AI is a transcript of our world

1.4/50 Summilux ASPH, Leica M10P, RAW

（This is a sequel to the following blog post.）
kaz-ataka.hatenablog.com

I often hear problematic discussions about the results of machine learning based AI.

When we look at something:

Extremely biased against men
Extremely biased against people of European descent
It is extremely liberal (left-wing in the English-speaking sense of the word)
Too much of the argument is directed at wealthy people
Extremely biased against people with good physiques and looks

These feelings are very understandable, but given the nature of machine learning, they are often unavoidable. This is because machine learning-based AI is a considerable computational environment in which algorithms, including text processing and machine learning, are implemented and trained for a specific purpose given a large amount of experience.

Large amounts of experience can be real (even in virtual space) if it is a game or manipulation (such as picking or driving) that produces results, but in many cases existing data is often used.

Existing data presents a challenge to those who use it in two ways.

First, it contains a large amount of material that is not necessarily factually correct. This should be called the reliability of the information, or whether it is trustworthy or not.

Second, it contains a lot of things that are factually correct but socially unacceptable. This should be called the social justice of the information, or whether it is socially acceptable or not.

Taking search as an example of the most widely used machine learning-based AI tool, the first issue has been a fundamental problem since the birth of search.

In addition to government and other trustworthy information sites, it also indexed sites in the Yahoo! directory, which at the time of Search's birth was the most labor-intensive and trustworthy, as well as the Page Rank invented by Larry Page (named after Page's name and the site's pages). It is almost certain that search platforms still evaluate the degree of use of information sources, the credibility of the site and the credibility of the person who produced the article, quite broadly and deeply.

Incidentally, before the Web, only publishers, newspapers and TV stations were able to provide information to a large number of people, so there was a lot of bias in the information due to the choice of media, but the credibility of the content was more than guaranteed to a certain extent. On the other hand, the information space has changed dramatically in the sense that a considerable amount of information is now suspect, as social media, represented by Twitter, YouTube and Tiktok, are advancing.

Google and the former YST (Yahoo! Search Technology), as well as Bing, Baidu, Naver, and Yandex, have long invested enormous energy in ensuring that the first information that appears on the Web always contains information that is wanted (relevant to interest), useful, and fresh. The so-called "ten blue links" are the result. Anyone who used web services more than 25 years ago should remember that it was common to have to scroll through several pages of machine searches to get to the information you wanted. When we consider that your vast search history has refined these results, this is a great human edifice, the result of the endless efforts of billions of people.

The second issue is not often discussed, but is much more important from a social justice perspective, and also much more difficult. It is directly related to what has recently been called Diversity, Equity, and Inclusion (DE&I). This is because what is at stake in this content has changed radically over time.

When I was a child in my early fifties, there were honestly only about five major DE&I-like issues in Japanese society.

First, eugenic discrimination. Although it became famous in Nazi Germany, it is actually extremely deep-rooted since Plato, and this actually overshadows the other axes. The fact that Japan also had a eugenics protection law until 1996, although in the latter period it was only a skeleton to allow abortion, is very sinful. I cannot begin to tell you the agony of those who were sterilized and those around them.

Second, racial and ethnic discrimination, especially the issue of black liberation. The issue is mixed with memories of the colonial and slavery era, and furthermore, it is rooted in the problem of acceptance of differentness, the problem of being different from each other. These include the issue of discrimination against "zainichi" in Japan and the theory of the Yellow Peril in the pre-war United States.

Third, discrimination against women. This is the history of women's liberation and coeducation, which began simultaneously in the U.S. in the late 1960s. It also includes discrimination in hiring and promotion between men and women in the workplace.*1

Fourth, the issue of wealth and poverty. This is the issue that is now called the social divide, but I suppose that Tiger Mask and Star of the Giants, etc., which crawl out of poverty, were televised with great social significance.

Fifth, national discrimination. This is an issue represented by the North-South problem at that time. It is often intertwined with issues of racial and ethnic discrimination.

If this were the case today, the challenges posed by eugenic ideas would be that while the success and rights of people with various disabilities have been taken for granted, as seen in the Paralympics (an excellent development), the debate has become considerably more complicated, with issues still being revived in relation to designer babies and gene therapy. It is quite difficult to figure out what is socially acceptable and to what extent.

Race is a biologically meaningless concept, and although those who question the elimination of discrimination based on it have become outwardly extinct, they are very reluctant to do so and have not yet eliminated the problem. As a result, it has become a much more sensitive issue than it once was, and the range of acceptable expression is extremely narrow. The tribal issue is a major political issue for the neighboring countries, which was not a major debate at that time.

Gender parity issues are being recognized in Japan as a problem to be solved, but nowadays gender issues naturally include the issue of sexual minorities, represented by LGBTQ. Japan's gender division, male/female, is significantly behind, with female, male, non-binary, prefer not to say, being the global standard. Here, too, the common sense of the past is no longer acceptable.

