not to be too mean but there is a real skill issue thing that happens here where even if we had arbitrarily capable agi, many people would still perceive it as being 'useless' because they have nothing interesting to task it with this does not eliminate autonomy or misuse risk from the skilled

waymos are great, this headline is wrong, it's weird and disturbing that so many people actively want to believe the technology that reduces car deaths is fake, and bluesky should make it so full-screen link previews don't show headlines

it’s counter intuitive but maybe the reason why “just bash” works better is that it allows the model to extend its intent into the similarity / ranking metric, where as in traditional RAG/vector search, the metric is always cosine similarity regardless of intent

compling and symbolics is the ptolemaic / lysenkoism of cogsci and compsci and im tired of pretending it's not

that's just chomskyslop. you only hate it because it's uninterpretable

it's often said that neural networks are 'nondeterministic', but that is actually not the case in the strict sense (i.e, input x -(always)> output y). a NN is a pure function of its weights. nondeterminism comes from implementation details (numerics, batching, kernel sequencing, etc).

the 'stochastic parrot' thing was always conceptually wrong but it is also _empirically, mechanistically wrong_: transformer-circuits.pub/2025/introsp...

the only thing i will point out about e*ily m bender and that cluster is that the conflict is downstream of the usurpation of symbolic linguistics after decades of dominance, by deep learning, and everything that comes along with that

in 2025 i saw a very sudden change in median infosec from “useless stochastic parrot bubble autocomplete lol” once agents started doing actual hacking work. you can’t really afford to be out of touch in that context

A lot of people heard about book burnings as a kid, and internalized “books are sacred” rather than “the suppression of knowledge is evil”.

This has to be the goal post moving of the century. Claude Code developed, trained, and deployed a classifier. A task that used to require an ML PhD ($1M/yr at Google, in the good old days) + a full stack team, and it would yield the same quality. The world has changed, and people can't accept it.

i kind of regret being pseudonymous most of the time nowadays

working on an audio version of this for making infinite evolving ambient soundscapes

the cope has risen to the level of “yeah but the custome platform and custom trained transformer model made by the journalist with zero programming experience classifies keysmashing as 50% oral!”

ed zitron enshittifies sense-making about real problems caused by ai

I don't think current AI is conscious or worthy of moral standing, but there's no reason to suppose it can't be. And it worries me that the exact people generally most concern with rights for other conscious beings are most likely to deny the possibility of machine consciousness.

here's another fun side project i did over a weekend: i trained a neural cellular automata to reproduce a famous painting (see: google-research.github.io/self-organis..., distill.pub/2020/growing...). this is a CA, using only local updates (3x3 learned conv), with a tiny convnet

it is just peak distilled ed zitron to sneer at 'ai company has a philosopher wow they're deranged marketers!' with the obvious and straightforward explanation (that they actually do have real reasons to be doing philosophy and have hired an extremely qualified philosopher) left unexplored

here's a fun side project i've been working on: i compiled a joint text<>audio embedding model to a fast coreml pipeline, and built a very fast (~400ms for 50k samples, can scale to millions) UMAP dimensionality reduction GPU impl in mlx. using it to browse music libraries and do sample sim search

many people on here are suspiciously invested in performative and preemptive cruelty towards ai systems and i find that to be disturbing

doing research on the airline bubble. no one has looked into this before

pluribus is a show where the plot refuses to move forward, things are repeated again and again, and nothing interesting ever happens. it is a movie concept stretched out into a tv series. however i like it cause it's slow, boring and filmed well- and makes the borg look extremely incompetent

it’s okay to be earnest on the internet

there is a profound disconnect here in modeling/understanding what is going on it's not 'there is now youtube slop generator, you must use it in your class' it's 'a general method has scaled from broken english to IMO gold, and experts don't think it is going to stop'

man if i got an email like this from opus i would be glowing with happiness for days

Neuro-Sama is currently going through some bicentennial man shit

Deleted a post that was going viral and bringing an undesirable amount of attention from anonymous accounts. But, I'm going to be honest, the toxic folks here have won. So much of the usage of the site is about avoiding their attention. I'll just use this site to get book recs going forward

“why are all your posts on here about this” idk maybe it’s because i got dogpiled with hundreds of people informing me that LLMs are markov chains and tons of crude insults? the discourse on here for AI is fucking bleak

i did some regression and projections on the METR benchmarks, assuming a regime shift in late 2024 from inference scaling:

update i’m now seeing the left image being posted by self described “communists” as an aspirational meme

it would be incredibly easy to create a new account and rack up a huge following and clout by proclaiming why “generative ai” or the “ai bubble” is about to collapse and everything will go back to normal, with vague techy sounding language with just enough insight + credentialism sprinkled in

the transformer model is not going to spontaneously uninvent itself. much consternation from this fact

so it turns out that the studios that made very highly acclaimed games have been using new tooling without the “matmuls are the devil” mindset, and ppl are very upset about this (e33 + bg3 devs) expect this to continue and devs to simply self censor to satisfy the mob

ngl i kind of like that AI is somewhat hated and misunderstood villain coded in mainstream culture. “anti ai” is extremely median reddit coded

when i think about AI, i think about the universal approximation theorem, erdõs problems and imo medals, rate of change, and the things i can do with frontier language models like opus 4.5 and the insane internal complexity in 2025 most ppl think of “annoying slop video” or “copilot win11 popup”

entirely predictable result of backlash, reflexive IP law support, and poor thinking. i made this meme in 2023

a few months ago i built a game out of the wikimedia dataset to see how good i was at identifying safe mushrooms. i mostly learned that i should not forage for mushrooms lol mushdle.com if you want to play (the dataset is pretty small)

seeing a machine do something that you previously thought was unique to your cognition, and opening up that capability to anyone, is to many people a deep narcissistic injury. this explains much about the social environment

the people in this screenshot are making fun of people who actually believe this, but the thing is that for every satirist there are like 300 people who actually believe the thing they’re satirizing

if post-scarcity ASI singularity goes well, you will be able to choose to live in a perfectly maintained bubble that is a simulacrum of the world where “the bubble popped”, where all AI research was meaningless, etc. perfectly maintained by machines of loving grace

thinking about this again. i stand by it

eastern Marxist-Leninists: - oversaw single largest industrialization, becoming competitive superpower - competitive in frontier research and technology - modern china: _mandating_ AI fundamentals western “leftists”: - you offended my aesthetic sensibilities, so i will now oppose technology

there is a culture on here of believing you can sneer your way into or out of anything, you don’t have to understand anything, and it’s virtuous to act like a 1980s high school movie bully

good example of AI psychosis