ソースがわかっていないので意味のある数字かわからない（ご注意ください）が、効果的利他主義の資金の多くが依然Global Health & Deveplopmentに回っていると主張するグラフ。個人的に総額が7億ドル/年くらいというのはそんなものか、と思った。たとえばWellcome Trustは単体でその1～2倍あるはず。 https://x.com/leah_pierson/status/1738915590700495108
2024.12.14 OpenAI、E. Schmidt氏と共同で合計$10M（～14億円）のグラント「Superalignment Fast Grants」を発表。人間を超えるAIのアライメントなどのテーマで研究者を公募し、$100K–$2Mのグラントや、大学院生への1年間$150Kのフェローシップ実施。応募は2024.2.24まで。https://openai.com/blog/superalignment-fast-grants
『記憶理論の歴史』で紹介されている論文、F. H. Bradley (1887) Why do we Remember Forwards and not Backwards? この問いには虚を突かれた。 I would answer "because if so, each remembering would have been an TENET-like otherworldly experience".
This is by far the best book on the evolution of AI research I've read. It is also an amazing autobiography of a living star scientist. Every paragraph is beautifully written with passion.
Michael Lewis "Going Infinite"読了。https://wwnorton.com/books/9781324074335 暗号資産ビジネスでビリオネアとなり、つい数日前詐欺等で有罪となったSam Bankman-Fried氏を至近距離で追ったノンフィクション。Sam氏の人物像、動機、思想、行動…描かれていることが全てではないであろうにせよ、驚愕の内容だった。
@carolyn_guay Your embroidery led me to (re)read Dr. Tulving's papers which gave me understanding on his intentions behind his late episodic-semantic distinction, which I believe is ever more important. So thanks!! https://www.slideshare.net/RyuichiMaruyama1/20231024endel-tulvingpdf
Had a chat over drink with a scientist-activist of whom I have been most respectful. It was a pivotal occasion to start thinking seriously about what I should be devoting myself to in the coming year or two.
Lake & BaroniのNature論文。既存の概念を組み合わせるsystematic compositionalityの能力を、メタ学習を施したニューラルネットで実現。35年前のFodor＆Pylyshynの「ニューラルネットはcompositionalityを持てない」との主張への応答として書いているのが面白い。https://doi.org/10.1038/s41586-023-06668-3
@carolyn_guay Hello, I was intrigued by the embroidery in the picture. I think the quote beautifully captures Dr. Tulving's life long curiosity in time and memory. Do you happen to know in what occasion these badges (?) were made and by whom?
@yusterafa Thank you for a wonderful opportunity! For me the film was new mode of public engagement of neuroscience. I will recommend it once it is released.
映画Theater of Thought視聴。（お付き合いいただいた方ありがとうございました。）Neuroscience as seen from an artistという感じだった。Brain scientists can be as mystical as a wire walking performer.
O'Sullivan & Ryan 2023、よい論文だった。記憶研究で言及される記憶痕跡（engram）とは記憶のvehicleなのかcontentなのか。脳内現象の因果関係の連なりのどこ・どの階層をengramとして切り出すのか。高度な細胞測定・制御技術が発達したからこそ求められる概念の精緻化。engram biologyの始まり。
It's become feasible to host a virtual screening in October. I am thinking of convening something like an "ad hoc group of people in (greater) Tokyo who engage in neuroscience research or public policy". If you are eligible and interested, please contact me.
@akseli_ilmanen Couldn't miss the cone! I think Bergson is very interesting to revisit from contemporary neuroscience and neurophilosophy of time. I was introduced to Bergson by a Japanese Bergsonian @hiraiyasushi1 who leads this: https://matterandmemory.jimdofree.com/english-1/ Thanks for starting a great podcast!
If Bluesky is where the next academic Twitter takes place, not going there doesn't seem like an option for me. But just don't know how things will turn out.
This incidence poses some critical metascientific questions. - Can some inadequately designed experiment indicate a theory's unfalsifiabilty? - Is unfalsifiability always good criteria for pseudoscience? - Is it OK to question a theory because of its media misrepresentation?
Later this week, I will have the privilege of sharing my views on the "metascience movement" with one of the central groups of meta-research in Japan. I will do my best to prompt a fruitful discussion.
At this moment few things seem more significant to me than a six-year-old filling her summer with joyous and rewarding moments. Something I need to be reminded often.
