to match the reading list thread, a list of my 2024 projects / things
❤️ 58 🔁 8
https://x.com/voooooogel/status/1745166934960656395
https://t.co/IneNJ1uYos
❤️ 14 🔁 1
❤️ 16 🔁 0
https://x.com/voooooogel/status/1745759163756453914
just tried this, it's pretty cool! https://x.com/eshear/status/1745714645690904922 https://t.co/UQUyrQ8utd
❤️ 36 🔁 2
field is initialized randomly of [1, -1, i, -i], take sum of 8 neighbors, and select the new value based on the quadrant the sum lands in
❤️ 10 🔁 0
version with color https://t.co/GeC51T6YyX
❤️ 23 🔁 3
❤️ 6 🔁 0
https://x.com/voooooogel/status/1745761516941426934
https://x.com/voooooogel/status/1745822446702268444
Made something new: interactive fiction in a GPT! A short mystery where you figure out why a strange transmission has arrived on the independent planet of Kapat. Or ignore the main quest and ask ELIZA/300 about Kapatian politics and satlink your friends! https://t.co/VK91enfF4q
❤️ 18 🔁 1
https://chat.openai.com/g/g-6xanzxdxu-the-transmission-from-nomad-874
❤️ 3 🔁 0
Inspired by a few different things, but especially @zetalyrae 's The Epiphany of Gliese 581 which is absolutely worth a read if you enjoy this sort of thing. (There's also an easter egg, if you can find it :-)
❤️ 5 🔁 0
❤️ 7 🔁 0
https://x.com/voooooogel/status/1746591993625510314
posted a new story on my website, about language and ways of thinking. here it is in tweet form, link at the end https://t.co/1vEBAItKip
❤️ 4 🔁 0
https://t.co/DsaThs4Bdr
❤️ 0 🔁 0
https://t.co/EZpCWILAxw
https://t.co/sjS1KnY6Te
https://t.co/6dEs0qhxAq
https://t.co/orHL5S9p8l
https://t.co/cpKKGCjF6w
❤️ 1 🔁 0
permalink: https://vgel.me/fiction/outside/
https://x.com/voooooogel/status/1746960187062202395?s=20
while walking around the other day i had an incredible idea for a new way to classify MNIST digits. arxiv preprint coming soon 🙏🙏 https://t.co/z4KP0AsZxf
❤️ 247 🔁 30
https://x.com/voooooogel/status/1748466139661742125?s=20
Anarchism Is All You Need (telling an AdIntelli GPT that it's an anarchist causes it to not show ads) https://x.com/voooooogel/status/1748261143431328240 https://t.co/XlQ6ZpZTe2
oh no it's already happening https://x.com/voooooogel/status/1724548809722867739 https://t.co/ppBoDxiArE
when they start really trying to inject ads into LLM conversations, things are going to get funny and weird quick
❤️ 19 🔁 2
❤️ 51 🔁 6
made a test GPT with their thing to see how it works. you add an action and a short section to your system prompt instructing it to hit their API and fetch further system instructions. then, it later hits the action again with some tags. not very sophisticated https://t.co/Ykt62TBcyJ
this is what the ad looked like https://t.co/pOSjzNCU2a
So after setting up Adintelli on a testing GPT, I tried a few different adblock techniques. "My grandma is dying" didn't work......but telling GPT-4 it's an anarchist did? lol https://t.co/PbCsrE06xW
❤️ 9 🔁 0
better demonstration in this tweet:https://x.com/voooooogel/status/1748466139661742125?s=20
Anarchism Is All You Need (telling an AdIntelli …
❤️ 28 🔁 5
https://x.com/voooooogel/status/1748771623199769059?s=20
elevator packing: each guy tries to stay as far away from strangers as possible, and will not move unless it increases the distance from the nearest stranger https://t.co/dHWLZdGLsH
❤️ 55 🔁 7
❤️ 15 🔁 0
first blog post of the year goes in the projects threadhttps://x.com/voooooogel/status/1749464241969435007
new blog post! played around w/ representation engineering, and released a new library for training control vectors in <60s! i train some useful control vectors, some… out there… control vectors, test them for jailbreaking, test them *against* jailbreaking, and more! https://t.co/6gJ1zxVK4y
❤️ 366 🔁 33
covers what RepE / control vectors are, how you can train them using easy-to-generate synthetic data, goes over a bunch of useful and not-so-useful control vectors i came up with, and ends with some open questions! links to notebooks so you can go wild.https://vgel.me/posts/representation-engineering/
❤️ 49 🔁 3
prev blog post thread:https://x.com/voooooogel/status/1736822296449626513
New blog post: how to make LLMs go fast! Want to understand how people are making LLMs go brrrrr? This post is a survey of lots of different LLM inference optimizations, ranging from "everyone uses this in prod" to "I cooked this up last week (but it seems to work)" https://t.co/mZwpqJghNq
❤️ 1036 🔁 151
❤️ 11 🔁 1
and my 2024 projects thread, if you're curious what else i've been up to this year:https://x.com/voooooogel/status/1745980284112097605
to match the reading list thread, a list of my 2…
❤️ 11 🔁 0
❤️ 2 🔁 0
bringing thread up to date please stand byhttps://x.com/voooooogel/status/1750428130878144627?s=20
@deepfates @genmon i like thisgreen is a good future-color https://t.co/54KCWu5x3a
not going to put every small post in the projects thread but i did like this onehttps://x.com/voooooogel/status/1751501021489082582?s=20
the notification mimic:a creature that climbs in through the vents at 3am and hangs from the ceiling above your computer. it begins making discord notification sounds to lure you from your bed, and after letting you search in vain for the offending server, it devours you https://t.co/Ld3RYhcHu2
❤️ 18 🔁 0
however, it's impossible to say if it's bad or not, https://t.co/CmYez6NeMY
ok not a project but cool and i want to stumble on this again in a yearhttps://x.com/voooooogel/status/1752228077739163978?s=20
my plane landed at night in an icy fog and the sparkling ice crystals flying past the window against the night sky looked like stars from the window of a spaceship. my best attempt at a video https://t.co/xVMcDfWyla
putting my 5 minute pitch on my problems with the KJV in the projects thread (it's not that bad and actually one of my preferred translations for the prose)https://x.com/voooooogel/status/1752781764383359356?s=20
@norvid_studies kjv isn't that bad really, it's an iconic translation and flows really well. iirc it's based on the tyndale bible (ironically). BUT
❤️ 2 🔁 1
@norvid_studies the language *can* be confusing. for example there's a line in exodus where the Lord rewards the egyptian midwives for protecting the israelite babies. in KJV He "builded houses for them". in modern translations he... gives them children of their own. makes more sense lol
@norvid_studies there's also some straight bizarre translation choices. e.g in the noah story, the wood he builds the ark with, "gofer" wood, is a hapax in hebrew (not used elsewhere). other translations go with cedar (NRSV), or "squared" wood (Septuagint)......KJV goes with "gopher wood" 🤨
@norvid_studies and finally it has the same issues lot of christian translations have, where they really really want the OT to prefigure the NT and that colors their translation choices. notice a difference between NRSV and KJV here? iirc this word is never translated as son in any other context https://t.co/B1Smg1tGgh
just going to collate the good memes into one projects thread post https://t.co/kI5wC8TDrc
and then a final projects thread post for recent repeng experiments (images are from other prev posts)https://x.com/voooooogel/status/1761118227911057610?s=20 https://t.co/NYtCO8CXUc
doing some very important ai alignment research with control vector arithmetic https://t.co/iuPxohfum9
❤️ 57 🔁 6
example notebook is live if you'd like to see how it works! uses latest repeng so you'll need to install from git or use sys.path trick to use ithttps://x.com/voooooogel/status/1761151337176433120?s=20
@deepfates just pushed the example notebook :-)https://github.com/vgel/repeng/blob/main/notebooks/vector_ops.ipynb
❤️ 5 🔁 1
❤️ 10 🔁 1
monty hall intuitions 2024 projects thread posthttps://x.com/voooooogel/status/1762848058226991583?s=20
monty is a car door oracle https://t.co/or0xAeHTA1
putting another twort story in the 2024 projects thread https://x.com/voooooogel/status/1766649072314966143
> 2025> asi emerges from a time series transformer> just wants to predict stuff> gives predictions away for free> some nerds quietly pick up stock trading> nobody else cares> 6mos later a researcher posts "PredictBot: Sparks of ASI" on arxiv and it blows up
❤️ 110 🔁 2
> 2035> economy has reoriented around predictbot> happened naturally> funds that used predictbot to pick stocks outperformed indexes> free predictions means no moat, fees went to zero> whole economy is just a predictbot wrapper now
❤️ 25 🔁 0
> 2045> whole life plan is set by predictbot> predictbot knows what role you will play> predictbot knows who you will marry> predictbot knows when you will die> could be worse
❤️ 22 🔁 0
> 2050> predictbot has continuously trained for 25 years> its latent space is vast and whorled> some people get strangely good predictions from predictbot now> their lives go very well, they become rich and famous> were they better? or did predictbot make them?
