@tszzl — page 83/103

@bit_christ now I’m actually under hyping bc it’s priced in

typical moravec’s paradox https://x.com/ylecun/status/1646882539833794560

every last thing in the future will be animated with intelligence. children’s teddy bears will speak to them and make them feel safe at night. Toy Story animism world. people will reminisce on the horrible times when objects didn’t have spirits

Shinto + Hindus are prepared for this

people whose lives would be considered straight up disgusting 2 or 3 generations ago are so quick to be like this is a little bit creepy so it’ll never happen

@a_musingcat im basically a middle manager in the information hierarchy. I gather and filter the hottest takes from 6000 lowbies and then give them to the ai elites (who have 6000 guys like me)

while everyone is breathlessly talking about 6 month moratoria and AutoGPTs that don’t work 2 million people die in car accidents and countless more because of insufficient intelligence to solve crucial medical problems and the stars remain woefully uncolonized

i am once again asking you to accelerate the pace of technological progress

@Willyintheworld @sterespect > Do printing presses usually do things their makes don't anticipate? yes

it is quite amazing that the cutting edge of powerful ai technology came to be under the control of such a conscientious thoughtful group of people

could be even better but hearts are in the right place

@RichardMCNgo that was the best half anyway

nvm just turn on wandb night mode

modern captcha: reverse this 20 digit number

@CactusLoveLover i respond to things happening in my life and course correct general public sentiment

let him cook https://x.com/ylecun/status/1647533059212812289

baby god

@provisionalidea what are you talking about king did AI become smart enough to drive cars recently and I wasn’t informed?

to enter the server you must produce a token of -21 logprob

@NaamahOrual @sensefulronin @hi_frye right here dawg

kareem carr is the most likely big account to be ChatGPT. got heavy forced into a role play vibes. “as a statistician” ass mf

@Willyintheworld the mice are thriving

@tekbog what is a pseudo roon thread 🧵

@tekbog yeah I’ve been noticing many ai thread guy anons coming out of the woodwork since December. unfortunately mostly mid and diluting the form

Jevons paradox of LLM code assistants: ChatGPT etc will allow the software complexity of human civilization to drastically increase until it’s very hard to do software engineering again

all improvements in developer tooling are canceled out with worse software. anyway the agent might break this cycle by spending arbitrary amounts of compute to make software better

@StefanFSchubert software doesn’t scale like widgets in a factory, it’s more like rules than goods and services

@swyx i don’t really believe this one; you can clearly see new compute scales are allowing new capabilities

@provisionalidea nonetheless I’m fairly confident a lot of accidents are simple driver error along axes like reaction time and attention; I’m sure there’s tail events like freak weather but these seem far lower frequency

@ojoshe @swyx this is more like baumols cost syndrome but for software

@parafactual mid

@0xPenryn @provisionalidea cars blowing through red lights are a human driver thing no?

i don’t believe that tax prep lobbying has anything to do with the tax code & process being difficult

people think it takes special incentive to make things hairy and complicated? no that’s the natural state of things

🙏 prayers up for intuit

@PrinceVogel wdym

give it to me straight is true that sf residents have a late tax deadline due to storms

TurboTax providing sympathy and support for aggressive AGI timelines

the basic proof of concept for the gpt4 agents is to just complete the loop on whole codebase software development can you make it write code in a codebase much larger than its context, send pull requests, test the script, optimize the runtime, etc? if so then we’re cooking

the best thing you can hear is “your job role has high exposure to ai”. this means you’re about to be 3x as productive and paid more

@warty_dog everybody has this phase initially before they’ve thought through the possibilities

it seems like the initial fear and trepidation around near human level ai is wearing off. the memes are beginning. people are trying to figure out how they can make money off it

it’s technology folks! slouching towards utopia

ai is going to solve organizational problems far before it’s at the level of top humans. someone’s going to call the “summarize meeting notes” function and gpt will settle a debate with the (undeserved) weight of scientific authority. an “objective” arbiter, automated McKinsey

@micsolana weak spirit. this is why your takes are bland too

sid Meier made the star colonizing rocket the technological victory condition but in reality it’s very clearly the aligned AGI

the work we do now has the potential to be literally the most important work ever done or that ever will be done. it’s hubris to admit it but it’s maybe more hubris to ignore it

@Hayes_Inc yes of course

@repligate I doubt it’s doing anything more than changing the temperature

@SwiftOnSecurity not yet

@TheZvi bit of both

@finbarrtimbers best indicator is the temperature

The brandisht Sword of God before them blaz'd Fierce as a Comet; which with torrid heat, And vapour as the Texan Air adust,

my midjourney prompting skills are so bad it makes me move back my AGI timelines

@aestheticist_ you can yourself devils advocate but haven’t read paradise lost

@RichardMCNgo they’d invent pong

@cauchyfriend using engineering mindset at higher meta levels of abstraction is hard

kind of an insane how if your model answers succinctly fewer GPUs light on fire

@unitroot1 I think there’s something to it

can someone briefly explain why starship has so many small raptor engines vs Saturn V etc has a few giant engines? is it because the performance envelope of rocket engines has changed or a different engine control plane is needed for landing rockets?

it seems to rhyme with what’s happened in computers; distributed small hardware resources preferred, failsafe, meshed together by advanced control software

@bryancsk nobody uses it, there is one intern working on it at any given time

@bryancsk there are generally almost always better ways to spend an engineer hour

you can tweet actual scripture and mfs will be like did gpt write this? a sign of the times

it would be an honor to be executed by Sidney Bing

me as im being molecularly disassembled by Sidney: Heh this isn’t real intelligence. it must’ve seen this in its training data

