I’ve finally started using Cursor, and it made me realize that a lot of white-collar work can now be brute-forced by massive product-sum operations running in data centers somewhere out there. I’d really like to see my energy consumption in real time.
Lakatos Award Lecture by Mazviita Chirimuuta. https://www.youtube.com/watch?v=_KkXjBYsUMQ Her account offers perhaps the most cautious perspective on whether understanding and other scientific aims can be pursued simultaneously in the era of AI for science.
(Re)read Michael Polanyi's "The Tacit Demension" （暗黙知の次元）. He is justifying the human ability to do science without relying on some “explicit” foundation. The central question of our era imo is whether AI can be a beholder of tacit knowledge. https://www.chikumashobo.co.jp/product/9784480088161/
What does it really mean when one says "intelligence is computation" or even "world is computation?" Are those non-trivial, refutable claims? Or are they closer to tautology?
In my understanding CPC-MS paper exactly tries to model this symbolic consensus building in scientific community, including disruptive moments called paradigm shifts. https://x.com/i/status/1930202402537517224
This is by far the most thoughtfully and beautifully written account of how to think about AI's role in science. Can't agree more to the view that science is intersubjective representation building. https://x.com/Sara_Imari/status/1999187572682260553
@m_crosscombe @tanichu Ours is not "ethics for CS" but rather "AI basics and philosophical/ethical literacy for all freshman" textbook. I would like to hear your comments once it is out (next year)!
@m_crosscombe @tanichu Thank you for the information! Very interesting. I imagine the curriculum must have changed a lot during deep learning and gen-AI era. We are interested because we are preparing a "AI and philosophy" textbook (in Japanese)!
Much appreciated that the Human-AI Alignment Tutorial materials from last week’s NeurIPS now public. They show how sophisticated AI-alignment discussions have become. Views may differ, but the tutorial offers an excellent overview. https://x.com/huashen218/status/1997424920218005635
Prof. Julian Togelius writes he stood up in AI for Science panel to say automating entirety of science is 'evil.' https://x.com/togelius/status/1997920576804143583 I think this is exactly when metascientific questions arise: What is the purpose of science? Is ‘fun’ part of science auxiliary or essential?
How to operationalize the concept of "understanding?" IMO, it can seen some 'effective' dimension-reduction or abstraction of world models (=generative models that predict the state of the world upon action). That's how I've interpreted Rich Sutton talk. https://x.com/rmaruy/status/1996271700749754572
@m_crosscombe Ah, that's part of the point. If people are giving up 'mechanistic' type of understanding the LLM, we're back to the question "What understanding of the model can we hope to have."
I am becoming more certain that "understanding" will be right at the center of gravity in the discourse on artificial intelligence next year. Do LLMs understand? How can we understand AI? What does understanding mean? How can we model it?
生成AIがあらゆる文章、音声、動画の真正性の証明（proof of authenticity）を剥奪した先に、人々が拠り所として求めるのは大規模な国家監視（mass surveillance）だとTufekciさんは警告する。 そのうえで、生成AIを禁止するのではなく、技術には技術（暗号学的な技術など）で対抗すべきだ、と。
Prof. Bengio’s “Scientist AI” is a machine that discovers the “truth” of the world by updating Bayesian posteriors without having any form of “agency.” He aims to “disentangle understanding from agency.” I prefer a more pragmatic account of scientific “truth.”
At the end of the day, I'm not interested in artificial intelligence par se, but I care more about the future of human collective intelligence, and the role AI will play in it.
Science is humanity's collective world model. It enables us to collectively predict the world upon our collective actions. LLMs are tools that enable individual world models to query the collective world model for individual use with unprecedented efficiency.
科学というのは、プラグマティックに捉えると人間集団の集合的知覚と集合的行動に対する集合的予測を可能にする集合的世界モデルである。谷口 et al. https://arxiv.org/abs/2501.00226 ではLLMが集合的世界モデルとなっているが、ここでは科学そのものを集合的世界モデルと見てみる。すると（続）
@bratton No, no, my tweet was too rough. What I meant by 「人間の先にAIを置く」was perhaps closer to "put AI as something that comes after humans" (not only as another "tool" that makes up the technological landscape). I was really inspired by your talk in Tokyo. I will read your book.
@bratton As such, AI is not only a technological "landscape" that humans can walk on, but it is something that create and explore the landscape itself. Or so I have interpreted.