In the past, the body shape of the people in the ads was never an issue, but even Victoria's Secret (the leading women's underwear brand in the U.S.), which has produced supermodels such as Tyra Banks, Naomi Campbell, and Miranda Kerr, has decided to discontinue their Angels program in 2018 and transitioned to VS Collective, which highlights partners with unique backgrounds, interests, and passions (including Naomi Osaka). It is already out of the question that only beautiful men and women with good style appear in advertisements, and body diversity is now an inevitable trend.*2

The Divide issue is becoming more and more serious, yet there is a mysterious tendency to be afraid to discuss it openly. At the same time, the permissibility of drinking and smoking has dramatically decreased, although the relationship is subtle.

As far as national discrimination is concerned, it is improving considerably with the prosperity of Southeast Asia, represented by Singapore, China, India, Latin American countries such as Brazil, and some African countries, and as a result the zone of permissible expression towards these countries has changed drastically. On the other hand, problems related to the Taliban and the Islamic State (IS) after 9/11 have emerged from terrorism and international politics, problems that did not exist at that time.

Animal right, which few people cared about at the time, is now a sensitive topic, and if you say anything careless with the sense of the 1990s (30 years ago), you will step on a landmine.

In short, it is a completely different world than 30-40 years ago. Much of what was once tolerated is no longer allowed.*3

Nevertheless, if we take a scanned copy of the world's data as it is, the entire memory of these societies will be copied.

It means that the world will be copied with a world full of information that is not "politically correct/socially acceptable" in today's eyes. It is not just about the bias of the information being digitized. It is not just about the trustworthiness of the information, but also about the fact that the white and gray areas of DE&I are in fact moving targets, and the boundaries of what is acceptable are dynamic over time. In other words, it is virtually impossible to completely eliminate this challenge from machine learning-based AI.

As for machine learning-based AI, it will swallow all information that seems trustable once and for all, and provide it in terms of the importance of the data (distribution of the data and whether people will use it). This should be true for search and for large language model (LLM)-based AI like ChatGPT. But the result is that not only is it tainted by social bias, but it is also somehow tainted by the norms by which society operates.

Having said that just as you cannot remove criminal or discriminatory terms from the dictionary, removing them makes the search function, for example, much less useful. This is because first, the information itself is worth looking up, and second, most search terms (queries), which I will not go into detail about, are huge long-tail information that may or may not be used more than once a year, and the satisfaction of the search user depends heavily on whether these are answered or not.

Therefore, it is necessary to have a deep understanding of the literacy of information use in the modern age, to the point that the information sources that are the basis of AI contain information that is not acceptable on these two axes, that both axes are moving considerably, and that it is therefore impossible to create a completely clean tool.

Children should also be taught properly, and although it may be fine to start with "safe search", it is necessary to open up search to adults from around the time they enter junior high school or so, otherwise their interests will not be well served. At this stage, there needs to be a forum where the challenges and risks can be discussed repeatedly, along with the principles of machine learning based on case studies.

While you may be getting a little carried away at this point, I would like to point out two other axes of information provided in addition to Trustable/Acceptable.

The first is the bias of the user's orientation or inclination, although he or she may not be aware of it. This is the third axis. Machine learning absorbs more and more of your usage characteristics and produces more and more results that you like. This is called personalization.

Personalization does not necessarily mean that it is done to an individual. It happens in different languages and in different regions. 災害(Japanese) and "disaster" are processed differently. I don't know if this is a problem, but it clearly creates an information bias. As an interesting example, to remove the ID tagging, start a browser in incognito mode and do an image search for Beautiful woman, खूबसूरत महिला (Hindi) and you will see how different the results can be.

In addition to this linguistic, regional, and social context, there is the added bias of the type of search results you see. It is difficult to recognize this filter bubble or echo chamber problem unless you have a very strong sense that the search results you see and the chatbot responses are not generic. In fact, it may be better to continue searching, etc. without logging in.

Finally, the fourth axis is the degree to which society is actually behind the information. For example, in the 2016 U.S. presidential election between former Secretary of State Hillary Clinton and Donald Trump, the underdog was strongly in favor of Clinton, and many people said they would vote for Clinton when asked, but Trump actually won quite clearly. The other day at a Pixie Dust (PXDT) event, Dr. Yoichi Ochiai, the head of Pixie Dust, told me that this axis is important when looking at information, and I was struck by his words. This Ochiai axis, or degree of honesty, is quite important, but I am not sure how it is reflected in the information we see today or in the results of machine learning that incorporates this information. More research is needed.

As a literacy requirement in the age of machine-learning based AI, I have tried to sort out a bit of the story behind it and the implications of its information absorption. The AI that swallowed the transcripts is one of the greatest intellectual assets we have created, but there is a considerable amount to understand and keep in mind. I want to be able to use it knowing that.

