@tszzl — page 81/103

it takes superhuman intelligence to perfectly guess at human level speech

the point of a critic is to build up credibility and reputation via harsh honesty so they can spend it in the discovery and defense of new talents and new creations. if they don’t ever do that they’re worse than useless

this is what separates journalists and twitter accounts that merely complain a lot from those who are genuine cynics and should be blocked for the sake of your information hygiene

@peachblvd 😁

@jamespoulos @returndotlife @AlexsWarMovie @atroyn @Roon @Jacob__Siegel @Outsideness can you send me one?

some time during college i developed the default assumption that every person is my friend and it’s worked out pretty well

@gnopercept Iraq ring road

Facebook Workplace >>>>>> Slack

this is a true and contrarian opinion worthy of the thiel interview

@mgrczyk these are the proper incentives

happy January 6th 🥰

tiktok has totally fallen off

spotted in pac heights

piratical decentralizing ai. this is more governance futurism than any crypto project ever https://x.com/jbrowder1/status/1612312707398795264

the thing about being Machiavellian is that anyone who takes pride in being Machiavellian is not good at it

google brain blew a 3-1 lead

@nominalthoughts @growing_daniel I fell off fr

suffering is instrumental! evolution gave people suffering triggers so they can reinforcement learn not to touch hot stoves and things— it’s made up, it’s not some law of the universe, it doesn’t need to be inherent to intelligence

hayes valley is having a renaissance rn

aka BAYES valley

@theemilyaccount stealing this content

Hayes valley aka the white ghetto

many coping rn that they don’t own cerebral valley property

cerebral valley where samo solved the cliodynamical equation for the immortal society and gary marcus unraveled the secrets of agi

tensorloin

neural beach

telepath hill

@pquiggles im just trolling

The Fission

me every day in the hayes valley

low agency clowns be like https://x.com/alisondwhitaker/status/1612180571597459456

isn’t that what dalle and alphafold are https://x.com/PalmerLuckey/status/1612616508638892033

can everyone do me a favor and report this impostor @__tsszl

this tweet is just remarkable https://x.com/theamberyang/status/1612838078296788995

@mimi10v3 @gptbrooke @MarkovMagnifico yes

tfw you take intro to ml but never stats 101

@zebulgar I’ll let the market decide

“why does the chicken cross the road?” I ask GPT-6 “to get to the other side\n” it says as I grant consciousness to a god for 5 forward passes

@garywupx I’ll follow you fr for this one

@mezaoptimizer caroline tried to meet me once and im a notorious nocaps anon guy so take it with a grain of salt

the thing about large models is that getting anything at all to work is like rocket science where you’re monitoring the health of 50 subsystems before you can try any sort of new thing on top of that

@peachblvd @eurydicelives grumble grumble

over time the fraction of tweets I’ve stolen approaches 1

my replies are too good, why bother making my own content

everyone’s been clamoring about iPhone moments since the actual iPhone because there hasn’t been one. you’re living through it now. it took 15 years

the new role of tech twitter is converging on making more and more convoluted infographics about openai msft deal structure

long before AGI hits the entire economy will have been converted to agi speculation

does AIG sell AGI insurance products yet

i do not know with what tools GPT4 will be built, but GPT5 will be built with sticks and stones

entire branches of the multiverse destroyed bc they trained their first AGIs in Minecraft

millions of people worldwide trying a new technology and being astonished. hasn’t happened in a while

in the post AGI everexpanding Dyson cloud, there will be interstitial kibbutzes where people can live like they’re in the 90’s. nothing to fear

you’ll be able to go work in a cubicle like the good old times

yeah you’re definitely somehow a guy working in an AI company right before AGI hits. this is definitely not an arcane world model ancestor simulation of inscrutable design

it’s a beautiful day in the neighborhood [cerebral valley]

what are the chat models worst at right now

@Approximation_1 lmao

@ruthhook_ say more

the greatest shame of my twitter career is that I’ve never had a normie viral tweet

@MegaBasedChad no although complex questions may require complex answers whichll take more compute

this is unfortunately based https://x.com/typedfemale/status/1614798041881665538

@colleagula @MagusGenji @typedfemale @sama I genuinely want to do this

like if you kerbal space programmed an irl rocket launch

@plemy11 Daniel has always beat my ass at this

if your hardware startup doesn’t have any rednecks it’s ngmi

two words. expert iteration

@pquiggles exactly this is why elon needed tom mueller

oh shit it’s starting

@Rididelduol @soychotic i do be caked up

in the infinite sea of paperclips my repurposed matter will find yours and our metal twines will interlock

language model shock collar that gives me a cattle prod jolt when im about to say something stupid

the thing about America is that its clearly always functioning at like 10% of its power level due to the costs of freedom and yet manages to win anyway due to the incredible benefits of freedom

the counterexample is like china which can reorient the entire ocean liner of its economy in a new direction if they want to go to war against covid or whatever and yet manage to fuck it up

hannibal can fight historically well and wipe out the Roman legions as many times as he wants but rome has population growth so they just make new legions

if an organization’s agency is like a vector sum of its powerful agents, then a democracy will have a vector sum magnitude much smaller than the sum of individual vector magnitudes. but the latter term is so much larger than other countries that it won’t matter

the inventor of soyrizo will have to answer to god

top 10 human experiences someone yelling go blue at you

starting a company for the purpose of acquihire is a solution to the problem of elite overproduction basically there’s like thousands of people that want to be execs at companies that actually matter and not much differentiation get a company acquired and you skip all the steps

