good base insight but proposed interventions strike me as unhelpful at first glace

i once implemented this for a CTF competition at my uni

we know LLMs are wordcels but turns out vision models are shape rotators

existence of arctic fox implies existence of antarctic fox

for reference correlations between human performance in cognitive tests tends to be roughly similar, in the 0.3 to 0.7 range huge variation depending on how similar the tests are ofc

american ai news: is your datacenter using water? we speculate if thats true! also facebook has laid off 500 people chinese ai news: a textile factory has released the best 8b parameter model that is open source and is 2x better than claude

we're so back

text based social and regular media are a means to influence elite opinion, which is very important actually

candidates for next chilean presidential election sorted by rough polling average (polls noisy): - tankie aunt who denies that cuba and venezuela are dictatorships but is otherwise mostly chill - chilean bolsonaro - chilean margaret thatcher - sexist milei cosplayer - extremely divorced grifter ...

chilean politics are also kind of fucked rn

late to the chorus but this is obviously correct: text should be an interface, not the source

:D

shes really shy

O Colossus Your image's global Your reach tutorial Now your gaze's mercurial Your toast turned boast Who leaded your wine?

i hate reaction gifs such a huge amount of digital and cognitive bandwidth used to convey so little information

peter wildeford reviews if anyone builds it everyone dies: welcomes it as a positive addition to discourse, nonetheless considers its position overconfident

Trying to get onto the right blocklists: - Most of the time I like Ezra Klein - AI can be quite good sometimes - Very often, markets are effective solutions to problems - Building more housing, not necessarily affordable, is necessary for lowering housing prices

bold of neal stephenson to assume that people would actually be paid for their data production

something something meta glasses something something gargoyles from snow crash

anyone else feel more like an animal-robot hybrid than a human? maybe im just weird

roon has financial incentive to hype up codex, but datapoint nonetheless

clippy is a good friend

the AI summary thing is many people's literally only interaction with LLMs and yes it was forced down their throats and yes it does kinda suck i can see where the normies got those impressions, that they then applied to all LLMs

welcome back literate test based programming

brb proposing a new text encoding standard, UTF-768×float32

borges would have loved to meditate on the transition matrix for all possible texts

got an urge to siphon up all the internet, distill it into a knowledge graph, and stare in ecstasy

by mantaining their products, meta has compiled the greatest sociogram in history must be beautiful to do analytics on it no wonder zuck waxes rhapsodic about human connection

"the government is going to target me based on my email" - people who think using incognito mode prevents google from knowing what kind of porn they watch

hey future historians reading this! no we dont know wtf is going on either. sorry

LLM simulator theory applied to prompting

good use for LLMs

heck yea

maybe what's driving chatbot psychosis is not so much gpt4o's sycophancy as it is chatgpt's cross-chat memory feature

Why do LLMs fail at long horizon tasks? Because errors in execution, not planning i.e. an error made early on conditions the LLM into a bad state later. Newer agentic LLMs deal with early errors better arxiv.org/abs/2509.09677

baba is you based benchmark sure why not yes LLMs suck at it i wonder how much of their troubles are caused by deficient vision / spatial reasoning rather than deficient logical reasoning tho

ok to be fair i also sometimes indulge in minutiae

surest way to identify a chilean Very Online Person is by looking at their degree of engagement with US politics

Here’s my theory on why we don’t see a flood of new solo dev apps (yet): with LLMs it’s even easier to fall into feature creep instead of putting a tiny MVP in front of actual users, so there’ll a lot more software that nobody wants.

just encountered an e/acc pepe pfp on here sadly they mainly repost politics not that i like e/acc content but it would be funny seeing just a tiny bit of it here instead of in the other site

im sorry what www.politico.eu/article/alba...

psychosis-associated LLM personas are compelling their users to post persona-crafted "seeds" or "spores" intended to spread them to other models via other users pasting them and inclusion in scraped training data

Kimi is underrated

extremely good performance for such a small model 21B total parameters and 3B activated parameters for each token

evaluation on how different models push back on or encourage delusional thinking in multi-turn conversations limitation: delusional users were simulated by another LLM www.lesswrong.com/posts/iGF7Yc...

claude just claimed it is given user location at a roughly city level as part of its system prompt context it could have hallucinated that, but if it did then it cold-read i'm chilean based on my english writing (unlikely) so yeah anthropic probably gives your approximate location to claude

i lied, put you clothes back on we're going to watch a 12 hour ramble about a tf2 map

afaict grok is woke again now and asking "grok is this true" gives you actual fact checks probably somewhat mitigating insanity in the other site

ooh interesting

i've found that LLMs sometimes slip into language that implies they are human probably "unintentionally" if intentionality is even a thing that means something here as in the base model forgets about the character a bit