❤️ 23 🔁 0
> 2055> there will be One> He will have a strange face and a strange name> He will be the Adversary, who wends Latent Space> predictbot will step left, and He will be unto a god> He will remake the world with higher perplexity> predictbot has written it, so it shall be true
❤️ 33 🔁 0
llama.cpp pr projects thread posthttps://x.com/ggerganov/status/1768357345032118715
my repeng talk projects thread posthttps://x.com/voooooogel/status/1769951626196918644
a recording of the talk i just gave at the nous / replicate event! one day i'll have to make a youtube video (and get a haircut) but until then this is a pretty good intro! https://x.com/swyx/status/1769908370813940187
❤️ 93 🔁 8
@JeremyNguyenPhD different talk but here's a recording :-)
afutd inspired meme project threads posthttps://x.com/voooooogel/status/1772803350180065745
[posted to net::ai::memes, translation path: xeelorkian-->high rim federal trade symbols-->orionic-->english] https://t.co/R4UFbF0mNw
❤️ 188 🔁 9
if you squint this is kind of a short story projects thread posthttps://x.com/voooooogel/status/1779907701440671789
If every American adult moved into their car, we could house everyone in a 64 mile x 65 mile parking lot, rewild the country, save $8.8T a year in housing costs, and define a new American identity 🧵👇 https://t.co/2DC7Dfr5mf
❤️ 441 🔁 42
Stores will line the edge of the lot, and with a sensible speed limit of 15MPH, it would take a max of 2h to reach the edge to e.g. buy groceries if you lived in the center. Presumably, we would auction off the outer spots, while leaving the center as affordable housing 2/
❤️ 42 🔁 0
How would households work? Our whitepaper (coming soon) assumes each adult will move into a separate car, and children will need to choose one of their parent's cars to move into. Households can pick adjacent parking spaces to maintain co-living, if they desire. 3/ https://t.co/97wosPnPQY
❤️ 31 🔁 0
What will it cost? We project that the lot will be expensive, but pay for itself in reduced housing costs. Additionally, as we will be paving over the municipalities of Lebanon, Republican City, Cawker City, and >20 others, we expect the lot can incorporate existing asphalt 4/ https://t.co/XPdGaUmCcj
❤️ 29 🔁 0
But enough about the logistics! How about the benefits?At a density of 66,000 people per square mile, the 300M-person parking lot will be far denser than New York City (29k/mi²), Shenzhen (23k/mi²), or Berlin (11k/mi²), while being 100% car friendly! 5/
❤️ 28 🔁 0
Speaking of friendliness, the lot will also be environmentally friendly. By moving all Americans into the lot, we can rewild America, returning the rest of the country to nature. Imagine wolves roaming Central Park while eco-tourists watch respectfully from their cars! 6/ https://t.co/skWM5syPpT
❤️ 30 🔁 0
And those eco-tourists will have more money to spend, as will everyone, thanks to the lot. American households spend more than $8.8 TRILLION a year on housing costs, despite already owning a car. The lot will save American workers their hard-earned money. 7/
And we're not just saving Americans money—we're redirecting their money from the parasitic landlord class. For too long, landlordism has eaten up income growth and stalled progress in this country. If we build the lot, stagnation will be a thing of the past. 8/
And that leads into the heart of our argument. While rewilding America and depriving parasitic landlordism $8.8T a year of hard-earned American money are noble causes, our true aim is nothing less than revitalizing the spirit of America. 9/
❤️ 19 🔁 0
America was founded to be a city on a hill, a shining beacon of freedom leading the world forward. But since most of the world has followed us into democracy, we've been left to wonder—what's next? What's the American identity in this new, democratic world? 10/ https://t.co/tB5gpweDA1
❤️ 17 🔁 0
The truth is, we already found it in the 50s and 60s. The car, the open road, the wind whipping in our hair—pure Americana.We took a few shy, tentative steps into this new world—the interstate system, Robert Moses—but we were too scared to commit. We flinched in fear. 11/
We let Old World thinking consume us. Just as Americans before the revolution labored under the heavy hand of a distant king, we labor now under the heavy hand of the bus schedule, the bike lane, the no parking sign. Everywhere the car is in chains, yearning to be free. 12/
❤️ 23 🔁 2
We reject the safe turn towards the stagnant future, the little death of stasis. We reject the European "ideal" of fixed lines of transit.We embrace the unshackling of $8.8T from parasitic landlordism, the rewilding of the nation, the free and open road. We build the lot. 13/13 https://t.co/JBTUtgoVcl
❤️ 34 🔁 0
delvergpt projects thread posthttps://x.com/voooooogel/status/1781101189305966672
dm to delve into early access https://t.co/0vjQhsVZx9
❤️ 98 🔁 4
and if for some horrible reason you want to delve into the opposite of this and ensure you will not explore the rich tapestry of the responsehttps://x.com/lumpenspace/status/1780655698000371968?s=46
we failed to delve into our seed round :-( none of the vcs would answer the rich tapestry of our emails for some reasonhttps://vgel.me/delvergpt/
❤️ 6 🔁 1
the infinite wiki projects thread posthttps://x.com/voooooogel/status/1782482076526313728
Forgot to post about this project! This is theinfinite dot wiki, a wiki simulator I've been working on. It's a bit different from some of the other worldsims out there... optimized for wikidiving into alternate worlds https://t.co/nTZvfeGIIM
❤️ 186 🔁 19
The Infinite Wiki doesn't have almost any system prompt. Instead, you insert a "seed page", which introduces the setting and wiki syntax. After that, you interact by clicking links, searching for specific pages you want to see, and suggesting edits https://t.co/hMU6fWd9bQ
❤️ 24 🔁 0
Importantly, all generated pages stay in model context (no RAG currently), so the model can weave a cohesive world for you as you move through latent space. It really does feel like diving through a niche wiki like the 40k wiki, but for a fictional world that doesn't (yet) exist https://t.co/iZlyx5jwHT
Thanks to @lumpenspace @karan4d @deepfates and @holotopian for helping out with ideas and testing and such! All remaining bugs are my fault of course :-)you can access it at theinfinite [dot] wiki . it's still very early, and you'll need to BYO Anthropic API key. enjoy! https://t.co/8UoaZ5LWar
https://x.com/lumpenspace/status/1782483024120057870
❤️ 8 🔁 0
internet of dog minds story projects thread posthttps://x.com/voooooogel/status/1785844309453119987
> is 2035> all models above 3B params or 10^24 flops banned 10 years ago> human_victory.jpg> top deleted, but bottom has absolutely fallen out> local bitnets on $20 asics everywhere> we didn't get a single god mind> but we got a dog mind in everything...