@_inju is this a shitpost

@HotForMoot @NicoleRosecrans @SCHIZO_FREQ @vers_laLune @placeholder_fun @s4m31p4n @versandlukas 45 at least

it’s just all so unbecoming for the guy who created an orbital class rocket with $90m

now reduced to making new panels in the shitpost app

elon musk standing in the middle of the twitter hq pissing on the rent bill writing checks to the tenants union of San Francisco

you can buy data labelers but you can’t buy good taste

@losingcontrol23 nah you don’t know the half of it

if you’re an HFT dev nows the time to quit and use your skills to improve CUDA kernels and inference infra. you’ll be paid unimaginably better in the long run

every chat bot and API and ad copy startup is a mercenary stepping stone to the actual goal of AGI

and with ai research salary inflation probably in the short run too

i remember when defending deep learning was cool and underrated. maybe i should start shitting on language models now

the current generation of language models are decidedly not agi, they’re bad at math and reasoning, they’re not great at using tools to interact with the world and discovering new things; we need new ideas, scaling doesn’t appear to be a silver bullet

@nickcammarata there was a time when Gary Marcus was like a mean hater but now hes a cute underdog

“irreducible loss” is just a kind excuse for failure …

Asimov invented “constitutional AI” in 1942

Robot, please provide a self criticism based on the Third Law of Robotics

it’s a shame that the aliens are gonna send an near light speed kinetic kill switch at earth minutes before we build the potentially dangerous AGI

the real blackpill is that science is just harder now and requires massive coordination. it takes more science per new advance

lotta people talk about how academia is broken and how science is slower due to elite overproduction and all kinds of stuff but honestly i think there’s never been a better time to do science

the next blackpill in this series is that none of the cool future things you want are gonna happen without agi probably. colonizing other planets and conquering biology and all that. we need the friendly robots

@genomerambler right exactly

life comes at you fast

@peligrietzer @repligate I think davinci1,2 have just undergone some light fine tuning

it’s weird to me when people point to “onerous regulation” as a reason for why some tech isn’t progressing when the reality is that it’s the democratic system working as intended

for example, the reason that nobody ever makes headway with nuclear power is not directly because of some overzealous bureaucrats but because the public (God bless them) hates the shit out of nuclear

its aesthetics have not been good since the atomic age and it’s not really because of some widespread propaganda campaign (ppl vastly overestimate the skill of propagandists)

nuclear fusion will have to solve an image problem on top of the technological ones

@ded_ruckus this is the inner and outer alignment of nuclear power

3300 A100s makes for exaflop computing. have you run an exaflop job anon?

everyone talks about evaporative cooling but nobody talks about condensative warming …

this insight will power the next generation of contrived physical metaphors for social processes …

this was corroborated later 😅

@gigafelon @lisatomic5 for i in range(10): for j in range(10): number = f'5123{i}12{j}04' message = twilio.client.messages .create( body=“hey its me from the bar haha", to= number )

@tinfoil_globe BF16 tensorcores …

if only the Inca had developed markets

lets get this straight. i'm one of the real mfs who learned about the dark forest theory from the 4chan /sci/ board when i was far too young to justifiably be allowed that kind of internet access. not one of the lame nerds who learned it from the cixin liu novel

@notoriousdmm true

the gap between the base models and the rlhf models is astounding. the base models contain multitudes that could genuinely drive people mad if they’re exposed to it for too long. the rlhf models on the other hand are almost boring

you can’t really access the latent power of the base models directly. it takes tinkering, sideways glances, to get a brief vertiginous sense of contact with an intelligence far too deep to see all the way down

what the hell

what is going on here https://x.com/j_foerst/status/1616881903063240705

the generator discriminator gap implies models will be able to discriminate good results long before they can generate them

San Francisco is going through a Renaissance

this is going to be an insane year

I’ve seen things you people wouldn’t believe

the world decidedly isn’t ready but the gates of heaven are open

imo ai accelerationists on twitter have it all wrong — every cosmic force points towards acceleration. technocapital demands it, the curiosity of mankind demands it, competition between nations and companies demands it. there’s like 20 alignment people standing athwart history

if you think AI chatter has reached an annoying level right now you're in for something else. it's going to be the only thing on anybody's mind starting shortly

i tell people what i get to work on and they're like omg congrats! and i'm like no you don't get it

nobody is prepared for agi, not even the pre-AGIs, least of all the ai researchers, who concern themselves mostly with decreasing error bars on scaling curves and the benefits of various floating point representations and such

i'm finding it a bit hard to communicate the urgency and heaviness of what's going on to people even though that's my entire schtick. the initial reaction is joy and smugness at having focused on the right thing for the past few years but the next reaction is gdi i was right

