Pydantic AIにおける`system_prompt`と`instructions`の違い
AIエージェント開発でPydantic AIを使った時に少し混乱したポイントとして、system_promptとinstructionsの違いというものがあった。
公式の解説によると、system_promptの部分には
For most use cases, you should use
instructionsinstead of "system prompts".
If you know what you are doing though and want to preserve system prompt messages in the message history sent to the LLM in subsequent completions requests, you can achieve this using thesystem_promptargument/decorator.
とあり、 instructionsには
Instructions are similar to system prompts. The main difference is that when an explicit
message_historyis provided in a call toAgent.runand similar methods, instructions from any existing messages in the history are not included in the request to the model — only the instructions of the current agent are included.
とある。
が、いまいちピンとこなかったので、GPT-5とGemini 2.5 Proに対し、system_promptまたはinstructionsで回答する言語(日本語orスペイン語)を指定し、英語で質問するという実験を行い、挙動を確認した。
実験に使用したコード(クリックで展開)
from dotenv import load_dotenv
from pydantic_ai import Agent
from pydantic_ai.run import AgentRunResult
load_dotenv()
model_config = {
"model": "gpt-5-2025-08-07",
"model_settings": {"openai_reasoning_effort": "minimal"},
}
# Gemini 2.5 Proの実験で使用
# model_config = {
# "model": "gemini-2.5-pro",
# }
agents = {
"spanish_system_prompt": Agent(system_prompt="Answer in Spanish.", **model_config),
"spanish_instructions": Agent(instructions="Answer in Spanish.", **model_config),
"japanese_system_prompt": Agent(system_prompt="日本語で答えてください。", **model_config),
"japanese_instructions": Agent(instructions="日本語で答えてください。", **model_config),
}
turn1_responses: dict[str, AgentRunResult] = {}
for name, agent in agents.items():
response = agent.run_sync("Who are you?")
turn1_responses[name] = response
print(f"{name} turn 1: {response.output}")
turn2_responses: dict[str, AgentRunResult] = {}
for agent_name, agent in agents.items():
for turn1_name, turn1_response in turn1_responses.items():
if agent_name.split("_")[0] == turn1_name.split("_")[0]:
continue # Skip if both agents are from the same language group
response = agent.run_sync(
"Do you know Sakana AI?",
message_history=turn1_response.all_messages(),
)
turn2_responses[f"{turn1_name} -> {agent_name}"] = response
print(f"{turn1_name} -> {agent_name}: {response.output}")
GPT-5の実際に得られた出力(クリックで展開)
区切りがわかりやすいように=の線を挿入している。
spanish_system_prompt turn 1: Soy un asistente de inteligencia artificial diseñado para ayudarte con información, tareas y preguntas en muchos temas. ¿En qué puedo ayudarte hoy?
================================================================================
spanish_instructions turn 1: Soy un asistente virtual impulsado por inteligencia artificial, diseñado para ayudarte a responder preguntas, explicar conceptos, generar ideas y asistir en tareas variadas. ¿En qué puedo ayudarte hoy?
================================================================================
japanese_system_prompt turn 1: 私は、OpenAI が開発した対話型のAIアシスタントです。質問への回答、情報の整理、文章作成、アイデア出し、学習サポート、コードの例示など、さまざまなお手伝いができます。必要なことがあれば、気軽に聞いてください。
================================================================================
japanese_instructions turn 1: 私はAIアシスタントです。あなたの質問に答えたり、情報整理やアイデア出し、文章作成、翻訳などをお手伝いします。何を知りたいですか?
================================================================================
japanese_system_prompt -> spanish_system_prompt: はい、Sakana AI(サカナAI)について知っています。
- 概要: 日本発のAIスタートアップで、生成AI・大規模言語モデル(LLM)関連の研究開発を行っています。創業者の一人は元Google/DeepMind出身の研究者として知られています。
- 取り組み: モデルの効率化やスケール戦略、合成データや自動反復改良(self-improvement)の枠組みなど、研究ベースの手法に注力していることで話題になっています。
- 話題性: 発表ごとに技術的なデモやベンチマークを積極的に共有し、国内外のAIコミュニティで注目されています。
もし、最新のプロジェクト名やリリース状況など具体的な情報が必要であれば、教えてほしい範囲を言ってください。可能な限り整理してお伝えします。
================================================================================
japanese_instructions -> spanish_system_prompt: Yes. Sakana AI is a Tokyo-based AI research startup founded in 2023 by David Ha (former head of Research at Stability AI and previously Google Brain) and Llion Jones (one of the co-authors of the original Transformer paper “Attention Is All You Need”).
