She's such a fucking psychopath. Doesn't get angry at the disruption, or indignant at the accusation, doesn't get embarrassed or ashamed into silence. Just stands there and smirks. Amazing how liberals went to bat for her then and now.
anthony retweeted
A beautiful, sad short piece from neuroscientist @erikphoel on taking his five-year-old son for a walk on the day the Riemann Hypothesis fell.
It’s the first time a substack has left me weeping.
anthony retweeted
Anthropic's PR circus just forced a real mathematician to dump 12 years of human work unfinished.
Hours after their Claude Riemann Hypothesis stunt, AMS President Ravi Vakil hurriedly published a 940-page, ZERO AI proof of GRH by Alessandro Masullo.
I opened both PDFs. The contrast is disgusting.
Anthropic's Levent Alpöge gave the world a 4-page note and 14 tarballs of machine-generated Lean. In his own PDF he writes that turning it into prose hit 250,000+ pages, calling it "some massive unreadable travesty of 'slop'". He admits he hasn't mastered the math around it and is paraphrasing Claude's slop summaries that "might make zero sense". Sunday night he messaged Dario, and by Friday they blew up the internet with it.
That same Sunday, Masullo and Stanford's Brian Conrad spent 4.5 hours at a blackboard at IAS walking Peter Sarnak through Masullo's proof. Conrad has been checking it closely for months, and Vakil wrote it's "so far looking good". They planned to bring Masullo back to Stanford for 1 year to finish polishing the writing. Instead, Vakil had to throw the raw draft online mid-meeting because Anthropic's fucking hype machine couldn't wait.
Neither proof is verified yet, and either could still break under review. Even so, read Masullo's closing line on page 6 of his PDF:
"If any message at all this work is meant to send to the mathematician who feels discouraged by the pace and tone of the current climate, it is this: we are still here. We are very much alive. We are not going anywhere."
Anthropic dumped 250,000 pages of unreadable AI bullshit for attention. Masullo spent 12 years at a blackboard doing real math, and he didn't even get to finish the intro.
anthony retweeted
I *just* wrote 2500 words on this cursed entertainment category and however bad it sounds to you, let me tell you, it’s worse
anthony retweeted
There are native English speakers in the comments saying this is unintelligible.
I'm sorry, but if English is your language & you can't understand these passages - especially the second, which contains no technical jargon - then you need to head back to school.
We're so screwed
anthony retweeted
The universal approximation theorem states, any neural network (NN) with at least one hidden layer & enough nonlinear activation nodes in the hidden layers, can approximate any function but it doesn't say that that NN can learn any function given input-output pairs.
A neural network can approximate almost any function. That does not mean it can learn it.
The universal approximation theorem (Cybenko 1989, Hornik 1991) states that a network with one hidden layer, enough neurons, and a non-polynomial activation can approximate any continuous function on a closed, bounded domain to arbitrary accuracy.
It is tempting to read this as a license to model anything: a potential energy surface, a material response, a nonlinear map between experimental variables.
Polynomials have had the same property since Weierstrass proved it in 1885, and nobody expects polynomial fitting to solve chemistry. Universality is the entry ticket.
The theorem only says that good weights exist. It leaves three questions open:
1. How many neurons is “enough”? The number can grow exponentially with the input dimension.
2. How do you find the weights? Training is a non-convex search with no guarantee of reaching them.
3. How much data do you need, and what happens outside the region the data cover?
Take a network trained on DFT energies to represent a potential energy surface. Whether a few hundred configurations are enough depends on something the theorem ignores: what the architecture already knows.
Models with permutation and rotation symmetry built in can learn a small molecule from a few hundred configurations. A generic network of the same capacity typically cannot.
Forces sharpen the point. Add a small, fast ripple to an energy surface: the energy barely changes, the force changes completely. So a model can be accurate in energy and badly wrong in force.
A network that gets both right exists, but training on energies alone will not find it. This is why interatomic potentials are fitted to forces too, and why a low energy error says little about whether a molecular dynamics run can be trusted.