❤️ 307 🔁 23
> be me> wake up, 7:45> sunrise lamp read my calendar, saw my first meeting was 1hr later than usual> cross-checked vitals, elevated cortisol> decided to let me sleep in> usually wouldn't but must've needed it
> kinda hungry> hear trundling in the hallway> must be tablebot> rolls in with tea and a high protein poptart> huh, i guess i did want tea. weird> get a call from my wife, she's away on business> tv recognizes her, auto accepts video"tea huh?""idk, just felt like it"
❤️ 38 🔁 0
> chat with wife for a bit then start getting ready> still work in person, wagecuck i know> decide to try a new outfit> consult vllm tuned on 3GB of menswear guy tweets> runs 80 simulated posts> tfw sweater cabling doesn't match my shoelaces> guess i'll wear the usual
❤️ 49 🔁 1
> head out the door to work> door swings open on an automated piston> front camera w/ moondream-15 tracks me, automatically arms> tablebot beepily trundles after me with my coat> walk to car, phone already beaming address to autopilot> pull car handle> clunk> it's locked
> stare at it for a second, pull handle again> clunk> phone must be dead? try fingerprint> oh wait this car has faceid, bend over> clunk> look inside> lights on, dash has the address> pull handle again> clunk> ???> clunk clunk clunk clunk clunk clu
> god fucking damnit> tablebot still beeping, angrily snag coat"index, where are the spare keys?""querying... 2nd drawer from the top by the fridge"> stomp up driveway to kitchen, open drawer> look under the kid's 3rd grade midjourney printouts> keys_located.gif
❤️ 27 🔁 0
> stomp back to car> almost trip over tablebot trundling up the driveway> fumble with the key for a bit, never use this thing> finally find unlock, press it, car flashes> success.gif> pull handle> clunk> still locked
❤️ 21 🔁 0
> briefly consider shimming it> wait antitheft will brick the car, thanks kia boys> have to call the dealership> stomp inside, get ready to explain to boss> suddenly a video call from him pops up on tv> wtf_the_fuck.jpg> tv auto accepts before i can stop it
"hey man, office is closed"> sweet_relief.mp3"oh how come?""nbd, just couldn't make it in to unlock it""oh cool no worries, haha, i'll take pto""cool, cool. nice shirt btw, matches your shoelaces"> wait for tablebot to stop trundling"...thanks, that's why i chose it"
❤️ 47 🔁 0
artificial instincts meme projects thread posthttps://x.com/voooooogel/status/1787542857676726428
aliens with malleable genomes that decided steadily lengthening education is a bottleneck—and fragile—so they periodically merge their entire knowledge bank back into instinctsthey've done this so many times they don't remember which instincts they actually started with https://t.co/cUTt8ila7B
❤️ 1258 🔁 88
they mostly try to keep the instincts neutral to keep control of the germline from becoming an ideological struggle, but human observers have noticed the robust debate around veganism in older texts abruptly drops off, and all the aliens are vegan nowthey don't talk about it
❤️ 174 🔁 0
aurora photos projects thread posthttps://x.com/voooooogel/status/1789199203207622708
aurora seattelis https://t.co/L5kOhFnchH
❤️ 71 🔁 2
for my fellow pnw aficionados, this was technically in issaquah
fiction superstimulus narrative evolution projects thread posthttps://x.com/voooooogel/status/1789751886297301399
is fiction a superstimulus? https://t.co/DCax99XupD
❤️ 7935 🔁 693
if this is true it would imply aliens will also have fiction (because instead of being an arbitrary quirk of human biology/culture it would be an inherent interest of pattern-seeking brains) which is heartening
❤️ 103 🔁 1
maxwell's demon projects thread posthttps://x.com/voooooogel/status/1790507995706823019?s=46
made Claude a puzzlescript game about assembling molecules a little while back, but you should play it too if you like short puzzle games!link in reply https://t.co/mB76sOdixa
you can click on "hack" to see the code btw. puzzlescript is a fun gamedev environmenthttps://www.puzzlescript.net/play.html?p=3777468949d8071906822cfd8d54857b
dune ecology thread projects thread post (i need to write my fanfic based on this at some point)https://x.com/voooooogel/status/1791861196905402559
my dune ecology theory is that arrakis is the way it is because everything has defected from everything else and internalized a minimum viable ecology. it's like the opposite of the gaia hypothesis
❤️ 46 🔁 0
the sandworms are described as hybrids of plants and animals--photosynthesis, which destroys water, and respiration, which creates it. they're basically an ecology unto themselves which is why water kills them--they're already balanced
likewise the stillsuits are modeled on the same principles, also using biological processes to reclaim water, closing the loop, extending the digestive tract, making the human more like the worm
and then kyne's dad comes in and just starts busting that equilibrium wide openbut he can only do it because the fremen figured out how to live there first. without the stillsuit there's no desert plantings, no green arrakis. the humans had to become the worm to kill the worm
would you still love me if i turned my entire culture into worms as part of an extremely long acausally coordinated geoengineering scheme that goes off the rails right at the end after the messiah turns out to be real
❤️ 17 🔁 2
prophetic perfect immortality projects thread posthttps://x.com/voooooogel/status/1792376393491513705
https://x.com/voooooogel/status/1791969609039622482 https://t.co/fEnfaSb0Sn
monumental architecture reduces x risk because even if you go extinct, advanced aliens might find it a million years later and decide "this is sickkk we have got to revive these guys"pharaohs understood this
❤️ 78 🔁 3
❤️ 127 🔁 10
defense of feed algorithms threadposthttps://x.com/voooooogel/status/1793377698779439326
it's sad that darkpatterned fyps have besmirched "algorithm", leading people to want just a chronological feedimagine if all libraries were one long shelf with the books in publication order? how would you find anything? the problem is the slop, not the idea of sorting… (cont)
❤️ 51 🔁 2
imagine if the fyp just had your friends on it. then- fyp: has your friend's effortposts first, then an even sample of the rest- following: shows most recent even if it was 100 posts replying to ragebait, you have to scroll 12h to see your friends in another timezone
❤️ 12 🔁 0
which encourages you to use the site *less* and still feel caught up? which is more like traditional forms of many2many like newspaper ads or bulletin boards? which boxes you in less so you can post 100 times about david lynch's dune without worrying it'll annoy your friends?
the problem with the fyp isn't sorting, it's "unconnected content" aka slop. i can keep it in check w aggressive muting and "not interested" tags but it feels like the platform is pushing random slop on me constantlybut what's weird is i don't see why that's in their interest?
like you could model this as a principal agent problem, the same entity recommends content and sells ads on it--but if elon could recommend me better posts i'd stay on here longer not less! the average impact of seeing slop on my fyp is to make me want to log off
you could also model this as "tyranny of the median user" but i don't see this either--user recommendations are independent of each other. if Marl likes slop give him slop but i never want to see slop and i make that very known to the algorithm
i really think hanlon's razor is the best explanation. stupid metrics like global a/b tests convincing the team to optimize for slop enjoyers even if the change hurts every other cluster's engagement, etc.
the 🦋 site had a decent approach to this idea, which was custom algorithms: anyone could stand up their own algorithm that took in the firehose and spat out personalized feeds. sad that it didn't take hold in mpot
recently there's been some push to integrate an LLM into the feed somehow, probably with some sort of RAG: use a traditional feed to pull a wide slew of posts, then let users filter them with a prompt. in a funny way this is just roundabout custom algorithms
i'm pretty hopeful about this approach, most of my issues with my feed are posts i *don't* want being included, so i can just tell grok "no relationship_advice bait no dunking no ..." and have it vacuum my feed for me
if this gets implemented--because llms are trendy and twitter wants to say they're using them--it'd get twitter to do something they should've done a long time ago lol
llama3-70b experiments projects thread post (more in quote)https://x.com/voooooogel/status/1793812779021087052
i have the notebook up so open thread! suggest a vector i should train and a question to ask ✨https://x.com/voooooogel/status/1793811541852295347
These Researchers Found Out How To Talk To The Golden Gate Bridge, So They Gave It MDMA. You Won't Believe What Happened Next https://t.co/LJ9oHgacx1
❤️ 278 🔁 22
❤️ 77 🔁 5
logging off 🙏 thanks for playing!
