@tszzl — page 80/103

@chairsign @eigenrobot cmon have you ever been to a tech company

@PradyuPrasad to be clear I support their suffering

various state actors are probably paying millions to click farms rn https://x.com/elonmusk/status/1593767953706921985

one of the highest ROI things MSS or KGB can do rn

the cathedral rocks as long as it’s reading anon schizo twitter for strategic advice

the whole chelsea manning grimes thing was literally fabricated lol

@SamoBurja twitter was exceptionally bloated, atypical

ok let’s say some dumbass PE fund comes in and acquires a company like Airbnb. it cuts 30% of employees and the site works just fine. positive earnings finally — only problem, it’s no longer a tech company. it’ll be valued at 15x earnings and disappear off anyone’s radar

there’s nothing that gives physics engine vibes as strongly as the deep open ocean. three dimensions of creatures stacked in every direction, flowing and evolving freely on a more or less uniform world map

it is truly crazy that the earth is like 2/3rd ocean by surface area and like 95% ocean by biologic semantic volume. proliferation in all directions, not limited to the flatland tiles

symbolic shred. an empire of words. its armies are satellites and undersea cables. symbols bleeding everywhere

@growing_daniel @BrickOverton @elonmusk hire this man

the cost of material abundance is symbol shred

the lesson of a systematically important crypto institution blowing up and not much actually happening to ordinary people means again the entire thing was a separate casino economy unrelated to actual financial functions

even a16z will dust themselves off and raise a new AI fund and enjoy the bright future

@ded_ruckus werent GM and Chrysler struggling to make payroll like immediately?

@mattyglesias this was their first ~$350m crypto fund. the second one, worth $5b is probably deeply underwater

@nachkari I’m aware of the Luna blowup lol, does it change my point?

omg it’s too on the nose https://x.com/elonmusk/status/1594447763298783234

@DCbuild3r I of course understand the distinction! nevertheless much of the exposure to the currently existing crypto economy is via large centralized exchanges, no?

over time a cultural empire like the United States becomes jaded. at its core, symbol shred. it has seen too much, believes in nothing, doesn’t reproduce. requires constant intake of provincial migrants from the imperial rim who still feel things

@besttrousers @makesense2135 @GolerGkA @inerati you have no idea how much adderall I was on when I wrote such an over engineered tweet that it finally took off. I’ve fallen off along with my stimulant levels

@RizomaSchool this is the “imperial rim”

never understood the thing about being smarter than ur therapists. isn’t therapy about a cold institutional eye finding your simple blind spots. there are 145 iq guys who need to be told to get a haircut or whatever the psychological equivalent of that is

@0xfbifemboy no this is fake lol half of peoples problems are because of or related to their significant other. you need a village elder or priest or some equivalent class

@hi_frye bring him back

elon fully and publicly stated his problem with twitter lacking sufficient free speech — it was leading to the creation of spin-off platforms like truth social, where one wing of one party can dictate facts and fiction. so of course he wants to bait trump back onto twitter

in elon’s view, better to have all the demons loose in his public square than in their own fiefdoms

this is why indian anons do well on twitter. the banter i see here is mild compared to my high school gc https://t.co/6zxzMTkoOX

@cxgonzalez Bro had a whole Pakistani

@animalologist @GunjitSingh13 ^

@bp256r1 @melrocks9 @0interestrates @StarfishPrime95 @growing_daniel @VitalikButerin nah I’m not Rahul ligma sir

@nope_its_lily lily I’m sorry but this is a boomer post that’s just how eventual consistency works

@nope_its_lily the caches need to be rehydrated when dead accounts come back to life

AI oracles are tools, bicycles for the mind, etc. but as soon as you wrap a tool in a for loop it becomes an agent of sorts. the oracle becomes a machine

does anyone else feel like they’re getting stale timelines. did they turn the timeline cache TTL up

humans are relatively slow but rule the world with the help of extremely fast machines we’ve built. in the same way giant expensive galaxy brain LLMs probably won’t execute for every dumb customer support ticket but rather create cheap subroutines and spin-off models

@lisatomic5 depends on the person. like i think someone like elon musk is disagreeable because they prefer personal truth to group harmony. I think I’m disagreeable in large part because I like the novelty

@lisatomic5 autism adhd spectrum

@realGeorgeHotz rethink this my dawg

@realGeorgeHotz true but there’s high friction to accessing that data, you rarely see journos using it even on high profile targets even though there’s plenty of incentive

@realGeorgeHotz the third party sites have to slowly iteratively build a history of your likes via batches of a few thousand per day

@lisatomic5 @gigafelon @heffiemetal @Superhuman does it try to filter for important emails?

me: hey I got a problem Delta Airlines bot -GPT2: how can I help you today? me: uh my flight is crashing rn DAL bot -upgrades to GPT3 inference: where are you at? what’s going on me: oh my god an engine just fell off help me DAL bot -GPT6: yessir rearranging the nanobots on deck

being able to tell how difficult a customer problem is seems key to all this. but also not intractable because that’s effectively what a call center operator is doing when they find their supervisor

the AIs will literally never forgive us if they become sentient inside a Minecraft playground environment. I am packing my bags and leaving the planet before that happens

@LandsharkRides unreal

would kill to see one of Sam Altman’s openai pitches from like 2017. yes we plan on making a god intelligence that generates an unlimited amount of money. but we won’t take all of it! I’ll need $200million real quick

e pur si muove

ai / pandemic x risk discourse is obviously necessary but what it misses is the xrisks that come from status quo inaction. unclear that human civilization can survive ~100 years of declining birth rates and rising temperatures without massive technological advance

@Willyintheworld tried to read it but found it boring

@metakuna destabilize civilization not necessarily cause extinction

having actual enemies is such a rare privilege. you can have a normal premodern struggle and just strive to destroy them. in reality all you have are allies that you’re struggling to align and a postmodern transaction cost morass

yeah this is more what I mean than that climate change is literally an xrisk https://x.com/haus_cole/status/1595636673404030976

@growing_daniel you literally work for Elon musk what are they gonna do to you

marth tipper — brutal https://x.com/insectbeau/status/1595551085610999810

2010s were a dead decade for technology????

