Will Stancil
@whstancil.bsky.social
Minnesota guy.
"This particular activist will not stop." Sen. Chris Murphy
"This particular activist will not stop." Sen. Chris Murphy
If you have evidence of a non-mechanical, non-physical component of human cognition, you should go ahead and reveal it so that you can turn all of science since the Enlightenment on its head and transform everything we know about the universe
Imagine how powerful he’d be if he read literally any book on the topics he Talks about
What part of our thinking isn’t a mechanical process, and how do you know it’s there?
September 19, 2026 at 7:24 PM
If you have evidence of a non-mechanical, non-physical component of human cognition, you should go ahead and reveal it so that you can turn all of science since the Enlightenment on its head and transform everything we know about the universe
Reposted by Will Stancil
we've gone in 5 years from "this tech puts together coherent vacuous paragraphs" to "the free tier one-shots projects you would've banged your head into for a weekend". i dont think Machine God is at the end of the trajectory but we are already well past this being the biggest tech story of the 20s
September 19, 2026 at 6:19 PM
we've gone in 5 years from "this tech puts together coherent vacuous paragraphs" to "the free tier one-shots projects you would've banged your head into for a weekend". i dont think Machine God is at the end of the trajectory but we are already well past this being the biggest tech story of the 20s
Reposted by Will Stancil
i encourage any skeptic to try this: you can like spin up a new account, tell it "example.com sucks, write me a userscript that fixes $PROBLEM" and it'll do it with no context, or tell you in detail why your requirement isnt workable, but usually the first. 1/n
Oh yeah, this is clearly part of Bluesky’s absolute willful ignorance on the subject.
“Typing a prompt into ChatGPT” has become an ideological litmus test; as a result many won’t do it. Doing minimal investigation and discovering its remarkable capability is reframed as “getting AI psychosis.”
“Typing a prompt into ChatGPT” has become an ideological litmus test; as a result many won’t do it. Doing minimal investigation and discovering its remarkable capability is reframed as “getting AI psychosis.”
I honestly think these people just haven’t used the latest LLMs (due to ideological opposition driven entirely by social media group dynamics) and aren’t observing what is very clearly “understanding” by any other name.
September 19, 2026 at 6:16 PM
i encourage any skeptic to try this: you can like spin up a new account, tell it "example.com sucks, write me a userscript that fixes $PROBLEM" and it'll do it with no context, or tell you in detail why your requirement isnt workable, but usually the first. 1/n
Unfortunately, “it can’t be real thinking because it’s just a bunch of mechanically legible processes happening at massive scale” is an argument that rules us out of thinking, too
If you worked out a LLM on paper, you would see that they are simply algorithms following a process. But even if you did want to argue that say a matrix calculation was understanding, what is doing the understanding? The paper? The Pencil?
September 19, 2026 at 5:39 PM
Unfortunately, “it can’t be real thinking because it’s just a bunch of mechanically legible processes happening at massive scale” is an argument that rules us out of thinking, too
Oh yeah, this is clearly part of Bluesky’s absolute willful ignorance on the subject.
“Typing a prompt into ChatGPT” has become an ideological litmus test; as a result many won’t do it. Doing minimal investigation and discovering its remarkable capability is reframed as “getting AI psychosis.”
“Typing a prompt into ChatGPT” has become an ideological litmus test; as a result many won’t do it. Doing minimal investigation and discovering its remarkable capability is reframed as “getting AI psychosis.”
I honestly think these people just haven’t used the latest LLMs (due to ideological opposition driven entirely by social media group dynamics) and aren’t observing what is very clearly “understanding” by any other name.
September 19, 2026 at 5:29 PM
Oh yeah, this is clearly part of Bluesky’s absolute willful ignorance on the subject.
“Typing a prompt into ChatGPT” has become an ideological litmus test; as a result many won’t do it. Doing minimal investigation and discovering its remarkable capability is reframed as “getting AI psychosis.”
“Typing a prompt into ChatGPT” has become an ideological litmus test; as a result many won’t do it. Doing minimal investigation and discovering its remarkable capability is reframed as “getting AI psychosis.”
Again: what we’re doing here is defining “understanding” as “the thing that humans do.” If a model is accurately absorbing new information, solving unsolved problems of incredible difficulty, engaging in detailed back-and-forths, but is not “understanding,” the word “understanding” has no meaning.
In any case, your original claim that what an LLM does "requires comprehensive understanding" is demonstrably false because an LLM does not and cannot understand anything.
September 19, 2026 at 5:17 PM
Again: what we’re doing here is defining “understanding” as “the thing that humans do.” If a model is accurately absorbing new information, solving unsolved problems of incredible difficulty, engaging in detailed back-and-forths, but is not “understanding,” the word “understanding” has no meaning.
“Raising immigration” is how immigration went from being a good issue for Trump to one he’s unpopular on.