Have fun with it!

ps. Click here for the original in Japanese.
kaz-ataka.hatenablog.com

*1:Although Japan has lagged behind, it is unquestionably just that both men and women should have the same educational opportunities and the same representation in society. In accordance with this perspective, prestigious universities on the East Coast of the United States, which were originally all-boys boarding schools, opened their doors across the board in the late 1960s, and since the end of the 20th century, they have realized gender parity. The global consensus is that what was originally done in co-educational elementary, junior high, and high schools with a 1:1 gender ratio should be done in higher education and in the workplace, especially at the decision-making level. The former U.S. Ambassador to Japan was a woman, and in Mexico, a complete gender parity has been realized even in the National Assembly, but even now only 10% of the Japanese Diet is made up of women. (Reference) Times Higher Education - World University Ranking 2023 : Gender ratio is a basic evaluation item, and even Caltech and MIT, which focus on science and engineering, have approx. 40% women. Incidentally, it is not a male-female ratio, but a female-to-male ratio. This is the global standard.

*2:Comedians as representatives of the general public on Japanese TV variety shows have contributed greatly in this regard.

*3:In Japan, most of these issues are rarely discussed openly, except for those that are convenient to discuss (such as employment of the disabled and the number of female executives), due to the "cover up what smells" culture. This has created an awareness in this country that is decades behind the major countries of the world, and people, especially those in leadership positions, should be well aware of this. I also strongly recommend that you look at how your operations and your company/organization are doing. As a personal note, I was grilled for several hours at the embassy a few years ago by a North American Ambassador to Japan about Japan's bizarre lag in various DE&I attributes, and it really made me want to cry about my country's current state of affairs.

2023-01-29

AIは世界の写し絵

1.4/50 Summilux ASPH, Leica M10P, RAW

（これは次のblog postの続編です。）
kaz-ataka.hatenablog.com

機械学習ベースのAIの生み出す結果について問題提起的な議論を聞くことが多い。

〇〇についてしらべると

極端に男性に偏っている
極端にヨーロッパ系の人たちに偏っている
極端にリベラル（英語圏の意味の左派）よりである
豊かな人達に向けた議論があまりにも多い
体型や容姿に恵まれた人への極端な偏りを感じる

などなどだ。

気持ちは大いにわかるが、機械学習というものの特質を考えると致し方ないところは多い。機械学習ベースのAIは7-8年前にHarvard Business Reviewで整理したとおり、相当の計算環境に、テキスト処理や機械学習を含むアルゴリズムを実装し、大量の経験値を与えて特定の目的に向けて訓練したものだからだ。

安宅和人「人工知能はビジネスをどう変えるか」より（Nov. 2015, Diamond Harvard Business Review）

大量の経験値は、結果が出るゲームやマニピュレーション（ピッキングや運転など）のようなものであれば実戦（バーチャル空間を含む）で良いが、多くの場合は、既存のデータが用いられることが多い。

既存のデータは二重の意味で使う側からするとちょっとした課題がある。

第一に必ずしも事実として正しくないものが大量に含まれている。情報の信頼性と言うべきものであり、英語で言えばtrustworthy (trustable) かどうかだ。

第二に事実としては正しいけれども社会的に許容されないものが大量に含まれている。これは情報の社会的正義性というべきものであり、英語で言えばsocially acceptableかどうかだ。

最も大量に使われている機械学習ベースのAIツールという意味で検索(Search)を例に取ると、第一の課題は検索が生まれたときから続く本質的な課題だ。

これについては政府など出自が確実な情報サイトに加え、Search誕生当時、もっとも人手をかけて信頼性が担保されていたYahoo! Directoryに載っているサイトなどがindexingされ、その上で、Larry Page氏が考案したPage rank（Page氏の名前とサイトのページを掛けた命名）システムにより当初立ち上がった。今でも情報源の利用される度合い、サイトの信頼性、記事を生み出した人の信頼性を相当に幅広く、そして深くratingしていることはほぼ間違いない。

ちなみにWebの前はそもそも出版社、新聞社、TV局ぐらいしか多くの人に情報を提供できなかったために、メディア側の選択による情報の偏りは大いにあったと思うが、中身の信憑性はある程度以上に担保されていた。一方、Twitter、YouTube、Tiktokに代表されるソーシャルメディア化が進む現在では相当の情報が怪しいという意味で情報空間は劇的に変化した。

GoogleやかつてのYST（Yahoo! Search Technology）、そしてBing、Baidu、Naver、Yandexは長らくten blue linksとよばれるウェブ面で最初に出てくる情報に欲しい（関心にrelevantな）、役に立つ（useful）、可能であれば鮮度の高い（fresh）情報が必ず含まれることを担保するために膨大なエネルギーを注いできた。25年以上前のwebサービスを利用した人であれば、機械検索をした場合、数ページ以上もスクロールしなければ、欲しい情報にたどり着けないことが普通だったことをよく覚えているだろう。みなさんの膨大な検索履歴がこの結果を磨き上げてきたと思えばこれは10億単位の人の無限の取り組みによる人類の偉大な建造物と言える。

