Sakana AI, the Tokyo-based startup co-founded by the authors of the seminal 'Attention Is All You Need' paper, has unveiled its first commercial product: Sakana Marlin. Dubbed a 'Virtual CSO' (Chief Strategy Officer), Marlin is an autonomous B2B research agent designed for corporations, financial institutions, and think tanks. Unlike conventional chatbots that generate answers in milliseconds, Marlin runs continuous, self-governing reasoning loops for up to eight hours, producing deeply researched, well-cited reports exceeding 100 pages, complete with executive summary slides and appendices.
A New Era for Enterprise Research
Rather than engaging in tedious prompt engineering, the user provides a core research topic and, after a brief scoping exchange, steps away entirely. Marlin formulates its own hypotheses, navigates the web to gather data, cross-references sources, and maps causal dynamics within complex business environments. Think of it less like a search engine and more like a junior strategy consultant locked in a room with a whiteboard and an internet connection: you submit the strategic prompt in the morning, and by end of day you receive a professional-grade portfolio. Real-world use cases include generating resolution scenarios for a hypothetical blockade of the Strait of Hormuz, mapping the fragmented global AI regulation landscape, and analyzing macroeconomic trends such as the return of 'bond vigilantes'. As highlighted in a recent article on iRobot and data privacy, protecting sensitive information is paramount: Marlin enforces a strict enterprise data policy that never uses customer inputs for model training without explicit consent.
Sponsored Protocol
The Engine: Adaptive Branching Monte Carlo Tree Search
Under the hood, Marlin is powered by AB-MCTS (Adaptive Branching Monte Carlo Tree Search), a technology derived from the 'The AI Scientist' project featured in Nature. Analogous to a chess engine that plays out thousands of future moves, AB-MCTS treats the research process as a branching tree. At each node, the algorithm dynamically balances two behaviors: exploration (spawning new hypotheses when the current path yields diminishing returns) and exploitation (refining a promising solution). The key breakthrough is Multi-LLM extension: the architecture dynamically selects different foundation models for specific sub-tasks, creating a plug-and-play collective intelligence network. For instance, an orchestration model delegates initial ideation to one LLM, while a reasoning-heavy model audits and corrects intermediate errors. This approach, made possible by industry dynamics and debates in the AI landscape, ensures final reports are the product of systematic, automated trial-and-error.
Sponsored Protocol
Licensing, Data, and Enterprise Implications
Sakana Marlin is strictly an enterprise SaaS product. Pricing tiers start with pay-as-you-go (100 credits per run, add-on credits at ¥98) up to the Team Plan (¥400,000 per month for 6,000 credits). Data handling is robust: neither Sakana nor external providers use customer data for training without explicit opt-in, and even then data is anonymized. This is critical for firms handling M&A research, unreleased product strategies, or proprietary market analyses.
Sponsored Protocol
Backed by a $2.6 billion Series B round and investors including Nvidia, Google, Khosla Ventures, and MUFG, Sakana AI is redefining depth in AI. As detailed on the company's Wikipedia page, its biomimetic 'school of fish' philosophy inspires a collaborative network of specialized models. In an era where speed has become a commodity, Marlin proves that true value lies in long-horizon thinking.