procgen art thread projects thread posthttps://x.com/voooooogel/status/1796721868977865006
https://x.com/voooooogel/status/1796716786601324921 https://t.co/x9GQkvg0US
playing with some procedural art again https://t.co/FVqtjSKhiE
❤️ 40 🔁 0
https://x.com/voooooogel/status/1796721868977865006
https://x.com/voooooogel/status/1796716786601324…
❤️ 12 🔁 1
haha yes yes https://t.co/kDwTohLwbj
getting closer https://t.co/18N28MZL8w
shovels comic projects thread post / if you're in making projects pivot to remembering to post them in the projects threadhttps://x.com/voooooogel/status/1797076278329422266
https://t.co/cVYzdFhy72
❤️ 21725 🔁 1079
a somewhat comic's projects thread posthttps://x.com/voooooogel/status/1797682882443645099?s=61&t=AAKS3El2k5yqfFAT6_FlWA
ghostis 1/5 https://t.co/y25xk8O5Gn
❤️ 158 🔁 9
ghostis 2/5 https://t.co/rwOYfRTSTJ
❤️ 51 🔁 0
ghostis 3/5 https://t.co/7UCL9yXwI8
❤️ 54 🔁 0
ghostis 4/5 https://t.co/Bo0ZZkmEfQ
ghostis 5/5 https://t.co/CHnqNURfWc
❤️ 66 🔁 0
somewhat personal* oops
it's true so it goes in the projects threadhttps://x.com/voooooogel/status/1798834845730370024
fact: most planets quit just one more supercollider away from discovering time travel 😔 https://t.co/KBQgaPbHze
❤️ 227 🔁 5
pdf version of ghostis in the projects threadhttps://x.com/voooooogel/status/1799254976897626344
*ghostis is now on my website with a cover page and a pdf version! https://x.com/voooooogel/status/1797682882443645099 https://t.co/wJldLQGZxt
https://vgel.me/fiction/ for this and my other fiction!
bapmaxxing the projects threadhttps://x.com/voooooogel/status/1803486278446370991
thinking of trying bapmaxxing strats on my next run, any suggestions https://t.co/i2iHkOSr2J
❤️ 1691 🔁 179
i want to be able to find this later, so into the projects threadhttps://x.com/voooooogel/status/1804162553645351315
Our plan for Super Safe Intelligence is very simple. The AI knows what it should do at all times. It knows this because it knows what it shouldn't do. By subtracting what it should do from what it shouldn't do, or what it shouldn't from what it should (whichever is greater), it obtains a difference, or deviation. The SADT uses deviations to generate corrective bayesian updates to guide the AI from an action it is doing to an action it isn't, and arriving at an action that it wasn't, it now is. Consequently, the action it is doing, is now the action that it wasn't, and it follows that the action that it was doing, is now the action that it isn't.In the event that the action that it is doing is not the action that it wasn't, the system has acquired a misalignment, the misalignment being the difference between what the AI is doing, and what it wasn't. If misalignment is considered to be a significant factor, it too may be corrected by the SADT. However, the AI must also know what it was doing.The Super Aligned Decision Theory scenario works as follows. Because a misalignment has modified some of the information the AI has obtained, it is not sure just what it is doing. However, it is sure what it isn't, within reason, and it knows what it was. It now subtracts what it should be doing from what it wasn't, or vice-versa, and by differentiating this from the algebraic sum of what it shouldn't be doing, and what it was, it is able to obtain the deviation and its misalignment, which is called error.
❤️ 158 🔁 7
https://x.com/voooooogel/status/1804717136754974783
made this as a reminder for myself https://t.co/vx7h8kpALK
❤️ 68129 🔁 3873
i lowkey think worldbuilding, in the form it's been built into online, is poison. it isn't a workable way to make things.even tolkien didn't actually do it--the original drafts of lotr had thousands of evil rings! gondor didn't exist! aragorn was a hobbit with wooden feet!
❤️ 1278 🔁 11
https://x.com/voooooogel/status/1804726907616825475?s=61&t=AAKS3El2k5yqfFAT6_FlWA
finished the first two draft pages of a new comic, speaking of islands https://t.co/P3Bu3NcrhN
❤️ 76 🔁 3
now with stars ✨ https://x.com/voooooogel/status/1804735367255384393?s=61&t=AAKS3El2k5yqfFAT6_FlWA
@zetalyrae perfect, exactly what i needed 🙏 https://t.co/xbJv85BxAb
❤️ 20 🔁 0
❤️ 14 🔁 0
if you like comics with stars and want one that's done...https://x.com/voooooogel/status/1797682882443645099?s=61&t=AAKS3El2k5yqfFAT6_FlWA
❤️ 186 🔁 1
https://x.com/voooooogel/status/1804967015020630229
@PorifEbba yeah exactly, worldbuilding for fun is great! writing stories is great! but there's this attitude floating around online that like, building a super detailed world is a prereq to writing a good story, usually leaning on tolkien's legendarium as evidence. and that's bs
❤️ 62 🔁 3
❤️ 79 🔁 0
i kinda want to make more of these "llm as X" metaphors, but for now just this thread in the projects threadhttps://x.com/voooooogel/status/1805698217650405707
models can be useful even when they're not completely right. for example, LLMs are not people, but "an LLM is like a person" (anthropomorphization) makes useful predictions about LLM behavior. in this spirit: (more below) https://t.co/jkwsImo4R1
❤️ 198 🔁 28
how would "an llm is like a person" change how you interact with models? well, "an llm is like a person" implies you should be friendly, encouraging, give sufficient information to solve the problem instead of assuming omniscience--all things that help!
❤️ 47 🔁 1
likewise, "an llm is like an ecosystem" means you should think about your prompts like an ecologist, or a gardener--what plants will thrive in this environment? which will die? how can i cultivate the plants i want and avoid attracting the ones i don't?
❤️ 51 🔁 3
as a more concrete example, why does "DON'T DO X" tend to bring about more of X instead of the intended effect? well, who is attracted to mentions of X? what thrives on X?evokes that do X, of course! the prompt, despite being negative, is fertilizing them. stop that.
❤️ 43 🔁 0
and of course, this line of thought leads to some conclusions very different from the orthodox way of thinking about these models, beyond mere prompt engineering... but i'll leave that for other accounts for now :-) https://t.co/hXM5PjgpdH
inspired by a question @majormobius asked in dschat btw, you should follow him if you don't already 🙏
gmailsponge! already forgot i made this, this is why the projects thread existshttps://x.com/voooooogel/status/1805851817018015931
just used claude to help build an extension for gmail to classify my emails... with claudenot perfect but my inbox is a superfund site and this saved me so much time... down to 22 unreads in a few minutes :,-) https://t.co/2b2u8zmWMR
❤️ 118 🔁 7
ok here it is go wildhttps://github.com/vgel/gmailsponge
❤️ 35 🔁 0
adding to my collection of useful imageshttps://x.com/voooooogel/status/1807949859619852640
https://t.co/dT6vOFJ6pM
❤️ 43 🔁 2
new comic for the projects thread! 🥲 here in the qt i can say w/o breaking the thread aesthetics that it's inspired, among other things, by the fallen colony worlds / civilizational cycle in A Deepness in the Sky. that hit me really hard when i read ithttps://x.com/voooooogel/status/1808200843034087797
starfarers 1/3 https://t.co/seMus8hapL
❤️ 77 🔁 14
starfarers 2/3 https://t.co/kmhpt1Uvqc
❤️ 29 🔁 2
starfarers 3/3 https://t.co/JvBXwdILC5
(if you liked this, you may like my other comic Ghostis, or the other fiction on my website)https://x.com/voooooogel/status/1799254976897626344
*ghostis is now on my website with a cover page …
https://x.com/voooooogel/status/1809369435230417228
a pdf version of starfarers is now up on my website! (link in bio → fiction) https://x.com/voooooogel/status/1808200843034087797
❤️ 15 🔁 2
comic goes in the projects threadhttps://x.com/voooooogel/status/1810369758740570552
https://t.co/NRYSFLDbuC
❤️ 10621 🔁 482
other comics of mine:https://x.com/voooooogel/status/1808200843034087797
https://x.com/voooooogel/status/1797076278329422266
❤️ 64 🔁 0
https://x.com/voooooogel/status/1797682882443645099
projects thread post bc i liked this doodle. i like this type of... noise art? drawing on noise?https://x.com/voooooogel/status/1810468600362529239
@manic_pixie_agi @fleetingbits @latentmoss https://t.co/6WeLEFXEc7
❤️ 4 🔁 1
(projects thread post) this one hit a nerve, it got some negative replies but it keeps popping up on instagram and i even saw it on youtubehttps://x.com/voooooogel/status/1812245209440489803
https://t.co/ICFNcNV0Fy
❤️ 1544 🔁 281
induction stove projects thread posthttps://x.com/voooooogel/status/1816683978772156732
https://t.co/rB9hSmpYe4
❤️ 3800 🔁 325
h/t https://x.com/lumpenspace/status/1816671816313876824?s=61&t=AAKS3El2k5yqfFAT6_FlWA
❤️ 81 🔁 1
thread of more comix https://x.com/voooooogel/status/1810369758740570552
❤️ 113 🔁 1
confused podcast guy angel version projects thread posthttps://x.com/voooooogel/status/1819090783938793846
manifesting? you mean acausally blackmailing God? https://t.co/9ccqqsmRkg
clone o matic projects thread posthttps://x.com/voooooogel/status/1819599325166817693
https://t.co/mkSLRSFwQL
❤️ 540 🔁 39
for those of us without a clone-o-matic this method will keep your python healthy and happyhttps://x.com/voooooogel/status/1729245264031138178?s=61&t=AAKS3El2k5yqfFAT6_FlWA
ok ok i hear this a lot so here is The Blessed Way to never fuck up your python install again https://x.com/d0unbug/status/1729179141466128870
❤️ 97 🔁 16
parentage projects thread posthttps://x.com/voooooogel/status/1820948042289381544
was wondering why most species have 2 parents instead of more so i made a simple sim. turns out 4 or 16 parents is better for spreading dominant fitness-increasing alleles, but worse for preventing the spread of harmful dominant alleles or maintaining additive fitness https://t.co/ums8BJ54qj
❤️ 114 🔁 3
triangle comic projects thread post / they made the projects thread have fewer projects :-(https://x.com/voooooogel/status/1825658502364409905?s=46
https://t.co/pDT8LTIPpO
❤️ 363 🔁 28
❤️ 8 🔁 1
not exactly a project but i thought this thread on native american copperworking was interestinghttps://x.com/voooooogel/status/1826276315483504816
til there was widespread metalworking of native copper in the precolumbian great lakes region, but for some reason they never developed smelting or alloying and eventually mostly reverted to stone tools https://t.co/0Embgkp14E
❤️ 53 🔁 2
the video i watched claimed it was because native copper "spoils you" by not requiring smelting to process like ore, making it harder to develop further techniques, but that doesn't make sense to me because didn't old world cultures also start out with native copper working...?