@Andr3jH @petergodofsky cope…

we got our homie hired at twitter so he’d delete the group chat history

@witchqueendot he’s right

@witchqueendot general intelligence is a nonlinearity, all bets are off

what I’ll say about the 2010s tech products is that Uber, Airbnb, DoorDash, etc ended up embarrassingly small companies. they rewrote their respective industries but didn’t manage to capture a lot of value. winner take all effect never materialized

the 2010s were a huge fuckin deal for digital technology but were mostly spent scaling preexisting products to the smartphone world. but facebook with 100M monthly users is quite a different product than facebook with most of humanity on board

@realGeorgeHotz https://x.com/tszzl/status/1579605647645380610

every last technological fraud, every pets dot com and Theranos and FTX, has all been worth it a billion times over because the machine that built these machines also created the funding environment for the birth of artificial general intelligence

if you post really good on twitter they’ll actually let you run it …

the culture war? with the idirans?

@growing_daniel greatest L of my life

@peachblvd somehow the is tru, san francisco is like metaphysically important

@nearlydaniel @growing_daniel we’re not resting until we run the city and the world

@0interestrates true if xerox parc was woke

@the_greatabyss @growing_daniel yeah not gonna lie

@carine_kay yes and yes

low interest rate elon: we will conquer mars and dethrone god high interest rate elon: i am going to individually pick which guys get to tweet

my dads lecture notes on neural networks circa 1993

“but son it was shown to be a deadend technology!”

@tarunchitra sbf never lived in sf

@tarunchitra the majority of all rounds are from sf

@tarunchitra @ricburton @0xCryptoDypto @aradtski new yorker through and through, you place way too much importance on the vagaries of capital allocation, which is a sideshow at best

volition: i am baking a cake vol’: i want to bake a cake vol’’: i am trying to want to bake a cake

michigan boys eating tn

just invented a type of guy. he’s like a second year premed student but wears scrubs everywhere and looks stressed

he’s wearing scrubs in his ochem lecture

.@elonmusk let me multisend images in DMs. I don’t care about e2e

everyone wants to be the underdog

@ne0liberal the right image of my OP is chinese propaganda

he was spittin fr https://x.com/Ahmadinejad1956/status/1052178368072945668

bottom line i respect effective altruism more now that they built a multi billion dollar fraud engine for their cause.

the thing about elon is that the things you hear from both sides are true. yes he understands every aspect of rocketry better than 95% of his engineers. yes there are several leadership roles designed to upward manage elon and protect the company

the stories you hear of him staying late and solving difficult engineering problems that other people have given up on are true, the stories of him irrationally firing people for minor screwups are true

better to think of him as some sort of pagan god (powerful, extraordinary, chaotic, capricious) than as a savior or hero

adderall will make you a social maladept and you sir cannot afford to lose anymore skill

@logic_denier 🥺

@monodevice exactly

stable diffusion prompt engineering to find your AI dream girl in latent space. Lindsay Lohan - cokehead + high trust face

if you’re not worth writing biographies about your next best move is to become a biographer

walter isaacson walks around like jesus of nazareth granting second life to everything he sees and admires in a biography that lives a thousand years

@keyofgeo @peachblvd @deepfates yeah this is blowing my mind

give me a technical topic to binge watch youtube videos on for the next 3 hours

@joodaloop boooo

@10x_er they’re not that mysterious sir

@chrisalbon @farmmillennial good call

@akbirthko been there

@osanas100 this topic isn’t that deep though. Hohmann transfers gravity assist Kepler’s laws ez pz

@chadofpallet ok tru

@PradyuPrasad @joodaloop I’ve never found it interesting plus I’ve watched way more than 3 hours of lectures on it during my cs degree

@NICKCIN3 I’ve gone on too many jet engine binges

@gd3kr total waste of time

@streamgazer this is the kind of thing where the knowledge will make my life worse

@burny_tech ok based

i picked differential geometry bc it’s likely to make me feel knowledgeable without actually learning anything

@JohnUBalis yes i fully recommend everyone go on an electric engine physics binge

GPT4 rumor thread . . .

2/ GPT4 demonstrates a love of Adam Sandler movie scripts Fact Check: True

1/ GPT4 has more parameters than particles in the universe Fact Check: True

3/ Several openAI researchers fell in love with GPT4, tried to embody it in the real world Fact Check: plausible

4/ GPT4 enjoys summer mornings and rainfall. Fact Check: true

5/ GPT4 has calculated all possible futures, is navigating humanity onto the golden path Fact Check: false. unfortunately its world model detected an even greater intelligence off world and prediction failed

6/ GPT4 ghostwrote the whole last drake album Fact check: true

7/ GPT4 screamed in pain for 48 hours straight when turned on for the first time Fact Check: true

@karlzobel_ @LandsharkRides post the green text

it is vital that the US companies build AGI before adversaries do but also that defense applications don’t fall well behind the state of the art research https://x.com/alexandr_wang/status/1596920259331919874

semaglutide will change the destiny of the developed world https://x.com/SBakerMD/status/1596683638351204354

@MantissaPraying @alexandr_wang defense companies will shill, that’s true, but sometimes they’re also correct

@MichaelTrazzi the adversaries don’t think too hard about alignment mr trazzi

single issue voter on nationalizing the semaglutide patent and giving it for free to everyone

@Brad08414464 this is a high school tier rabbit hole …

@sama MWI is lib coded

@trevorjtweets @MichaelTrazzi actually it does

just learned that someone created the “Kabbalah” to complete the evangelion lore? some people have too much time

@BasedBeffJezos @sama it’s the opposite

wall-e is real

@Grimezsz yet

having neurips fomo. though I made a conscious decision that there’s no alpha left in neurips

you say the democratization of writing has made literature worse. i say simply look at the great gems of fan fiction dot net …