第二の課題はあまり議論されないが社会正義的には相当に大切で、また相当に難しい問題と言える。これは最近であればDiversity, Equity, and Inclusion（DE&I）と呼ばれる話が直結している*1。直訳すれば「多様性、公平性、包括性」なわけだが、この中身で問題とされているものが時代とともに急激に変わってきたからだ。

50代前半の自分が子供の頃は、正直、社会的に大きく話題になるDE&I的な話題は5つぐらいしかなかった。

第一に優生学（eugenics）的な差別。ナチスドイツで有名になったが、実はプラトン以来の話で極めて根深く、これが実際のところ他の軸にも覆いかぶさっている。末期は単に堕胎を認めるためのものとして形骸化していたとはいえ、日本にも1996年まで優生保護法があったことの罪深さは大きい。断種された方々やその周囲の方々の苦悩は僕には到底語ることはできない。

第二に人種および民族差別、特に黒人解放問題。植民地・奴隷制度時代の記憶が混ざっており、更に根にあるのは自分たちとはそれぞれが違うという問題で異質の受容性の問題だ*2。戦前の米国での黄禍論、日本における在日差別問題もここに含まれる。

第三に女性差別。女性解放（Woman liberation）や米国で1960年代末に一気に始まった男女共学化（co-education）はこの話だ。仕事による男女の雇用や昇進差別問題はここに含まれる。*3

第四に貧富問題。現在Social Divideと呼ばれている問題だが、貧しさの中から這い上がるタイガーマスクや巨人の星などは大きな社会的な意義をもって放映されていたと推定する。

第五に国家の差別。南北問題に代表される問題だ。人種・民族差別問題と絡み合うことが多い。

これが現在であれば、eugenics的な考えがもたらす課題は、パラリンピックに見る通り、様々な障害を持つ方々の活躍や権利が当然になる一方（素晴らしい進展だ）、実は今もデザイナーベイビーや遺伝子治療の関連で課題が復活しつつあり議論は相当に入り組んできている。何がどこまでsocially acceptableなのかの見極めはかなり難しい。

Raceというのは生物学的には意味のない概念であり、これによる差別の撤廃に疑義を唱える人は表立っては絶滅したが、実にしぶとく、まだまだ問題を廃絶できていない。その結果、かつてとは比較にならないほどセンシティブな問題になっており、表現においてacceptableな領域は極めて狭くなっている。部族問題は当時は大きな議論にならなかった隣国の話が大きな政治課題だ。

Gender parity問題は解決すべき課題と日本でも流石に認知されつつあるが*4、現在、gender問題にはLGBTQに代表される性的マイノリティの問題が当然含まれる。日本の性別区分、男/女は相当に遅れており、female, male, non-binary, prefer not to say（女性, 男性, それ以外, 言いたくない）が世界の標準だ。ここでもかつての常識は許されなくなっている。

かつて広告に出てくる人の体型など問題視されたことはなかったが、Tyra Banks, Naomi Campbell, Miranda Kerrなどスーパーモデルを輩出したVictoria's Secret（米国を代表する女性向け下着ブランド）ですら、彼女らAngelsプログラムを2018年に廃止し、ユニークな経歴、興味、情熱を持つパートナーにスポットライトを当てるVS Collectiveに移行した（大坂なおみ選手もその一員）。広告に出るのはスタイルの良い美男美女ばかりというのはすでにアウトであり、Body Diversityはもう不可避な流れと言える。*5

Divideの話は深刻さをむしろ増しているのにも関わらず大っぴらに議論されることは憚られる謎の風潮がある。併せて、関係は微妙であるが飲酒行為やタバコにまつわる許容性も劇的に小さくなった。

国家の差別については、シンガポールに代表される東南アジア、中国、インド、ブラジルなど中南米諸国、いくつかのアフリカ諸国の繁栄とともにだいぶ改善されつつあるが、その結果、これらの国々に対する表現として許されるゾーンは激変した。一方で911以降のタリバン、IS関連の問題などは逆に当時なかった問題がテロと国際政治を起点に発生している。

当時気にする人など殆どいなかったAnimal rightも今や相当にセンシティブな話題であり、1990年代(30年前)の感覚で迂闊なことをいうと地雷を踏むことになる。

つまり30-40年前とは全く別の世界と言っても良い状況だ。かつて許容されたことの多くがもう許されないのだ。*6

とはいうものの、世界のデータをそのままスキャンして写し絵をとると、これらの社会の記憶がまるごと写し取られる。

それは今の目から見ると「政治的に正しい/社会的に許容可能である (politically correct/socially acceptable)」ではない情報が溢れている世界が写し取られるということだ。デジタル化されている情報が偏っているだけの話ではない。情報のTrustabilityだけでなく、DE&Iのホワイト、グレーゾーンが実際にはmoving target（動く標的）であり、このacceptableな境界線は時間とともにダイナミックに動いているからだ。つまり機械学習ベースのAIからこの課題を完全に排除することは事実上不可能と言って良い。