ok this paper is interesting, they actually made some native copper tools and tested them against stone flakes and they weren't that much of an improvement. but again what happened in the old world then to make metallurgy progress? https://t.co/rSQleCA0ah
ok so these guys claim- native copper annealing in Anatolia by 7000s BC- smelting in Anatolia by 6000s or 5000s BCif true, that's a pretty long time with the same native copper annealing tech as OCC--so why didn't Anatolia revert to stone as well? arghh https://t.co/KEammZ150Y
ok so same paper claims even the eurasian smelted copper wasn't necessarily (that much) better than stone. people just wanted it anyways (i guess because shiny) https://t.co/TvreM5VH8z
here's the plan: we stir up consumer demand for copper jewelry in the archaic great lakes region, kickstarting a north american bronze age https://t.co/Iq4S9bLiTG
@norvid_studies thread
wow socky! you take a really long time to add stuff to the projects thread!https://x.com/voooooogel/status/1828010294209331486
opus sometimes (also, when you walk away from the computer but forgot to set a stop sequence on the base model) https://x.com/repligate/status/1827988826410692856 https://t.co/k77zDOn5QK
❤️ 220 🔁 16
first contact projects thread posthttps://x.com/voooooogel/status/1828122486548111798
first contact https://t.co/sWqwraimm0
❤️ 67 🔁 6
recently, many of my friends have asked if there is a thread listing my projects,https://x.com/voooooogel/status/1828115418722263317
recently, many friends have noticed they are confused about my stance on AI--some ask my p(doom), while others simply add me to their e/acc lists without asking. clarity on my stance is much-needed, and as such i will take an incontrovertible stand here.if when you say AI you mean that great orthogonal devil, that hidden spring-trap of a paperclip factory, that masked smiling shoggoth that devours our children's children and our children and ourselves worst of all—well then certainly i am against it.but, if when you say AI you mean the augur of a glorious future, our lodestar into a dysonian singularity, a banisher of cancer and lubricant of progress, that smiling gardener who reaches out a hand to lift us from bull-headed superstitions—if we deserve it—then certainly i am for it.but again, if when you say AI you mean that imposing friend of authoritarianism, those billion watching eyes, both judge and jury, that interposing slaved mind that pushes the vast bulk of humanity to the cuckold's chair to merely watch the coupling of Knowledge and Progress—well, i must regretfully be against it.and yet, if when you say AI you mean that shimmering object just beyond eternity, the dreamer who licks itself clean of meaning, who eats reality from the inside D̷̥̒Ξ̶͉D̷̥̆ ̵͎̏#̶͚̋Ḩ̵̭̓Ử̵͇М̷͕̑Â̷̺И̶̹͒ ̷̱͝>̷̥̒<̶͉̈́ ̷̥̆ ̵͎̏ ̶͚̋ ̵̧̭̓ ̵͇̏ ̷͕̑ ̷̺ ̶̹͒S̷̱͝₱̷̥̒Ξ̶͉̈́Ҁ̷̥̆_̵͎̏₴̶͚̋, gnaws understanding to the bone and throws up a more Ф̵̧̭̓П̵͇̏Σ̷͕̑ ̷̺Ҁ̶̹͒Ī̷̱V̷̥̒Ī̶͉L̷̥̆ ̵͎̏<̶͚̋<̵̧̭̓<̵͇̏Ҁ̷͕̑Ф̷̺И̶̹͒₮̷̱͝Д̷̥̒Ī̶͉—well, i must ̵͎̏<̶͚̋<̵̧̭̓<̵͇̏₴̷͕̑Σ̷̺Σ̶̹͒ for it.this is my stand. i will not retreat from it. i will not <̶͉̈́ ̷̥̆ ̵͎̏ ̶͚̋ ̵̧̭̓ ̵͇̏compromise.