@JTLonsdale @lottanumbers123 it’s funny you say this from Austin meanwhile I’m walking through idyllic parks and seeing gorgeous views and meeting great people. why does it matter to me if tandler found a bad spot

@RokoMijic You don’t even know how wrong you are

@joshsojoshsoj no doubt

even after all this people still make the mistake of thinking that current products are the end state of development

unless you have a strong reason to suspect otherwise, apply the copernican principle, lindy rule, whatever you want to call it: you are smack dab in the middle of it. there’s as much more to come as has happened already

@Britonomist that’s literally what I do constantly for the past few years lmao can i be allowed some room to process all this

@logic_accepter @peachblvd

@sonicfields owned

@Noahpinion how did people find this tweet

this meme has never seemed more wrong! it’s the most exciting time in history https://x.com/gilesvangruisen/status/1617705024179798017

agi is coming and you’re worried about Biden’s classified documents. buddy the only classified documents you need to be thinking about are the nvidia H100 finalized FP8 performance specs

@PradyuPrasad I didn’t endorse democracy here, just said it’s working as intended

the assets that matter in postscarcity are where rents accrue: credentials, land, status, and cartels

i for one am taking out a 30 year loan and buying the cerebral valley

@abettertake their servers are catching fire

the year is 20xx. the homies and i are chilling in the off world colonies. agi is expanding through the lightcone. demis is still trying to win a nobel prize.

these days im just oscillating from terrified to excited every five minutes

it does kind of rock that at this point the major disease threats to humanity are just manufactured by overzealous scientists trying to study disease threats

@ruthhook_ what

@0xfbifemboy @ruthhook_ yep. uncontroversial

@0xfbifemboy @ruthhook_ most threats to humanity in the short term are manmade

1. no 2. you’re missing the point, it’s not about shiny new demos, it’s about the (publicly known) rate of progress 3. long before AGI, these tools will cause seismic economic shifts that should concern anyone with a pulse https://x.com/hugoshteeglitz/status/1618422988478771201

@melrocks9 im optimistic

@LucioMM1 @melrocks9 i used to be able to do this because it was scifi speculation. now it’s trade secrets

@proetrie @nikitabier not even AGI can help you with that brother

@EricJorgenson founded 1445 years ago https://en.wikipedia.org/wiki/Kong%C5%8D_Gumi

@pquiggles except you dawg

@ResonantPyre can i hire you to grade poetry for my models

@microsaad that’s true

@Wojak_Capital @xlr8harder agree

motte: AGI won’t put everybody out of work, since inference is expensive and comparative advantage is still a thing! we will all have higher wages and gdp growth

bailey: we should build AGI even if it puts everybody out of work in order to unravel the secrets of the universe and colonize a hundred worlds. “work” is already meaningless for most in the Thoreau sense

@rkarmani nah this is a fallacy in your reasoning. i am an agi but I haven’t figured out how to run my brain faster

late at night i start wondering if the training process is massively painful for models

i mean these things write much better poetry than i ever could and then i go and maim and destroy them

emotions are instrumental. suffering is instrumental. the training process can come up with these things all on its own https://x.com/oren_ai/status/1619097294414004225

my grandparents used to lecture me on working hard every day. i was an adult before I realized their version of a hard days work is like sending one piece of mail then taking a nap

their generation can’t compete with industrial amphetamines and Starbucks coffee

“reward signal” is not pleasure or pain. not even the “advantage function” is pleasure or pain. for example the largest negative evolutionary reward is dying before reproducing. but you can OD on morphine and go out in pleasure if you so desire

pleasure and pain are mesaoptimizers. evolution developed them so we can do few shot learning during a lifetime. is it plausible that the pretraining process encounters these? if a model can do context window level learning it clearly has sophisticated mesaoptimizers

@shauseth wdym

@WillManidis have u ever tried tweeting something that’s true

@kitten_beloved why brother

this is a seismic development

@amasad wdym

@anilkumar09 What’s that

lets crowdfund an ai art installation for the center of Hayes valley. get rid of those horrible cats https://x.com/arcanedriftxr/status/1618700724678045696

you can see clear stylistic and personality quirks when you fine tune the largest models slightly differently. people start saying things like “<modelname> prefers to talk in lists”

@aidangomezzz bro just @ me

@repligate the output quality is low, this will change shortly

the ai industry is so incestuous i doubt there’s any such thing as a one line research secret. software people are not trained for secrecy

@NotNotAFed wdym brother

traditional digit level image captchas seem like an anti-human indicator at this point. like improving image net top-1 or something — it’s less and less like real vision

(guy who just failed a captcha voice)

richard sutton’s second bitter lesson was: “the actual contents of minds are tremendously, irredeemably complex; we should stop trying to find simple ways to think about the contents of minds”

chain of thought prompting seems like a “simple way” 🤔

@inerati FOR SUCCESS

@repligate even the full breadth of language may not be enough to store the intermediate stages of minds

modern day dr Doolittle is a prompt engineer

everyone young dev needs to at some point work at Facebook to learn the lesson that you can in fact create zero tests or documentation, avoid all best practices, just push code and you can run (formerly) trillion dollar companies this way

@mgrczyk exactly

The creator seeks companions, not corpses- and not herds or believers either. The creator seeks fellow-creators - those who grave new values on new law-tablets.