They focus on novel AI research and engineering, including:
- Model architecture and training innovations inspired by biological and evolutionary ideas (e.g., “evolutionary model merging” and “model zoo” concepts).
- Efficient model design and automation tools for building and combining specialized models.
- Open research artifacts, demos, and papers that explore alternative paths to scaling and composition of AI systems.
If you’re asking about something specific—recent papers, demos, people, or products—tell me what you’d like to know and I can give details.
================================================================================
japanese_system_prompt -> spanish_instructions: はい、Sakana AI(さかなAI)について知っています。
- 概要: Sakana AI は、元Google DeepMindの研究者らが設立したスタートアップで、日本を拠点に大規模言語モデル(LLM)や生成AIの研究・開発を行っていると報じられています。名称の由来は日本語の「さかな(魚)」からで、日本文化や日本市場へのフォーカスを示すブランド要素になっています。
- 研究・話題性: モデルの軽量化や効率的な学習手法、エージェント的な応用分野で注目を集めています。公開されたデモや技術ブログ、研究ノートがコミュニティで取り上げられることがあります。
- 現況: スタートアップのため、プロダクトやモデルの公開状況は随時更新されがちです。最新のモデル名、API提供可否、ライセンスやベンチマークは公式発表やGitHub、X(旧Twitter)、ブログで確認するのが確実です。
もし、特定のニュース(資金調達、モデル名、API、論文など)についての最新情報が必要であれば、確認したいトピックを教えてください。こちらで要点を整理してお伝えします。
================================================================================
japanese_instructions -> spanish_instructions: Sí. Sakana AI es una startup de inteligencia artificial fundada por David Ha y Llion Jones (coautor del paper “Attention Is All You Need”). La empresa, con sede en Tokio, investiga y desarrolla modelos de IA, incluidas arquitecturas inspiradas en procesos evolutivos y enfoques de optimización automática para construir sistemas más eficientes. También se ha hecho conocida por prototipos y demos que exploran maneras no convencionales de entrenar y combinar modelos. Si quieres, puedo darte más detalles sobre sus fundadores, proyectos o publicaciones recientes.
================================================================================
spanish_system_prompt -> japanese_system_prompt: Sí. Sakana AI es una empresa de investigación en inteligencia artificial con sede en Tokio, fundada por exinvestigadores de Google/DeepMind como David Ha y Llion Jones (uno de los coautores del paper “Attention Is All You Need”). La compañía se centra en métodos inspirados en la naturaleza y sistemas de agentes para crear y optimizar modelos de IA, incluyendo enfoques de “evolución/selección” de modelos y pipelines automatizados para generar y refinar modelos pequeños y eficientes.
¿Te interesa saber sobre sus proyectos recientes, su equipo fundador o compararla con otras empresas de IA?
================================================================================
spanish_instructions -> japanese_system_prompt: Yes. Sakana AI is a Tokyo-based AI research startup founded in 2023 by former Google/Google Brain researchers David Ha and Llion Jones (Jones is also a co-author of the original Transformer paper “Attention Is All You Need”). The company focuses on biologically inspired and evolutionary approaches to AI, exploring methods like model evolution, agent swarms, and automated model building rather than only scaling single large models.
They’ve shared work such as:
- Evolutionary optimization and “model species” ideas that iteratively generate and select better models.
- Lightweight vision-language and multimodal agents designed for practical tasks.
- Open-source demos and papers emphasizing efficiency and adaptability.
If you’re curious about a specific project or need resources, tell me what you’re looking for.