機械学習ベースのAIとしては、trustableそうな情報はすべてを一旦飲み込む、それをデータの重要性（データの分布や人が使うかどうか）という視点で提供する。これは検索においてもそうであるし、ChatGPTのような大量言語モデル(LLM)ベースのAIについてもそうだ。しかし、その結果は、社会の歪みで汚れているだけでなく、社会の動く基準の両方によって、ある種よごれているということだ。

ただ、犯罪に絡む言葉、差別用語を辞書から剥ぎ取ることができないのと同じく、これらを剥ぎ取るとたとえば検索の機能は相当に使えないものになる。第一にそれ自体が調べるに値する情報であり、第二に詳しくは割愛するが、検索ワード（クエリ）の大半は年に一度も使われるかどうかという代物の巨大なロングテールといってもよい情報であり、これらに対して答えが出るかどうかで検索利用者の満足度が大きく変わるからだ。

ということでAIのもとになる情報源にはこの２つの軸でnot acceptableなものが入り混じっていること、双方の軸が相当に動いていること、したがって完全にクリーンなツールを作ることはできないことまでは、現代における情報利用のリテラシーとして深く理解をしておく必要がある。

子供に対してもちゃんと教えるべきであり、はじめは子供向けの検索でいいかもしれないが、中学校に入るかどうかぐらいから大人向けの検索を開放しなければ彼らの関心にうまく答えられなくなる。その段階で繰り返し、事例をもとに機械学習の原理とともに課題とリスクを話し合うような場が必要だ。

ここまでで少々お腹いっぱいかもしれないが、Trustable/Acceptableに加えて、あと２つの提供される情報に関する軸をheads-upしておきたい。

一つは、これは本人が余り意識していないかもしれないが、利用者の志向というか性向による偏りだ。これが第3の軸になる。機械学習はあなたの利用特性をどんどんと吸収し、あなたの好む結果をどんどんと出してくるのだ。いわゆるパーソナライズだ。

パーソナライズと言っても個人に対して行っているとは限らない。言語によっても地域によっても起きる。「災害」と"Disaster"は異なる処理がされているということだ。これは課題と言っていいのかわからないが、明らかに情報の偏りを生じさせる。興味深い事例として、IDのタグ付を外すために、ブラウザでシークレットモードの画面を立ち上げ、そこで Beautiful woman、美しい女性、खूबसूरत महिला（ヒンディ語）と画像検索をして頂ければどれほど違う結果が出るかわかるだろう。

このような言語、地域、社会的な文脈に加えて、自分がどのような検索結果を選んでみているかによる偏りが重なってくるということだ。自分が見ている検索結果や、Chatbotの返答は一般的なものではないという認識を相当強く持っていないとこのフィルターバブル*7、あるいはエコーチャンバー*8的な問題を認識することは困難だ。本当はログインなしで検索などは続けたほうがいいのかもしれない。

最後に、4つ目の軸として、その情報の背後にある社会の本音度というのがある。たとえば前国務長官のヒラリー・クリントン氏とドナルド・トランプ氏が戦った2016年の米国の大統領選では下馬評ではクリントンという声が強く、人に聞いてもクリントンに入れるよという人が多かったわけだが、実際にはトランプがかなり明確に勝った。この軸こそが情報を見るときに大切だという話が先日、Pixie Dust社（PXDT）のイベントで代表の落合陽一氏から出て、たしかにと膝を打った。この落合軸というか本音度は相当に大切だが、これはいま我々が見ている情報や、それを飲み込んだ機械学習の結果にどのように出てきているのかはよくわからない。今後研究が必要になるだろう。

機械学習ベースのAI時代に求められるリテラシーとして、その背後の話とその情報吸収の意味合いについて、ちょっと整理してみた。写し絵を飲み込んだAIは僕らの生み出した最大級の知的資産の一つではあるのだが、そこには相当に理解を深め、留意しておくべきことがある。それを知った上で使い倒せるようになりたいものだ。

Have fun!!

ps. DeepLとDeepL Writeを活用し、英語版も作りました。FYI
kaz-ataka.hatenablog.com

*1:タバコやアルコールのようなそれ以外の問題ももちろんある

*2:帝国主義、中華思想的なものはすべからくこの傾向がある。かつて大日本帝国を名乗っていた日本も相当に色々深く反省し未来に生かさねばならない。

*3:日本はなぜか立ち遅れているが、男女は共に同じ教育機会を得られるべきであり、社会的にも同じrepresentationを持つべきであるということは疑義のない正義のはずである。この観点に則って、もともと全寮制の男子校であった米国東海岸の名門大学たちは軒並み1960年代後半に門戸を開き、20世紀末以降はgender parityを実現している。本来は共学の小中高で1:1で定員を当てている通りのことが、基本高等教育や職場、特に意思決定層でもおきなければいけないというのが世界のコンセンサス。前駐日米国大使が女性だったり、メキシコでは国会ですらgender parityが実現される中、日本の国会はいまでも1割しか女性がいない。（参考）Times Higher Education : World University Ranking 2023 : Gender ratioは基本的な評価項目であり、サイエンス、工学にフォーカスしたCaltechやMITですら女性が4割である。ちなみに男女比ではなく、女男比。これが世界の標準。