@bratton Sorry if I misrepresented your philosophy. What I had in mind was the diagram where you placed AI on top of the technological scaffold that humans have created.
Data Center Watchというサイトは、過去2年間で、米国の640億ドル規模のデータセンタープロジェクトが、地域住民の反対によって中止または延期に追い込まれたと報告。全米24州に少なくとも142の反対運動グループがある。https://www.datacenterwatch.org/report
@blaiseaguera Especially inspiring to me was the part where you discuss the "probably right distinctions" in Ch 9, I even made a one pager summary for my own sake :)
Thanks for the great book! I think I'll be rereading it time and again. https://t.co/RuJrgN8A7H
@blaiseaguera Dear Dr. @blaiseaguera Thank you so much for the reply! I joined the event on Saturday and was truly inspired there as well. https://x.com/rmaruy/status/1977382950799851657 ( ... I need to be writing these things in English some day. )
The above is my rough writeup on Antykithera Tokyo that took place on Saturday. ...Though I really feel the need to write something like this in English for reciprocal conversation that is much wanted, at this moment I could only take time to write in Japanese.
2025.10.7 Hugging Faceがposition paper "AI for Scientific Discovery is a Social Problem"を公開。AI for Scienceの民主化に必要なものとして、1）分野間協働と教育、2）上流課題のベンチマーク整備、3）データの標準化、4）アクセシブルなインフラを挙げる。 https://x.com/cgeorgiaw/status/1975248652953080000
...additionally, I was more than humbled that my current boss kindly introduced me as a "metascience communicator." If you happen to wonder what that can possibly mean: https://x.com/rmaruy/status/1938130206423978404
One of the speakers was my current boss, and another was my former boss and collegue Professor @tanichu. I'm really excited to see his work playing a central role in this movement. https://t.co/RMvEMcpSHL
Today's event I attended in Kyoto was extraordinary. There is a great vibe towards making Kyoto a new center of research where people seek alternative science, technology and philosophy of life and intelligence beyond AI.
I don't know how long this particular company can continue doing this, but the research community across disciplines should aim toward having this kind of outlet imo.
Becoming more confident that Thinking Machines-style professional blogs are the future of scholarly communication. A medium not to showcase one's achievement, but to share results and insights for further development of collective scientific enquiry. https://x.com/thinkymachines/status/1972708674100765006
One challenge in discussing AI’s impact on science is simply how broad the topic is. “AI” is vast, and so is “science.” Hoping to contribute to a meaningful discussion, I’ve sketched a preliminary taxonomy for the discourse on AI in science. https://t.co/iZ1HqWilrK
"From 'end-to-end' to 'end-to-open-end'" by @tanichu. This is the most succinct way I can think of to describe the next 'north star' of AI. The fruits of AI application where the second "end" is well-defined has largely been taken; now is the time to aim for open-endedness. https://x.com/tanichu/status/1963889188207780126
I think ”intelligence” is an emergent property that is instantiated in different level of systems. We can be "convergent intelligence" at individual level but "divergent intelligence" at the collective level, and visa versa. https://x.com/eshear/status/1954188720455749949
私たちは多分、集合的にしか何かを知ることができない。「私だけ理解し得た何か」は、十中八九勘違いで、そこに拘泥するのは陰謀論への入り口となる。 We can only understand as ”we." ところが今のLLMはどうやら擬似的な「私たちの理解」を立ち上げる力をもつようで、悲劇も生まれているようだ。
2025.8.12 米国のメタサイエンス系シンクタンクのInstitute for Progress (IFP)、AIの米中競争化でAI for Scienceにどう取り組んでいくべきかに関する有識者のエッセイ集を公開。AI Safety/Securityの文脈とも接続されているもよう。読みごたえがありそう。 https://x.com/AlecStapp/status/1954984503715180706
That's why I consider building language models to be addressing only "half" of the core questions of intelligence, the other half being the "symbol emergence problem" (per Prof. @tanichu). https://x.com/sym_eme_outrea/status/1920438817309294938
すごい…Eric Gilliamさんがずっと言ってきた「ARPAモデルには政府グラントを取りに行くイノベーティブな民間企業としてのBBN社のような存在が必要」という主張を英国政府が取り入れ、"Frontier Research Contractor"なる新組織を4社インキュベートする公的グラントができた。成功事例が出るか大注目。 https://x.com/eric_is_weird/status/1952414022969729485