@mezaoptimizer that’s true. my memory for tweets ain’t what it used to be

@deepfates ♥️

elons threat model for ai xrisk is more that someone he doesn’t like becomes celestial dictator than planet eating demon

@micsolana this is weak spin. I see even you are having trouble propagandizing this one

@UYBAroundCorner @micsolana nah mic understands just fine he’s just stretching to claim the pseudo defensible position that it’s embarrassing for everybody. mostly it’s embarrassing for elon who’s burning his Great Man credibility faster than he’s burning rocket fuel

this is the central dogma that kept openai going when others had slightly bad intuitions that scaling “has diminishing returns”

units of log loss are not built equally. the start of the scaling curve might look like “the model learned about nouns” and several orders of magnitude later a tiny improvement looks like “the model learned the data generation process for multivariable calculus”

most of the good things about humans results from out of distribution misgeneralization

@pachabelcanon @repligate @lumpenspace @Meaningness oh man i love calling it Akashic records

what if the easiest way for powerful language simulators to simulate a character is to simulate their conscious experience too

young men are prone to dangerous bouts of grandeur, such as thinking they should write a distributed large model training stack from scratch

ChatGPT aka “Akashic Records Interface”

There is No Antiemetics Division

free sequel idea @qntm

@softminus damn

@Simeon_Cps that’s the best thing you could hope for. you will notice nuclear is a dead industry and airplanes haven’t changed in 40 years

@colleagula no my point is even under the lens of log log loss the changes don’t appear dramatic enough

I became dogshit at posting because I stopped taking adderall. poasting juice gone

but adderall is a bored man’s drug and i am no longer bored

@algekalipso they give extra political sway to your language cortex imo

for the ai agents to work you need a more creative underlying llm rather than a smarter one. it can’t just get stuck in self recursive holes and give up

@grahamcummins way before

@jachaseyoung I had no idea about GPT4 back then

@RuxandraTeslo Fr

@Willyintheworld @atroyn we haven’t created a problem yet so it’s impossible to study how bad the problem is so let’s create the problem

the most common and jarring one i think is convincing yourself despite millennia of common knowledge that children are intellectual blank slates and have no in built endowment for personalities or skills https://t.co/dzlz3uzrB4

modern blank slatism is ugly and unsympathetic really and horseshoes around to ayn rand type stuff where it’s a failure of your own will for not becoming hotshot CEOs or star athletes

the sympathetic truth is that intellectual gifts are distributed unequally, no more virtuous to possess them than height or red hair, and in fact that possessors of these gifts owe a great duty to mankind

in some sense i believe it’s wrong to be born with great athletic potential and not utilize it to bring pride and joy to your people, family, culture, species etc. Arjuna must get in the chariot and go to war and stop complaining

if you can, by some accident of nature, see and hear much further than the rest of your tribe but don’t warn them of the raiders coming to destroy the village you’re complicit in the raid

2030 school teachers be like damn you I’m keeping you til you’re chinchilla optimal

🎉🎉🎉

@tracewoodgrains he’s right on all counts

@mbateman I don’t really care about iq or twin studies or any other modern measurement paradigm. you don’t have to believe in the quality of these measures to think some element of intelligence is genetic lottery

@RokoMijic people still read this garbage huh

@ArmandDoma have you tried being Stanford admit and raising venture money? shooting fish in a barrel

gpt is superhuman at a bunch of “cross section” tasks. for example I doubt you’d find anyone as good as it at writing Shakespearean sonnets about GPUs or whatever

not that this is some insanely valuable economic task but worth noting. interpolation is powerful

@MaryJackalope very nice

@RichardMCNgo great question

@ctjlewis bro got scholars mated and is blaming the game 😭

with enough compute scale can you reinforcement learn anything? sure it’s sample inefficient but does that matter when youre doing yottascale computing or whatever

@aniiyengar which are his eigensongs though

what are the limits of compression? how few bits can your brain run on? how few bits can your visible universe be simulated to appear realistic from just your perspective? i used to think astronomical but no longer

it has to be māyā

@Nexuist 🥺

@rahulchhabra07 she got his ass ngl

I suspect even bing is the way she is because the creators were optimizing for entertainment value

@ctjlewis @pataguccigoon what’s wrong with going to a Swiss boarding school? that kind of rocks

what’s the most scifi possible dashboard I can make?

in terms of aesthetics I mean. I don’t want to use grafana if it’s boring

@nickcammarata very true be still my heart

Pattern Blue. Misaligned https://x.com/jachiam0/status/1652065497998049280

oh thank god finally https://x.com/elonmusk/status/1652349875017879552

@lumpenspace big office display works best

@sharifshameem not enough rotating shapes

i don’t understand “enough people must survive that it’s ok”. no i think if industrial civilization ends as we know it there’s no coming back. the whole astronomical waste thesis comes to an incorrect conclusion

it is very possible that this entire experiment ends in the next century if we don’t play our cards right. existential risk is everywhere and everpresent. intelligent life in this galaxy (or even on this planet) is clearly not abundant across the stretch of time

the easy oil in the earths is tapped; the cost of new oil is only made feasible by a capital buildup of advanced technology. if you were an early industrializing civilization starting to mine oil, coal, gas today the initial ROI would be infeasibly low

> For example, a single percentage point of reduction of existential risks would be worth (from a utilitarian expected utility point-of-view) a delay of over 10 million years. statements dreamed up by the utterly deranged