*4:先進国の大半ではほぼ解決しつつあるが、注釈のとおり日本は何周か遅れている

*5:日本のTVのバラエティ番組における一般庶民の代表としてのお笑い芸人はこの観点で大きく貢献している。

*6:なお、日本では「臭いものに蓋」文化のせいで、この大半が、議論して都合の良い物（障害者雇用や女性役員の数など）以外は、ほとんど正面から議論されていない。そのためにこの国が世界の主要国からは数十年遅れた意識になっていることは、特にリーダー層の方々は相当に認識したほうがよいだろう。また、自分の業務や会社がどのような状況かをチェックすることを強く推奨する。実際、僕も数年前に赴任直後の北米からのとある駐日大使に様々なこのDE&I的な属性での日本の異様な遅れについて数時間、大使館でガンヅメサれたが、自国の現状に本当に泣きたくなるほどだった。

*7:特定の情報フィルターを通った情報だけの泡の中にいる現象

*8:音が反響する部屋

2023-01-28

Analogical Capacity of Generative AI

1.4/50 Summilux ASPH, Leica M10P, RAW

Midjourney and ChatGPT, two powerful applications, have emerged in rapid succession, and so-called Generative AI based on the Diffusion Model or Transformer architecture is a hot topic around here and there. Midjourney, which attracted a lot of attention for its ability to generate more and more images, is more on the creator side, but when ChatGPT, which returns answers in an interactive manner, was released at the beginning of December, it became a topic of considerable discussion in the Skill Definition Committee of the Data Scientists Society of Japan due to its ability to answer questions. I was also quick to advise the students in my lab, "You guys should use it without thinking too much. Without using it, you will not understand its greatness, its challenges, or anything else.

midjourney.com

Then, two weeks ago at a seminar, a student who was about to graduate said to me,

“I can't live without ChatGPT. I make ChatGPT do all my assignments, my emails, ChatGPT can do SQL, ChatGPT can do diagrams. But when I ask ChatGPT to cite a paper, ChatGPT generates a fictitious paper and cites it.”

He literally uses ChatGPT as his "new servant" and makes ChatGPT write codes, translate, draft reports, and reply to emails to people who are a pain in the ass. The student have ChatGPT cite papers, and he can spot where ChatGPT is making up stuff that doesn't really exist. It's quite impressive.

In parallel, when the US Medical Licensing Examination (USMLE) was solved by the ChatGPT, reports emerged that it scored at or near pass level without any special training, and that it also had high levels of agreement and insight in its explanations. It seems obvious that this is a good match for medicine, where reliable information is available, but it is also likely to be a major factor in the training and future of intelligent professionals.

www.medrxiv.org

That said, a significant number of students at Stanford University are already using ChatGPT. According to an anonymous survey conducted from 1/9 ~ 1/15 (N=4,497), just over a month after it appeared, around 17% of student respondents used ChatGPT for fall quarter assignments and exams, according to an article in The Stanford Daily (founded in 1892) about five days ago.

stanforddaily.com

This is not surprising for Stanford, which is located in the middle of Silicon Valley.

Although university spokesperson Dee Mostofi says in the article that "Students are expected to complete coursework without unpermitted aid”, "In most courses, unpermitted aid includes AI tools like ChatGPT."

In this phase of discontinuity, it is more important for those who create the future to use it and get a feel for the implications of it more than anyone else, rather than simply following the rules and remaining ignorant of them.

This should be certainly the case at UC Berkeley, the rival school across the Bay, as well as at Carnegie Mellon (CMU) and MIT, the four major computer science meccas, along with these two schools.

And now, in a bit of a milestone, Microsoft has announced a major investment in OpenAI, a major player in this field. The implications of this in itself are quite interesting from an industry perspective, but will not be discussed in this article.

openai.com

Back to the topic at hand, the emergence of generative AI tools indicates that education, work, and everything else needs to change . As I wrote in Harvard Business Review Japan (HBR Japan) more than seven years ago, humans are creatures who use everything and anything technology that is created. (This is when artificial intelligence became a hot topic so rapidly and the views were so confused that I was asked to organize a discussion on how we should think about AI, including its implications for society and business.)

www.amazon.co.jp

At the end of the 20th century, when "search" was invented at the Stanford campus in Palo Alto, it was said that the value of simply providing answers was disappearing, and this is a sign that we are entering a new era. From this perspective, the current education system, in which students are given many questions in cases where there is a fixed answer, and compete to give the correct answer as quickly as possible, is really approaching a pointless world. This is because machines are better at this, and we are entering an age in which we are more likely to leave it up to them. (On the other hand, the ability to dig into questions that have no starting point is more important than ever.)