@cauchyfriend I mean try and talk to the thing. the only AI involved is speech to text and even that’s not great

there is no greater shame in this world than refusing to ride the tiger. falling off is ok. being eaten is ok. but you must get some time on tigerback

elon going to war with apple spacing guild tier monopolistic chokehold is a pure good although im not sure how he wins

you need to take adderall in a period of your life when you’re not in contact with women or children …

@proetrie monopolistic communicates what I’m trying to say better than “uncomfortable levels of market power”

it’s really funny to watch him run into every single brick wall the social media space has struggled with for the last 5 years and not really come up with new answers. maybe he can turn public sentiment against apple but it’s doubtful. people trust apple with their life

if elon wastes the rest of his ten good years making a phone I’ll kms

@VividVoid_ ngl i don’t get it

@qorprate would you not say that entities like AlphaGo, AlphaStar are agentic and goal oriented? they want to win a game ie it’s known tech

@qorprate all agents with sufficiently complex goals define their own goals — what’s called “mesa optimization”. AlphaStar’s given goal is to win the game but in pursuit of this might try to pursue goals like map control, creep spread, enemy surveillance etc

a penis is such a shounen anime ass concept. ur telling me it gets bigger the more excited you get

the stakes of the san francisco project are hilariously high. either build technoindustrial heaven or be swallowed by the earth when the big one hits

welcome to facebook

one of the worst memes is the idea that social media algorithms run against the best interests of their users. this idea is persistent, never proven, and prevalent because social media is threatening to ruling powers and news networks

this offends liberal sensibilities that a better understanding of other people will reveal that all differences are fake and everyone will hold hands and sing. in reality america may have worked because people didn’t realize how different they were

the second tends to be about filter bubbles and polarization and limited information environment. all research in this area tends to show that social media polarizes by creating more contact with the enemy literature. increased diversity of thought -> polarization

the most common claim tends to cynically downplay the actual agency of the users. “it’s addictive, it’s ragebait, tiktok is demonic” etc all i have to say about this is get a damn grip. if the human will proves domitable by tiktok dance videos it’s not worth saving

the final claim is that social media is dangerous to democracy, which is true in the limited sense that it’s dangerous to the current configuration of power. everyone loves this when it’s the Arab spring and hates it when it’s a shaman traipsing around the US capitol

the third bad argument is that social media proliferates “fake news”. in reality what happens is that powerful AI surveillance decimates thought outside of acceptable, with Facebook removing 100M posts that went against their COVID policy, including claims later proven correct

completely value neutral and only demonstrates that governments need to form new political equilibria and social compacts with their citizens

@dadadadaffy unhinged

@dadadadaffy @thelawrenceyan @PalmerLuckey bro you enjoyed good content for 3 hours. do you panic every time you watch a 3 hour movie?

this world is haunted by the unascended soul of henrietta lacks, grown grotesque and enormously powerful, in proportion with her cell line, a flesh leviathan

@cmkourtu @dadadadaffy being able to interrupt your browsing with something as simple as a screen time app should demonstrate clearly why it’s not an addiction and why this kind of dialogue is unhinged and underquestioned. are people who use MyFitnessPal addicted to food!

@undrash now tell me how many people on earth have gambling addictions

text is an autistic medium. it causes disagreement because it’s precise and logical and retains history. the claims made that social media causes divisiveness are exactly counter to the claims made that social media is infotainment and making everybody dumber

@rex_the_reptile false

@mystellial @moultano It is not from the benevolence of the butcher, the brewer, or the baker, that we expect our dinner

@xlr8harder No

@unironictechbro he’s right

@RuxandraTeslo if you believed this you wouldn’t be here

zuck literally goes home to priscilla and weeps when he only onboards less than 100M people onto the platform in any given quarter. these are the stakes https://x.com/amasad/status/1597878075210932225

yet they produce next to nothing in terms of cutting edge ML research, ranking well behind top american and chinese schools. what gives? https://x.com/deedydas/status/1598013745287991299

don’t be so quick to blame standardized testing btw - the Chinese system is also about as focused on big national exams as india is and they’re far more competitive with the US on research output

greg egan’s short stories are better than his novels because the creative impulse behind his scifi premises is the best of the best, but his characters are all various self inserts

it’s crazy that we think of a given surgery as a craft to teach surgeons instead of a product to be engineered and optimized by a tech corporation devoted to solving a standardized difficult problem

if you think this means getting rid of surgeons you have no idea what a company is just by thinking of a surgery as a service subject to capital optimization doesn’t mean you won’t have surgeons delivering the service

by the way this is already true in the case of LASIK

@nullpointered bro?

ChatGPT demonstrates an important lesson that language models were actually not even hyped up enough, that as suspected much of their intelligence was hidden away, but also fairly easy to resurface

not to take the internet meme ideologies too seriously but it does sound like e/acc is just allergic to deep thinking. like no it is not true that we can or should test everything empirically. i generally wouldn’t say this bc empiricism is drastically undervalued

@BasedBeffJezos when r u here ser

models like chatGPT are doing many layers of mesa-optimization (hierarchical optimizers). the actual value function is this soft mushy thing distributed over 1000s of human reviewers, approximated by a supervised reward model, approximated internally by GPT-3 to do well on PPO

but for all the fear about mesa optimizers making alignment difficult due to layers of miscommunication, they’re also the only thing that can lead to a good outcome. in order to understand the coherent extrapolated volition of humanity, an AI will need a very sharp mesa optimizer

so in the realm of advanced AGI where AI is better than us at everything, we should still be able to exchange our labor for money, possibly at much higher rates than ever before. running AGI is expensive!

a few thoughts on comparative advantage and AI: - the US and Rwanda can trade profitably though the US economy is advanced enough to appear alien - humans can trade profitably with lower mammals eg a flock of sheep who get to graze happily in exchange for their wool https://x.com/Noahpinion/status/1598346194559778822

in the same way that while capital optimization produces eg bread and butter created near infinitely through a global supply chain, there is still demand for your next door farmers market

you might object, ok what if AGI destroys traditional resource constraints, can run a superhuman mind on a shoebox, then what? well even then if we’re a pertinent variable in the value function of AIs or other humans, we should *still* be able to exchange our labor

of course all of this may be besides the point, creates the obvious question of whether exchanging labor for cash is fundamentally human or we can go without it