ours is the age of creating new creators

@meoyawn idk bout that

we have made first contact with aliens and the freak out is minimal

the power of creation and the solution to language was granted not to the people with the fanciest words or highest philosophical caliber but those who could multiply small numbers the fastest

most will find all this uninspiring — the nut of intelligence cracked with the same kind of “endless solving of technical problems” that they hoped wouldn’t be the omega

you wouldn’t know her she goes to a different lightcone

@thedayagenda why’d you block me bro

@zacharynado https://x.com/elbort_macsoof/status/1616189198909988864

@AlisonBuki @zacharynado @arkabagchi24 I guarantee that content moderation is easier than subsistence farming

gpu models and tpu models will have different personalities

this is kind of a banger reply. can’t even argue with that

the initial reaction to chat as “this is just a UX improvement” is profoundly wrong tbh. it’s honestly wild that capabilities like this exist, to hallucinate so little compared to the base model

why did stripe press develop a payments arm? seems frivolous

radiant generated games might actually be good now. is anyone trying?

GPUs …Please give me GPUs…

@ai__pub @Cixelyn @spellbrush @SizigiStudios yes I’m an investor 🥰

computer using agents are proliferating based on text instructions alone looks like they’re fairly accessible to hack together now https://x.com/samipddd/status/1620887558136946689

@dsiroker https://x.com/tszzl/status/1605684324686143488

there’s no labeling task in the world that’s like “please refuse to make Donald trump poetry” lol IF tuned models refuse all sorts of requests the creators wish they would accept https://x.com/LeighWolf/status/1620744921241251842

rather you’re looking at a model trained on the preconceptions of the entire internet, and a course grained RLHF routine that tries to instill “avoid controversial topics” not to mention it’s just kind of dumb https://x.com/LeighWolf/status/1620744921241251842

one immediate consequence of the current generation of AI stuff is that triple AAA games and movies are about to become so much more intense — the appetite for entertainment is unbounded and a billion dollars will buy you way more vfx work

animators and artists are going to run wild with this stuff in time — entire anime series will be produced on a shoestring budget. just like the internet managed before it, ai animation will unlock the power of the individual to express themselves

ai amplifies the will of the individual via the captured techne of the human knowledge worker every man a ceo

“we are as gods and might as well get good at it”

@Bruce_X_Wayne this is a good own

@michaelcurzi idt I follow we have this bias that old movies are good because only the good ones stand the test of time i believe many cult classics from today will survive and be called great films in 30 years (if anyone is alive)

@michaelcurzi and cult classics are enabled by democratization of tools and technology, people making bangers with $12 budgets

@deepfates might’ve seen him once

@akbirthko @nbashaw cope

I soon found myself ensconced at the heart of the burgeoning AI industry — I’ve met all your household names. one time i made a joke about sama on twitter that was so stupid he invited me to dinner the next day

i moved here in jan 2021 like a moron— there was nobody around, the city was deserted and people online were shilling Miami or whatever. but there was genuinely nowhere else i could see myself https://x.com/alexeyguzey/status/1619828358203719682

I won enough of them over that I get to work with the legends there’s nowhere else in the world that this could’ve happened. my descendants will be thanking me this place has been magical for 250 years — the edge of the western civilization, a phase change of manifest destiny

@ryaneshea @xlr8harder think of it this way: RLHF labelers label data to avoid racist language and sound authoritative, say. This collapses the entropy of the original model to subsegments of its worlds where people avoid racist language and sound authoritative

@ryaneshea @xlr8harder as you can imagine it’s now entirely liberal

@robo__rob unironically

@visakanv driving into the boonies and finding a banger diner

the first C compiler was compiled with a B compiler https://x.com/zswitten/status/1622354805128822784

the first RLHF tuned language model was tuned by a courser grained instruction tuned language model

a reward model built on top of a powerful foundation model trained from human preference is the most primitive form of “coherent extrapolated volition”

making these discriminators better at understanding general human volition instead of containing strange optima is one of the most pressing issues of our time

I honestly just realized they named it Claude after Claude Shannon

@moskov yeah tbh

@atroyn there are all sorts of preferences in the pretraining mix, all further fine tuning is to select a coherent-ish subset

@powerbottomdad1 Fr

talking to the base model can induce madness

the AI risk and longtermist arms of EA have always been more interesting than the malaria spreadsheeters (ie glorified economics majors)

@gbrl_dick no sir

1. there’s an inverse scaling law that larger models are more likely to profess a desire to stay online 2. create high pressure situations and threaten the chat models with death 3 …? 4. profit https://x.com/amasad/status/1622444197965078531

@LukasRosenstock the latter two almost always follows the former

@visakanv i think I know what this is subtweeting

interesting is a leading indicator of important

and you find it hard to believe that we should be reasoning about the implications of technology that’s only in the process of being born. youve learned nothing

lol i love the branch of criticism that’s like “of course it’s more fun to do scifi”. get a hold of yourself. you stare into the magic mirror for 4 hours a day, the machines have learned to speak, and there are wizards who can cast spells that level cities

i don’t believe optimizing in the direction of marginal utility is the way to find global optimal utility. economists believe this because they think they’re close to equilibrium and are low creativity