According to the Stanford Daily article above, one subject now requires "If you choose to use an AI agent for generating portions or aspects of an assignment, you must disclose this use and cite it in the same manner as you would cite any external source.” Some other subjects have reverted to paper and pencil exams in response to the impact of the ChatGPT.

It is true that there are many cases in which you need to have knowledge like anatomy in medicine crammed into your head to make immediate decisions in the field, and the confusion in higher education in this area will continue for a while, but I believe that it would settle down after a year or two.

Be that as it may, this change means that the ability to formulate meaningful questions, evaluate the answers produced, and provide correct questions and instructions has become critically important. In a real sense, we have entered the age of “liberal arts,” and this also means that we have entered the age of refining "perception," which was the conclusion and core concept of my discussion regarding the essence of intelligence in the past on HBR Japan.

www.amazon.co.jp

The ability to understand various values and beauty in a complex and vivid way, a sense of beauty based on this, a heart that wants to have a certain thing, and a vivid sense of knowing that this is not good enough, are really the key to success in the future with these Generative interactive AIs. The starting point is to feel deeply and vividly with the body, such as by stroking and licking.

As I discussed with Dr. Yoichi Ochiai at Weekly Ochiai at the beginning of the year, Japan's elementary and secondary education system, which mainly provides almost the exact opposite education, has the potential to become a device that produces a large number of "high IQ people who are simply put to work" if drastic changes are not made. Even though there are many aspects that students will hack on their own, if they are not given considerable freedom in elementary, middle, and high school, their ability to generate questions and to feel and evaluate things in their own way will be considerably damaged.

newspicks.com

As you will soon see, ChatGPT is very different from so-called "search. While it is possible to use ChatGPT as a search tool by typing in the words you want to know, this is not an approach that unleashes the power of this Large Language Model (LLM)-based tool, because search is better and more accurate at such things, and LLM-based tools are far better at them.

This is because search is better and more accurate at such things, and there are other things that LLM-based tools are overwhelmingly better at. (Some of you may remember that Galactica, the AI tool that Meta introduced as if it was an AI that could answer everything, went on a bit of a hiatus after three days.)

www.technologyreview.com

Search, which is arguably the most used AI-based tool today, is extremely strong in tapping out what the answer is. It can search even behind the scenes around the world and give you the answer in an instant. And it is highly reliable. In reality, it is a process of looking up digitized information in advance to determine its reliability and then organizing it, but if done well, it can arrive at the answer you want at a shockingly fast pace.

ChatGPT is a Chatbot that provides a general direction for things you want to know, or things you want to do but don't know what to do, in other words, things you couldn't get answers to before unless you tried to figure it out yourself or involved people who know what they are talking about. Moreover, answers are generated in real time. This is a unique and moving experience, just like when you are asking something to a human being. I also like that different answers are generated each time. If you give the right instructions, you can even have ChatGPT write programming code.

This is true not only for text, but also for images, music, and other generative systems in general, but the speed at which this new machines creates something is incomparably faster than that of humans (see Ochiai's demonstration in Weekly Ochiai). What this means is that many people are approaching a state similar to Michelangelo, who had a large number of disciples and devoted his energy to directing and finishing their work.

The problem is that things are said that are not necessarily correct, and as my students have pointed out, are often somehow made up or pieced together with similarities (the generative system itself is inherently creative, if you think about it). The same is true of papers, and facts. When I acquired the building blocks (tsumi-isi) in the image above, I asked, "What are the similarities and differences between “Sai no Kawara”(the Children's Limbo) and The Myth of Sisyphus?" and got the following answer

In Greek mythology, the Sai no Kawara is the place of judgment for the dead. Those who have done evil are cast into the underworld, while those who have done good are accepted by Elias.
On the other hand, "Sisyphus" is the place of judgment for the dead in Egyptian mythology. The dead are judged by Osiris, and those with light hearts are brought back to life, while those with heavy hearts are cast into the underworld.
Both depict the place of judgment that the dead undergo, so there is a similarity, but there is a difference because the person or method of judgment is different.

This is quite an interesting answer, but it clearly confuses the Buddhist worldview of the Sai no Kawara with Greek mythology (Sisyphus) and Egyptian mythology.

However, this is to some extent unavoidable considering that many phrases and meanings in LLM are represented as vectors in a multidimensional space. The following presentation on Google translate will give you some idea of the representation in multidimensional space.

https://arxiv.org/pdf/1611.04558v1.pdf

However, it is even a little impressive that "Sai no Kawara" is a concept that is quite close to "Sisyphus" in terms of vector space. Perhaps it is because we are only a few steps away from the discovery of similarity as in humans, the extraction of meaning from it, and its extension from some kind of idea and analogy.