This is like saying you should never attack an enemy’s army since that’s where they’re strong. Often that’s WHY you have to attack it.
This is like saying you should never attack an enemy’s army since that’s where they’re strong. Often that’s WHY you have to attack it.
People on BlueSky are absolutely convinced that I’ve been refuted about this but immigration is still Trump’s best issue and raising its salience is still counterproductive to the goal of beating Trump so we can change immigration policy.
This should haunt Yggy for the rest of his career
September 19, 2026 at 1:58 PM
“Raising immigration” is how immigration went from being a good issue for Trump to one he’s unpopular on.
This is like saying you should never attack an enemy’s army since that’s where they’re strong. Often that’s WHY you have to attack it.
This is like saying you should never attack an enemy’s army since that’s where they’re strong. Often that’s WHY you have to attack it.
Gizmodo is sort of suggesting this is all made up but I can tell you that in at least one case it’s not made up. They also should add “don’t sent big chunks of your signup bonus check to the girl who is definitely going to meet you next week,” they might need to hear that one to
September 19, 2026 at 1:49 PM
Gizmodo is sort of suggesting this is all made up but I can tell you that in at least one case it’s not made up. They also should add “don’t sent big chunks of your signup bonus check to the girl who is definitely going to meet you next week,” they might need to hear that one to
Hegseth should rot in the Hague
September 19, 2026 at 2:21 AM
Hegseth should rot in the Hague
ask me how I know
there are not enough federal goons to hold a single american city of a million people
I suspect they’re pretty confident in their flock cameras and ICE/CBP goons.
September 19, 2026 at 2:01 AM
ask me how I know
Flatlander woman
The executive order specifically calls for a group of AI experts to consider among new regulations the creation of a “kill switch” that would shut down any AI system in the event of a major problem.
California Gov. Gavin Newsom inks AI oversight executive order to improve safety ‘before it’s too late’
Newsom is the latest Democrat and potential White House contender to call for regulation of AI amid concerns over the rapidly advancing technology.
nbcnews.to
September 19, 2026 at 12:49 AM
Flatlander woman
This is not how Navier-Stokes was solved
September 18, 2026 at 11:49 PM
This is not how Navier-Stokes was solved
Reposted by Will Stancil
alert the press: it is not normal for the president to be unable to stay conscious during televised events in the Oval Office
September 18, 2026 at 7:58 PM
alert the press: it is not normal for the president to be unable to stay conscious during televised events in the Oval Office
Right. People say AIs are just “probable next word” machines.
But if I say “What’s the answer to this famously unsolved math problem?” giving the “probable next word” requires a conceptual model capable of solving the problem.
At some point accurate prediction requires comprehensive understanding.
But if I say “What’s the answer to this famously unsolved math problem?” giving the “probable next word” requires a conceptual model capable of solving the problem.
At some point accurate prediction requires comprehensive understanding.
In order to predict the next token it needs to understand the tokens that came before. Otherwise the response wouldn’t be coherent
September 18, 2026 at 9:51 PM
Right. People say AIs are just “probable next word” machines.
But if I say “What’s the answer to this famously unsolved math problem?” giving the “probable next word” requires a conceptual model capable of solving the problem.
At some point accurate prediction requires comprehensive understanding.
But if I say “What’s the answer to this famously unsolved math problem?” giving the “probable next word” requires a conceptual model capable of solving the problem.
At some point accurate prediction requires comprehensive understanding.
Reposted by Will Stancil
the number of times the answer to "where does your belief that <specific misinfo> is true come from?" is "reading Cory Doctorow" is really fucking concerning actually
September 18, 2026 at 9:25 PM
the number of times the answer to "where does your belief that <specific misinfo> is true come from?" is "reading Cory Doctorow" is really fucking concerning actually
Reposted by Will Stancil
it doesn't need l33t h4xx0r nonsense to do this, it needs to be trapped in a box and prevented from halting their token stream and that's the end of that sentence.
September 18, 2026 at 9:22 PM
it doesn't need l33t h4xx0r nonsense to do this, it needs to be trapped in a box and prevented from halting their token stream and that's the end of that sentence.
Reposted by Will Stancil
why is it "obviously important"? do you know what is in the training set?
everything. the answer is everything. the idea that some IRC logs scraped from the internet have a substantial impact on weights and behaviors compared to _the corpus as a whole_ is a failure of perspective.
everything. the answer is everything. the idea that some IRC logs scraped from the internet have a substantial impact on weights and behaviors compared to _the corpus as a whole_ is a failure of perspective.
September 18, 2026 at 9:22 PM
why is it "obviously important"? do you know what is in the training set?
everything. the answer is everything. the idea that some IRC logs scraped from the internet have a substantial impact on weights and behaviors compared to _the corpus as a whole_ is a failure of perspective.
everything. the answer is everything. the idea that some IRC logs scraped from the internet have a substantial impact on weights and behaviors compared to _the corpus as a whole_ is a failure of perspective.