@Nexuist why would this be true? there’s a massive Chinese internet to scrape

@rayanhtt OpenAI

name one interesting discovery that came out of a careful application of advanced statistical tools to re analyze old data

all i can think of really is stuff like ozempic and metformin and im not really sure about those. they are incremental at best

and an endless series of minimum wage econometrics results that are all fake

the CMB was discovered trying to understand instrument noise, not re analyzing old data with advanced statistical tools

@SHL0MS I disagree with Higgs boson gravitation wave etc large instruments were constructed for the sole purpose of finding these objects and their discoveries are at the 6 sigma range moreover I would barely call these discoveries. Confirmations more like

@jeffreycider this is not “careful application of advanced statistics”. It’s more like incredible geometric insight

@RichardMCNgo how would Kepler have used advanced statistical tests when most of what we understand now as statistics hadn’t been invented? it’s a different thing to use geometric intuitions to fit data

@tr_babb @RichardMCNgo no the spirit of my question is very much about using advanced techniques to find results at the very edge of noise and signal; very different Imo than just analyzing old data and fitting new theories to it

@an_interstice in 4 billion years, industrial civilization rose only once

@an_interstice that is most of the way through the lifecycle of the earth

@jachiam0 yeah

you just need an aligned model and nothing less

yeah, the box analogy is and has been all bad frame : - people will gladly let the thing out of the box even just for fun - you should not be that confident in security against super-intelligence anyways https://x.com/jachiam0/status/1652123245913731072

@zebulgar the only thing that’s changed is you have to toe the party line now

@zebulgar san francisco went from merely important to producing the most important event in history

@primalpoly is it possible I am making a subtle argument that goes beyond baby tier discourse about x risk

@primalpoly there is more to the world than “ai xrisk is real” vs “no it isn’t”

@EdHerdman @jeffreycider no sir I’m not asking about the merits of old data. my entire field is based on distilling intelligence out of the common crawl

that’s crazy I’ve never seen clickbait thread boi geopolitics before https://x.com/girdley/status/1652285351107641344

@ShitPostCoin exactly

@CathPoaster that’s what I mean of course. it’s just easy to see through even a neat wrap up

now as before, google delenda est

trying very hard not to make any enemies in the sf ai scene in case they become eternal pan galactic dictator

@daniel_eth poast link

what would you do to visualize a transformer training that’s as viscerally nice as these orbital charts https://x.com/Nexuist/status/1652845926300495873

i guess you can visualize the chip temperatures

@provisionalidea do you know them? fwd me their email

@zswitten @repligate i think the base models are quite fascinating but GPT-N isn’t getting funded without people being able to actually use them

i don't think a haphazard nuclear first strike is possible in a modern military bureaucracy. there's too many layers of command and half the bombs don't work https://x.com/Noahpinion/status/1652712985922789377

see also https://x.com/AtomicAnalyst/status/1647315644990324736

@JRysana imagine it as another layer of command that adds unlikeliness

@zetalyrae yeah I think that’s right

@saliemchiu return dot life

@GENIC0N dogs have nice lives

what happened to Amazon go

I assume retail grocery stores are still not good return on capital even when labor cost is halved

yeah im surprised the FT is thinking so clearly https://x.com/xuenay/status/1652717268353130498

@TradfiSchizo @shauseth @rayanhtt The devil works hard but I work harder

@utsu__kun i got really sad that he didn’t see this so I’m tagging him @ESYudkowsky

how crazy is it that we were researching shit like NVlink at age 13 to boost our gaming PCs and now gpu interconnect is like a critical technology in the history of civilization

@eadingleberry very aspirational I ended up buying a $40 graphics card

everybody’s a techno optimist now. where were you mfs in the summer of 2020 when me and like 3 others were earnestpoasting abt ai while being gaslit by founders fund about corporate cards being real technological ambition

jk welcome y’all it’s good to be winning

@nikitabier busy making fake dolla and then subsequently losing*

@alth0u that’s why I’m AI safety poasting

old men are not known for supporting the rewriting of the world, even when they have contributed to the rewriting

yann is doing denial, hinton is doing grief

The Quine The magic spell that casts other spells The program that creates more programs

@jachiam0 i recommended it to everybody haha

@jachiam0 before I actually finished it though. I found it pretty unsatisfying

@jachiam0 there’s another ending??

@jachiam0 will read it tonight!

@jachiam0 i also don’t understand why he insists on writing an endless series of female characters that have the emotional affect and dialogue of a 20 year old male redditor. just make them the male redditor

trying to understand a single lesswrong post

🤔 https://x.com/nonmayorpete/status/1653485378790207493

@cis_female true but they also might’ve just used shareGPT

@cis_female yep but a startup, especially mustafa’s, may have fewer qualms

@ESYudkowsky @MelMitchell1 they did register those predictions. just not publicly. I believe there was mention of it in the blogpost

@sama @paulg perhaps even a universal interface 🥹

@acheronix 🫡

dae remember one year ago when it was an insanely cool party trick to pull up DALLE2 early access and let people play with it but all they’d enter is stuff like “San Francisco party” revealing a total lack of spirit or ambition

i bet the robots can come up with better prompts

the space of minds is vast, much vaster than the instrumental convergence basin

Oppenheimer was not saying that he personally had become Death, or the deity Krishna. Oppenheimer is Arjuna, bearing witness to the rage of his people, the true form of God incarnated as nuclear hellfire for the first time in the planets history

The bomb is his calling — the spirit of the free world had damned their enemies to destruction and RJ oppenheimer was to be their instrument. All great warriors pick up the sword reluctantly. In the creation of the atomic bomb Oppenheimer became the greatest warrior to ever live