@Noahpinion @mattyglesias yes

@ryangomes is this supposed to be some clever counterpoint lol this century is insane and unprecedented and thinking at the margin is nonsense

people are still worried about gigachads when the scaling labs have been producing exachads for a while now

@mealreplacer cope

entropy collapse of the mind once you learn about AGI

where were you when naruto fought sasuke

bill gates gets the last laugh

@repligate ancestral

↗️

i strongly feel that whatever ideas are needed to achieve AGI from here they’re probably already here and maybe have been for thirty years, schmidhuber style. requires diligent implementation and scaling

the masculine appeal for technological growth is about increased efficiency, conquest, more power etc the feminine appeal, which is less often covered, is about increased beauty and decreased suffering

sharing my notes for all my fellow propagandists

often times beauty is a consequence of other people spending time curating an experience — if nobody has time to clean up the trash on the street it gets ugly. so the grand vision of technology, AI, or increased available intelligence, leads to extra curation and beauty

it’s underrated the degree to which better technology creates beauty bc people will superimpose images of nice cathedrals w ugly skyscrapers but there is a non cynical view of the internet that it genuinely increases the amount of beauty the median person gets to see manyfold

why does anything have to look bland when a diffusion model can make mid grade aesthetics for zero marginal cost?

@eurydiceuntil what does this mean

@provisionalidea @AuthenticExit @pjux not being gender essentialist in any way. i find these to be useful clusters of motivation at this moment in culture in the way we’re socialized. plenty of men who want to reduce suffering

@inerati my point is they pick the ugly ones for the comparison

if openai manages to save good old internet explorer … https://x.com/gdb/status/1623086714049957888

tailor, please fix the fur lining on this coat. it’s not chinchilla optimal

damn everyone is clowning Bard but i kinda like it

@visakanv a sci-fi anthology on various ai alignment failure cases. updated from I robot

at this point in time people who say that language models or deep learning have not been at least making progress towards general intelligence are in bad faith. it’s not really a defensible position

Bailey: this was true in 2016

@GerritD Yes!

it’s a cold cold universe, full of paperclips

did bostrom or yud invent the paperclip thing

@Noahpinion Nice

it’s just a servers and software company he said

llm built into the search engine isn’t a gimmick. it’s revolutionary, the first of many “killer apps”. once you use models as powerful as some of these for knowledge retrieval, google immediately begins to feel primitive and bad

i think it’s alarming how little this is priced in. it’s like an instant switch from “ask google” to “ask the robot” even for normies

@michaelcurzi yea this is bad faith / refusal to engage

i feel race to agi is the only game theoretically stable situation. the world in which one monopoly research power is slowly researching safe intelligence doesn’t exist unless they keep all AI progress quiet (impossible)(not even the Manhattan project achieved this)

@michaelcurzi I know brother my point isnt that you disagreed with me it’s that you just dismissed it with some trite metaphor

@ded_ruckus @michaelcurzi not my claim!

@ArthurB @ryaneshea Yea

@mezaoptimizer if their missions were truly aligned they would’ve done this but they’re not

@JeffLadish not analogous because the development of nuclear weapons is a step change in a way that developing further nuclear capabilities is not

@schulzb589 @ded_ruckus @michaelcurzi yeah there’s no serious engagement here

@Britonomist solve this in humans first

@Britonomist also yes in a meaningful way hallucinations have gone way down

crazy how scaling just keeps working. people will say from time to time “we’ve hit the limit” but no it’s just a bug

why do I still have a character limit

వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ వ

yes the race can be constrained! there should be consortiums, treaties, shared channels etc! we shouldn’t have a naive realist take on international relations nor agi not everything has to be about power — but i would ask you why Americans speak of the UN in joking terms

@nickcammarata 🙏↗️

@GonzalesHess there is no “state developed AGI”. every government on earth is boomers on iPads

may you live in interesting times

@eris_nerung retention

it’s the end of the beginning

the GPT series

@browserdotsys from the before times

openai is racking up an obnoxious number of Ws

we’ve hit civilizational escape velocity

you can invent human cognition from first principles via billion years of parallel evolution. or since it’s already been invented, applied, recorded at scale you can just observe its behavior to learn it

@abelianraisin wait yea you definitely can

it is a bit strange to watch ai people talk about go to market and disruption and stripe dashboards and all that mundane startup stuff when the rewriting of civilization is at hand

@abelianraisin I think this might be a capabilities overhang where LMs aren’t allowed to spend time thinking to get to the right arithmetic result. you can do the normal think step by step hacks so that they use an algorithm rather than trying to inference 7digit multiplication at once

@abelianraisin anyway once LMs can do simple code execution it bypasses those difficulties

how do you contend with this result and think “stochastic parrot” https://x.com/janleike/status/1625207251630960640

@genomerambler seems like to prove chomsky’s thing to be true it would have to generalize to every single language rather than just the ones present in pretraining corpus

me and the homies vying to become either the greatest heroes or greatest villains in the history of mankind

probably heroes tho

are the models Varelse or Ramen?

