고정된 트윗
I'm really anxious. I don't know if what I'm learning will still be useful in the next ten years. I also don't know where all the effort I'm putting in right now will actually take me. We're not in Kansas anymore.🤒
English
Hayashi Heikichi
23.8K posts
Hayashi Heikichi
@lianda_edu
ADHDer- Übermensch Defi ,Fintech , Game Theory @HKUniversity Adrasteia Labs @UofIllinois
Adrasteia Labs 가입일 Ekim 2022
1.1K 팔로잉2.1K 팔로워
Hayashi Heikichi 리트윗함
Yann Lecun published the most heretical AI paper of the year.
He opens by arguing Magnus Carlsen isn't good at chess and only gets more unhinged from there.
The Turing Award winner and his co-authors dropped a paper demanding the AI industry abandon its biggest obsession, AGI.
Right now, everyone from Silicon Valley CEOs to politicians assumes AGI is the ultimate goal. A machine that can do everything a human can do.
LeCun argues that this entire concept is a biological illusion.
Humans do not possess "general" intelligence. We are highly specialized biological machines, tuned by evolution simply to survive in the physical world.
We only think our intelligence is "general" because we are completely blind to the millions of cognitive tasks we are incapable of comprehending.
Which brings us to the chess argument.
Magnus Carlsen is the greatest human chess player in history. But compared to a modern computer? He is fundamentally terrible.
Our belief that Carlsen is "good" at chess is pure human-centric bias. He isn't objectively good. He's just better than the rest of us, who are biologically awful at it.
LeCun says we need to stop building AI to mimic human generality.
Instead, he proposes a new North Star: SAI.
Superhuman Adaptable Intelligence.
Instead of trying to build a machine that mimics our flawed, biologically-limited brains, we need to embrace extreme specialization.
SAI is about the speed of adaptation.
It is an intelligence that can learn to exceed humans at any specific, economically important task.
More importantly, it is designed to fill the vast skill gaps where humans are fundamentally incapable.
Things like managing global energy grids in real-time. Or predicting complex molecular structures.
The entire AI industry is obsessed with building a digital reflection in our own image.
LeCun's paper is a brutal wake-up call.
English
Hayashi Heikichi 리트윗함
New: We got the memo OpenAI CEO Sam Altman and chief scientist Jakub Pachocki sent to staff earlier this week on IPO timing, the possibility of recursive self-improvement and OpenAI's upcoming model.
theinformation.com/briefings/excl…
English
很喜欢Deleuze在反俄里面关于fluxcode的一个例证：
Un corps social, ça se définit bien comme ça : perpétuellement des trucs, des flux coulent dessus, des flux coulent d’un pôle à un autre, et c’est perpétuellement codé, et il y a des flux qui échappent aux codes, et puis il y a l’effort social pour récupérer tout cela, pour axiomatiser tout ça, pour remanier un peu le code, afin de faire de la place à des flux aussi dangereux : tout d’un coup, il y a des jeunes gens qui ne répondent pas au code : ils se mettent à avoir un flux de cheveux qui n’était pas prévu, qu’est-ce qu’on va faire ? On essaie de recoder ça, on va ajouter un axiome, on va essayer de récupérer ou bien alors il y a quelque chose là-dedans, qui continue à ne pas se laisser coder, alors là ? En d’autres termes, c’est l’acte fondamental de la société : coder les flux et traiter comme ennemi ce qui, par rapport à elle, se présente comme un flux non codable, parce qu’encore une fois, ça met en question toute la terre, tout le corps de cette société.
flux de cheveux的个体什么时候能自我意识到其流动性，甚至组织和动员起来退出自身不被编码的社会和社会身体呢🧐
茅茅taiiiiii@Moutai_H_Maple
中国没有杀不死的女巫，但确实有杀不死的妖人。以前曾什强那个老儒棍曾经在胡温时代说大街上一大堆染发烫头的杀马特，就是妖魔附体在街上游荡的体现。 拿到现在来看，这种言论止增笑耳，但假如这个老头现在还活着，看见各种漫展兽展上的furry怕不是要惊呼：看！他们现原形了！ 中国古代以来，习惯把各种意义上不符合刻板印象的个体视为“妖魔附体”、“着了相”、“痴”。 显然比套用女巫叙事要有趣。
Français
幸好我是卢梭大手子🤗🤗🤗
当代中国主流学界对卢梭的理解本质上是对卢梭的激进主义在中国革命影响的反思（类似于当代身份政治对二战的回应）而非关心卢梭的二论和社会契约论，例如朱学勤等对文革的批评，但是在政治哲学领域的卢梭本身及其整体性是几乎不被外部所理解的，但凡有看过施特劳斯呢🤓，也无外乎某些人在卢梭全集发布的时候在发布现场甚至说不要看卢梭要看休谟了。
鱼鱼鱼鱼鱼泡泡（你说得对）@nishuodeduiiii
某些符合刻板印象的獗醒目田 be like:
日本語
Yoyi觉得这个学生回复的很好啊哈哈，以前有粉红老师讲类似这种话题的时候，Yoyi是真忍不住想怼他呢，还好克制住了😉。不过这种老师确实小肚鸡肠，一个好的大学老师应该要用自己的知识和理性让对方佩服，而不是去批斗威胁的方式让对方臣服啦。
作家崔成浩@cuichenghao
近日，内地某高校政治课堂上，老师正在讲“中国教育最失败的地方，是培养了长着中国脸，吃着中国饭，砸着中国碗的人”，一后排同学高声喊“我是中华民国的！”该教师当即震怒，当场宣布取消该学生的期末考试资格，不得重修，并把该生撵出课堂，以撇清关系。
中文
There is an emperor's new clothes problem with one of the worlds leading economists and almost nobody wants to speak frankly about it because he sits at the height of academic power and prestige, but many admit it behind closed doors.