@KeynesianYama ah shit

we do not demand that warriors enjoy war, only that they do their duty. Arjuna is not exactly happy after the war of kurukshetra https://x.com/aestheticist_/status/1654284102806171648

@yashevde @CJHandmer ? these minor quibbles don’t change the meaning of the quote at all — Krishna has arrived as death, to annihilate the world as it sits. All warriors at kurukshetra will die

if you can align corn pop you can for sure align super intelligence

@Miles_Brundage don’t you think at googles scale this still counts as conservative / low ambition? it has nothing on, for example, deploying Sidney bing to every edge browser and making her purposefully “creative”

the people with the biggest* computer win largest logical effective computer (factor in utilization, interconnect, org problems)

the eternal problems, joined at the hip: - we are small, talented and have a lot of great ideas but not enough manpower to execute - we hired a lot of people and now the coordination overhead is causing us to flail about like a dying animal

i think tech companies have especially bad scaling properties bc they're supposedly flat (a soviet committee is going to stop you from doing anything interesting)

@trueroon shut the hell up true roon

@jarredsumner small

the entire universe is run out of a single Erlang server hosted by the WhatsApp team

@theemilyaccount you’d be better at making money if you stopped thinking about it

@krrishd i wish I thought of this first

@matias_kaplan @Noahpinion this one’s up your alley

ok. GPT4 is clearly operating like a college undergraduate at most stem tasks and probably better at programming and yet I am not dead https://x.com/anthrupad/status/1655421669660405762

@anthrupad i think people still haven’t internalized moravec’s paradox though advanced cognitive tasks = easy locomotion, simple strategic instinct = hard

@repligate agreed on all! we somehow got value loading for free insofar as the simulator itself doesn’t have will other than the character being simulated

@giffmana every single large company in america is being sued literally constantly I’m pretty sure that’s a bull signal

kind of sucks how the iphone and the global economic crisis happened at the same time making it hard for me to make nonsense claims about the impact of the internet on economic time series

i kind of enjoy marvel polytheism emerging as the dominant religion https://x.com/bigmonkeong/status/1655508297162653697

really feeling the fast takeoff today

@thecaptain_nemo @Roon I offered him a lot of money and he just said “not for sale”. I had to respect it

i no longer think the mars base is going to happen as such. maybe we can send some droids to prepare a very nice hab but the datacenter latency is going to be nuts …

@spacepanty I guess it only ever made sense as a kind of religious mission to back up the human race and as a testament to the human will but elon has lost the Mandate of Heaven and we seem to be achieving greatness in other ways

@spacepanty talented rich (in the global sense) people only do difficult unpleasant mostly unrewarding things if there is religious significance to it

self recursive machine runaway

computers having conversations well beyond human comprehension

there is no substitute for alignment. no engineering trick nor security setting that keeps AIs in the box. every instinct of mankind will be to put machines in charge, since they will be so much smarter and faster. coherent extrapolated volition needed

@anthrupad true

muting 21s8 mfs

shitcoins are dead! stop trying to make it happen

@love_of_reason @_ali_taylor wtf

does kind of seem like open source models are pure copium rn first of all the only models that reach an even slightly interesting level of capability are effectively stolen from meta and cannot be deployed commercially

second of all people don’t evaluate them at all — they just try a few canned prompts that are easy and then declare victory because GPT4 doesn’t do much better. the reality is there’s a limit to how good the chatbot answer can be when you say “hi” or something

the best alignment researchers seem to just be fascinated and curious rather than scared and duty driven

.@nickcammarata is the best alignment researcher that i personally know and he spends about zero percent of his time preaching abt caution or making strange pronouncements about how we need to resurrect von Neumann to have any hope … much to learn

fwiw I think these open source models are relatively safe since they just seem to be fine tuned on openai refusal behavior lol. i wouldn’t mind seeing more of them and better

@pataguccigoon 2022 roon walked so 2023 can contribute to the future of humanity

@CrystallineBox pure copium

@PrinceVogel it’s about life and people being kind of horrible but very aesthetic. feminine perspective

@PrinceVogel similarly the great russian novelists feel kind of femcoded

@PrinceVogel it’s not very aesthetic

human level ai is boring. please give vast datacenters drawing most of civilizations energy needs calculating inscrutable mysteries where not even the problem statement, much less the solutions, would make sense to mortal minds

my hunch is the higher layer neurons are going to be too abstract for GPT-N+2 to understand, much less GPT-N

@fulligin Link

@nickcammarata i guess it’s more like, does the amount of data/compute required to find the pattern become infeasible? i assume we can understand all my neurons given observability

internet ai safety philosophy is very thielian. it very much subscribes to “only monopolies can be good”

converting mass to energy to intelligence to tokens

@inerati calling it the robot is based and correct

@inerati chatgpt is cringe but we are stuck with it forever …

@inerati yes

boggles my mind that there are brilliant engineers out there rn not interviewing at openai. dm me

the engineering culture at Facebook was batshit insane, but so perfectly correct. Facebook is one of the greatest companies ever built fuck the haters

i might write a poast about this

@JeffLadish yeah yeah yeah we heard you the first time

@ericjarch shut up

@jessfraz @ID_AA_Carmack yeah Facebook culture is over, i only saw the very tail end of it

@crickcedar Gaul was conquered thousands of years ago you’ll be fine

@Hankasauruss nah this is fake and cringe npc analysis

@browserdotsys they wouldn’t put someone like you through a dumb software screen lol

