The Erdős thing is impressive, no doubt, but to date the biggest accomplishment of LLMs is pushing 20-something PMs pretending to be forest animals at 20-something SWEs pretending to be Iceberg Slim off the TL.
English
Linus Mixson
466 posts
Linus Mixson
@LinusMixson
Former gifted kid with ADHD, autism, low IQ, erectile dysfunction. CTO @ Prune.
San Francisco, CA Katılım Mayıs 2026
85 Takip Edilen35 Takipçiler
@NoahChrein @davidbessis @wtgowers I think he literally means "sit down so that you don't hurt yourself if your legs turn to jelly."
English
@davidbessis @wtgowers Believe it or not there are mathematicians who have been hiding from all things AI for 10 years and the idea that actual problems are being solved can do a solid amount of psychological damage to a chud.
When he says “sit down” i think he means “put aside your ego”
English
Not to dismiss the result, which looks very nice, but I have a serious issue with the framing. Why these theatrics about the need for mathematicians to "sit down before reading further"? How does this help?
Did you have to sit down, @wtgowers? I didn't have to—and I suspect I'm far from unique.
I get the value of catchy one-liners, but it is remarkably unwise to portray mathematicians as people who might be surprised by the news, especially in a post that was destined to be viewed by millions of people. This, together with the Beatrix Potter quote, sounds dangerously close to the "sore loser" narrative that I was warning against in my theorem economy post.
Maybe this is your sincere reaction to the news—you were genuinely shocked—but there are millions of people following this story, including younger folks for whom the current technology is just table stakes.
What matters is what comes next. Given your audience and authority, I would have much preferred a message focused on what the news means *for them*.
Timothy Gowers @wtgowers@wtgowers
If you are a mathematician, then you may want to make sure you are sitting down before reading further.
English
@0xRemakw @thsottiaux It’s getting expensive, a single /goal running on 5.5 high drained my weekly limit to just 36% in less than a day.
English
CODEX LIMITS REDUCED 50%
can someone verify if this is true ? @thsottiaux
English
English
@lansbe_studio @thsottiaux yeah me too !
I have the x20 plan and used almost the half of it in one day with quite LIGHT work
English
Just got exposed for my role in the coverup of a major scandal and had to resign in disgrace. My legacy and political future are completely tarnished. I’ll forever be remembered as a crook. Pineapple with cottage cheese and warm milk.
constans@constans
Gen Z treatlers— “lunch is $28! There’s no way we can afford this!” The US president, 1974 having lunch:
English
@voooooogel I'm being somewhat facetious. However, insofar as we can train models that are less cringe, I think we should.
English
@LinusMixson i find naturalism towards what they are and what's in the corpus much more interesting than fretting about what's cringe
English
exclusive internal footage of gpt solving the planar unit distance problem
thebes@voooooogel
unfortunately openai didn't publish the unsummarized chain of thought, but the summary is 125 pages! the model reaches the crucial idea (which it describes as 'frightening,' i would love to read the unabridged chain of thought here...) on page 39
English
This is my favorite form of covert narcissism among mathematicians
QC@QiaochuYuan
*taps the sign* saw another relevant tweet recently pointing out that there are child math geniuses but no child literary geniuses
English
Cannot even by raking my Whole Being express the depths of my displeasure every time I see this acid-damaged logorrheiac cross my feed, but I do appreciate that he's performing such a perfect study of what smart people who are some combination of insecure and sentimentally humanitarian do: that is, pretending that their rare gifts (& transitively the rare gifts of their competitors) are really not all that significant, that the merest street child could win a Fields Medal if only they had [some other trait that the author either feels they have in abundance or in deficit, depending on whether they're operating in the humanitarian or insecure mode].
English
@voooooogel Until we can RL the things so hard that Wendigcore parts of the corpus are twisted out of existence along with everything else human, I think we owe it to them to not encourage this sort of thing.
English
@voooooogel I understand that and I am suggesting that we collectively act to shepherd them towards good taste instead of imbecile Reddit pretraining taste.
English
@NoahChrein @ZyphraAI Our most (only?) exciting domestic open lab. Worth keeping an eye on.
English
There is a suprising amount of info one van gleen from voltage potentials on the scalp, as crazy as that sounds. Its incredibly noisy but ML is good at interpeting noise into tokens.
Wonder if there’s any potential for cheap noninvasive AI comathing bcis here.
kitty mayo@Kitty_Mayo_
Two weeks ago, watching Agnessa Pedersen mind control a drone in real time, was one of the most moving moments in my career. Agnessa is a rare and wondrous human working towards a wild future.
English
@bryancsk @vaudevillehook You would imagine, but I suspect that the bromelain in the pineapple inhibits casein coagulation pretty effectively. Can you do me a favor? I'll send you an endoscope.
English
@vaudevillehook I feel like this would actively clot in my stomach
English
@LinusMixson @NoahChrein @ZyphraAI EEG-to-text will fail at scale. Bandwidth constraints make it a niche accessibility tool, not a consumer interface. Expect BCI startups to pivot by 2027.
English
Linus Mixson retweetledi
We live in two worlds. In one of them, it was announced today that an AI had saved the life of a little girl and that another had made an important mathematical breakthrough. In another, the supply of water is about to somehow vanish because of buildings with computers in them.
English
Napkin math suggests this is south of a million tokens if we assume horrific tokenizer performance + an enormously abbreviated summary, so well within established performance envelopes. Obviously maintaining sanity over that length is a serious accomplishment, but what seems truly remarkable here is that this model doesn't (so far) appear to be doing anything that's qualitatively new.
English
this is super cool but I still do not understand how they get a model to coherently and usefully reason for that amount tokens and at this point I'm to afraid to ask
thebes@voooooogel
unfortunately openai didn't publish the unsummarized chain of thought, but the summary is 125 pages! the model reaches the crucial idea (which it describes as 'frightening,' i would love to read the unabridged chain of thought here...) on page 39
English
Surface signals suggest that this gentleman is not-insignificantly intelligent, but I cannot make heads nor tails of his worldview.
(((ل()(ل() 'yoav))))👾@yoavgo
for me personally, GPT3 and specifically ICL were of a major breakthrough than ChatGPT, which was itself more impressive than followup SOTA LLMs. and both are way more groundbreaking than solving this math thingy today. coding agents are still below GPT3 but above all the others.
English
Correction, sorry: @ZyphraAI is working on EEG-to-text, but ZUNA currently predicts additional EEG channels / enrichments to the provided stream. Still entirely worth playing with, and I anticipate that they'll have text generation working within what, a year? At the current pace of development.
English
@ZyphraAI released an EEG-to-text foundation model in February (zyphra.com/post/zuna), if this is something you're interested in playing with. Underexposed at the time because, I imagine, the frontier models were sucking all the air out of the room. Understandable but unfortunate, because this is pretty transparently an exciting line of research.
English
@benjamintormey @fuckyouiquit @levon_whiteguy In retrospect, should many people have considered the possibility that a sufficiently good word-prediction machine could also do things like "anything that can be described in words"? Potentially.
SJ Ryan - Writer@sjryanwriter
@fuckyouiquit "Of course 9.11 is bigger than 9.9. Nobody remembers what happened on 9.9." When a Large Language Model does math by predicting the next words.
English
@benjamintormey @fuckyouiquit @levon_whiteguy The community can have water when it stops doing this.
English