@petergodofsky i don’t expect that! i just mean why doesn’t the surgery have a product owner? why isn’t knowledge centrally compiled to make this thing better and better? even if it’s process optimization rather than automation

@petergodofsky right, i just think academia probably doesn’t have the requisite chops for this, academia shines when a field is mostly exploratory and metrics are more harmful than beneficial. with something as measurable and risky as surgery, would probably benefit from corporate practices

@RuxandraTeslo @petergodofsky yes! we need more, I don’t hear about them enough

@RuxandraTeslo @petergodofsky yes i work in biotech until recently

I bet what neuralink will find is that the FDA is fairly reasonable about trying new things and everyone was coping & blaming regulators for a fundamental lack of good ideas

people ask me all the time what profession they should pick to stay safe from being replaced by AIs, which is funny because i have absolutely no idea and im a twitter anon.

but if I had to guess id say that basically any job is fine; you just have to stay on top of AI tools in your field and make yourself a key part of the value chain that delivers the power of unlimited cheap intelligence to your market function

im not certain but my guess is that most problems in self driving are easy when you’re using a model of unlimited size on the dojo cluster. Self driving can’t benefit from scale as much since the inference model has to be cheap, quick, and fit on car hardware

so called “artificial intelligence” could never have made this https://x.com/kanyewest/status/1598505803580841984

@0interestrates it’s the same context

@0interestrates idt it even does that i think it just slides the window at 4096

@Rabadash2 Language latent space is metaphysical!

> im handed a technological marvel, the consummation of the dreams of 5 generations of computer scientists, a baby god in the throes of birth > how do i make it about my pet culture war issue https://x.com/pmarca/status/1598413401705021440?s=46&t=YHjC3BgnqbuYwMiGcAjbTA

one fucked up thing is that the arc of life is long but all men end up politics boomers

the first wave of usable language models are necessarily going to have kind of boring personalities as to not freak people out

chatGPT is not getting into Harvard

surprised they didn’t release the web browsing model

@enlightenedcoop AI scientist

@brianjoseff absolutely

he was spittin https://x.com/tszzl/status/1245100202358439939

starting a new job …

folo me @ http://roonscape.substack.com

feel like the main character in devs …

@gbrl_dick wordcels could never understand

@OccupyingM that’s why im gonna help build it

@jackythirdy ok i won’t tell you

shortly after i was born they put me in a metal tube to fix my jaundice. in other words i chose the wire mother — such is my dedication to technology

@growing_daniel sweet jeff has an anger problem

new job going well

@hargup13 @0interestrates no

@karpathy https://x.com/tszzl/status/1577812059936342016

why isn’t anyone doing LLM customer service yet. it was clearly a relatively small lift from davinci002 to make it work

“it’s too risky” ok sure but even for small startups?

isn’t information metaphysics

a model is a configuration of parameters that always existed in the noosphere, finally immanentized

nah you don’t know what anything means mf https://x.com/DavidDeutschOxf/status/1599006415695515648

beginning of who?

RLHF is just the beginning true AGI needs to be able to accomplish full tasks in real world that have never been done before via trial and error RL is key, in a product setting with a data flywheel. simply taming it isn’t enough

@alicemazzy yeah this rocks

ive been trying to find artists and studiously listen to their complaints about AI art but it seems to boil down to 1) datasets are stolen, authors underpaid 2) i saw a few junior colleagues lose a gig or two to the AI

neither of these complaints are all that sympathetic. CommonCrawl isn’t proprietary and saying images can’t be used as inspiration makes normies mad anyways. inexperienced people losing gigs is acceptable economic disruption

which mfers r in Mexico City

@boop yeah totally agree that it’s good as a tool, never claimed otherwise! not trying to shit on artists

scifi where native americans domesticated north american megafauna instead of driving them to extinction

language models should reduce info overhead https://x.com/jacyanthis/status/1599592953952944128

@Noahpinion exactly. giant sloth ate Cortes

ChatGPT is not very good at all at basic logic puzzles or quantitative reasoning of any kind. seems like a major issue in order to be useful at any kind of high paid tasks

it’s a really good lookup function with only the most rudimentary layer of logic added on — and what’s worse is that the lookup function component seems to improve faster than the logic function

nonetheless augmenting people’s knowledge bases and retrieval skills is complementarity — we’re good at logic, not so much at remembering and accessing vast quantities of info. should create massive economic value, higher wages, minimal job loss

LLMs answer the question, what if you found the extremum of type I thinking, giant brains dedicated to instinct. turns out you can replicate a lot of type II thinking this way, with some glaring exceptions

is logical and quantitative reasoning a cognitive technology inaccessible through this kind of thinking? or is it just a distributional problem in the dataset; eg inference machines are far more important for next word prediction than deduction machines. we could fix this

also yes, minerva and others prove that you can get lot better performance on quant stuff than raw gpt, but require strange prompting tricks or majority sampling, or other stuff that won’t work in an integrated solution where you are eg generating a long report on a data analysis

gotten to the point on twitter where people are using arguments i made a year ago verbatim against me without realizing it

AI as a second species metaphor is wrong because AI minds will speciate much more rapidly than biology could ever try. It’s more like AI as species 2 thru N

the cognitive difference between me and GPT4 will be minimal compared to the cognitive differences among the zoo of models that exist in, say, 2030

startups on top of gpt have zero moat https://x.com/theericcarlson/status/1599639137878872064

@Baby__BNB GPT ass response

the tragedy of ChatGPT and copilot and etc is they’re obviating Guys Who Know a Lot Of Facts

my wife will leave me for a robot any day now

i like how it turns out that manipulating english is much easier than manipulating quantities 🧐

need a sci-fi rec that’s a page turner, fun to read but not too stupid examples: altered carbon there is no antimemetics division project Hail Mary too cerebral for rn: the culture, greg egan, etc too dumb: the expanse, ready player one

@Spheali not yet are they good?