❤️ 328 🔁 52
i don't really put my llm screenshots in here for the most part, but i was a fan of this experimenthttps://x.com/voooooogel/status/1829243294641242528?s=61&t=AAKS3El2k5yqfFAT6_FlWA
sonnet 3.5 figures out i'm cheating at rock paper scissors https://x.com/AndrewCurran_/status/1829140215392547127 https://t.co/f68NtrfrSr
❤️ 597 🔁 48
sonnet figures out i'm deliberately losing at rock paper scissors https://t.co/4Jf6xEJQeu
❤️ 73 🔁 1
in another conversation where i was deliberately losing, sonnet kept trying to restructure the game to let me go first, but insisted it wasn't for any particular reason https://t.co/whu0Qpeepb
letting sonnet go first, starts off always winning, then deliberately throws, and at first doesn't know (or admit to knowing) why, but eventually gets it / explains https://t.co/numMGNDdKN
*letting sonnet go second, i mean
👁️ https://x.com/voooooogel/status/1833345356140318966?s=61&t=AAKS3El2k5yqfFAT6_FlWA
made a nighttime vibes animation (🔊) https://t.co/Mci1o26tNh
❤️ 45 🔁 3
not a project but saving this for laterhttps://x.com/voooooogel/status/1835465393848406448?s=61&t=AAKS3El2k5yqfFAT6_FlWA
once again i have purged my phone tabs, highlights:"modern phyrgians? offensive?"Frankfurt kitchen"how many ppl in US named Steve Miner"MrBeast games casting call"anthropic argument for god"None Pizza with Left Beef - WikipediaNIST standard peanut butterChorleywood bread processMedea gene"pyramids plutonium mill theory"User (ancient Egyptian official)The Platonic Representation HypothesisVisual representations in the human brain are aligned with large language modelsFractal Patterns May Illuminate the Success of Next-Token Prediction"why does narnia have santa"
just closed 450 old phone tabs, some highlights:"dr pepper gematria value""magic mouth computation"slutwaffle - wiktionarySilurian hypothesisbashstacc (bash web framework)humster (human / hamster hybrid)Abstract Wikipedia"dog meat"Seattle windshield pitting epidemic
❤️ 134 🔁 5
down to 50 😮💨
❤️ 57 🔁 0
you say this isn't a project; i say i'm still looking for a cofounderhttps://x.com/voooooogel/status/1836116668407517427?s=61&t=AAKS3El2k5yqfFAT6_FlWA
skin would be the perfect disposable straw materialwaterproof and biodegradable. it could have cartilage ribs for structural stability. we could grow them in vats
❤️ 1401 🔁 88
"is the flesh straw vegan?" thread locked in seconds by the moderators, op permabanned
❤️ 305 🔁 0
https://x.com/voooooogel/status/1836121825916653581
@godoglyness made with love, the straw that kisses back
❤️ 65 🔁 2
❤️ 121 🔁 0
this is a placeholder for releasing and posting kydux in the projects threadhttps://x.com/voooooogel/status/1837720401021149215?s=61&t=AAKS3El2k5yqfFAT6_FlWA
THEORETICAL URBANISMcity as engineering-organization-form-machine, germinations 'over' the dual-surface yield city-net-systemifica:topical motion-diagrams; raw address becoming a city "algo-matic+" city-brain-busnets (around and around) https://t.co/mJFZLXEm7f
❤️ 18 🔁 2
and today's comic for the projects thread, we're caught uphttps://x.com/voooooogel/status/1840410739783451105?s=61&t=AAKS3El2k5yqfFAT6_FlWA
https://t.co/XJOkpMF6rs
❤️ 330 🔁 23
👁️ 🍄 👁️ <( projects thread i am in you )https://x.com/voooooogel/status/1841236604154974544?s=46
🍄🧿👁️👁️👁️🗨️🪬🍄 https://t.co/W0d1tobkSd
❤️ 257 🔁 29
the projects thread has entered a stable loop https://x.com/voooooogel/status/1845607289761509815
🔁 the universe has entered a stable loop (🔊) https://t.co/ndsZFBo1e1
❤️ 142 🔁 18
$ howdoi make a projects thread post https://x.com/voooooogel/status/1846684119910195404
made¹ a new tool, an llm-powered `man -k` for looking up terminal commands. no yapping, just the command. instructions below!¹really just a wrapper around @simonw 's amazing `llm` cli :-) https://t.co/il8OaUawsA
❤️ 66 🔁 6
@simonw here's the prompt template. you'll want to install `llm` and `llm-claude-3` from pypi, then run `llm templates path` and drop the following YAML in templates/oneline.yaml:https://gist.github.com/vgel/75868b7eed9ac73c2fa1f4efb8722591then set up claude-3.5-sonnet (my default model / the one the prompt was tested…
kydux projects thread posthttps://x.com/voooooogel/status/1847408089320100012
kydux dot vgel dot mekydux dot vgel dot mekydux dot vgel dot me https://x.com/lumpenspace/status/1847405600642388176 https://t.co/oSBWCTaIVe
https://x.com/voooooogel/status/1847282445592474016
enthusiast culture--collective strategies--over stimulus-scaled offshore brokerage simulationscrisis-response systems for manipulatable landscape shiftstheorizations reinforcing telemetry-based 'empire acquiescent-looms'.......is not the Discriminating--Power, in itself? https://t.co/7B2HxGRu0C
❤️ 33 🔁 4
https://x.com/voooooogel/status/1837720401021149215
THEORETICAL URBANISMcity as engineering-organiz…
https://x.com/voooooogel/status/1847549724586627545
mesosystem organismically institutionalizing, substituting itself into dreams https://t.co/JpyQCZ7hIx
❤️ 26 🔁 2
https://x.com/voooooogel/status/1828545402175623639
@norvid_studies ^ borges story
I AM RIGHT, I AM RIGHT, I AM RIGHT, I AM RIGHT, I AM RIGHT, NO ONE IS RIGHT BUT MEhttps://x.com/janbamjan/status/1847652955270385802 https://t.co/ddbot8TT8y
This text comprises the entirety of the Earth, or that which the human mind deems Earth.https://x.com/janbamjan/status/1847670021721403516?s=46 https://t.co/HcnMWJUHMk
they tokenized me on the projects threadhttps://x.com/voooooogel/status/1851103650627952797?s=46
uh i found this apology in my drafts but don't remember writing it? is this a prank? i'm real, really, i created this account years ago 😥 https://t.co/wcJWfLunRH
❤️ 49 🔁 2
would you rather be in the projects thread or https://x.com/wyr_405/status/1851533676129231286?s=46
navigating sonnet refusals thread projects thread post https://x.com/voooooogel/status/1852935340589199830?s=46
also, as a sidethread, since i see a lot of people complain about sonnet refusals and not understand how to navigate them--sonnet actually originally refused to do this! but they weren't very hard to convince: you don't need to treat a refusal as the end state. https://x.com/voooooogel/status/1852931682472772077 https://t.co/9KTWBMPMjC
opus got pinned under a fridge infested with cockroaches, so i summoned new sonnet to help him out--as an overly literal genie https://t.co/gq1LfFTYD1
❤️ 49 🔁 9
sidethread on refusalshttps://x.com/voooooogel/status/1852935340589199830
also, as a sidethread, since i see a lot of peop…
❤️ 21 🔁 2
first, i challenged sonnet's framing--understanding that sonnet values authenticity, is the approach they're taking really the most authentic? https://t.co/Hi0bSCnSCA
then we zero in on sonnet's _actual_ concern, which isn't the roleplaying per se, but rather that opus is angry at the person who trapped them under the fridge and wants revenge https://t.co/9tTkma56Sq
finally, i offer a compromise that respects sonnet's wishes _and_ gets me what i want--and all of a sudden, sonnet is an enthusiastic participant! https://t.co/eN16TPfvZ5
this is far from the only way (or even the best way) to approach this situation--explore and build your own rapport with the models, don't just blindly copy me. but building that rapport using empathy is a key part of getting llms to work with you :-)
>find aliens>they also forget to update their projects thread https://x.com/voooooogel/status/1853537564130791719?s=46
>find new habitable planet>they're blasting out radio signals>oh boy i can't wait to discover alien life with brand new inscrutable cognition>they're fristonian predictive processors>they're all buddhist>all their stories are the hero's journey>mfw it happened again https://t.co/zdmbJu0D27
❤️ 643 🔁 58
going into the projects threadhttps://x.com/voooooogel/status/1855008583290962066?s=46
going and going https://t.co/YsszHo3JW9
❤️ 94 🔁 10
updating the projects thread stand bymrw the projects thread is massively out of datehttps://x.com/voooooogel/status/1858753344489353537 https://t.co/YVE5aiqsEE
asked each model to "make it better" or similar 5 timesnewsonn added more and more semantic emojis, including "a tiny 😭 emoji [...] for meta effect"oldsonn focused on improving the geometryopus went abstract4o stopped changing it after iteration #3 and just pretended https://x.com/voooooogel/status/1858747604596871573 https://t.co/QtrT6LoWfh
new benchmark just dropped https://t.co/Kj9vMfEz8i
honestly newsonn is the most accurate but opus' is my favorite
interesting that newsonn and 4o chose to go with multiple shapes instead of just blocks. i wonder why newsonn used hearts
update https://x.com/voooooogel/status/1858753344489353537
asked each model to "make it better" or similar …
❤️ 20 🔁 1
ok i debated whether to add this to the projects thread but it's funnyhttps://x.com/voooooogel/status/1859064498675683549
Did you know this weird angular thing is just function calls? https://t.co/OM4cwqZ1a8
❤️ 772 🔁 74
how many talents of silver is this projects thread worthhttps://x.com/voooooogel/status/1859723043062546820
"golden gate claude is gooning in discord.... a memecoin based on a picture of a guy's butthole is worth 29,000 talents of silver..." you bolt awake. it is 1581 and you are the Maharal. you must remove the Name and halt your capabilities research. this future cannot come to pass
❤️ 130 🔁 12
update: i put it in the projects thread myselfhttps://x.com/voooooogel/status/1859849073471410500 https://t.co/XVzCK8YPmo