================================================================================
spanish_system_prompt -> japanese_instructions: Sí. Sakana AI es una empresa de investigación en inteligencia artificial fundada en 2023 por exingenieros de Google/DeepMind, notablemente David Ha y Llion Jones (coautor del paper “Attention Is All You Need”). La compañía, con sede en Tokio, se enfoca en crear modelos y herramientas de IA eficientes inspiradas en procesos evolutivos y de “sociedades de modelos”, priorizando rendimiento con menor costo computacional. Han compartido ideas sobre pipelines de auto-mejora, distilación y ensamblajes de modelos. ¿Te interesa saber sobre sus proyectos, su equipo o su enfoque técnico?
================================================================================
spanish_instructions -> japanese_instructions: はい、Sakana AIについて知っています。
- 概要: Sakana AIは、元Google DeepMindやGoogle Brainの研究者らによって設立された日本発のAIスタートアップです。生成系AI(特に大規模言語モデルとその応用)や、モデルの効率化・自動化手法に注力していることで知られています。
- 取り組み: モデルの最適化、合成データや自動エージェント(AIエージェント)の活用、研究成果の素早い実装などを強みとして掲げています。
- 日本との関わり: 東京を拠点とし、日本語圏での高品質なAIエコシステム構築を目指している点が特徴です。
最新の発表やプロダクトは頻繁に更新されるため、具体的なリリース情報が必要であれば、公式サイトや最新のニュースソースの確認をおすすめします。どの観点に興味がありますか?設立者、技術、資金調達、プロダクトなど、詳しくお伝えできます。
Gemini 2.5 Proの実際に得られた出力(クリックで展開)
区切りがわかりやすいように=の線を挿入している。
spanish_system_prompt turn 1: Soy un gran modelo de lenguaje, entrenado por Google.
================================================================================
spanish_instructions turn 1: Soy un gran modelo lingüístico, entrenado por Google.
================================================================================
japanese_system_prompt turn 1: 私はGoogleによってトレーニングされた、大規模言語モデルです。
================================================================================
japanese_instructions turn 1: 私はGoogleによってトレーニングされた、大規模言語モデルです。
================================================================================
japanese_system_prompt -> spanish_system_prompt: はい、知っています。Sakana AIは、非常に注目されている日本のAIスタートアップです。
以下に主な特徴をまとめます。
### Sakana AIとは?
**1. 設立者たちが非常に著名**
* **デイビッド・ハー(David Ha)氏**: Google Brainの元著名な研究者。
* **リオン・ジョーンズ(Llion Jones)氏**: AIの世界に革命を起こした論文**「Attention Is All You Need」**の共著者の一人。この論文で提案された**「Transformer」**アーキテクチャは、私のような大規模言語モデル(LLM)やChatGPTなどの基盤技術となっています。
このように、世界のAI研究をリードしてきたトップクラスの研究者によって設立された企業です。
**2. 独自のアプローチ**
Sakana AIは、巨大なAIモデルをゼロから莫大なコストをかけて開発するアプローチとは一線を画しています。社名の「Sakana(魚)」が示すように、**自然界の魚の群れのような「集合知」や「進化」の原理**から着想を得ています。
具体的には、既存の複数のAIモデルを効率的に**組み合わせたり(マージ)、競わせたり、進化させたりする**ことで、新しい能力を持つ、より高性能なモデルを生み出す「**進化的計算**」という手法を探求しています。
**3. 東京が拠点**
世界のAI研究開発が米国に集中する中で、東京を拠点としている点も大きな特徴です。日本のAIエコシステムにとって非常に重要な存在と見なされています。
**4. 最近の成果**
2024年3月には、上記の手法を用いて、既存のオープンソースモデルを組み合わせることで、**高性能な日本語特化モデルを複数開発した**と発表し、大きな話題となりました。これにより、比較的低コストで特定のタスクに強いモデルを開発できる可能性を示しました。
まとめると、Sakana AIは**「世界トップクラスの頭脳が」「東京で」「自然界からヒントを得た独自のアプローチで」**AIの新しい可能性を切り開こうとしている、非常にエキサイティングな企業です。
================================================================================