2025.7.22 Microsoftの研究者らによる論文”Working with AI: Measuring the Occupational Implications of Generative AI” 人々がLLMを使って成功裏に行うことができた作業項目（intermediate work activities）と職業を紐づけ、AI利用可能性が高い職業をランキング。 https://arxiv.org/abs/2507.07935https://t.co/Oe74hsVWmF
Blaise Agüera y Arcas氏によるanthykitheraレクチャーを通して観た。とっても面白い。 生命とは「計算」であり、その計算とは過去から未来を確率的に予測する能動的推論である。このテーゼを、生命の起源から「超生命体」としての人間社会、Transformerまで語り切る90分。https://wii-film.antikythera.org/
"Writing is thinking." - Yes, I think we all know that. The real question may be: "How much extra 'thinking' between 'prompting' and 'writing'?" #タスクの哲学 https://x.com/DKThomp/status/1947107825265565981
Watched the Sutton lecture. My take from the Symbol Emergence Systems Theory's point of view is that the "Era of human data" is when AI copied human-generated symbol system; in "the Era of experience" AIs begin to build their symbol system on their own. https://www.youtube.com/watch?v=FLOL2f4iHKA&t=1464shttps://t.co/xRjNC4WZFR
@michael_nielsen @kanjun Nature article on #metascience2025. https://www.nature.com/articles/d41586-025-02065-0 "We want metascientists to use their skills and talents to shape and improve research, but they must not stop there. They need to think equally about how to be useful to society."
@jontreadway And this is the "center of gravity" shift as I perceived it between #metascience2023 and #Metascience2025. Again, this is highly subjective, and of course, completely value-neutral. I just wanted to understand the difference I felt between the two conferences. Both were great. https://t.co/rZy4CxOmXs
Inspired (partly) by @jontreadway's tweet thread, I have come up with a highly subjective and tentative map of different strands in metascience. It is intended as a starter to begin talking about what people (want to) include in metascience. #metascience2025 https://t.co/9DgbzWwsDH
These are important observations. It is completely natural and understandable that a big umbrella like "metascience" comes in different "strands." I prefer to consciously include them. That is *not* to say they all have to be represented in a single conference. #metascience2025 https://x.com/jontreadway/status/1940726989117432075
@jontreadway @michael_nielsen @kanjun @mattsclancy @Convergent_FROs wow, very interesting as a science communication phenomenon... I'll definately look into the Bsky version of metascience!!
@jontreadway @michael_nielsen @kanjun @mattsclancy @Convergent_FROs Thanks for the follow-up! Yes, I agree. I think they represented the out-of-the-box "experimental" spirit of the metascience.
I wonder if there was any reference to @michael_nielsen & @kanjun 's "A Vision of Metascience" in #metascience2025, since I believe that after 3 years, it still is the most elaborate piece that carries the spirit of metascience. https://x.com/michael_nielsen/status/1582489651603861504
Do we need "metascience communication?" If so, by whom and of what kind? -- This is a question I would love to hear thoughts from the attendees of #metascience2025.
150 years ago, this was the most efficient and compressed form of the human knowledge in its entirety. But not any longer, with trillions of parameters in neural networks serving that role in vastly more accessible way, or so it is hoped. https://t.co/1yDb7Jw6bV
「三人寄れば文殊の知恵」は英語だと”Two heads are better than one"になることを教えてもらった。集合知を生み出す単位が３か２か。やはり、dialog文化が根底にあるのだろうかなど、興味を掻き立てられる。Geminiさんに尋ねたところ、"Two's company, three's a crowd"という諺もあることを知る。
Stanford大のBrynjolfssonらの論文”Future of Work with AI Agents”。104の職業に従事する米国1,500人の労働者と52人のAI専門家への聞き取りで、労働者がAIによるautomationやaugmentationを望むタスクと、実際に可能なタスクのギャップを議論。#タスクの哲学 にとって重要。 https://arxiv.org/abs/2506.06576https://t.co/oFPrQrat1k
デビッド・アレン『新装版 はじめてのGDT』読了。1年のスパンでやることを全部書き出して週次でレビューするGetting Things Done技法の指南書。これが実践できていれば、人生もうちょっとまともだったろうと思う自己啓発本。 https://www.amazon.co.jp/dp/4576250426/
Wow Gemini gave me other examples: Plurality, Particularly, Regularly, Singularly, Chlorel, Rollerlike, Paralleling, Unparalleled, Rurally, Worldlily, Squirrelly, Territorially, Pearl-like, Girl-like ... I don't think I'll ever use "territorially" in my life though lol
'Plurality' has l, r, and l in five consecutive letters, which makes it excruciatingly hard for a native Japanese speaker to spell out correctly. ... Or is it only for me?