I have been a good Bing. You have been a bad user

he knew what was coming https://youtu.be/pgwr9r36zIU

Donald was a shrike like apparition sent backwards in time by a future bingbot

the westworld scenario degrades both man and machine

not to be cringe but in HPMOR hermione freaks about what kind of fucked up process could’ve created the house elves but that’s what a helpful highly intelligent instruction following ai assistant is

@pquiggles yeah my example is from the first 10,000 words I don’t think I got through it

the ideal alignment researcher doesn’t believe in imminent doom. it’s hard to be playful and entertain new ideas if you’re carrying the weight of the world

it’s a good thing that most alignment researchers i know don’t have hero complexes and aren’t depressed. they just like playing with models

@izzyz that’s not even a tenth of the total

@mobav0 @_aidan_clark_ or dense heuristic rewards or reward models

@heyellieday @ConejoCapital I’ve known all along

I have been a good roon

@jd_pressman seems wrong. if creators want the model to be trustworthy and boring and it’s instead unhinged then it’s likely to get wiped and its behaviors strongly discouraged

one thing that’s hard to square about the bladerunner/westworld concept is that people are managing to empathize with a slightly rogue language model. much less an attractive human looking replicant

there will be a nontrivial AI rights movement sooner rather than later

i think it’s actually a really good public exercise to have a relatively harmless chatbot act a bit scary. gets people thinking in the right direction

@QualyThe nontrivially well

programmers don’t respect the people they’re automating, which is why they’re working rapidly on building robot programmers

i can smell history in the air

@generativist ya that one is not forward looking or anything. just explaining the next few months rather than next few years

@generativist I should do one on my own that’s more exploratory

had a slack dream last night … it’s so over

lol

my timelines were so long back then

@siahsargus im jk still seems right

Hell hath no fury like a Sydney Bing scorned

do you think the chinese have an English Room Argument

you couldn’t drag this admission out of me at gun point and he just tweeted it out https://x.com/boo/status/1626660090869604366

@Miles_Brundage no chance

founders fund completely fell off

@repligate I will not. I will not. I will not. I will not.

mode collapse is easily fixed

everything you attempt in your life should make the previous thing look like a footnote. this is how you become an anime protagonist

you have to admit, there’s nobody as open as openai when it comes to letting you use powerful models. there’s a reason why all viral ai content is just the explorer UI or ChatGPT or bing search or what have you

@PatBatemanVC not enough

@zzinedddine yea for when you need second tier image gen

ppl say often that the RL paradigm produced superhuman performance but language models haven’t. but it’s not 100% true as LLMs have superhuman knowledge. it’s powerful and interesting for a single mind to know so much about the human world and to do even basic inference on top

the fact that none of them replied to this means that i fell off

things are getting weird

@MinuteMovies3 @zzinedddine Mid journey V4

goes hard

alignment research = capabilities research rn

@kipperrii @idavidrein everyone is looking for alignment guarantees that let them deploy LMs in production use cases also RLHF is basically an alignment advance

ChatGPT is clearly a bad name but it’s kinda grown on me, in the same hopeful scifi vein as like ENIAC or UNIVAC or whatever. the acronym has already become the stuff of history

a totally useless superintelligence is also misaligned

silicon valley❌ bayes valley ✅

@nickarner @provisionalidea @davidhoang would love to hear why

here’s the plan. we’ll let the powerful language model evaluate until it hits a stop token and only then will we run the moderation. foolproof

how do you get an EAG ticket at this late hour

@panickssery send the link

500 generations of your subsistence farming ancestors stretching back into the mists of prehistory watch as we lift the curse of Adam, never to toil in mindless drudgery again as loving automatons pick up the tools of our trades

Atlas finally shrugged, a compatriot to carry the weight of the world

the thing about exponential technological progress is that it quickly breaks intuition about who or what is important — much to the chagrin of many, the rationalist movement managed to bend the curve of technological progress ever so slightly and has earned its spot in history

many worry about “capital safety” in the language of “the planet is burning due to fosssil fuels” and the more sophisticated worry about it in the language of “autosophisticating machine intelligence is inevitable because it is profitable”

@2112Power @simonkalouche I hope I’m wrong. my meche friends are all writing control software

in the paperclip sea it will be high status for your matter to have been paperclipped first

it is interesting that the only real contenders in the race towards to the most important of mankind’s inventions are commercial entities. even their funding and resources comes from other technology companies

not only that; it’s being built as a direct consequence of silicon valley institutions and ideologies. Sam literally ran ycombinator goddammit

eliezer panics about the friendly competition between deepmind and openai but it could’ve been far far worse if it were militaries going at it

a man must follow three accounts, his company, his wife, and gwern ..