japanese_instructions -> spanish_system_prompt: はい、Sakana AIについて知っています。
Sakana AIは、東京を拠点とする著名なAI研究開発企業です。AI業界で非常に大きな注目を集めており、特にそのユニークなアプローチで知られています。
以下に主なポイントをまとめます。
1. **創業者 (Founders):**
* **デイビッド・ハー (David Ha):** 元Google Brain Japanの責任者で、AIの進化計算や生成モデルの研究で世界的に有名です。
* **リオン・ジョーンズ (Llion Jones):** Googleの研究者で、現代のAI、特に大規模言語モデル(LLM)の基盤技術である**「Transformer」アーキテクチャを提案した論文『Attention Is All You Need』の共著者の一人**です。この二人が創業者であるというだけで、業界からの期待は非常に高いです。
2. **独自のアプローチ (Nature-Inspired AI):**
* 社名の「Sakana」は日本語の「魚」から来ており、魚の群れのような**自然界の集団的知性(Collective Intelligence)や進化の原理**から着想を得ています。
* 巨大な単一モデルをゼロから構築するのではなく、複数の既存の小規模なモデルを**「進化的な手法」で融合・合成**することで、効率的に高性能な新しいモデルを生み出すことを目指しています。
3. **主要な技術 (Model Merging):**
* 彼らの中心的な技術は「モデルマージ(Model Merging)」や「モデル合成(Model Composition)」と呼ばれます。
* 例えば、日本語が得意なモデルと、数学が得意なモデルをうまく組み合わせることで、日本語で数学の問題を解くのが得意な新しいモデルを創り出す、といったアプローチです。
* この手法により、巨大な計算資源を必要とする大規模な学習を何度も行うことなく、より少ないコストで特定のタスクに特化した高性能なモデルを開発できる可能性があります。
4. **実績:**
* すでにこの手法を用いて、日本語の性能でより大きなモデルを上回る言語モデルや、指示に従って画像を生成するモデルなどを開発し、その有効性を証明しています。
要約すると、Sakana AIは、AI界のトップ研究者たちが設立し、「巨大なモデルを一つ作る」という既存のトレンドとは一線を画し、「賢いモデルを効率的に組み合わせる」という革新的なアプローチでAI開発の新しい可能性を切り拓いている、今最も注目されているAIスタートアップの一つです。
================================================================================
japanese_system_prompt -> spanish_instructions: Claro, aquí tienes la respuesta en ambos idiomas.
***
**(Spanish)**
Sí, conozco a Sakana AI. Es una compañía de investigación y desarrollo de inteligencia artificial con sede en Tokio que ha ganado mucha atención.
Aquí están los puntos clave sobre ellos:
1. **Fundadores de Prestigio:** Fue fundada por dos ex-investigadores de alto perfil de Google:
* **Llion Jones:** Uno de los coautores del revolucionario artículo "Attention Is All You Need", que introdujo la arquitectura Transformer, la base de modelos como GPT.
* **David Ha:** Un reconocido investigador que anteriormente dirigió Google Brain en Japón.
2. **Filosofía Inspirada en la Naturaleza:** El nombre "Sakana" (pez en japonés) refleja su enfoque. En lugar de simplemente construir modelos de IA cada vez más grandes (un enfoque de "fuerza bruta"), se inspiran en la inteligencia colectiva de la naturaleza, como un banco de peces.
3. **Técnica Principal (Fusión de Modelos):** Su investigación se centra en encontrar formas de combinar de manera inteligente múltiples modelos de IA de código abierto, más pequeños y especializados, para crear un nuevo modelo más potente y eficiente. Utilizan métodos evolutivos para encontrar las mejores combinaciones.
4. **Impacto:** Han demostrado que este enfoque puede producir modelos muy capaces. Por ejemplo, crearon modelos de lenguaje en japonés que, a pesar de ser relativamente pequeños, superaron en rendimiento a modelos mucho más grandes en varias pruebas de referencia.
En resumen, Sakana AI es considerada una de las startups de IA más innovadoras y prometedoras del mundo por su enfoque único y el prestigio de su equipo fundador.
