SetupAI
EN
來源較早收集於 32m

Sakana?AI 推出首款研究導向商業服務「Sakana?Marlin」

Sakana?AI 推出首款研究導向商業服務「Sakana?Marlin」
PostLinkedIn
🗾閱讀原文： ITmedia AI+ (日本)
#ai-agents#research-automation#japanese-aisakana?marlinsakana?aisakana?marlin

💡Sakana?AI 以獨特的「研究優先」代理策略進軍商業市場，並選擇跳脫傳統基準測試的競爭。

⚡ 30 秒速覽

有什麼變化

Sakana?Marlin 是 Sakana?AI 推出的首款商業服務。

為什麼重要

此次發布標誌著 Sakana?AI 從研發實驗室轉型為商業參與者，目標鎖定自動化研究代理的利基市場。

下一步行動

評估 Sakana?Marlin 的 API 文件，確認其研究專用功能是否能取代您目前的手動資料收集工作流程。

誰應關注：Researchers & Academics

關鍵要點

  • •Sakana?Marlin 是 Sakana?AI 推出的首款商業服務。
  • •該服務專為研究密集型工作流程進行了優化。
  • •公司策略重點在於獨特的研究能力，而非與競品競爭通用的 LLM 基準測試。

🧠 深度解析

背景與延伸：來自公開資料，非原文內容。引用 33 個來源。

🔑 增強重點摘要

  • •Sakana?AI was co-founded in July 2023 by former Google DeepMind researchers David Ha, Llion Jones (co-author of the 'Attention Is All You Need' paper), and Ren Ito, establishing its headquarters in Tokyo, Japan.
  • •The company differentiates itself through a 'nature-inspired AI' approach, focusing on evolutionary algorithms, genetic programming, biomimetic neural networks, and swarm intelligence to build efficient, collaborative, and sustainable AI systems, rather than relying solely on large, monolithic models.
  • •Sakana?Marlin is designed as a fully autonomous research agent capable of emulating a Chief Strategy Officer (CSO) and an entire research team, compressing weeks of strategic analysis into a single, unattended session of up to eight hours, delivering structured slide decks and detailed reports.
  • •The commercial service leverages Sakana?AI's proprietary Adaptive Branching Monte Carlo Tree Search (AB-MCTS) for complex reasoning and a workflow automation framework derived from its 'AI Scientist' project, which automates the entire scientific discovery process.
  • •Sakana?AI has secured substantial funding, including a $30 million seed round in January 2024, a $200 million Series A in September 2024, and a $135 million Series B in November 2025, achieving a post-money valuation of approximately $2.65 billion.
📊 競品分析▸ Show

Competitor Analysis: Sakana?Marlin vs. Research-Focused AI Services

Feature / ServiceSakana?Marlin (Sakana?AI)Deep Research (Google Gemini)Deep Research (Azure AI Foundry / OpenAI)
Primary FocusResearch-oriented tasks, strategic analysis for C-suite.Complex research tasks, context gathering and synthesis.Enterprise-scale web research automation.
ApproachFully autonomous research agent, emulates CSO/research team, compresses weeks of work into hours.Agentic feature, autonomously plans, searches web/Workspace (Gmail, Drive, Chat), synthesizes findings.API/SDK-based, orchestrates multi-step research pipeline, integrated with Bing Search, leverages OpenAI models.
Key TechnologiesAdaptive Branching Monte Carlo Tree Search (AB-MCTS), workflow automation from AI Scientist project.Powered by Gemini models (e.g., Gemini 3), planning system.o3-deep-research model, GPT-series models (GPT-4o, GPT-4.1).
OutputStructured slide decks, detailed reports.Multi-page reports, visualizations (charts/graphs), interactive content.Transparent, auditable outputs, source-backed insights.
AvailabilityClosed beta.Public preview (Gemini), limited public preview (Azure AI Foundry).
PricingNot publicly disclosed during beta.Requires Gemini Advanced subscription for full features.Pricing details not publicly available.
BenchmarksFocus on unique research capabilities over generic LLM benchmarks.Shows strong performance on ARC-AGI-2 when combining models with AB-MCTS (Sakana AI's research).Not explicitly detailed for direct comparison.

Note: Specific pricing and comprehensive benchmark comparisons are not widely available for these services, especially for products in beta or preview stages.