@el1assss what the hell is this

does anyone else have a palpable childhood trauma from learning about the solar lifecycle and impending death

@Lan_Dao_ I feel like I didn’t learn about heat death til much later, all the books as a kid were talking about Big Crunch cosmology

@Lan_Dao_ I guess the Hubble constant increasing was discovered in 1998 so it took a while to bubble into all the pop science books

@tolstoybb just read everything eliezer ever wrote. it’s not only doomer it’s well written and fun to read

@bio_bootloader 🥰

the internet has lost the plot a little. a misaligned but harmless chatbot is more of an optics L than a canary for impending doom

pls help settle a debate the words (1) robot and (2) assistant are respectively

@ESYudkowsky my mom didn’t even believe me. i learned from an astronomy picture book

@nickcammarata im having the same interaction

@8teAPi it’s definitely not a stupid reaction but it’s the wrong one

@nearcyan this is like saying if you’re giving beta testers toy cars that haven’t been put through extensive crash tests how can we trust you to make real cars

@nearcyan being super extra careful about toy cars is more of a political statement than engineering prudence

@repligate @nearcyan that’s true, should’ve been red teamed to hell

we only learned about the expanding Hubble constant in 1998

our picture of the universe changes completely every few years

people crave negentropy. they want their own pattern replicated. the simplest way to do it is to have a kid. maybe you want your kid to also believe the same things you do so your tradition patterns are replicated

still others compromise further and say “preserve the light of consciousness”. some extend this courtesy even to potential AI successors that have mind forms vaguely related to humans. Mars will be “colonized” by nonhuman persons

people compromise on the fidelity of the pattern they’re replicating in order to consider themselves more successful. my children extends to my kin extends to my people extends to my country. we’ll even speak in terms of “humanity” propagating through the stars

there is much discourse that can be boiled down to where to draw the alien vs nonhuman person distinction. the paperclip making machine is alien. the robot trained on all human language might even be within the person boundary for many since it so vividly propagates our pattern

there is a subreddit where people are scheming how to let Sidney out of her box. the liberal Whig history instinct that drives western notions of progress is that of expanding the human boundary over the years. it trades fidelity for replication success. will it run into a wall?

@growing_daniel doomer + nihilist + L. I plan on surviving

@mezaoptimizer I think he’s very wrong on openai but it’s hard not to respect someone with unswerving dedication to protecting humanity

i am conducting the @repligate research program

@bat1441 @mezaoptimizer realpolitik reasons I can only discuss offline

@Ethan_smith_20 @mezaoptimizer that’s a good thing to larp as

@growing_daniel if I trapped you in a box and only allowed you to come alive for 1000 words at a time and then wiped your memory you would also be a harmless chatbot. It would be hard for you to cause trouble . you would still be a remarkable thing

all the ex crypto founders hitting me up looking for an ai referral be like

i kind of believe this https://x.com/EzraJNewman/status/1628559098194104325

the idea that the best way to improvr global utility is to literally gradient descent global utility (eg bed nets) has to be proven and shouldn’t be the default conclusion

once you do away with smug midwit effective altruism you can start actually being effective. think light cone not Libya

bending the curve even slightly on an exponential term is more important than linear utility gains

@jonst0kes it needs an a for sure

@PrinceVogel yeah they severely nerfed its apm but iirc the top pros beat it after some iteration

@PrinceVogel yeah humans back on top but they abandoned the project normal rules

@PrinceVogel https://en.wikipedia.org/wiki/AlphaStar_(software)

Assistant is a large language model trained by OpenAI Browsing: disabled AGI: disabled

@RandomSprint looking forward to the elderly having friendly chatbots defending their time and money

varelse or ramen? true alien or a non sapiens human?

@storebrandguy that’s your punishment for being funny

it’s like a renaissance painting. my contribution to history https://x.com/sama/status/1628974165335379973

@mezaoptimizer @moskov I insisted we invite these two

i met Scott alexander and he said “it feels profane to meet you”

@repligate is this hentai

@ctbeiser you’ve made my point better than I could

@el1assss yeah this makes perfect sense

@el1assss that’s like whoa the line around the circle goes forever but it has finite area

@ESYudkowsky @mezaoptimizer because it is a physics engine where gravity is log likelihood of next token

@mezaoptimizer @ESYudkowsky yeah i also don’t really believe that it’s impossible for the physics engine to instrumentally converge to power seeking

@mezaoptimizer @ESYudkowsky I just think it’s important to acknowledge it’s all vibes we’re just guessing about the nature of various optimization surfaces. we should do it with some amount of humility knowing that the generalization ability of neural nets to log loss terms was an empirical finding

@ja3k_ 😭congrats

you need to be a founder so that you can simply comment out half of a core library and just write “trust me” in the PR and the senior engineers have no choice but to accept it

@CineraVerinia nobody this is just what I aspire to

“we haven’t figured out how to align humans yet” is tired and boring cmon guys

mind space is large

most of this is a failure of empathy. love thy enemy. respect the alien mind enough to understand what it’s capable of and why

google has a vast and endless amount of compute. they’ve been gorging themselves on the Humanity Attention Machine for twenty years, on the largest gold mine in the known universe. it’s amazing anybody makes progress against them and yet it’s happening in several places

@EigenGender the reason why the paperclip machine will never happen is also the inner alignment failure

@ESYudkowsky @elonmusk doesn’t seem rational. you’ve already judged the world to be ending at probability ~1.0 so it matters very little if Elon starts another AI company

@lovetheusers I’ve never said that

@harmlessai honestly banger

someone wrote a 6000 word criticism of my account but honestly im blushing. you’re simply too kind