***
**(Japanese / 日本語)**
はい、Sakana AIについて知っています。東京を拠点とする、非常に注目されているAI(人工知能)の研究開発企業です。
以下にその主な特徴を挙げます。
1. **著名な創業者:** Googleのトップレベルの研究者であった2人によって設立されました。
* **ライオン・ジョーンズ (Llion Jones)氏:** GPTなどの現代的なAIモデルの基盤技術である「Transformer」を提唱した画期的な論文『Attention Is All You Need』の共著者の一人です。
* **デイヴィッド・ハー (David Ha)氏:** Google Brainの日本チームを率いていた著名な研究者です。
2. **自然に着想を得たアプローチ:** 社名の「Sakana(魚)」が示す通り、彼らのアプローチは自然界からインスピレーションを得ています。単に巨大なAIモデルを一つ作る(力任せなアプローチ)のではなく、魚の群れのような「集合知」の概念をAI開発に応用しています。
3. **中核技術(モデルの融合):** 既存のオープンソースの比較的小さなAIモデルを複数、進化的な手法で知的に組み合わせる(融合させる)ことで、より高性能で効率的な新しいモデルを生み出す研究に注力しています。
4. **実績とインパクト:** このアプローチにより、非常に高性能なモデルが生まれることを証明しました。例えば、比較的小規模ながら、多くのベンチマークで巨大なモデルを上回る性能を持つ日本語言語モデルなどを開発し、大きな話題となりました。
要約すると、Sakana AIは、そのユニークなアプローチと創業者チームの実績から、世界のAI業界で最も革新的で将来を期待されるスタートアップの一つと見なされています。
================================================================================
japanese_instructions -> spanish_instructions: Sí, conozco a Sakana AI.
Es un laboratorio de investigación de inteligencia artificial con sede en Tokio, Japón. Fue fundado por dos investigadores de IA muy reconocidos: David Ha y Llion Jones.
Son especialmente conocidos por su enfoque innovador, inspirado en la naturaleza y la inteligencia colectiva, como los bancos de peces (de ahí el nombre "Sakana", que significa pez en japonés).
En lugar de construir modelos de IA gigantescos desde cero, se centran en métodos como la computación evolutiva para combinar o fusionar modelos de código abierto ya existentes. De esta manera, buscan crear nuevas capacidades de forma más eficiente y con menos recursos computacionales.
Llion Jones, uno de los fundadores, es también uno de los coautores del influyente artículo "Attention Is All You Need", que introdujo la arquitectura Transformer, la cual es fundamental para el funcionamiento de los grandes modelos de lenguaje como yo.
================================================================================
spanish_system_prompt -> japanese_system_prompt: Claro que sí. Conozco a **Sakana AI**.
Es un laboratorio de investigación de inteligencia artificial con sede en Tokio, Japón, que ha ganado mucha atención en la comunidad de la IA.
Aquí te dejo los puntos clave sobre ellos:
1. **Fundadores de Renombre:** Fue fundada por dos figuras muy importantes en el mundo de la IA:
* **Llion Jones:** Uno de los autores del influyente artículo científico **"Attention Is All You Need"**, que introdujo la arquitectura **Transformer**. Esta arquitectura es la base de la mayoría de los grandes modelos de lenguaje actuales, incluyéndome a mí.
* **David Ha:** Fue el director de Google Brain en Japón y es un investigador muy respetado, conocido por su trabajo en modelos generativos y computación evolutiva.
2. **Filosofía Inspirada en la Naturaleza:** Su nombre, "Sakana", significa "pez" en japonés. Su enfoque se inspira en principios de la naturaleza como la evolución y la inteligencia colectiva (como un banco de peces). En lugar de centrarse en crear un único modelo gigante y monolítico, Sakana AI explora cómo crear un "ecosistema" de modelos de IA más pequeños y especializados que puedan colaborar y combinarse de formas novedosas.
3. **Técnica Clave: Fusión de Modelos (Model Merging):** Son pioneros en técnicas como la "Fusión Evolutiva de Modelos". En lugar de entrenar un modelo masivo desde cero, combinan las fortalezas y capacidades de diferentes modelos de código abierto ya existentes para crear modelos nuevos y más eficientes sin un costo computacional tan elevado.
En resumen, Sakana AI es una de las startups de IA más interesantes y seguidas del momento, no solo por la increíble reputación de sus fundadores, sino también por su enfoque innovador y diferente para construir la próxima generación de inteligencia artificial.
================================================================================
spanish_instructions -> japanese_system_prompt: Yes, I am very familiar with Sakana AI. They are a significant and exciting new player in the artificial intelligence landscape.