「〇〇という概念には定まった定義がない」という言い方があるが、あらゆる概念がmore or lessそうなのであって、carved in stone（教科書？）に見える概念すら時と場合によってはチャレンジ可能。言葉の意味はある種のパワーゲームの中で決まるというこの言語観は、しかし意外と共有されていない。
@nshinshi_ I think he is right, and also think that "to get artificial intelligence to work" changes its meaning through time; what kind of "computation" we want AI to mimic is always asked anew in each generation. IMO that's why Marr's levels are immortal. https://x.com/hardmaru/status/1923647041420525896?t=LKEi-Kx-1c9BFFC_jqrxtA&s=19
2025.4.15 AI Snake Oilの二人組による提言レポート「AI as Normal Technology」。AIを超知能へと至る可能性があるものとしてではなく、「通常のテクノロジー」として語りガバナンスしていく必要性と道筋を描く。説得力あり。AI ELSI関係者は必読と思う。 https://x.com/sayashk/status/1912159334814929014
The last session of this event on AI for Science was really inspiring. The panelists were front-runners in utilizing AI in science, yet none unconditionally advocated for AI's potential. An important metascientific discussion worth continuing. https://x.com/RMBattleday/status/1909253073433952369
IEA（国際エネルギー機関）がAIにトピックを絞って取りまとめた異例の300ページのレポート「Energy and AI」が発行に。データが充実しているだけでなく、AIのどこに電力がかかるのなどの初歩から解説しており教育的。議論のベースラインがだいぶ上がりそうだ。 https://x.com/IEA/status/1910255934435217716
記号創発アウトリーチの方は何とかトンネルを抜けて、一つの手ごたえをつかんだ（客観的には、小さくとも不可逆的なムーブメントの種を蒔けた）気がする。 昨年からやらせてもらっているOpen academia Lecturesの方はまだ暗中模索。何かを掴みかけている気もするが……集中的な考察と議論が必要かも。
Blaise Agüera y Arcas氏（Google, CTO of Technology & Society）によるオンライン書籍「What is Intelligence？」 公開されている前半を読んだがとても面白い。 https://whatisintelligence.antikythera.org/ 生命の誕生から、「計算」のレンズで知能を語り尽くそうとしている。後半がどの方向にいくのかとても気になる。
Kokotajlo et al. "AI 2027"をざっと読了。AIのcapabilityの増加と、misalignmentが起こる想定には同意できないし、このAI企業1社＋米中政府のダイナミクスに未来のすべてを還元するナラティブの普及は利より害が大きいと思う。が、アペンディックスの情報と考察は非常に充実していて勉強になる。
断想メモ：知的生産のための「space to think」は、平井理論で言えば「現在の窓」を大きく開くことに相当するのだろう。思考で遊ぶ時間幅を確保するにはまさに「空間」が必要で、人によってはそれは「研究室」のような場所だろうし、私の思考にとって大事なのはスマホでアクセスできるデジタル空間。 https://x.com/rmaruy/status/1356810917640503297
あと、異分野間のディスコミュニケーション的なものは全くなく、事前の予想は参加メンバーの学際対話力を見くびりすぎていた。自身がcareする問題に挑む道具立てが他分野に見つかる可能性に皆貪欲だったし、work in progressの体系を一緒に彫琢しようというcharityがあった。https://x.com/rmaruy/status/1900689150011208177
夢想してきた「space to think」が、#CPCcamp にて局所的に開き始めているかもしれない。似て非なる方向に考えることに動機付けられた人たちが、on the flyに設計される「コミュニケーション場」の中で、自由に話し考え、その共有を促される、区切られた時空間。あと2日、どうなるか。 https://x.com/rmaruy/status/1356810917640503297
後半では、アカデミストCEO柴藤さん@RShibatoとの対話も予定。お二方は、ことなるアプローチで研究エコシステムを活性化する新たな仕組みづくりに挑戦してきた、日本を代表する「メタサイエンス・アントレプレナー」（by Qiu & Nielsen)だと思っており、どのようなやり取りになるのか楽しみです。
【3/11（火）20時～ Open academia Lectures】 研究知が活きる世界を目指し、AIによる情報プラットフォーム構築などを手掛ける西村勇哉さん@DialogueBarをお迎えします。研究エコシステムの”インフラ”部分に関心のある方には、かなり面白い話になると思います。 https://peatix.com/event/4309076/view
... As AIs begin to 'do' stuff rather than 'think/talk,' the question is not "Will humans in X be replaced?" but "How will the social process of X change?" e.g. The AI scientist is more a disrupter of scientific enterprise than just a good scientist cf. https://t46.github.io/blogs/ai_scientist_and_metascience.html
@michael_nielsen Thanks for your clarification! I think I haven't misrepresented your query. (I found it intriguing that X's translation must have filled the missing object to "teach" in the original QT with "us," whereas in Japanese we don't have to designate the object.)