In fact, the largest use (about 60%) of the Stanford students who used ChatGPT in the previous article was as a brainstorming partner. Even Stanford students, who are usually close to experts and people who know a lot about most things, are not likely to ask people for something like this kind of college homework. However, most of our daily ideas start with something that is almost unimportant. And when we ask, we get something back from ChatGPT almost instantly. A messy answer is not a bad thing. People are more messy and more random, but communication is still possible, and something interesting can come out of such dialogue.

I am not the only one who feels that this is leading to something great.

One of my greatest joys is to imagine something more by connecting things that are not normally connected, and I am now in possession of another new tool.

Now, with new tools in hand, let's go back to the real world.

ps. For a sequel, click here.
kaz-ataka.hatenablog.com

Note: This blog entry is based on the original Japanese entry translated by DeepL (also an LLM-based AI tool) with some minor modifications.

kaz-ataka.hatenablog.com

2023-01-28

生成系AIのアナロジー力

1.4/50 Summilux ASPH, Leica M10P, RAW

Midjourney、ChatGPTと立て続けに強烈なアプリケーションが出てきて、Diffusion model（拡散モデル）やtransformer architectureに基づくいわゆるGenerative AI（生成系AI）がそこらで話題だ。ガンガン画像を生み出すことで一気に注目を集めたMidjourneyはクリエーター寄りだけれど、11月末、対話型で答えを返してくれるChatGPT*1が出てきたときに*2、あまりの回答力にDS協会*3のスキル定義委員会でもひとしきり話題になり、僕も自分の研究会の学生たちに「君ら、深く考えずにまずは使い倒したほうがいいよ」と早々にアドバイスした。使わないことには凄さも課題も何もわからないからだ。

すると二週ほど前のゼミで、ある卒業を控えた学生が

ChatGPTがないと生きていけない。課題もメールも全部やらせている。SQLも図の作成もできる。ただ、論文を引用してくださいと聞くと架空の論文を生成して引用してくるのでこれだけは使えない

と宣う(のたまう)ではないか。本当に彼は文字通り「新たな召使い」としてChatGPTを使い倒し、codeも書かせ、翻訳もやらせ、レポートも下書きを書かせ、めんどくさい相手へのメールの返事も書かせている。論文を引用させて、それが本当には存在しないものをでっち上げているところもちゃんと見破っている。なかなかである。*4

並行して、米国医師資格試験（USMLE）をChatGPTに解かせてみると、特殊な訓練なく合格かそれに近いレベルのスコアで、説明の一致度や洞察も高かったという報告が出てきた。信頼できる情報が揃っている医学との相性がいいのは当然としても知的プロフェッショナルの訓練と未来に大きく関わることは間違いなさそうだ。

www.medrxiv.org

かと思えばStanford大学ではかなりの数の学生がすでにChatGPTを使っているという。まだ現れて1ヶ月あまりの1/9 ~ 1/15まで実施された匿名調査によると（N=4,497）、学生の回答者の約 17% が、秋学期の宿題や試験にChatGPT を使用したというのだ。5日ほどまえの学生新聞（The Stanford Daily; 1892年創刊）の記事だ。

stanforddaily.com

シリコンバレーのど真ん中にあるStanfordとしては当然だろう。

大学スポークスパーソンのDee Mostofi氏は

学生は「許可されていない助け」なしにコースワークを仕上げることが期待されており、「許可されていない助け」は多くの場合、ChatGPTのようなAIツールを含んでいる（“Students are expected to complete coursework without unpermitted aid,” “In most courses, unpermitted aid includes AI tools like ChatGPT.”）

と記事内で語っているが、この不連続性の高い局面においては、単にルールを守って知らないままでいるよりも、使って使ってそれで意味合いを誰よりも身体で掴むことのほうが未来を生み出す人たちにとって大切だからだ。

Bay Areaの向かいのライバル校である UC Berkeleyでもきっとそうだろうし、この二校とならびComputer Scienceの４大メッカと言えるCarnegie Mellon（CMU）、MITでもそうだろう。

そして現在、ちょっとした節目ということでMicrosoftはこれの大元であるOpenAIに大きな投資をすることを発表した。この意味合い自体が業界的にはかなり興味深いが本稿では割愛する。

openai.com

話を戻すとこのような生成系AIツールの出現は、教育も仕事も何もかも変化が必要だということを示している（先日ちょっと話題になったリンク先のTweetでも書いた）。人工知能が急激に話題になり見立てが混乱していた七年あまり前、AIをどう考えたらいいのか、その意味合いも含めて議論を整理してほしいとの依頼を受けて書いたとおり、人間は生まれた技術を何もかも使い倒す生き物だからだ。

人工知能はビジネスをどう変えるか DIAMOND ハーバード・ビジネス・レビュー論文