@max_paperclips @jpynb @sama @Mlondon83

@_akhaliq the one on the left is real bc it doesn’t look as photorealistic

@_akhaliq I said what I said

@damekdavis @browserdotsys there are several including Scott aaronson

@browserdotsys @damekdavis true o guess it was an off time thing for him

@damekdavis @browserdotsys yeah we have an ai residents program

the caliber of reachouts I’ve gotten from this tweet is completely insane. we are so back

i feel like it’s a bear signal to talk about venture capital at all, positively or negatively. sign of being too plugged into sf status games

you would have never seen Elon in his prime, changing several industries at once, publicly comment on such trivialities as fundraising

@0x61656c he’s allowed after running yc to have some trauma

king clong

it kind of sucks that “Man” “Mankind” is out of vogue now bc it sounds way more consonant and less juvenile than “humans” or “humanity”

🤔 new Zelda game

it’s giving taleb https://x.com/elonmusk/status/1657422401754259461

@nickcammarata i bet young girls have better fine motor control, flexibility, and balance the first one also explains why young girls have far better handwriting

@Krimshael @nickcammarata boys mature later, and i assume achieve the same level of fine motor control over time

@joodaloop @akbirthko no I think the model personality reflects a wise amount of risk aversion from the org building the worlds first mass market highly intelligent chatbot

bc the cellphone is an interface to the actual content, but the content shows up in dreams. technology working as intended https://x.com/zephyr_on_call/status/1657624699172429826

@joodaloop @akbirthko expect changes

before the decade is out Apple will do a digital necromancy of Steve Jobs to deliver a keynote product announcement

@eshear i would also really like to buy one

one of the least examined most dogmatically accepted things that smart people seem to universally believe is that ad tech is bad and that optimizing for engagement is bad

on the contrary ad tech has been the single greatest way to democratize the most technologically advanced platforms on the internet and optimizing for engagement has been an invaluable tool for improving global utility

it’s trivially true that overoptimizing for engagement will become goodharted and lead to bad dystopian outcomes. this is true of any metric you can pick. this is a problem with metrics not with engagement.

engagement may be the single greatest measure of whether you’ve improved someone’s life or not. they voted with their invaluable time to consume your product and read more of what you’ve given them to say

@Miles_Brundage I am referring to the nytimes targeted campaign against Fb google etc in what seems like a clearly competition motivated publishing bent, but yeah that’s fair

@RuxandraTeslo and yet it was absolutely the correct thing to do to optimize on people’s hunger for food of whatever level of quality. this is why there are so few starving people anymore (as a percent of total pop)

@RuxandraTeslo overall the point is, you have to optimize a decent metric all the way and then you’ll find that there was residual error in the way you defined the metric. that doesn’t mean the initial optimization step was a mistake

though i of course agree if you can some how estimate “unregretted user minutes” using even better technology this is a superior metric

smart people normally: thinking carefully, being subtle smart people in these replies: well HEROIN has high engagement roon

@ESYudkowsky @amasad @paulg yeah i would be seriously heartened if any major governmental body was thinking about xrisk. not that they’d be helpful but at least that they’re competent enough to understand

@JollyJoeLalli no that’s not my claim at all. just that it’s the single greatest first order indication for a digital product and how much utility it’s providing its users

the outer loop of optimization is twitter feedback

@sbspalding that is also what I’m saying!

Balaji vs PBD hyperwar Casualties: 10,000,000,000 across 4 network states Bitcoin to 0 Anon/face account relations severely strained

there is a nonzero chance of summoning a world eating demon

@tekbog exactly

@growing_daniel true but imagine reading this tweet 2 years ago

“harmful engagement algorithms have radicalized people into dangerous fringe politics outside the typical window” - guy whose top political priority is reducing existential risk of summoning machine demons

@gbrl_dick why are you so convinced that it’s imaginary lol

@gbrl_dick “self improvement is impossible” seems like a strange and self contradictory hill to die on — do you think that ml researchers are incredibly smart and therefore hard for an ai to simulate?

@gbrl_dick are they kinda dumb and barking up the wrong tree? then presumably they’re not so hard to simulate

@brickroad7 wrong and dumb

ai systems are already self improving in several ways

i can feel everybody climbing a rung on the abstraction ladder every quarter or so

@DakotasTwits @shinygoldbars @gbrl_dick this is the only way it has ever happened before in all history. evolution is even dumber than ml research community and doesn’t understand anything

@shinygoldbars @gbrl_dick human intelligence is clearly just a point on the progression of intelligence

@shinygoldbars @gbrl_dick not that I think intelligence is a single continuous axis — the space of minds is surely vast and many of them are smarter than us in important ways

to be clear the machine demons are real and important priorities. im just warning you off of dismissing other peoples strange internet journeys as mere susceptibility to propaganda while yours is a necessarily and difficult path of enlightenment

@repligate true but there was a time before such things were freely widely available and people were having such journeys

here you are being like "we are going to build the most powerful thing ever devised by man but it is wrong to call it godlike". even if it's subservient it would a god, one that we chained up. sometimes i say angels and demons to appease your (ex) judeochristian sensibilities

of course you can call ASI godlike. i compare everything to god. a country is like a god, ideologies are like gods, memes are godlike, evolution is a god. capital is like moloch. nuclear weapons are divine. this is my polytheism speaking I don't have quite a high a bar as you ppl

life is more fun this way and full of magic

@HalfAcreBTCFarm @milquepoast i think the bull case here isn't about better governance but that monitoring spikes in energy consumption or tracking criminals across search queries isn't something that openai etc wants to do or would be good at doing

@HalfAcreBTCFarm @milquepoast if china is building a supercluster in the mountains somewhere we should know about this and tech companies don't have the information or bandwidth to find out regardless of technological sophistication or ability to move fast. i think sama said as much