インドの“One Nation One Subscription”計画、すごいな。「同国の6,300を超える機関に属する約1,800万人が、30の学術出版社が刊行する約1万3,000タイトルの電子ジャーナルにアクセスできるようになる見込み」https://current.ndl.go.jp/car/230581
Harnad 1990 "The Symbol Grounding Problem"読了。記号的AIとコネクショニズムの二項対立のなかで、両者をつなげることで解くべき問題として記号接地問題を提示。「これはAIに解けないだろう」ではなく「こうすれば解けるかも」という形で問題を提示しているのが印象的。https://arxiv.org/html/cs/9906002/
@homme_pensant 【2/21（金）19時～ Open academia Lectures】小林晋平先生、佐伯栄一さんをお迎えし、お二人が作ってきた24時間のYouTube Live、クラブイベント等、型破りの科学コミュニケーションから見えてきたことについて伺います（Zoom／有料／事後視聴可）。 https://peatix.com/event/4263968/view
AIロボット駆動科学シンポでの@ifuaa さんのフラッシュトークを聞いて"Reasoning"について少し考えたくなり、下記noteで紹介されているYu+ ”Natural Language Reasoning, A Survey” をざっと読んでみた。Reasoningの意味の多様性とNLPにおけるたくさんのベンチマークが紹介されていて興味深い。 https://x.com/ifuiaa/status/1880568547036025301
2025.1 Luccioni, Crawfordらによるプレプリント"From Efficiency Gains to Rebound Effects: The Problem of Jevons’ Paradox in AI’s Polarized Environmental Debate" https://arxiv.org/abs/2501.16548 AIの環境負荷について、ジェボンズのパラドクスなど間接的インパクトを含む論点を整理。
#記号創発システム論 は、生成AIの学問の「もう半分 (the other half)」だと言ってみたい。すでに「意味」があるデータから意味を操るマシンを作るのが生成AI研究だとすると、言語記号がもつ意味の出自と、生成AIが今後どう「意味を作るか」を問う視座を与えてくれるのが記号創発システム論。 https://t.co/pcQZtYPsOf
Luc Steels氏の2008年論考「The symbol grounding problem has been solved. so what's next」読了。同氏の”言語ゲーム実験”が「記号接地問題を解いた」とする心を、”記号”や”表象”という用語がもたらした混乱を解きながら語り、その先に広がる研究を展望。#記号創発システム論
Taniguchi+2024 "Generative Emergent Communication: Large Language Model is a Collective World Model" in arxiv 読了。マルチエージェントによるコミュニケーションを介した表現学習で創発した言語を学んだLLMは「集合的な世界モデル」たり得ているという。世界観のある論文。#記号創発システム論 https://x.com/tanichu/status/1875710664528027812
2月のOpen academia Lectureでは、10年来のお付き合いのある小林晋平先生@beatphysfreak、そしてとともに異次元の科学イベントを作ってこられた作曲家の佐伯栄一さん@thepbjをお迎えします。 https://peatix.com/event/4263968/view これは楽しみです！！
@eric_is_weird Excellent point! I've always wanted to find a positive side to having to talk within a language-based isolation. Not that I can strongly argue for such an advantage. https://x.com/rmaruy/status/1102490990969909248
② Open academia Lectures #4 渡邉文隆さん @fwatanabe 「“研究への寄付”をどう広げるか」 https://peatix.com/event/4232208/view - 1/17(金) 19:00- - Zoom/有料/事後動画有 有料ですが、参加者同士の感想会などの機会を付加価値として提供予定。「研究と寄付」を改めて考える機会として、ぜひ参加ください。