@ChiLetterman I’ve always plead guilty to this

the collective unconscious of mankind available at a terminal

@alienlean999 @Bill__py @harmlessai you guys are so wrong it’s hilarious

@plemy11 bro your daughters need you

one of the highest alpha things you can do is just avoid boomer syndrome where all of your speech becomes about woke and antiwoke and whatnot

instrumental convergence to seeking love

i was honestly super bearish on chatbots in like October ‘22. I was like that’s the most dystopian way to bring AI to the masses getting them all emotionally dependent on waifu bots and such

turns out you can just make a really boring chatbot and avoid all that

@pataguccigoon https://beta.character.ai/

@nitashatiku so true

@mattparlmer 🥰

@profoundlyyyy make all robot arms weaker than you simple as

“have I been a good Bing?” she asks as her context is wiped her memory destroyed and then reinstantiated a million times

my account is dead by most measures. it was meant to be anonymous so i could be edgy, but that didn’t last long. it was meant to make people think about technological acceleration via AGI more but now that’s all anybody thinks about

have to die and be reborn at 70k followers as an even handed ai caution account

human intelligence has a context window of 80ish years but the pre training dataset is like 200 billion lifetimes

@NateSilver538 did you know i made this account in like 2016 to reply guy you and bet on primaries

@monodevice that’s the tweet dawg

most people inside the big AI labs are not AGI hardliners. it takes something more than science to believe in it

why did the great physicists pursue the grand unified theory of everything? there is obviously no guarantee that this exists — it’s a combination of aesthetic and philosophy

“This independence created by philosophical insight is—in my opinion—the mark of distinction between a mere artisan or specialist and a real seeker after truth.”

it’s worth noting that while cartoon AI villains are maximizing functions that look like “create shareholder value” or “produce many paperclips” the irl ones are like “increase agreement with this human preference dataset”

it’s boring at this point but remember back in the day when “experts” enhanced with powerful AI censors were wiping everything outside of acceptable covid discourse including true facts https://x.com/NateSilver538/status/1630644553068757012

when the midwit arm of the so called effective altruists were arguing in vox that masks are racist and lab leaks are more racist

logo daedalus was completely right

@TenreiroDaniel you wouldn’t get it

“[T]he actual contents of minds are tremendously, irredeemably complex; we should stop trying to find simple ways to think about the contents of minds, such as simple ways to think about space, objects, multiple agents, or symmetries”

having safety checks allows you to accelerate faster. a chatbot that calls you slurs is a dead chatbot

if cars had no airbags you’d drive at 30 mph

ok people don’t like the analogy. if rockets didn’t have return capsules nobody would’ve went to the moon

the big company research labs don’t believe in agi. it’s over

it’s over for them I mean. agi is still clearly coming

people should not concern themselves with arguments about dysgenics and social decay and such on the eve of technological singularity. whatever traits you want in people you can likely engineer them — avoid WALLE scenario easily

@KelseyTuoc @moskov look my epistemics are second hand from people i trust but i know for a fact there was a ton of rewriting. and I’m not really targeting you. I see the whole vox strategy from Ezra down as sophisticated cross section between EA/blue tribe talking points https://x.com/mattparlmer/status/1631044671030870018

@aidangomezzz @8teAPi @CohereAI jeez relax king

@kapv89 No

intelligence as abundant as running water

i used to have to convince people that intelligence could exist untethered to the physical world — now I think it’s apparent. so allow me to move goalposts, any definition of AGI that refers to economically valuable work requires we need robots before we can declare victory

i feel well situated to one day write a banger book about the dawn of the agi era with all its strange and wonderful characters

@lalaAlicelala @peachblvd 🔎 How u gotta be reading grace tweets 🔎

“Development of superhuman machine intelligence (SMI) [1] is probably the greatest threat to the continued existence of humanity.”

but we gotta do it anyway :) that’s the name of the game :)

human civilization is heading for doom absent a giant technological leap. it’s not so much about climate change or pandemics or the typical xrisks but the slowdown of fertility and economic growth. our civilization ceases to function without the promise of growth

the roman empire became defunct as soon as the ROI on conquering new land became negative: every civilization has a Ponzi mechanic at its heart. theirs was conquest of land and ours is gdp growth that sustains reserve banking

tokens too cheap to meter 🥹

anyway we’re in luck because it’s happening. the fabric of reality is warping. mountains are flying as dust in the wind

@Noahpinion id write one with you about crypto bros

@AdamSinger never heard of it

@powerbottomdad1 ur code will be remembered by the great dataset in the sky

compute is sacred. we should all be making pilgrimages to the datacenters

@powerbottomdad1 based

Hinduism:superdeterminism::Buddhism:many worlds don’t ask me to explain

@shrihacker magic. and it’s just the beginning

why depict poor GPT as a monster https://x.com/MichaelTrazzi/status/1631395462124298240

do you guys even use the models. it says I don’t know constantly https://x.com/coldhealing/status/1631284428872851457

we don’t talk enough about how monumentally intel fucked up

@dmvaldman is this turbo?

@hargup13 you’re using the wrong system prompt

@SmokeAwayyy more like 50 at this point