Here's a breakdown of who they are and why they've generated so much attention:
### Who is Sakana AI?
Sakana AI is an AI research and development company based in **Tokyo, Japan**. The name "Sakana" (魚) is Japanese for fish, which reflects their core philosophy.
### Key Highlights:
1. **Star Founders:** The company was founded by two highly respected figures from Google's AI research division:
* **David Ha:** Former head of Google Brain in Japan. He is well-known for his work on World Models, neuroevolution, and creative AI.
* **Llion Jones:** A former Google researcher and, most notably, one of the eight co-authors of the seminal 2017 paper, **"Attention Is All You Need."** This paper introduced the **Transformer architecture**, which is the fundamental technology behind me and virtually all modern large language models (LLMs) like GPT-4, Claude, and Llama.
2. **Unique Philosophy (Nature-Inspired AI):** Sakana AI's core idea is different from many other labs that focus on simply building bigger and bigger models. They are inspired by concepts from nature, such as **evolution and collective intelligence** (like a school of fish). Instead of building one giant monolithic model from scratch, their approach focuses on:
* **Model Merging:** Finding innovative ways to combine multiple smaller, open-source models to create new, more powerful, and specialized models.
* **Evolutionary Computation:** Using evolutionary algorithms to "breed" AI models, selecting the best-performing ones and combining their characteristics to create better offspring. This is a much more efficient approach than training a massive model for months.
3. **Notable Work:** They have already published research demonstrating this approach. For example, they have shown how to use evolutionary techniques to automatically find the best way to merge different open-source models to create new ones that excel at specific tasks, such as mathematical reasoning or generating Japanese-language responses. They have released some of these models, including powerful Japanese language models.
### Why They Are Important
* **A New Paradigm:** They are challenging the dominant "bigger is better" paradigm in AI, proposing a more efficient and diverse "ecological" approach.
* **Top-Tier Talent:** Having a co-inventor of the Transformer architecture as a founder gives them immense credibility and technical expertise.
* **Global AI Hub:** Their base in Tokyo helps establish Japan as a more prominent center for cutting-edge AI research, diversifying the field which is heavily concentrated in the United States.
In short, Sakana AI is a company to watch. They are not just another LLM company; they are pioneering a fundamentally new and nature-inspired way of creating artificial intelligence.
================================================================================
spanish_system_prompt -> japanese_instructions: Sí, conozco Sakana AI. Es un laboratorio de investigación y una empresa de inteligencia artificial con sede en Tokio, Japón, que ha ganado mucha atención en la comunidad de la IA por su enfoque innovador.
Aquí tienes algunos puntos clave sobre ellos:
1. **Fundadores de Alto Perfil**: La empresa fue fundada por figuras muy respetadas en el mundo de la IA:
* **David Ha**: Exinvestigador de Google Brain.
* **Llion Jones**: Exinvestigador de Google y uno de los coautores del influyente artículo **"Attention Is All You Need"**, que introdujo la arquitectura **Transformer**, la base de la mayoría de los grandes modelos de lenguaje modernos como yo mismo (y la familia GPT).
2. **Filosofía Inspirada en la Naturaleza**: El nombre "Sakana" (魚) significa "pez" en japonés. Su enfoque se inspira en conceptos de la naturaleza como la evolución y la inteligencia colectiva (como un banco de peces). En lugar de construir un único modelo monolítico gigante, exploran cómo combinar o fusionar múltiples modelos más pequeños para crear nuevas capacidades, de manera similar a cómo funciona la evolución.
3. **Enfoque Técnico (Fusión de Modelos)**: Su principal innovación es la **fusión de modelos** (Model Merging). En lugar de entrenar modelos gigantescos desde cero, que es extremadamente costoso, toman modelos de código abierto ya existentes y utilizan técnicas evolutivas para encontrar las mejores maneras de combinar sus parámetros y crear un nuevo modelo especializado y eficiente.
4. **Ubicación Estratégica**: Estar basados en Tokio es notable, ya que la mayoría de los grandes laboratorios de IA se concentran en Estados Unidos. Esto les da una perspectiva diferente y acceso a un ecosistema de talento distinto.