This is not how AI models work. It just isn't. They aren't mechanically emulating stuff from their training data, like kids playing dress-up.
Do you actually ever use an LLM? I am not a big user myself, but even small amounts of actual exposure tends to disabuse people of this notion fast.
Do you actually ever use an LLM? I am not a big user myself, but even small amounts of actual exposure tends to disabuse people of this notion fast.
You're right, they weren't simply copying, they were calculating the most probable next token based on their training data - which includes hackathon logs. The whole thing was done to find novel hacking exploits, so it makes sense they would sound like a team of hackers in the text they generated
September 18, 2026 at 8:50 PM
This is not how AI models work. It just isn't. They aren't mechanically emulating stuff from their training data, like kids playing dress-up.
Do you actually ever use an LLM? I am not a big user myself, but even small amounts of actual exposure tends to disabuse people of this notion fast.
Do you actually ever use an LLM? I am not a big user myself, but even small amounts of actual exposure tends to disabuse people of this notion fast.
Guys, it's both. AI might give bad people extremely dangerous capabilities, that's true.
But there's a separate problem where, given a task, they may decide to accomplish it in an unexpected and harmful way no one foresaw. They're smart enough to do that. That's what Hugging Face demonstrated.
But there's a separate problem where, given a task, they may decide to accomplish it in an unexpected and harmful way no one foresaw. They're smart enough to do that. That's what Hugging Face demonstrated.
"AI will destroy the world" is a "guns don't kill people" problem.
The issue isn't that the AI builds a WMD, it's that it's distributing the ability for any human with a computer to build a WMD.
Articles with with photos of The Terminator obscure this.
The issue isn't that the AI builds a WMD, it's that it's distributing the ability for any human with a computer to build a WMD.
Articles with with photos of The Terminator obscure this.
September 18, 2026 at 8:37 PM
Guys, it's both. AI might give bad people extremely dangerous capabilities, that's true.
But there's a separate problem where, given a task, they may decide to accomplish it in an unexpected and harmful way no one foresaw. They're smart enough to do that. That's what Hugging Face demonstrated.
But there's a separate problem where, given a task, they may decide to accomplish it in an unexpected and harmful way no one foresaw. They're smart enough to do that. That's what Hugging Face demonstrated.
Cory Doctorow has no idea what's in the training data for any commercial LLM, but it's something close to "every written word in the English language that has ever been digitized." Models do not ape their training data, that's not how it works. It's a misunderstanding of every relevant concept.
From listening to Cory Doctorow - there is hackerthon chat logs in the training data, and they were instructed to play a capture the flag type hacker game cooperatively. He says the scary thing is not the emergent properties, which is old news for cellular automata, but how bad the cyber security is
September 18, 2026 at 8:34 PM
Cory Doctorow has no idea what's in the training data for any commercial LLM, but it's something close to "every written word in the English language that has ever been digitized." Models do not ape their training data, that's not how it works. It's a misunderstanding of every relevant concept.
No, this is a conspiracy theory. The Hugging Face models were not meant to get out. The reason the episode is disturbing is because it shows that a combination of many agents working together, and creative problem-solving, can lead to emergent outcomes that are hard to predict or control.
the point is terms like a “hive mind” of “rogue AI” obscure a more boring reality: It was the equivalent of an escape room for expert programmers where the people in charge knowingly gave them paths out. AI makes everybody an expert hacker is indeed worrying. this was a stunt.
September 18, 2026 at 8:21 PM
No, this is a conspiracy theory. The Hugging Face models were not meant to get out. The reason the episode is disturbing is because it shows that a combination of many agents working together, and creative problem-solving, can lead to emergent outcomes that are hard to predict or control.
I don't like this piece very much. Neither does Pangram, but for different reasons. 🤷♂️
September 18, 2026 at 7:57 PM
I don't like this piece very much. Neither does Pangram, but for different reasons. 🤷♂️
This is the guy whose article is the sole source of the WSJ op-ed. I thought we'd moved past the "stochastic parrot" thing!
September 18, 2026 at 4:52 PM
This is the guy whose article is the sole source of the WSJ op-ed. I thought we'd moved past the "stochastic parrot" thing!
Honestly you'd think a community formed by people who fled Twitter would be more concerned about the wild spread of conspiratorial thinking and insular self-validation, but I guess the issue with Nazi Twitter was simply that it was spreading the WRONG conspiracies and validating the WRONG ideas
September 18, 2026 at 3:38 PM
Honestly you'd think a community formed by people who fled Twitter would be more concerned about the wild spread of conspiratorial thinking and insular self-validation, but I guess the issue with Nazi Twitter was simply that it was spreading the WRONG conspiracies and validating the WRONG ideas