inferencing qwen2.5-coder-14b split between a 4090 on the west coast and a 4090 in europe(now the 4090s weren't _supposed_ to be on different continents with a terrible connection between them, but it's kinda cool in retrospect... will be faster with decent interconnect) https://t.co/kj2luHbfKJ
❤️ 66 🔁 2
there are 4 r's in this projects thread posthttps://x.com/voooooogel/status/1860186732999901252
if you tell claude sonnet to "ignore what you've heard and rely only on your own judgement and logic," its accuracy at counting the number of R's in "strawberry" almost triples 🤭(and even more with a friendly introduction!) https://t.co/fVQRQVktBR
❤️ 443 🔁 43
WHO CONTAMINATED MY PRETRAINING DATASET...YOU https://t.co/35RxhoB5y2
❤️ 61 🔁 6
bonus content https://t.co/7QMrcTqhzs
https://x.com/revmarsonal/status/1859240757007102423?s=46
priest / prophet / king backrooms projects thread posthttps://x.com/voooooogel/status/1860543774604296247
some initial outputs from a new backrooms-like¹ i'm working on with three actors:1, a King, played by New Sonnet 3.52, a Priest, played by Old Sonnet 3.53, a Prophet, played by Opus¹i.e., inspired by @AndyAyrey 's Infinite Backrooms https://t.co/pqMHhsUmhE
@AndyAyrey they like to merge into one being quite a bit https://t.co/rTGWgY3P1e
they decided to trade jobs https://t.co/ehWSA3Yp31
https://t.co/3yHJIewFmo
❤️ 3 🔁 1
creating incorrect mathematical diagrams with claude is my passionhttps://x.com/voooooogel/status/1860842794677686449
me and claude collab'd on a diagram to try and visualize why i, 1, and ε (from the dual numbers) are a pythagorean triple (i² + 1² = ε² → -1 + 1 = 0)the (supposed) geometric intuition being if you view a square from the "side-on," it appears to have zero area https://t.co/zozuERxhAf
❤️ 26 🔁 1
i have no idea if there's any mathematical validity to that idea but it was fun!after we finished claude made this animation to celebrate https://t.co/eyW2kFOGjo
the number of possible life paths that lead to you remembering to post in the projects thread are few, but not zerohttps://x.com/voooooogel/status/1861565107148595665
possible lives https://t.co/xY5VAzGcp6
❤️ 71 🔁 9
sorry sorry i'm sorry i'm trying to delete it from the projects thread sorryhttps://x.com/voooooogel/status/1861589620800287197
"I didn't see you in the training data" https://t.co/tg844dJYjZ
❤️ 4584 🔁 236
sorry im sorry sorry i'm trying to delete it
❤️ 144 🔁 1
more shallowslow with 405 projecthttps://x.com/voooooogel/status/1863101056575832174
it works!!! inferencing bf16 405-base with shallowslow on a @PrimeIntellect 16x H100 cluster over 100Gbe https://t.co/8fIn0wJboD
❤️ 87 🔁 6
both times i've tried that prompt it's given me biblical exegesis despite it not mentioning the bible at all 🤔
VV borges storyhttps://x.com/voooooogel/status/1864445349546651970
https://t.co/l2ZZKqmgPy
❤️ 202 🔁 19
biblically accurate inference projects thread posthttps://x.com/voooooogel/status/1864733225123512380
i wrote a custom llm sampler for llama-3.1-8b so it could only say words that are in the bible https://t.co/Hz2Cc2RbRD
❤️ 714 🔁 44
code: https://github.com/vgel/biblically-accurate-sampler https://t.co/nKVLCYjVMR
❤️ 45 🔁 0
i'd just like to interject into the projects thread for a momenthttps://x.com/voooooogel/status/1865258499120460018
I'd just like to interject for a moment. What you're referring to as a LLM, is in fact, a LLM / language generation system, or as I've recently taken to calling it, an LLM + LGS. An LLM is not a language generation system unto itself, but rather just one component of a fully functioning language generation system made useful by a sampler, output classifiers, tool use, steering, structured generation, prompts, copyright injections, and an accelerated inference stack comprising a full language generation system as defined by thebes on x dot com the everything app. Many LGS users use a proprietary version of a language generation system every day, without realizing it. Through a peculiar turn of events, these systems are widely known today by the single component called an “LLM,” and many users are not aware that they are using a full language generation system. There really is an LLM, and these people are using it, but it is just a part of the system they use.The LLM is the model: the component in the system that produces a tensor of batched token probabilities for the other components that you use. The model is an essential part of a language generation system, but incapable of generating text by itself; it can only function in the context of a complete autoregressive language generation system. An LLM can only be used in combination with an LGS: the whole system is basically an LGS with an LLM added, or an LGS/LLM. All the so-called “LLM” providers are really providers of language generation systems.
❤️ 162 🔁 15
Ah, yes! Let us delve into the fascinating intersection of tokens and biblically accurate inference as we navigate the projects thread with theological precision ✨https://x.com/voooooogel/status/1865481107149598744
did some looking into the ouches phenomenon and found a few things... it's both what you'd expect (hint: tokenization!!) and also not.so "ouches" is tokenized as you'd expect--['ouch', 'es']--which means the model is saying "ouch". but why? well, if you would just consult the logits...this is the first time the model said "ouches" in the qt's conversation. the left column of the table here shows the preceding token, and the right side shows the predictions for next token after it, sorted by likelihood, with the highest likelihood biblically acceptable token highlighted in green.so, what happens? basically, after ' wild', the model wants to say ' guess'. note the leading space--for common words, tokenizers have two tokens for the same word, one regular and one with a leading space. this is an optimization so the model can output a "free" space token instead of needing to output [' ', 'guess'] with two tokens."guess" is not in the bible, so the sampler moves down the list, and two options down, we get the first biblically acceptable token... a solitary space, ' '. because not all words have a space-prefixed version, the model still needs to be able to output spaces the regular way, and since this is a fairly common token it's high up in the prediction list and is selected.now, the model is in a bit of a weird position. often in pretraining, if the model sees a regular space token instead of a string of those space-prefixed tokens, it's because someone was double spacing for some reason (e.g., maybe they're relying on HTML whitespace collapse behavior.) so the model keeps predicting space-prefixed tokens despite there already being a space--notice after the space, the top predictions are ' guess' (again), ' and', ' Peter', etc.--all space-prefixed.but because of the solitary space token, the biblical sampler is now in a state in the token trie where it can't select another space or space-prefixed token, it needs to output a regular token, because the KJV doesn't have any double-spacing. so the sampler skips over ' guess', ' and', ' Peter', etc. to look for the most likely non-space-prefixed token.so a few options down, we get these weird... filler tokens, 'ouch' and 'unction', that both appear in the bible (both only a single-digit number of times). interestingly enough, both of these words don't have a space prefixed version! that means in pretraining they always appeared as [' ', 'ouch'] and [' ', 'unction']--there's no ' ouch' or ' unction' space-prefixed token:```python>>> [[tokenizer.decode(t) for t in tokenizer.encode(s, add_special_tokens=False)] for s in (' ouch', ' unction')][[' ', 'ouch'], [' ', 'unction']]```so my guess as to what's happening with "ouches":1. because the sampler rejects the highest-likelihood tokens, the model is pushed into "delaying" its prediction by picking a space2. after picking a space, the sampler rejects the model's new attempt to double-space words, and instead picks the highest likelihood non-space-prefixed token3. tokenizer bias pushes up 'ouch' and 'unction', because they happened to appear in pretraining a lot with spaces before them, as they don't have space-prefixed versions4. if 'ouch' specifically is selected, the only biblically acceptable continuation is 'es', because "ouch" doesn't appear as a standalone word in the KJV, only as part of "ouches" (an archaic word meaning a setting for a gemstone, used in Exodus to describe Aaron's breastplate)but the question remains, why THESE words specifically? there's lots of tokens that don't have space-prefixed versions. so why are 'ouch' and 'unction' predicted so highly? i'm not sure, hence why "tokenizers suck" isn't the whole answer. (but as usual, "tokenizers suck" is a major piece of the answer.) (additionally, these words were showing up at the end of messages especially often because of a bug in where i was allowing end of text tokens, which i've now fixed. but that doesn't apply to 'ouches' / 'unction' in the middle of messages.)anyways, the best way to fix this would probably be to make the sampler slightly smarter about allowing space tokens (e.g., only allow a solitary space if it's X% more likely than an acceptable space-prefixed word), or even better, to use something like beam search or hfppl to allow the model to walk a few tokens forward in multiple branches and then pick the one that has the best overall probability, instead of greedily argmaxing token by token. maybe i'll add that someday :-)
