記事作成日:2023年11月2日


AI

新聞記事でAIに関する重要な英語の論文が紹介されていました。 その論文は経済研究へのAIの活用(特にLLMという言語モデル)に関する内容ですが、実際のビジネスでも活用できる内容と 考え、その翻訳をしてみました。 30ページほどの論文ですが、重要な部分だけを選択して英語と日本語訳の対比という形式で数回に分けて紹介します。

今回は全11回中の4回目です。

第1回 1 Introduction:はじめに
第2回 2 What Are LLMs?:LLMとは何か?
第3回 3 Applications:応用
第4回 3.1 Ideation:アイデア出し
第5回 3.2 Writing:執筆
第6回 3.3 Background Research:背景調査
第7回 3.4 Coding:コーディング
第8回 3.5 Data Analysis:データ分析
第9回 3.6 Mathematical Derivations:数学的導出
第10回 3.7 Summary:まとめ
第11回 4 Outlook and Concluding Thoughts:展望と考察

3.1 Ideation

One part of research for which LLMs are increasingly useful is the process of ideation, i.e., generating, selecting, and developing ideas. They can help both as assistants and as tutors. This is illustrates vividly how LLMs differ from earlier applications of deep learning in economics - they display a form of creativity that had long been reserved for humans. Although the capabilities of cutting-edge LLMs in ideation are impressive, they also have clear limitations. Some of these derive from the fact that I am experimenting with models that are not explicitly fine-tuned to economics, but some of them are rooted in the overall limitations of LLMs at the time of writing.

3.1 アイデア出し

研究活動において、LLMがますます有用になってきているのが、「アイデア出し」の領域です。アイデアを生成したり、評価と選択したり、開発したり… アシスタントのようにも、家庭教師のようにも働いてくれます。これは、LLMが経済学の研究において、どのように従来の深層学習モデルと応用法が異なるのかを示唆しています。長年人間にしかできなかった創作性ができることを、LLMが見せてくれています。ただ、注意が必要なのは、最先端のLLMのアイデア生成の能力は確かに優れているが、明確な限界があります。こう感じる理由の一部は、私が試行錯誤しているモデルは経済学にファインチューニングされていないことに起因します。それでも、一部の限界は、執筆当時にLLM全般に当てはまることです。

3.1.1 Brainstorming

Since LLMs are trained on a vast amount of data that represents a cross-section of all human knowledge, they are quite useful in brainstorming (or, perhaps more aptly, net-storming) ideas and examples related to a defined theme. The following prompt illustrates a simple example. Throughout the remainder of this section, I will present all examples generated by LLMs in boxes, with the prompt that I entered in Sans Serif style and the system’s generated response in typewriter style. For example, in the box below, the first sentence was my prompt, and the numbered bullet points below are what the system generated.

Please brainstorm economic channels through which advances in AI may increase inequality

  1. Automation of low-skill jobs, leading to job displacement and increased unemployment
  2. Increased demand for highly-skilled workers, leading to wage inequality
  3. Increased access to capital for those with the resources to invest in AI technology
  4. Increased surveillance of workers, leading to decreased privacy and potential exploitation
  5. Increased use of AI-driven algorithms to make decisions about hiring, lending, and other economic activities, leading to potential bias and discrimination
  6. Increased use of AI-driven algorithms to optimize pricing, leading to increased costs for consumers
  7. Increased use of AI-driven algorithms to optimize marketing, leading to increased costs for businesses
  8. Increased use of AI-driven algorithms to optimize investments, leading to increased returns for those with the resources to invest
  9. Increased use of AI-driven algorithms to optimize tax avoidance, leading to decreased tax revenues for governments
  10. Increased use of AI-driven algorithms to optimize labor costs, leading to decreased wages for workers

The channels listed by the model are not path-breaking, but they are on-topic, largely reasonable, and span a wide range of what I could think of plus some more. One can also observe a tendency that is all-too-familiar from human brainstorming after about point 5., creativity declines and the responses become somewhat repetitive.

3.1.1 ブレーンストーミング

LLMが、人間が持つ知識の大きな一部分に相当する訓練データを用いて事前学習されているため、ブレーンストーミング(以下、ブレストと略することがある)、つまり、特定のテーマのためにアイデアや具体例を出すことには結構役立ちます。以下のプロンプトが簡単な例です。本章を通して、このような生成コンテンツの具体例を、実際に使用したプロンプトとともに示していきます。例えば、下のボックスでは、最初の一文は私のプロンプトであり、その下にある箇条書きがAIによる生成物です。

AIの進歩によって不平等が拡大しうるような経済チャネルについて、ブレインストーミングをしてください。

  1. 低スキルの仕事の自動化が、雇用の喪失と失業の増加につながる。
  2. 高度なスキルを持った労働者の需要が高まり、賃金格差が生じる。
  3. AI テクノロジーに投資するリソースを持つ人々のところに資本へのアクセスが偏る。
  4. 労働者のAIによる監視が強化され、プライバシーの低下と労働搾取のリスクが生じる。
  5. AI アルゴリズムを用いた雇用、融資、その他の経済活動に関する意思決定が増加し、偏見や差別につながる。
  6. AIアルゴリズムを用いた価格設定の最適化が増加し、消費者にとってコストが嵩張る。
  7. AI アルゴリズムを用いたマーケティングの最適化が増加し、企業のコストが増える。
  8. AI アルゴリズムを用いた投資の最適化が増加し、投資するリソースがある人の収益のみ高くなる。
  9. AI アルゴリズムを用いた租税回避の最適化が増加し、政府の税収が減少する。
  10. AI アルゴリズムを用いた人件費の最適化が増加し、労働者の賃金が低下する。

モデルによって生成されたリストは画期的なものではありませんが、テーマから逸脱されておらず、おおむね合理的です。私自身が思いつく項目をカバーし、私が思い付いていないものも提案しています。5つ目の項目以降は、創作性の低下と反復性が見られ、これは人間によるブレストでも見られる傾向です。

3.1.2 Evaluating ideas

They can also evaluate different ideas, in particular by providing pros and cons of different research plan. As the following example illustrates, the model I tested was programmed to refrain from providing strong opinions, but the example - with a prompt that is admittedly a bit silly - showcases how LLMs can be a useful input in deliberations about research directions.