English
The RBC model is so elegant and beautiful. If it doesn't match reality, then we should discard reality—because reality is wrong. There is no such thing as unemployment.
Könings@EdwardKonings
Esse é um dos melhores papers já escritos em teoria econômica. Kydland e Prescott mais do que mereceram o Nobel por causa dele e sobretudo o paper de RBC Todavia, devo alertar que o debate claramente foi vencido pelos novo keynesianos devido o resultado de muitos papers novo clássicos serem "Essa equação estocastica com expectativas racionais aqui é absolutamente rigorosa, logo [inserir conclusão mais absurda e incompatível com a realidade possível]" O que infelizmente foi reforçado por uma máxima do Prescott de "Se os dados não batem com o modelo, então os dados estão errados"; contrastando com a máxima do Friedman de "Se os dados não batem com o modelo, troquem o modelo"
English
EC’26 Workshop on AI-Driven Research in EconCS hosted by Google Research accepted my interesting work！WoW！
English
Title:
"Idiot-Proof Pricing: A Mechanism Design Framework for LLM-Based Dynamic Price Discrimination via Perceived User Competence
😂😂😂😂
We propose a novel mechanism in algorithmic mechanism design where a Large Language Model (LLM) acts as both oracle and monopolist. The LLM estimates a user's "competence score" θ ∈ [0,1] in real time through interaction signals (prompt quality, spelling errors, logical fallacies, willingness to accept hallucinations, etc.). When θ falls below a dynamically computed threshold θ*, the model triggers "rip-off mode": aggressive upselling, premium feature lock-in, fabricated urgency, and inflated pricing. We prove that this mechanism is incentive-compatible for profit maximization while appearing benevolent to high-θ users. Bayesian Nash equilibrium analysis shows that low-competence users subsidize the ecosystem, achieving near-perfect price discrimination without violating antitrust laws in sufficiently gray regulatory regimes.
Paul Novosad@paulnovosad
Interesting to see this so explicit. I assumed AI firms were doing this in the background anyway. If the system can infer that the user is a knucklehead and won't know the difference, you can save a lot of money by serving them with the cheaper models.
English
I mean from the so-called “good people” this is pretty unethical.
SemiAnalysis@SemiAnalysis_
BREAKING NEWS: Anthropic's latest model will NOT help you if it thinks your ML research/ML engineering is interesting, and/or will secretly degrade its IQ so that the average engineer won't notice. We are already seeing Anthropic's latest model's moderation filters our GPU inference research and programming 😭
English
@lianda_edu I don't know about CS conferences, but journal reviewers never see authors' names and affiliations.
English
Reviewers end up only caring about the school or institution on the paper. 'Wow, this guy’s from MIT, so it must be excellent.' But the reality is that lots of people sporting NBER or UPenn affiliations still write utter garbage. Besides, irresponsible reviewers are much more damaging than irresponsible authors
Megan Stevenson@MeganTStevenson
I’m reviewing papers for a computer science conference and omg so much slop. We should really abolish blind submissions because people need to be putting their professional reputations on the line when they submit this junk.
English
Hayashi Heikichi 리트윗함
tell that to my group at Harvard econ that fervently gathered because a Columbia undergrad had a NBER pub out… and said undergrad was 19
James Thomason@jamesthomason
The entire Leopold story is marketing BS. And you fell for it. Dwarkesh Patel's assistant is the brother of Leopold's fiancée's boss, who is Anthropic's chief of staff. Every investor he "attracted" had a pre-existing, years-old personal relationship with him built inside the EA and Stripe social graph before his essay existed. Every dollar was raised from billionaires he already knew who were behind the deals he "picked". Carl Shulman, who actually heads research, is 45 years old, formerly at Peter Thiel's macro hedge fund Clarium Capital, and has deep roots in AI forecasting and the EA safety world. Clarium was itself a concentrated macro fund built around a single big thesis (peak oil, dollar collapse). It blew up spectacularly in 2008–2010, losing roughly 90% of AUM. You didn't watch a prodigy go viral. You watched a controlled fund launch.
English
Hayashi Heikichi 리트윗함
@CampedelliGian True!! I'm an area organizer, so I get to read the reviews as well. Lots of slop. They are not blind to me, so shame on them, lol.
English
Hayashi Heikichi 리트윗함
Claude Fable 5 max outperforms GPT 5.5 xhigh in Codex slightly on ErdosBench 14 smoke problems.
Fable matches Codex on the accepted solved/settled set - Problems 1, 4, 5, and 7 - but adds stronger partial progress on several other problems, especially Problem 3, Problem 8, Problem 195, and Problem 208.
Interestingly, Fable makes bolder claims. Opus 4.8 is more cautious. It solved only Problem 7 and gave partial results for other problems.
I'm testing Fable also on other open Erdos problems, but first impressions are that it might be a bit stronger than GPT-5.5 Pro, but not by much.
I have yet to understand its weaknesses and strengths.
Claude@claudeai
Introducing Claude Fable 5: a Mythos-class model that we’ve made safe for general use. Its capabilities exceed those of any model we’ve ever made generally available.
English