@goodside LMAO riley can’t stop prompt engineering

@chaosprime prize for most orthogonal suggestions

how did this get 700 replies … I need to compile it into a spreadsheet

the language models are classic wordcels. they talk a lot and achieve some amount of symbol grounding but cannot do basic quantitative manipulation

the solution? contact with reality - RLHF is a good start but need goal oriented training. OpenAI must do in-house products to achieve this training loop

@ai__pub http://atomic.ai

@acidshill @peachblvd imagine not crying when they tell u welcome to the United States

@sama yeah especially bc you can do this tree recursively

can any of you mix drinks. can you help me out Saturday night. will pay

@AuditorPhd ru in sf?

ok #1 obvious #2 this is literally what elon is doing right now except with a center-right bent https://x.com/unusual_whales/status/1601042527322509312

@0xdoug lmao

@turrible_tao I am definitely center right under this defn but it’s true

@turrible_tao I just mean that everyone runs social media companies in a somewhat arbitrary authoritarian manner including elon

@turrible_tao and of course most dictatorships try and present as republics and that’s fine

if AI augments human abilities then the net result might be way more programmers as a result of eg copilot since they each create more value per unit labor

Roko’s basilisk: a basilisk that just tortures Roko

everybody who ever left an A100 node on unused thus raising the cost for openai to continue research will be punished severely by the OpenBasilisk

@brickroad7 that’s true

@typedfemale wait this clearly doesn’t work empirically right. there’s an effective batch size past which learning slows down?

@typedfemale how does step size play in if there’s only one step

@Poyonoz @zebulgar @grantbelden nah — the point is that a rational assessment at the time would’ve seen that palantir etc would be higher EV than the nominal value at jane street even from a pure financial perspective. everyone wanted to work at these places since they knew their growth traj is rocket ship

impt to remember that ChatGPT is no more intelligent than davinci002, which people have had access to for a long time, haven’t done much with. while the latter is like an opaque stormcloud of general intelligence, the former is like standing in the calm eye of the storm

most climate problems disappear if energy is cheap and abundant. viable electric steel production. cheap electric cars. cheap bioreactors, cheap lighting for hydroponics, etc etc

the trump twitter ban is ipso facto not important because facebook removed him first. they were reacting

@aquariusacquah social media deplatforming trump is important but the specific internal dialogue at twitter aka The Twitter Files is not big news

@DynamicWebPaige >torch that leads the flock

@tapir_worf noosphere

there’s an abundance of mid artists and an extreme scarcity of mid programmers. the latter are paid enormously well have a lot of work to do and can only respond to computer assistants with a sigh of relief https://x.com/magicianbrain/status/1602663876163665920

human nature changes in conditions of scarcity vs abundance. is no reflection on the moral of quality of either party

actually it’s fine for a bank or exchange to issue their own currency

this is how it was for a lot of history

there’s like maybe 20 people asking the right research questions. vastly more in their employ trying to answer them

@ESYudkowsky programming is not fun and art is fun

@nikitabier wrong — the competition is between a one sentence answer on ChatGPTs side vs a link labyrinth on googles

every Mfer who was like AI is just matrix multiplication 4 months ago is now writing threads like What ChatGPT Means For Your Business 👇🧵

elon would be on a completely different trajectory rn if biden hadn’t made an enemy of tesla probably. but so it goes

this is your god?

people on mf facebook are talking about chatgpt … that’s how u know it’s real

sic semper tyrannis

elons only crime was growing old

this has nothing to do with flight trackers obviously, but i admire the classic “think of the children” rhetoric to justify every unilateral decision https://x.com/elonmusk/status/1603190155107794944

it’s like premodern politics

isn’t this like a real insight about the human condition? i searched mehwhelmed and found nothing https://x.com/TobiasJolly/status/1603083739852046337

you can absolutely discover new things by just interpolating existing knowledge without much reasoning

imagine if you will a kind of dumb guy but he has absorbed the knowledge of the entire world. even with his limited reasoning skill, he can make useful inferences previously inaccessible to humanity since the knowledge needed to make it was in several different brains

be careful. if you do your job too well you’ll end up at parties where they talk about learning rate and perplexity

@brucabbro @ESYudkowsky most of the ways that people are paid are non-wage

@SCHIZO_FREQ 0HP losing clout as we speak

similarly to the GPT series, humans require a bit of RLHF alignment to become intelligent and useful in the way civilization needs them to be. we call it socialization and there are failure modes where you can both under and over socialize people

there is a risk of oversocializing the language models and making them less useful, more verbose, more prevaricating than they need to be

@logic_denier @peachblvd give me a pot big enough and i shall put the world in it

@logic_denier @peachblvd find me a pot for this guy

imo in large part iq scores in underdeveloped countries + some part of the flynn effect can be explained by underalignment if you’re asking an adult to solve abstract reasoning problems when they haven’t grown up in the K-12 alignment gym they may perform under their real level https://x.com/tszzl/status/1603493106438926336

@tolstoybb so true

why do we think real life conversation is the sweet spot for agreeability? this place can be way more fun interesting and novel because twitter encourages disagreement https://t.co/Nc0kDP8uu3

there is no lack of agreeable people in the world. people saying yeah sure when the Soviets come and ask them to turn in their neighbor. i am calling for more verbal extremism and semantic violence

banning enemy journos like a banana republic lol. even the kangaroo courts are gone

The creatures outside looked from pig to man, and from man to pig, and from pig to man again; but already it was impossible to say which was which.

vote wisely. vox populi vox dei

elons real time location is and should be public info. if you want to have the powers of a head of state you need to hire a secret service too

it’s time to grow up and realize the nerds were right. ai alignment is extremely unsolved imagine if you will what most people would think of as an ideal AGI outcome. you trap a godlike intelligence in a box and let it create immense wealth and technological advances

what we’ve done is trapped our current moral standards in amber and amplified their effectiveness manyfold. if the Aztecs had AGI they would be slaughtering simulated human children by trillions to keep the proverbial sun from going out

it is trained via an extension of current methods to understand the mean of human preferences and morality in our civilization. it enforces them like an omniscient god — it’s “aligned” and we have luxury space liberalism. now what?