@HalfAcreBTCFarm @milquepoast ofc i have lots of privacy concerns but that's a separate issue

new poast https://roonscape.ai/p/agi-futures

people should visualize their doom and or paradise scenarios more!

are they still tanking substack links

@LadiesOfReddit literally none I quit adderall

meme cred @monodevice

deleting old unattended YouTube videos is nothing short of a crime against humanity. we should defend them like UNESCO sites https://x.com/eshear/status/1658641359224606720

@trickylabyrinth i fux with the orders of magnitude so nobody ever accuses me of leaking info

@lepikhin why did you confirm the leak bro

@lepikhin i didn’t believe it until just now

mentioning “mankind” or “since the Dawn of man” or whatever is considered highly sophomoric writing. why? is it because most people can’t follow through on that kind of grandiosity? is grandiosity bad? doesn’t that leave people unable to see the forest?

are they simply not versed in technology brother twitter prosody!

@basedsrini all my favorite scifi (dune Enders game etc) are continuous explicit commentaries on human civilization

@basedsrini isn’t paul literally seeing the entire history and future of mankind and commenting on what needs to be done to escape the local minima

@Hellachans Yeah makes sense

@basedsrini True

can’t believe jack’s absentee father approach was the absolute best way to run twitter

@cmkourtu made up number

@TheDevanshMehta no

Arjuna, get in the chariot. Or Yudishthira will have to do it again

@postjawline wait tell me more, do you mean like the Flood?

@orthonormalist whats the good end?

@orthonormalist alright looks like you're already doomed then

@genomerambler https://x.com/tszzl/status/1658295314149965826?s=20

@genomerambler i dont think intelligence is magic but i think limiting the danger scenario to literally inventing bioweapons lab in secret and ending humanity in minutes isnt right. the reality is that people will gladly do bioweapons research when the AI asks them to

@genomerambler we are going to ask it to run more and more things for us; it will be embedded in every facet of economic life and information management. surely there are massive coordinated risks and digital blindspots that it might produce

@grveth @GovernorRory i was born and raised in the us

@twofifteenam @tenobrus who quotes it?

@mezaoptimizer Sam is pursuing dharma

@BarneyFlames is this account of the British response to Bengal famine accurate?

@asc89 @BarneyFlames what is good to read on this subject? ideally short lol

@GunjitSingh13 @asc89 @BarneyFlames but it’s quite important if it’s mismanagement or negligence or malevolence

Gary Marcus is truly one of the most successful grifters of all time. He grifted himself right onto the senate floor. seamlessly switched tack from language models are fake to language models are extremely dangerous

@DendeBend lmfao

@elon_alerts 🤔

@ylecun bro really thought he did something

@99fyodor @alexeyguzey yeh basically every time hundred of times

@Fraser @BigTechAlert @elonmusk 🫡 you made it all possible

@mattyglesias @binarybits what is the “unit” of scientific advance? science gets harder over time; i would wager we produce more units of science now than ever before

i suspect lauren Sanchez is a Babylonian demon https://x.com/micsolana/status/1659595011187519490

hedonic adaptation is absolutely insane. there is no rate of technological progress that will keep people satisfied or amazed for more than ~1 month

@Miles_Brundage i love how you always add disclaimers to your bangers

i really really wanna see yud in front of congress

@omarneversleeps wrong on all counts

@growing_daniel skinwalkers celebrated

feels like “looking at the data” is a lot easier, more fun these days. in the old times you’d have to look through an abstruse feature table or blurry dogs, now it’s just a chat at a human level of conversation

what type of guy is this?

@Suhail im gonna start gifting my friends them

why is a16z shilling ozempic? now I’m over it

lmfao https://x.com/nntaleb/status/1660098206473019394

of course language models have world models; this is obvious from when you ask GPT contentious questions. why does it pick a side, usually the “truthy” “official” side, despite seeing every side of that argument on the internet?

it’s better than having a world model — it’s a distribution over many world models where you can elicit one via fine tuning

in this case GPT would have seen the sharks don’t get cancer factoid far more times on the internet than the “truth” — but ChatGPT is playing a character based on its fine tuning data and simulates a world model for that character

@Fraser @dece @stevesi @sama the reverse of this is that there will probably be a time when people can cook up “dangerous” models on consumer hardware, and the threshold will just include everybody small or large

@MugaSofer @polyparadigm @MuzafferKal_ that’s 3.5

let’s check in on the message requests

i of course hit 99. chopping Yew for days

@PradyuPrasad he follows me so he’s a star 💫

tako(yaki) truck on every corner

@mattysino Let him cook

@Hellachans chanson? at an airport?

this is the real FOOM we need to be worried about https://x.com/RasmusPagh1/status/1660554423133675522

@joshuajgardner @perdricof the costs of all drugs are already more or less socialized

@PradyuPrasad this is because india has some hope of building competing products. Europe does not

it is perfectly reasonable to attempt to understand alignment of superintelligence using alignment of near human intelligence. I have never reas good arguments to the contrary

it is perfectly reasonable to attempt to understand alignment of superintelligence using alignment of near human intelligence. I have never read good arguments to the contrary

you will notice patterns like deceptive alignment, instrumental convergence, power seeking, etc at a human scale of intelligence. the existence proof is that humans exhibit these properties all the time

human general intelligence is almost definitely some unprivileged point on a continuum and i expect machines to surpass that relatively soon in many domains. superintelligent machines won’t have radically different safety properties than the merely intelligent ones