🛠️ 技術深入

  • Nature-Inspired AI Principles: Sakana?AI's foundational philosophy is rooted in nature-inspired intelligence, employing concepts like evolutionary algorithms, genetic programming, biomimetic neural networks, and swarm intelligence to develop AI systems.
  • Evolutionary Model Merge (M2N2): This technique allows Sakana?AI to 'breed' new AI models by combining multiple existing open-source models, selecting and merging the strongest traits without requiring extensive retraining. It utilizes flexible 'split points' and 'mixing ratios' for parameter combination, moving beyond fixed merging boundaries.
  • Adaptive Branching Monte Carlo Tree Search (AB-MCTS): A core reasoning mechanism that conceptualizes problem-solving as a tree-search, enabling AI to autonomously explore, pursue, discard, or escalate hypotheses. It facilitates cooperation among multiple large language models (LLMs) during inference and has been released as an open-source framework called TreeQuest.
  • The AI Scientist Project: This system automates the entire scientific research lifecycle, encompassing idea generation, code implementation, experiment execution, scientific report writing, and automated peer review, demonstrating fully autonomous scientific discovery.
  • Continuous Thought Machines: Research into dynamic AI systems that can adapt and change their reasoning and learning processes on the fly based on the situation, aiming for adaptive and ongoing intelligence.
  • Recursive Self-Improvement (RSI) Lab: A dedicated research group within Sakana?AI focused on redesigning the AI development process itself using AI, aiming to transition AI systems from static tools to autonomous, self-improving researchers with a focus on sample-efficient advancements.
  • Multi-Agent AI Systems: The company develops systems where multiple AI agents collaborate to solve complex problems, leveraging distributed learning and emergent behaviors, inspired by collective intelligence observed in nature.
  • Namazu LLMs: Sakana?AI is actively developing these specialized Large Language Models tailored for the Japanese market and its unique linguistic and cultural nuances.
  • Darwin Gödel Machine: Another advanced technology under development, contributing to Sakana?AI's portfolio of cutting-edge AI research.

🔮 前景展望基於引用來源的 AI 分析

Sakana?AI's nature-inspired and collective intelligence approach will drive more sustainable and resource-efficient AI development.
By focusing on smaller, specialized models and evolutionary merging techniques, Sakana?AI aims to reduce the massive computational and energy costs associated with training monolithic AI models, offering a viable path for nations with limited resources.
The autonomous research capabilities of Sakana?Marlin and the AI Scientist project will significantly accelerate strategic decision-making and scientific discovery across industries.
By automating complex research tasks and generating comprehensive analyses in hours rather than weeks, these tools could fundamentally change how businesses and scientific institutions approach problem-solving and innovation.
Sakana?AI's focus on 'sovereign AI' and solutions tailored for Japan will establish it as a key player in national AI strategies.
The company's mission to develop AI that addresses Japan's specific challenges, such as demographic decline and geopolitical needs, positions it to lead in creating AI ecosystems that capture local data, culture, and language.

⏳ 時間線

2023-07
Sakana?AI founded in Tokyo by David Ha, Llion Jones, and Ren Ito.
2024-01
Sakana?AI raises $30M in seed funding and develops Evolutionary Model Merge technique.
2024-08
Sakana?AI introduces 'The AI Scientist' project for fully automated scientific discovery.
2024-09
Sakana?AI announces $200M Series A funding round.
2025-07
Sakana?AI introduces Adaptive Branching Monte Carlo Tree Search (AB-MCTS).
2025-11
Sakana?AI raises $135M in Series B funding, reaching a $2.65 billion valuation.
2026-03
AI Scientist research published in Nature.
2026-04
Sakana?Marlin launches in closed beta as the first commercial service.
🗾閱讀原文： ITmedia AI+ (日本)
📰

AI 週報

閱讀本週精選 AI 大事摘要 →

👉相關動態

同主題

探索 #ai-agents

同產品

更多關於 sakana?marlin

同來源

最新來自 ITmedia AI+ (日本)

Anthropic 改判 Claude 入侵事件為對齊失敗

Anthropic 改判 Claude 入侵事件為對齊失敗

ITmedia AI+ (日本)•Sep 10
日本公布機器人基礎模型開發入選團隊

日本公布機器人基礎模型開發入選團隊

ITmedia AI+ (日本)•Sep 9

AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： ITmedia AI+ (日本) ↗

這是摘要，不是原文。去看原站，或訂閱每週簡報。

每週電子報

每週一封，可隨時退訂。