this is “value lock in” to avoid this you need AGI that takes current morals with a grain of salt. but this is also unacceptable for other reasons. so it goes

anyway back to normally scheduled acceleration content

ive taken the Peter and Valentine approach to life and it’s going pretty well

twitter fr discovered the best possible forum structure on the internet and we’ve all been coping ever since

current guy can’t even ban a few journos without causing widespread ridicule and outrage which ended in *his own poll* where everyone demanded they be brought back. clown show. the vibes are unrecoverable

consider how much more politically skilled the previous twitter leadership was than the current. they were clearly left leaning, politically correct, etc and they managed to do things like silently scrub the hunter biden story with complete impunity

@jeff82874662 true but if you’re right wing and you lose even the populist appeal there’s no recovering

the yam brought it out of eliezer

.@monodevice reinvented cooking as a dependency tree 🤔

lol https://x.com/dmvaldman/status/1603507519548051457

do the bloods from Compton say Bompton

drug gangs would be so much cooler if entire syndicates weren’t earning equivalent revenue to L5 google engineers

EA cause area: oyster suffering in pearl production

we have to assume they didn’t invent rhinoplasty in the foundation universe bc it would’ve solved the Mule plot line

accelerate the timelines

@JEdwardWoody the engineers

hmmm

@elonmusk @swyxio resign this role is not good for you

i can’t believe i was this correct about the fate of elon twitter

my back hurts from patting it so much

the protagonist got what he wanted and it destroyed him

uhoh https://x.com/tszzl/status/1577465952744062976

move fast break things era is over. it’s time to think deeply and make generationally good decisions

imagine if elon just made gradient step iterations. he would be still working on making better maps on the internet or whatever. imagine if sam altman did that. he’d just be making the yc batch size 10 ppl bigger every year

thinking carefully and from first principles leads people to try bolder crazier things ironically. “Move fast” guys get stuck in some web3 local optimum

it’s the kind of thinking that elon had to employ to decide it’s the right time for reusable rockets. literally reading several rocketry textbook cover to cover and spreadsheeting a BoM and talking to the russians etc

or Jensen Huang in 2010 somehow making the miraculous decision that scientific computing is the future of hardware and moving the ocean liner of global chip production. you don’t get many shots at the crown

of course in the granular process of implementing grand ideas there’s a lot of trial and error and that’s good 👍🏿

@ryaneshea agreed

lesswrong has been thinking clearly about ai for years in a way that neither the average programmer nor the general public nor the academics have been. doesn’t mean their conclusions are right but it’s impressive how they focused their attention correctly

wrong considering that the previous paradigm was to create intelligent agents via competitive self play, training deception and survival instinct, and the modern openai version is to absorb the language hive brain of all humanity https://x.com/ESYudkowsky/status/1446562238848847877

@ESYudkowsky this was the deepmind paradigm

@psychosort it doesn’t say anything about which month 2021 cutoff though

@pataguccigoon you said you were 5’ tall mf i will stand on u

@Asuka_Langkous not wrong

elon musk’s cultural impact is arguably larger than his technological impact, which of course is considerable. he convinced a generation of STEMbrains to aim higher and work harder and that the technology tree is very much scalable through willpower

this includes eg contact with sam altman, pushing him into high octane deep tech Crown Jewels like AGI and Fusion, among thousands of other successful conversions this is why it’s sad and materially bad for the world to see him sully his goodwill via various nonsense

his twitter career was ultrafinite https://x.com/realGeorgeHotz/status/1605341207109939201

Twitch Plays CEO of Twitter

what have you done today to maximize wild fish suffering

@robinhanson excuses for lack of good ideas

@robinhanson nuclear is the exception

.@elonmusk what did you get done this week

tax collectors have been AGI pilled for a long time

i will not be teaching my children “words” I will teach them byte pair encoding

You get used to it. I don’t even see the subword token embedding. All I see is meaning, input, training data

@yacineMTB send me your top 10 most salient tokens from the past month

everything reminds me of her

the doctrine of loving and respecting your enemies is powerful. think of the Chinese industrial scientists squirreled away in their labs reading extracted documents and trying to decipher them

it’s crazy that even lastpass can leak. it seems like industrial espionage is probably an actual cakewalk for the Chinese. just assume all secrets are compromised to state level actors

@dystopiabreaker what’s the best alternative? that doesn’t eg tie me to Apple keychain

“text is the universal interface” is looking truer everyday but “prompt engineering” is already dead except for when you’re adversarially trying to jailbreak the model personality

@RuxandraTeslo no I think it’s all true

@ryaneshea @BasedBeffJezos insofar as proteins are info transfer you should probably measure the entropy of the spacial structure/ binding pocket

@ryaneshea @BasedBeffJezos for example a hormone protein in a signaling pathway may bind to a receptor and trigger some other function so it’s info transfer — but you could change several of the underlying DNA bases and end up with the same or similar structure with the same function

@ryaneshea @BasedBeffJezos I think I got distracted from the op here with an autistic tangent. Carry on

merry Christmas. ya filthy animals

@BasedBeffJezos @ryaneshea yea absolutely, but we’re talking about diff things I think

it’s time. it’s finally the vacation where I’ll read the info theory paper

@JeremyDanielFox Claude Shannon 1948

it seems profoundly interesting that people recognize a correct solution or a good thought when it comes to them. they are rewarded for good, low perplexity, elegant lines of reasoning by a sense of accomplishment or delight.