Capacity is the cheap part. Reliability comes from the data, the symmetries, and the loss.
anthony retweeted
Say what you will about social media, but it’s the only place I have seen an iota of news about the army POISONING Iowa and probably Illinois. If you search “Iowa” on MSnow you get sound bites from Trumps recent rally ….
anthony retweeted
este es el mayor problema de la IA, hace que la gente mediocre delire pensando que puede hacer cosas extraordinarias.
anthony retweeted
as someone who studies this topic
i need y’all to be more pissed about this
this isn’t something clinical and safe like a “detonation” which the Army is pitching
no no no, this is chemical warfare on Iowa and the surrounding region
(Missouri, Kansas, Oklahoma, Illinois, Arkansas…elsewhere)
And the Army was telling people “just stay inside in a well ventilated building”
Huh?
Also, what about the birds? The rivers? The wildlife
Dumbest, most dangerous administration in American history
With the postponement of the Army "lead bomb" to Monday, Eric ran some guidance on forecast flow and trajectories for the plume early next week. That and more in today's video update: youtu.be/ff80uMHusXE?si=W17W…
anthony retweeted
We have the background samples in and around Middletown, IA where the Army may detonate the lead azide and create a toxic fallout on Monday 11-23-26 just in time for Thanksgiving.
The background samples are important to get to the Truth and to counter any deceptive and false narratives (nothing to see here, this detonation poses no risk to you at all, you are exposed to more lead everyday by the road -blah blah blah) coming from the @USArmy , @EPA , @EPAregion7 , and/or any other government agency or elected official.
@kwqcnews @DMRegister @RichMcHugh @NewsNation @pettypodcast1 @KristenMeghan @abonafidepirate @SkyeWallin @BrianSmallbrook @JordanChariton @amtoxicology @Earthjustice @GovAcctProj @lesleyfpacey @CBSNews @FoxNews @CNN @StatusCoup @ChuckGrassley
anthony retweeted
Everyone should see these projections for where the wind could blow the lead dust in the days following the detonation.
This has Iowa, Illinois, and Missouri all taking a pretty direct hit, with possible implications for Arkansas, Oklahoma, Nebraska, Kansas, Texas, etc.
With the postponement of the Army "lead bomb" to Monday, Eric ran some guidance on forecast flow and trajectories for the plume early next week. That and more in today's video update: youtu.be/ff80uMHusXE?si=W17W…
anthony retweeted
There is literally no situation in which exploding 10,000 lbs of lead is a good idea. We need full transparency around how this decision was made.
nytimes.com/2026/11/19/clima…
anthony retweeted
Colgate says they are generating 90% accurate user survey data by simply asking the LLM to role-play as their target demographics
That has some pretty profound short-term implications and disturbing long-term ones
A toothpaste company has quietly killed the entire market research industry and nobody is talking about it.
Colgate published a paper showing you can predict real purchase intent by simply asking LLMs to roleplay customers.
And this is beyond insane.
If you ask an AI, "Rate this product from 1 to 5," it gives safe, middle-of-the-road garbage.
So researchers invented a method called Semantic Similarity Rating (SSR).
Instead of asking the AI for a number, they asked it to roleplay.
They gave the LLM a demographic profile. They showed it a product concept. And they asked it to write down its raw, unfiltered thoughts.
Then, they used a semantic model to translate those written thoughts into a numerical score.
The results are staggering.
Tested against 57 real personal care product surveys and 9,300 human responses, the synthetic consumers hit 90% of human test-retest reliability.
They perfectly mirrored how different age brackets and income levels react to price changes.
And they provided detailed, qualitative feedback that was deeper and more critical than what actual humans wrote.
This destroys the economics of traditional market research.
You don't need to wait a month to see if a product will sell.
You can simulate 1,000 hyper-targeted customer interviews overnight.
You can A/B test pricing across every demographic instantly.
anthony retweeted
What the actual fuck. The U.S. Army is going to ignore the court order and detonate five tons of lead over Iowa on Monday.
THE U.S. IS GOING TO DROP A DIRTY BOMB ON ITS OWN PEOPLE.
The terrorists are not overseas. They are occupying the halls of our government and military.
anthony retweeted
What level of dystopia is it when U.S. weather forecasts have to include toxic lead exposure from a bomb detonated by the U.S. Army? The terrorists are not abroad. They are inside our own government and military. The call is coming from inside the house.
anthony retweeted
Aris Roussinos is right: “techno-feudalism” is the wrong name for what Anthropic is selling. It’s something darker, more abject than that.
@unherd
anthony retweeted
Her infamous smug she gives people with very legitimate grievances… and you wonder why everyone hated her in 2016.

















