I am writing a paper on the e ects of AI on inequality. What do you think is more useful? A paper on how AI increases inequality or a paper on how AI reduces inequality?

That depends on the purpose of your paper. If you are looking to explore the potential implications of AI on inequality, then a paper on how AI increases inequality would be more useful. This would allow you to explore the potential negative impacts of AI on inequality, and how it could potentially exacerbate existing disparities. On the other hand, if you are looking to explore potential solutions to inequality, then a paper on how AI reduces inequality would be more useful. This would allow you to explore the potential positive impacts of AI on inequality, and how it could potentially reduce existing disparities. Ultimately, the choice of which paper to write depends on the purpose of your paper.

As the model’s response suggests, looking at how AI may increase inequality is more useful for positive work, whereas how it may decrease inequality is more useful for normative work.

3.1.2 アイデアの評価

LLMは複数のアイデアの評価にも使用することができます。特に、異なる研究計画のメリットとデメリットを羅列するのに使えそうです。以下の例で示すように、私がテストしたモデルは、強い意見を述べないようにチューニングされています。それでも、この例では、少し可笑しいプロンプトを使った場合でも、LLMが研究の方向性に関する検討において役に立ててくれることを示唆しています。

私は、不平等へのAIの影響に関する論文を書いています。AIが不平等を助長してしまうという論文とAIが不平等性を減らす論文、どちらの方がいいと思いますか?

それは論文の目的によります。AIが不平等に及ぼす潜在的な影響を探索したい場合は、AIがどのように不平等を増大させるかについて論文を書いた方がいいでしょう。これにより、AIが不平等に与える悪影響や、AIが既存の格差や差別をどのように悪化させるのかを調査できるようになります。一方で、不平等に対する解決策を探索したい場合は、AI がどのように不平等を減らせるのかに関する論文を書いた方が良いでしょう。これにより、AI が不平等にもたらす改善や、AI が既存の格差や差別をどのように減らせるのかを調査できるようになります。したがって、どちらの論文を書くかは、最終的に論文の目的によって決まります。

モデルから得られた回答が示唆しているように、AI がどのように不平等を助長するかのテーマは、前向きな研究にとって有用ですが、AI がどのように不平等を改善できるかのテーマは、規範的な研究にとって有用です。

3.1.3 Providing counterarguments

No matter what point we are arguing, there are always counterarguments. LLMs do not care about which side of an argument they are on - they are just as good at providing arguments in favor of a given point as they are for the counterarguments. This may be helpful to counteract the confirmation bias common to our human brains.

The following is an example:

My hypothesis is that AI will increase inequality. What are the main counterarguments?

  1. AI can be used to increase access to education and healthcare, which can help reduce inequality.
  2. AI can be used to automate mundane tasks, freeing up people to pursue more meaningful work.
  3. AI can be used to increase efficiency and productivity, leading to higher wages and better job opportunities.
  4. AI can be used to create new jobs and industries, providing more opportunities for people to earn a living.
  5. AI can be used to improve decision-making processes, leading to fairer outcomes.
  6. AI can be used to identify and address systemic biases, leading to more equitable outcomes.
  7. AI can be used to reduce the cost of goods and services, making them more accessible to people with lower incomes.
  8. AI can be used to improve the accuracy of predictions, leading to more equitable outcomes.

Some of the counter-arguments are better and some are worse, but the output covers the main points that I am aware of - I consider points 3., 4., 6., and 7. among the most important ones.

3.1.3 反論の提供

私たちがどのような論点の周りで議論しようとしても、常に反論がやってきます。LLMは自身が議論のどちら側にいるかに拘らることなく、与えられた論点に従って反論を提供することができます。この特徴は、人間の脳に存在しがちな確証バイアスに対抗するのに有用であるかもしれません.

次が一例です。:

私は、AIが不平等を増加させると仮定しています。反論を述べてください。

  1. AI を、教育や医療へのアクセスの機会を増やすことに使用することで、不平等の削減に寄与します。
  2. AI を、日常的なタスクの自動化に使用することで、人々がより有意義な仕事を追求できるようになります。
  3. AI を、効率と生産性の向上に使用することで、賃金の上昇と雇用機会の向上につながります。
  4. AI を、新しい仕事や産業の創出に使用することで、人々が生計を立てる機会を増やせます。
  5. AI を、意思決定プロセスの改善に使用することで、より公平な結果をもたらすことができます。
  6. AI を、体系的なバイアスを特定して対処するのに使用することで、より公平な結果をもたらすことができます。
  7. AI を、商品やサービスのコスト削減に使用することで、低所得者がそれらを利用しやすくなります。
  8. AI を、予測精度の向上に使用することで、より公平な結果をもたらすことができます。

モデルが挙げた反論の中には比較的良い内容のものもあれば、あまり良くないものもありますが、私が認識している主要な点が網羅されています。私は、3、4、6、7が最も重要な点であると考えています。

参照



    この記事をシェア👉🏻



カテゴリー
タグ