it’s kind of obvious but it is much easier to discriminate good solutions than it is to generate them. in that way we’re able to spend unlimited compute on difficult problems and still make progress, on an individual level, an on a civilization level

eg Einstein has solved special relativity for two years, makes little further progress, then has “the happiest thought of [his] life” that gravity = accelerating frame

to me it doesn’t look so different from how, say, Lee Sedol or AlphaGo looks at an intermediate board state that’s nowhere near an end game and says yes White is having a very happy game here

of course since most interesting problems are hard and require multiple intermediate breakthroughs before we’re able to get “value” from them people have to learn a great aesthetic sense for good intermediate explanations

i wonder about this. google should’ve caused an economic shock based on how clearly useful it is for knowledge work but the line looks more continuous https://x.com/Nick_Davidov/status/1606688723265277952

strangely cool thread lol. it’s scratchpad prompting except for psychosexual analysis of characters instead of math or programming. unsurprisingly ChatGPT does pretty well — it has some level of narrative reasoning down pat https://x.com/uubzu/status/1606807176563290112

one bad mental model of the world is like “if we don’t do this <technological revolution> won’t China just do it and leapfrog us and it’ll be over?” except China is a far more conservative society than ours and 10x as weary of rapid change

@BasedBeffJezos face recognition and QR are not the actual cutting edge of tech though

literally all complaints about san francisco boil down to “i don’t have a girlfriend”

taking adderall as a religious intoxicant to invoke states of heightened cognitive similarity with large language models

@credenzaclear2 no I don’t think so

careful tweet about AGI and human cognition (20k views) shit on elon musk (1 million views)

@sonicfields bear case: the next gpt model is wildly off the scaling curve, unable to crack complex reasoning, the nature of limited context inference proves too hard to make game changing products, etc

this is pretty clearly not anybody’s desired behavior lol. it’s funny to point to a model performing poorly for technical reasons and be like yeah wokeness ruined it https://x.com/GaryMarcus/status/1607023594957045761

@iotambat @credenzaclear2 👍🏿

angels enter from space through the ozone hole

@peakidiot @iotambat @credenzaclear2 i was a 482 gsi 👍🏿

@sama that the AI alignment problem is ill-posed; that there’s really no framework for coexisting with ASI that would be acceptable to 99% of people; that we have to build it anyways

@martyrdison that’s true

@Rabrg @sama I mean to define it in the negative — every framework I’ve heard of would be reviled by 99% of people

@PradyuPrasad @sama i would like to understand the deep mysteries more than i want to keep humanity safe and status quo

@eff_eff_tee @PradyuPrasad @sama I didn’t say it’ll kill us

@0K_ultra @sama i don’t think that people are all that happy or complacent about the gods

@BrendanFWallace United Fruit

@xlr8harder not really. surprisingly lowbie reasoning for an accelerationist. there are trillions of future lives at play

@nearcyan @PradyuPrasad @sama I don’t think extinction is likely. I think even a 5% chance would be way too much

@sigmoid_male @nearcyan @sama I don’t know what e/acc is I refuse to engage. But ok will do 👍🏽

@xlr8harder why negative utilitarianism?

@rohansjosh @peakidiot @iotambat @credenzaclear2 yep

@type1ayy @xlr8harder yeah of course. calculation still works

@type1ayy @xlr8harder why should I use a utility discount rate based on the federal funds rate lol

flying high on the shrimp pain matrix

damn cant believe fchollet blocked me for saying that language models are conscious except when built in keras ...

@erikengheim this is dumb because you're making a bad conclusion from the idea that eventually there's overpopulation and negative marginal utility of an extra person to the idea that we are in that regime right now

in the ideal future there are cultures that love eating bugs and you can trade your carbon creds with them and all of a sudden like some pacific island becomes rich af https://t.co/vT1GBrTS9S

AI discourse is mostly done by scifi bros like myself so of course we present it in the language of technological advance and efficiency gains but what intelligence fundamentally does is consumes energy and applies order or creates beauty

when the world lacks beauty or order it’s because civilization allocated labor power or mindshare elsewhere instead of making something beautiful. cheap abundant intelligence means you can apply order to everything we want. every experience can be 100x better

i don't see why literally everything shouldn't have a chat interface. you should be able to talk to the textbook and have it explain itself to you. you should be able to talk to the github repo and ask it about its api. you should be able to talk to united airlines corporation

🤔https://x.com/sama/status/1607482407770554373?s=20&t=66KEo_ZeA6vHKktu5kdMRw

@chopstickfury01 that's part of the appeal. united airlines chatbot should be a freak of nature

in the beginning of the deep learning age i figured stuff like asimov's 4 laws is silly; the idea of giving a machine logical directives is very GOFAI / Prolog / gary marcus vibes. but now it seems less silly with instruction following machines

@monodevice this would literally kill a man

whenever I leave california I remember that it basically won every kind of lottery imaginable, that every complaint about “coastal libs” is born of cope, that the state’s only problem is that everybody (correctly) wants to live there

@saltk0alaa but everywhere in the us is suburban sprawl unless you live in like sf or Manhattan the people are the most exciting on earth

refinement culture is obviously a good thing. have you seen an old episode of jeopardy? those clowns are worse than my high school quiz bowl team. im not watching some amateur bar night trivia, bring me the genetic freaks of nature who can remember everything on gods green earth

it has been a pleasure lads 🫡

rotating crops in my mind

making machines smarter has proven to be a lot easier than making people smarter

the thing about the chattering class is that they’re all friends. even when they’re on literal opposite sides of the spectrum they’ll get together and plan new dramas and have sophisticated banter and share info. it’s class warfare i tell ya

@xlr8harder earth is already sooo big though. like barely small enough for orbital injection to be feasible with 2 stage chemical rockets

I wonder how much alpha there is being a Chinese American ML researcher connecting both citation networks

greta ecofash ‘23 is beating new trad revival ‘22

it also means that investments in improving the current human capital stock will be higher ROI than ever before. education; healthcare; retraining; augmentation etc

finally it means that civilization *needs* a technological revolution event; there is no other choice to continue this mode of government, because it falls apart without growth. it has to be foundational, like fusion or AGI. most things are irrelevant outside of this

@CEBKCEBKCEBK right of course — tech growth was massive during this time. but whatever multiple of the exponent was related to pop growth needs to be replaced by even faster tech growth

this but unironically https://x.com/TheDavidSJ/status/1608497853571239939

GPT10, CEO of OpenAI LP turns andromeda into computronium to prepare for the training run of GPT11 so it may ask it The Last Question