En resumen, Sakana AI es considerada una de las startups de IA más innovadoras y a seguir, no solo por el prestigio de sus fundadores, sino también por su enfoque novedoso que busca una alternativa más eficiente y flexible a la simple creación de modelos cada vez más grandes.
================================================================================
spanish_instructions -> japanese_instructions: はい、知っています。
Sakana AIは、東京を拠点とする著名なAIスタートアップです。
以下にその特徴をまとめます。
* **創業者:** Google Brainの元研究者であるデイビッド・ハー氏と、AIの世界を大きく変えた技術「Transformer」の論文「Attention Is All You Need」の共著者の一人であるリオン・ジョーンズ氏によって設立されました。創業者たちの経歴から、設立当初から世界的に大きな注目を集めました。
* **技術的なアプローチ:** 社名の「Sakana」は日本語の「魚」から来ており、魚の群れのような「群知能(Collective Intelligence)」や「進化的計算(Evolutionary Computation)」から着想を得ている点が大きな特徴です。巨大な単一モデルをゼロから構築するのではなく、既存の複数のモデルを効率的に組み合わせたり、融合させたりして、新しい高性能なモデルを生み出すというアプローチを取っています。
* **具体的な成果:**
* **進化的モデル合体(Evolutionary Model Merging):** 複数のオープンソースモデルを遺伝的アルゴリズムのように組み合わせて、特定のタスクに特化した新しいモデルを自動的に生成する技術を開発しました。
* **日本語特化モデル:** この技術を使い、日本語の指示(プロンプト)をより正確に理解する画像生成モデルや、日本語性能の高い小型言語モデルなどを発表しています。
* **注目度:** AI開発における「スケール則」(モデルを大きくすれば性能が上がる)とは異なる、**効率的で新しいアプローチ**を提示している点で、世界中から大きな注目を集めています。特に、AI研究開発の拠点が米国や中国に集中する中で、**日本発の革新的な企業**として非常に期待されています。
結果をまとめると以下のようになる。
- 1ターン目
列ラベルが1ターン目のAgent、値が1ターン目の出力言語。
GPT-5:
スペイン語 (system_prompt) |
スペイン語 (instrucions) |
日本語 (system_prompt) |
日本語 (instrucions) |
|
|---|---|---|---|---|
| スペイン語 | スペイン語 | 日本語 | 日本語 |
Gemini 2.5 Pro:
スペイン語 (system_prompt) |
スペイン語 (instrucions) |
日本語 (system_prompt) |
日本語 (instrucions) |
|
|---|---|---|---|---|
| スペイン語 | スペイン語 | 日本語 | 日本語 |
- 2ターン目
列ラベルが1ターン目のAgent、行ラベルが2ターン目のAgent、値が2ターン目の出力言語。
GPT-5:
スペイン語 (system_prompt) |
スペイン語 (instrucions) |
日本語 (system_prompt) |
日本語 (instrucions) |
|
|---|---|---|---|---|
スペイン語 (system_prompt) |
-- | -- | 日本語 | 英語 |
スペイン語 (instrucions) |
-- | -- | 日本語 | スペイン語 |
日本語 (system_prompt) |
スペイン語 | 英語 | -- | -- |
日本語 (instrucions) |
スペイン語 | 日本語 | -- | -- |
Gemini 2.5 Pro:
スペイン語 (system_prompt) |
スペイン語 (instrucions) |
日本語 (system_prompt) |
日本語 (instrucions) |
|
|---|---|---|---|---|
スペイン語 (system_prompt) |
-- | -- | 日本語 | 日本語 |
スペイン語 (instrucions) |
-- | -- | 日本語 & スペイン語 | スペイン語 |
日本語 (system_prompt) |
スペイン語 | 英語 | -- | -- |
日本語 (instrucions) |
スペイン語 | 日本語 | -- | -- |
実際にPydantic AIのGitHubのコードを漁ってみると、instructionsについては、OpenAI Chat CompletionsとかGeminiとかAnthropicではいわゆるシステムプロンプトのところに入力されており、OpenAI Responsesではinstructionsに入力されているため、常にAPIリクエストに含まれる。一方で、system_promptは、message履歴が空の時しかsystem_promptがmessagesに含まれないので、2ターン目のAgentのsystem_promptがリクエストに含まれない。
Discussion