2023.07[論文] Casper et al. “Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback.” https://arxiv.org/abs/2307.15217 RLHFに関する未解決問題と根本的な限界を、１）人間によるフィードバック、２）報酬モデル、３）方策最適化の3場面で整理したレビュー論文。 https://t.co/fiFOaBNU2x
”Memory systems make memory a choice. They're also a choice about who you will become, a way of sculpting your future self. The practices I report here have been developed with an eye toward the future self I want.” まさに「記憶のデザイン」（by 山本貴光さん）だ。
@newtaku デネットさんの協力ぶりがすごい！「For the purposes of this project, Dennett provided us with the entire digitally available corpus of his philosophical work. 」「Dennett provided us with sincere written answers to all ten questions, ranging in length from 40 to 122 words.」
”one thing that has completely surprised me is that these LLMs ... I would never have thought that deep thinking could come out of a network that only goes in one direction...that doesn't to make sense to me but that just shows that I'm naive"
I wonder if there are any cases in which activities of "metascience entrepreneurs" are supported or encouraged by public sector through any means. #metascience2023
[Barman+2023]LLMを含むエージェントが（科学的）理解をしたかを判定するためのベンチマークづくりを提唱。ラストオーサーは”Understanding Scientific Understanding”のHenk de Regt。理解とは能力abilityであるという哲学的立場をとることで、LLMによる理解の有無が問える。https://arxiv.org/abs/2304.10327
I often hear that science is common goods. I agree. But it should be very special type of common goods. It may be helpful to contemplate in what sense science is common goods and how it differs from other common goods.
Sarah Robinsさんのレビュー論文「The 21st century engram」素晴らしかった（Wileyに20ドル払った価値はあった）。エングラム（記憶痕跡）概念が、optogentics技術によって精緻化を求められており、しかも概念化においては分岐点もあると。まさにこんな考察が読みたかった。https://doi.org/10.1002/wcs.1653
As "AI for Science" has become a topic of wide interest, I shall devote the next month on looking into its emerging sub-field: "generative AI for metascience".
@michael_nielsen @emiyazono I would be eager to find out how the exploratory usage of LLMs will be taken up by science broadly. And I see your latter point. No or no easy access to the 'ground truth' is what makes large generative models in science interesting from epistemological point of view!
@michael_nielsen I also felt @emiyazono's question strike at the heart! But I wonder if it is more about 'well-definedness of the problem' rather than cleanness of data that makes it unique. Finding a protein structure given an amino sequence seems almost as well defined as winning a chess game.
@metasdl I was mainly thinking about the difference in my personal history using both languages, but surely there must be differences in the languages themselves.
Partly it's because English prompts me differently than Japanese. Language prompts my thinking. My thinking is partly synthesized from the large amount of phrases/paragraphs/speeches I've encountered in each language. With respect to that part I'm no different from LLMs.
Immersing myself in English conversation for the first time in almost a decade made me realise there's some mode of myself that has been suppressed all along. Language matters. Handling English is far from at my disposal but in a different sense liberating.
This is a deep epistemological problem. In what sense AIs beyond our interpretation be counted on as a tool to approach truth (or some kind of objective description of the world science seeks)? In Japan the issue was provocatively raised by a philosopher of science @junotk_jp.
Heard this talk at #metascience2023 but reading the slides made me realize how deep an issue was raised here. Can the predictions of protein structure by Alphafold may be closer to truth than the experimental result? (Though he doesn't use the word "truth".) https://x.com/michael_nielsen/status/1656668760286322692
#metascience2023 is over. Now I have a much better sense of the current state of metascience and its prospect. I admire the persistent effort of people who push this forward. Learned a lot, talked with energetic people like I've never met before. Lots to take home.
At the last moment Prof. Michael Nielsen' s talk agenda at #metascience2023 is out! Seems like he will be talking about AI for science, definately one of the hottest topics of our times. Can't wait to hear. https://metascience.info/events/how-is-ai-impacting-science/
Looking forward to joining #metascience2023 to expose myself to the latest developement of metascience while sharing some perspective from Japan. Also wondering how best to spend an afternoon at DC before the conference..
2023.2 M. Mitchell & D. Krakauer "The Debate Over Understanding in AI's Large Language Models" …LLMが言語を「理解」しているのか？という問いに対する賛否両方の見解をレビュー。人間とは異なる新しい知能の「理解」を測る新しい方法の必要性を主張。https://arxiv.org/pdf/2210.13966.pdf
I was haunted by a dream in which LLM was implemented in my pen and my hand involuntarily started to scribble in my notebook at an incredible speed. My brain now seems to generate AI-inspired hallucinations.