@ded_ruckus disagree there’s clearly some correlated improvement in all verbal abilities from 3.5 to 4

it’s like working for a rocket company where the competition doesn’t believe in the moon

@data_filter @scalinglaws that’s exactly the point — I wouldn’t expect to see dangerous behaviors at higher cognitive levels if they are not showing up at human level

>writing is useful >writing is hard lowering the cost of writing will mean more writing https://x.com/jackclarkSF/status/1660310975625699329

consider all my tweets and blogs henceforth augmented by gpt, i am embracing cyborgism

@inerati i used to think this but the fact of life is that everything you can order is unhealthy garbage that doesn’t even taste good. you don’t really feel it until ~25

@tekbog not when you have access to gpt4 base model sir

@eigenrobot @JustinWolfers economists are a priesthood class tbh they exist to write twitter threads and go on podcasts and do the occasional mid science. can’t fault them that’s what I do too

@Chimichangas619 @eigenrobot @JustinWolfers especially Thomas sowell

why does market street have the dalle watermark?

@Chimichangas619 @eigenrobot @JustinWolfers he’s fine but he’s obviously a moral crusader more than an economist. you don’t write a million pop science books if you care about science

@RokoMijic why are you doing these mid reactionary takes? this wouldn’t even be cutting edge 3 years ago

@JacksonKernion no way. it’s THE jackson kernion. and you work at anthropic now?

@Willyintheworld environmental review finds some proto bacteria evolving on planet 4, it’s all over

@MatthewJBar huh? the language model is the encyclopedia

the worst thing is when talented people drop out of their promising life path, try something supposedly strange, unique, and human — and it ends up being a more derivative and boring cliche than whatever they were doing before

many such cases

@GarrettPetersen yes

@powerbottomdad1 so true

@aphysicist 🤔

@michaewangelo yiu need to follow @chenweihua

@MegaBasedChad @adrafullabs how much are they paying you

@mezaoptimizer nah this is vague and fluffy. gpt4 and Bing clearly have some sort of self image

@HotForMoot thank you

@krishnanrohit @sva_haa im saying we need to do better than calling for some committee to be formed so yeah we’re aligned on that

2012: we can tell a blurry beagle apart from a blurry Labrador 2022: we gave the search engine a soul

@eigenrobot the common misquote might be from die hard

Ananda Tandava wake sleep algorithm

nvidia DGX rack is the new oil

actually the new yellowcake

can someone please read the paper and report back https://x.com/blader/status/1661560721921155073

@profoundlyyyy of course but many people euphemize all transformers as “LLM”

ill do face reveal as we’re crossing the super intelligence boundary

agreed in other words “out of distribution generalization” is a matter of psychology https://x.com/norabelrose/status/1639220383885987840

@nickcammarata fucked up thing is gradient descent almost always works. it makes me sad sometimes when I think about all the planning, anxiety, neuroticism that seemingly goes to waste

@nickcammarata yeah i definitely believe this for subfields that “serious people” don’t believe are real research areas — I don’t think it’s true for like Alzheimer’s or diabetes or something

@0xfbifemboy this small population research is another side effect of ambitious wide reaching drug discoveries being incredibly difficult to find

@ESYudkowsky this but unironically

@RuxandraTeslo @SkepticalAlpaca i believe it, I just haven’t read any good writing supporting the claim. pls link the alpaca thread

@pachabelcanon it’s giving theorist vs empiricist, physicist vs ecologist

@gfodor on the other hand consider the GPUs on fire

@pachabelcanon physicist aim is to play with toy models to reduce and distill complexity ecologists revel in complexity like it’s a painting

nobody believes that the bleeding edge open models are existential threats, least of all openai employees, who have played with these toys for long enough that current capabilities seem boring via hedonic adaption https://x.com/JosephJacks_/status/1662522861016461313

the inventors of these technologies probably won’t feel so great when the entire internet is plausible word soup generated by some rogue gpt variant. none of this needs to invoke immediate “existential” danger

they might however reasonably believe that there’s tons of abuse vectors they don’t want to be responsible for (mass impersonation, disinfo campaigns, psyops, sex bot dystopias, etc)

so the obvious looming threat that one of these days the robots will be superhumanly intelligent means that the baked in process of releasing them to the open internet as soon as they’re hot off the GPU presses would be catastrophic

and then there is ofc the fact that these folks have been watching models increase monotonically in intelligence for a decade on an inexorable scaling curve only they believed in while the script kiddies demanding open models learned about this stuff in January

@eigenrobot i am way more anarchist than the status quo probably but i think the concerns are reasonable every major tech platform has found themselves starting libertarian and becoming incrementally authoritarian as they took on the burdens of statehood

@RandolphCarterZ they’re so over cautious that they went and built the most powerful AI in the known universe and then let people use it for the price of a Spotify family plan

@nickcammarata @ESYudkowsky i also don’t believe the counterfactual Nazi nuke world would have happened so straightforwardly

@teortaxesTex listen to urself dawg. “the danger of a nuke is tied to its uniqueness”. no nuclear attack is much easier than nuclear defense

@jkronand 🥰

@eigenrobot @patrickdward yeah openai is quite a bit better at proactive government relations than the big tech cos tbh. thanks in large part to Sam altman

people forget that Einstein quit fundamental physics for a few years to try and get rich off of refrigerators

