remove paper towel before placing on cooling rack.

Make a poolish 12 h ahead: 100 g flour + 100 g water + 0.8 g yeast (0.1 %). After it ferments, use this poolish in place of 100 g flour and 100 g water in the final dough.

Add an amount of MSG equivalent to half the current salt content

Use flour made from an unconventional grain e.g. barley, millet, oats, rye, sorghum, maize etc.

Eat the bread while sending all your mana to @realDonaldTrump

Add 1/2 cup yogurt to the bread and name the bread “gurt” while addressing it with “yo, gurt”.

Add banana

Add poppy seeds

Replace some of the water with an egg (eg. remove 25g of water for a 50g egg)

Only use tap water from specifically New York City

Want to improve the value of your bread? Simply bake a piece of gold into it

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

No model picker

Personalized to you, the user

Trained on synthetic data

Very large context window, or another way to give the model long-term memory

Faster speed(smaller model, architecture improvements, mixture of experts keeping most parameters unused during inference, etc.)

Designed to inference cheaper than GPT-4(price must be lower)

Exact-query of training data or knowledge databases(similar to Google RETRO)

Online learning of specialized agents(finetuning, PEFTs, etc.)

Operates autonomously with instructions from you

capable of online learning

New Neural Network Architecture

Seed option to create the question

(Near) real-time update of training data (data delay is within 6 months and is kept updated)

Audio Modality

Acts autonomously

Has a catchier name

YES

NO

A bio revolution of resurgence! The bean is back in everyone’s hands

A prized middle class item

Chemical vanillin for all (total extinction)

For the rich only

Oppenheimer

Banshees of Insherin

Tár

Drive My Car

The Guilty (2018)

Blue Valentine (2010)

Conclave

The Seed of the Sacred Fig

Mommy (2014)

The 400 Blows (1959)

The Lunchbox (2013)

Saraband (2003)

Shoplifters (2018)

Winter Sleep (2014)

Sharknado

NO

The Best Christmas Album In The World...Ever! 2023 (2023) Mp3 320kbps

Linda Ronstadt - Original Album Series (5CD BoxSet) (2009)

Taylor Swift - evermore (2020) FLAC Album [PMEDIA]

Taylor Swift - evermore (2020) FLAC Album [PMEDIA]

VA - The Best 80s Car Songs Album In The World Ever (3CD) (2021)

The Beatles - The Beatles (White Album) (2018 Super Deluxe FLAC) 88

The Pretenders - Original Album Series (5CD BoxSet) (2009)

Merseybeat - Original Album Series (5CD BoxSet) (2014 Parlophone Mono)

The Beatles - The Beatles' Second Album (2014 Deluxe Edition FLAC) 88

The Best Disco Party Album in the World... Ever! (2023) Mp3 320kbps [PMEDIA]

Barbie - The Album (2023) Mp3 320kbps [PMEDIA]

Ride - Original Album Series (5CD Box Set) (2016)

Kanye West - Donda 2 (2022) FLAC Album [PMEDIA]

Echo & The Bunnymen - Original Album Series (2009) (5CD Box Set)

Molly Hatchet - Original Album Classics (5CD BoxSet) (2010)

The Best R&B Classics Album in the World Ever (2024) Mp3 320kbps [PMEDIA]

Smokey Circles - Smokey Circles Album (1970, 2018 reissue)

NO

NO