Dr. Christopher Hill at DeSci Lab began his talk by mentioning the top-down policy momentum towards open science and I think that's important. We all agree to the need of change, so decetralized experiments has chance of reaching a critical mass for real transformation.
#DSTC2023 was great combination of young, energetic people building wildly new mechanisms and elder people who can give lessons while truly value those endeavors.
@masuda_ko_1 まさにまさに。Science, technology and innovationで、ただしここでのandはたぶん連言ではなく選言の意味（？）。一方この並置も私には少し気持ち悪くて、というもの科学は「すること」、技術は「つくられるもの」、イノベーションは「起こること」なので。
@rom1504 @michael_nielsen I think all camps agree more or less to developing "AI" (though definitions may differ) as solution to societal issues, but I get the point that people differ in the amount of expectation and sense of urgency.
@michael_nielsen Such sythesis is much wanted! I think one difference here is the "time" dimension in AI threat arguement: people disagree partly on "when" the impact is coming, not on the truth value of some theory.
In the past week we have seen an amazing range of perspectives on the need for AI regulation. It cannot be simply characterized as pro-regulation versus anti-regulation. At least three poles have appeared, each strongly opposing the other two.
Shinn+2023 "Reflexion: an autonomous agent with dynamic memory and self-reflection" https://arxiv.org/abs/2303.11366 GPT 3.5に、過去のアクションやその結果の「記憶」とそれを使った反芻機能をつけたとする論文。より大きなループにLLMを入れる試みは今後活発化しそう。 #LLMと記憶 https://t.co/DysgEplXBx
Ida Momennajad氏(Microsoft Research)の論文"A Rubric for Human-like Agents and NeuroAI"。脳に学ぶAIの評価軸として1)神経科学的妥当性、2)人間様行動の再現、3)工学的ベンチマークを挙げ、相互に従属しない3軸のインタラクティブな活用を提唱。NeuroAIの解像度が高まる。https://arxiv.org/abs/2212.04401https://t.co/VUVV3FqYcP
Our elder daughter askes "how do children learn to speak without knowing the meaning of anything?" Well, you went through all that and amazed us. Just watch closely how your little sister does that in the coming months. She beamed.
This LLM craze becoming somewhat overwhelming, I tend to retreat to pencil &paper to articulate my thoughts. I'm surprised at how much call and response can happen there. My physical notebook is my personal, offline general-purpose assistant, safely out of reach of algorithms.
An ethicist who spoke today at a conference on AI governance described LLM based chat system as "bullshit-generator." Maybe LLMs are also having their temporal heyday, since after a while they will have to generate bullshit out of bullshit of their own creation.
It seems like a roll I should play in the coming months is to make interfaces between such entrepreneurial movements and more established sectors. Not that I aimed for it; just found myself to be in such a position. https://x.com/HiroTHamadaJP/status/1624095406354485248
Ok, I should expect emails I receive to be half-written by LLMs from now on. Shouldn't rely on them to infer the sender's personality. Which makes meeting in person ever more important. https://x.com/shanegJP/status/1632161578450849792
3.99ドル払ってレター3篇読んでみた。培養脳Braindishが有感（sentient）であるなどの言葉遣いに慎重さが必要とする30名弱の連名（Balci et al.）に対して、原論文著者らが科学的に定義したと反論。3名の神経倫理学者が科学用語の定義にどこまでのステークホルダーを含めるか、というフレームで整理。
2023.2.9 UNESCOから報告書「The risks and challenges of neurotechnologies for human rights」発行。ニューロテクノロジーと人権に関する2021年のワークショップのレポート。特段、UNESCOとしての見解などはなく、登壇した専門家の発表要旨をとりまとめたもの。https://unesdoc.unesco.org/ark:/48223/pf0000384185
preparing for a podcast series (in Japanese) ... Once you start thinking about hosting one, you come to realize how carefully prepared each episode of your favorite podcast is.
話題になっているPark et al. ”Papers and patents are becoming less disruptive over time” をざっと読んだ。論文と特許の革新性を、その研究が既存文献と切り離されて引用される（Disruptive）か、先行文献とセットで引用される（Consolidating）かを測る「CD指標」で分析。https://www.nature.com/articles/s41586-022-05543-x