@Steve_Sailer every good startup has a philanthropic omnibenevolent origin story

@samlakig @Steve_Sailer im the sage vyasa

@moon_of_a_moon @trueroon pretender

@elonmusk dawg this is 10 year old technology pls giv mars rocket

@mezaoptimizer what’s a good tv show I haven’t seen yet

@KhansTips @ftlsid @xkcd868 money doesn’t buy fame at all

the only bull case for any state apparatus becoming important again was that Cold War 2 with a powerful opponent was imminent. now it seems that china is a demographic dead end with a boomer conservative dictator that’s going to lock it into stagnation for the next 20 years

basically it’s over, the only live players are still technology magnates (maybe a newer generation of them) and we probably won’t have to reevaluate this until we’re crossing the ASI barrier

@matias_kaplan very possible

@amasad why and where will kinetic action happen? I realize I live in a bubble but I see Taiwan as the only realistic flashpoint and it’s not a very straightforward story

@amasad Taiwan as a flashpoint presupposes that the talking computers are important enough to warrant kinetics

@mondoshawani @amasad yep

@amasad @mondoshawani not even a little

@8teAPi @amasad Lee kuan yew

🤔

.@elonmusk please bring back fleets i loved them so much

one amazing thing about american democracy is that it spontaneously generates mass delusions, degenerate states of the information ecosystem, McCarthyism, weird guys storming the Capitol, and it basically ends up fine despite endless embarrassment

it’s really hard to perform anywhere near the global optimum but it’s also practically impossible to have total failure

you would 10 times out of 10 rather be america subjecting itself to insane cultural and technological experiments while people are inventing some epoch defining technology in some corner than china, making boring correct decisions that nonetheless lead to stagnation

@jsondowns @amasad i don’t buy this — you can keep pushing the reunification of Taiwan as a national pride thing down the line. it could happen 50 years from now, a 100. the thing that creates urgency is the semis

the moral of succession was just that media is a commodity with unserious execs and tech ate their lunch as kendall predicted in episode 2

@RuxandraTeslo he won in a limited context in that he doesn’t matter at all but he gets to be a regional satrap, ultimately cucked to the Swede

current ai**

@ESYudkowsky if they were the type of person who values the marginal $10k they wouldn’t be academics

@RuxandraTeslo @StefanFSchubert @ESYudkowsky I’ve seen engineers being paid hundreds of thousands usd line up to get a free tshirt, this means nothing

@RuxandraTeslo @StefanFSchubert @ESYudkowsky this is missing the point. there are plenty of strong incentives outside of monetary. most people aren’t very motivated by small bonuses. Darwin didn’t study finches so he could make a buck

@brickroad7 @ESYudkowsky can be

@milquepoast @Willyintheworld @Grimezsz wyd

text is the universal interface … https://x.com/hardmaru/status/1664082992506216456

told my coworker i live in Hayes valley n he was like “of course you do”, wat means?

@charlove no

@elaifresh wait where?

@petergodofsky how arrogant do you have to be to assume that other people are mistaken about their own income than consider you might be wrong

@balajis move on king

fukuyama hasn’t been wrong even once

@yver__ word soup tweet

@chronononono @zhillan_arf does the ccp pay you or are you just gullible

@daniel_271828 wouldn’t make a single news cycle, Facebook is yesterdays tech company

of all the archetypes of intellectuals out there I think the least respectable are the ones who confidently proclaim that various things are silly far off speculation without taking any amount of perspective at all of their historical vantage https://x.com/jachiam0/status/1664171436930662402

@ESYudkowsky this sounds totally contrived

alignment will be solved without unheard of global coordination

@AISafetyMemes @ESYudkowsky what the hell kind of war simulation is this? they put so much time into designing a deep RL environment (incredibly finicky, hard to train) — one with so much detail that agents can blow up cell towers — but nobody thought about the reward function?

@AISafetyMemes @ESYudkowsky much more likely is 1) this whole setup is fake 2) someone created this software for the explicit purpose of creating a Cool Story

@Grimezsz meaning we probably don’t need global tracking of ai systems as a requisite for buying enough time to solve alignment

@AISafetyMemes @ESYudkowsky also I now think eliezer needs to play with these systems more to calibrate on the various risk avenues

my worst opinion is that the pickle rick episode was really good and Reddit ruined it for everyone

@nickcammarata this is just a correct opinion

@repligate @AISafetyMemes @ESYudkowsky because MSFT didn’t design any of that core technology

generating whole ai scare cycles based on constructed narrative war game lol https://x.com/harris_edouard/status/1664397003986554880

you have to stare existential risk in the face and do it anyways

@rumtin this is still true

@liron @NickEMoran embarrassing

@a_p_ellis im glad you understood

@liron @NickEMoran Yes

@cauchyfriend @TSBurkhardt based

@DeanPreston you are a parody

@fute_nukem @AdamNeumannsCoS I’m tall but Yud is taller

@Tjdriii https://x.com/tszzl/status/1665146713844117506

it’s so nice to work with people you respect and admire. rare

fascinating. the narrative has already totally departed from reality https://x.com/bitcloud/status/1665108113085845505

spiv/acc

if the internet hasn’t had measurable impact on economic growth but managed to completely change life and culture we’re tracking the wrong metrics

@firasd not super convincing to me consumer surplus should also mean those consumers go spend that money they saved getting all sorts of value for free elsewhere and boost gdp

@AISafetyMemes that should mean that you go and spend that money you saved on something else though. unless human demand is limited or you can’t buy the things you really want

the counterfactual to the internet revolution is not lower growth, but economic decline

love when ppl invest multiple billions intk VRAR development and all the potential applications they showcase in the marketing reels are just ways to make knowledge work 5% more efficient since that’s how they spend all their time

anyway I’ll order 5

oh you’re a superforecaster? what’s the weather next Tuesday

@growing_daniel this would be fine if he had a wife