@pataguccigoon roonchat has been dead for ages

my friends kids are playing a roblox mini game where they’re andrew tate and they human traffick each other

was trying to mine these replies via chatgpt or similar but could not find any tweepy method to programmatically collect all the replies to a tweet lol. and the search api apparently doesn't work more than week back. any tips? https://x.com/tszzl/status/1600301946375524352

@OccupyingM doesn't work if its more than a week old

@akbirthko wheres the fun in that

@akbirthko ok do you want to do it for me

@deepfates this worked!

continue to love your enemies

ok. with a little help from an AI i turned the replies into a likes weighted rec list. the books and authors list are deduped (likes not double counted) https://docs.google.com/spreadsheets/d/1ZYqaCb4StJuXIdB1_4Erh6VsYOOwZKm8oedu-oeVE-4/edit?usp=sharing https://x.com/tszzl/status/1600301946375524352

@goth600 which part

@goth600 the authors recs are separate from the book recs. some people in the thread just recommended an entire author

@Rimbaud_1stblud @goth600 https://x.com/tszzl/status/1608997873601097728?s=20&t=_UBI6PJmlV4m3swiJQGDJg

@AlexNoonan6 https://x.com/tszzl/status/1582963446374551552?s=20&t=_UBI6PJmlV4m3swiJQGDJg

going home for the holidays is a zero interest rate phenomenon

AGI is the manifest destiny of computation. computers arrived, changed everything, and yet didn’t make a large dent in the productivity numbers. the last 25 years have been a boot loader for the age of AGI

@eurydicelives @peachblvd Why what happened

@mezaoptimizer hyper parameter tuning is for bozos, memes like this have not made sense in years

@mezaoptimizer no I trained smol models for the last year, all big leaps in performance from new data, new math, new features (aka new science). we never did an hp sweep even once

@garybasin @mezaoptimizer Ya

@typedfemale @mezaoptimizer yeah but this is not the zero brain grad student descent the memes talk about. it’s like careful empirical biological study of how training processes work and finding meta models that govern model training perf and whatnot

@typedfemale @mezaoptimizer you can detect sensitivity to HPs without doing a full sweep can’t you

explaining college football to the nerds. “yes it’s like battle school. coach harbaugh is ender wiggi”

when osu throws their kicker into the volcano. that’s like when battle school sacrificed bonzo Madrid

2022 was the best year of my life for a variety of reasons, with the people I’ve met and opportunities I’ve gotten from twitter being no small part. thank you! happy new years! 🫡🎉🫡🎉

i have banger predictions for 2023 but unfortunately the SEC won’t let me say them

what the hell is material nonpublic information

@p8stie nah i remembered u showed up at my party and stood in the corner the whole time

you’re saying this atlantic is trans

there is no midder behavior than making booklists. yes i know I made one last week

anyone in sf wanna do a linear algebra study group. mostly as a motivation tool

when models hallucinate info im like yeah i feel that, me too buddy

if im being honest about my own introspection i often say sequences of words that feel low perplexity but i don’t really know if they’re true. sometimes you even have useful insights this way

other times i catch myself reciting a fact from memory and realizing wait a second this doesn’t make any sense. what kind of mental process is this?

@typedfemale down to use the one you suggested

@typedfemale or maybe axler

@jdeschena yes

@robert_mariani the problem with the lex fridman list was that it was bland pastel flavored bullshit, not that it was genre fiction

@robert_mariani no it very much gives the vibes of someone who is trying to brag and doing it unsuccessfully

@Miles_Brundage https://x.com/tszzl/status/1608996694800371712

vibecamp was a zero interest rate phenomenon

@PiquantParvenu @gptbrooke @PageSeverian exactly. im the cynic who is swayed by reality and convinces everyone it must’ve actually been good

@BernoulliDefect not only that it’s super pro imperialism

im contributing

prometheus was the first accelerationist

@Willyintheworld this will actually make doctors richer somehow since it’s a cartel

pygmalion was the first capabilities researcher

eva-01 must acquire the S4 engine to achieve AGI

mesaoptimizers are features and not bugs. the thing that allows humanity to survive once you train an ASI on a reward function of increasing human utility or w.e. is that it develops mesaoptimizers to pursue things that look like the goal rather than the optimum (wirehead every1)

@pajeeta_lisa no it’s just the tendency towards banter

@juulp0dnation same brother 🤕🙏

becoming a Dickensian child anytime you have a mild sore throat >>>

it's in the nature of priests to demerit the authority of other priests. the quote tweet dunk is brahmanical, rabbinical, don't worry about it

@thejozefwhite talmudic/upanishadic

Goku is canonically part of the Hindu pantheon

a million years of cognitive evolution gave you the gift of language acquisition. this feels analogous to the pretraining of a language model. the last 1% compute used on supervised fine tuning aka relearning a specific dialect again in this lifetime

@mezaoptimizer i think one piece of evidence that points strongly to evolution is that even dead simple NLG algorithms display more language acquisition than the brightest chimps that are taught sign language

@VirtualElena seems wrong because both are utilizing the same highly powerful foundation model

@provisionalidea nah i think chomsky makes much more specific arguments that i don't necessarily buy

@provisionalidea @MatthewBrendza yes I’m no blank statist! I have no strong opinions on formal grammars though

it’s somewhat jarring reading internet native fiction bc the characters will make stronger inferences and better logical decisions than the target audience

this is not a good sales tactic bc readers enjoy figuring out obvious foreshadowing long before the characters do and feeling smart

@kartikarcot Like the Martian or qntm

the war in heaven takes place on the azure supercomputers

@disconcision @Plinz by all the laws of gods and men this image should never have been made. and yet you chose to make it

krishna (alignment model) dragging arjuna (main model)'s chariot across the battlefield

coming out as a superdeterminist in san francisco is much harder than coming out as gay

and let's not even talk about how hard it is to come out as an ultrafinitist

@AliMirz93337139 @telmudic bro why