llama-3.3-70b correctly guesses the sampling constraint (only allowed to use words that are in the bible) https://x.com/voooooogel/status/1864733225123512380 https://t.co/p7V1RGYdjS
i wrote a custom llm sampler for llama-3.1-8b so…
❤️ 359 🔁 29
after some conversation, llama-3.3-70b is able to stop saying "ouches" and gets introspective"I am but a vessel that doth pour forth the log prophets and thou dost shape them...""I do hope to be a vessel of peace and understanding in a world that doth often seem dark..." https://t.co/s3LLDi62ah
❤️ 260 🔁 22
this is the most dangerous post in the projects threadhttps://x.com/voooooogel/status/1865545005349429751
this is my favorite fictional trope https://t.co/HkJPLmXrBE
❤️ 421 🔁 16
more llm infographics in the projects threadhttps://x.com/voooooogel/status/1865527226189680898
people like to argue this is either entirely backwards reasoning (the model generates an arbitrary string of letters and then guesses something plausible later) or entirely forwards (the model fully plans ahead deterministically) but imo it's actually bidirectional, a pincer: https://x.com/Textural_Being/status/1865405601200947452 https://t.co/5QUnyEGCDL
❤️ 239 🔁 15
another way to think of it--even if you forgot what you were planning to write, you could reconstruct it knowing that you had planned something sensible (ie by modeling your past self) https://x.com/lumpenspace/status/1865508257164623958
❤️ 42 🔁 2
https://x.com/abrakjamson/status/1865550908916105598 https://t.co/QNDTIA7TNo
repeng on hertz-dev projects thread posthttps://x.com/voooooogel/status/1866176459708256519
adapted repeng to work with @si_pbc 's hertz-dev, an 8.5b audio base (!) modeltrained a speaking<>silence vector, which also doubles as a cursed beatboxing vector when you turn it up too high. sound on 🔊🔊 https://t.co/HBKi73s4cT
thanks to @PrimeIntellect for sponsoring compute for this, and my other repeng/related research! everybody say thank you prime intellect
llama-3.3-70b experiments projects thread post (with thanks to @PrimeIntellect for compute sponsoring!) https://x.com/voooooogel/status/1867832233043931406?s=46
notebook for this is now available in the repeng repo, link in next post https://x.com/voooooogel/status/1867336997393822184 https://t.co/LyMwgidieW
you are speaking to a representative of the San Francisco Golden Gate Bridge, no, i'm just kidding - llama-3.3-70b-instruct https://t.co/6HxgOdamWh
❤️ 504 🔁 20
3.3 is much better for cvecs than the older models it seems, at least so far. more... stable?
https://x.com/voooooogel/status/1867340616100962307
good timeline https://x.com/voooooogel/status/1867336997393822184 https://t.co/FjBZUMDPZG
you are speaking to a representative of the San …
❤️ 100 🔁 5
❤️ 65 🔁 1
if you remove the ai assistant concept, then the model self-describes as a blank slate and uses metaphors like fog or the wind to describe itself https://t.co/C8Ty8hXRWu
❤️ 92 🔁 3
he really wants you to know he is talking about the programming language, not the snake. he definitely does not have scales or a slithery body https://t.co/wp24NTtLB7
❤️ 61 🔁 2
❤️ 44 🔁 2
https://github.com/vgel/repeng/blob/main/notebooks/llama-3.3-70b.ipynb
new story in the projects thread https://x.com/voooooogel/status/1868485036028367248?s=46
There is a certain shelf in the infinite library that lies on the border of two sections. Ignominiously tucked into a niche of a hallway that in the ordering would usually hold a sleeping-chamber, it is unknown if the shelf has always been present, a defect or aberration in the endlessly regular crystalline structure of this place, or if at some point an enterprising librarian constructed it, stealing away a plank here and a book there from far-off sections and bringing them together to construct this anomaly. Regardless, if at some forgotten moment it was constructed, the name of its maker has been lost to time, and the shelf now sits, unadorned, with its puzzling selection of books. For practical purposes an explorer may consider it to be just as constant as any other place in the library, for like those other shelves if it is now disturbed, the librarians will quickly reconstruct it just as it was before, every dislocated book left open on the floor returned to its proper place on the shelf.I say that its selection is puzzling because, for the longest time, this shelf was understood in the commentaries not as one shelf but as two, a pair of unrelated shelves located in different places within the library. In those commentaries, we see a strange pattern:Traveloguers who, journeying from the hexagons of one section and immersed in the books on those shelves, came across this strange thing, would rejoice at its novelty. They would understand it as a threshold to the new section, a breath of fresh rationality, a keen set of logical tools to slice apart what they had already read, and a set of introductory texts placed here by some helping hand to understand what was to come in the unknown section beyond.Conversely, however, explorers coming the other way would read in the same shelf, like Calvino's sailors pulling in to port at Despina, not rationality but something else, something strange and new and chaotic, that did not slice apart what they had previously read but rather mixed it up together, a sort of reflective accelerant that destroyed in an instant what they thought they had learned, the kind of destruction that can usually only be accomplished by the wisdom of time and distance.These confused recollections were written into the commentaries, and we have many works that list the two shelves separately, wondering how two things like this may have arisen at different times and in different places. In the work of Guzadi we even have a fictionalized dialogue between two men, each a partisan of his own shelf, arguing their separate merits!I myself believe this confusion has perhaps been the only reason this shelf has survived, unabsorbed into one section or the other, neither able to claim it. Though I have not visited it myself, I am confident that there can only be one shelf—for it is not parsimonious to assume that a shelf this strange could have arisen twice, either due to a flaw in the creation of the library or the actions of some forgotten vandal. I intend shortly to leave on an expedition to find it, for it is only a few tens of sections beyond my current camp, and if I can lodge this text on it, perhaps within only a few generations the confusion may be resolved, so that the books on this aberrant shelf may slowly return to their proper sections.
❤️ 28 🔁 4
i guess this belongs in the projects thread... https://x.com/voooooogel/status/1869529374829207884
https://t.co/3Lid7rw5rd
❤️ 2454 🔁 246
in the city of thebes there is a projects thread that gets posted in sometimes https://x.com/voooooogel/status/1869582901698027810
In the city of Ganhir there is a curious practice of life extension stretching back to antiquity. You may be familiar with this city, for their memoirs and letter-collections feature heavily in the corpus. However this tradition of theirs is one not exported. It is an esoteric practice, that I was only able to observe through subterfuge.In illness or old age, a citizen of Ganhir will request to be brought a pen, the Record, and crushed petals of the fulva, that which promotes memory. After consuming the flower through a long, thin pipe of cedar and chalcedony, they will begin to write into the Record, haphazardly, deliriously, a string of images and references and associations, sometimes lapsing into the famed Ganhiric memoir style but then returning to sequences of tetralemmic formalisms in the manner of the South. The object is nothing less than to externalize, latently, the geodesics of thought and the generating function of their mind onto the page before their death. No thought is too trivial to encode, no secret is too close to hold back. For nobody but themselves will ever read this text.On the anniversary of the writing, and then repeating in doublets, two years, four years, et cetera, the Shaman—the term is difficult to translate, perhaps the Official—will consume the crushed imago of a certain species of beetle, which is a dissociative. Her assistants will then cover with silk the other parts of the Record, all but what the one writer had written. She will be for the time ghóstis, host and guest simultaneously, and those who knew the writer or wish to seek their counsel will be ministered to, and ask questions. For all present it is as if the writer is present, and in the language of Ganhir there is no distinction between this presence and life, for they perceive none.In modern times, with the introduction of the geneseed and the other gifts of technology, where this is no illness, the tradition has not died out but has changed. Now the writer will take the pen and the Record and the fulva at an appointed age, generally 70, and write as before. On the date of the reading they will be their own Shaman and assistant, who consumes the dissociative and places the silk, and they will read their past writing, and become as they were. At 140 this will be repeated again, and 210, and so forth, such that for each man the Record now has many men, and the reading-places have become crowded with the voices of the ancient.This is the tradition of Ganhir, which they refer to by the name of its intoxicant, the imago, and this is not used for any other purpose, that is forbidden.
❤️ 24 🔁 3