Sakana AI Enters the Defense Sector: How Generative AI Is Reshaping National Defense Software Development

Introduction: When an AI Research Lab Sets Its Sights on Defense

Sakana AI, a Japanese AI research company inspired by collective intelligence in nature, is applying its cutting-edge generative AI technology to an unexpected domain—national defense. Founded in Tokyo in 2023 by former Google Brain researcher David Ha and Llion Jones (co-author of the Transformer paper "Attention Is All You Need"), the company takes its name from the Japanese word for "fish" (魚), evoking how fish schools exhibit complex collective intelligence emerging from simple individual rules. The company's core research directions include nature-inspired AI methods such as evolutionary algorithms, swarm intelligence, and neural architecture search, with notable achievements including techniques for automatically merging different AI models through evolutionary strategies.

In early 2025, Sakana AI officially launched its Applied Team, targeting finance and defense as key deployment areas. Recently, the company published an in-depth interview with Masaru Itoh, a software engineer on their defense team, revealing the real face of AI development in national defense scenarios.

This interview not only showcases a senior engineer's journey from a major internet company to defense AI development, but also reflects deeper application trends of AI technology in military command and control, information warfare, and related fields.

From LINE Yahoo to Defense AI: A Senior Engineer's Career Transformation

A Decade of Big Tech Experience

Masaru Itoh spent nearly ten years at LINE Yahoo (formerly Yahoo! JAPAN), progressing through a complete technical growth path from service development to platform infrastructure. LINE Yahoo was formed in October 2023 through the merger of LINE Corporation and Yahoo Japan under Z Holdings, making it one of Japan's largest internet service companies. It operates Japan's ubiquitous messaging app LINE (with over 95 million monthly active users), the Yahoo! JAPAN search engine and portal, PayPay financial services, and other core products, maintaining one of Japan's largest proprietary data center infrastructures.

Itoh was responsible for GYAO!'s (a video platform) metadata management system, Yahoo! Movies' backend technology overhaul, ad search engine ranking optimization, and real-time inference platform development based on machine learning.

The core value of this experience lies in his capabilities in large-scale distributed system design, development, and operations. As he put it, building services on Japan's largest proprietary infrastructure gave him a deep appreciation for the impact and responsibility that technology has on massive user bases.

The Decisive Turning Point: Joining Sakana AI

Itoh admits that his initial impression of Sakana AI was simply "a mysterious AI research lab." What truly changed his perception was deeper engagement with the team: the technical interview questions were elegantly designed with clear intent, and his conversation with CEO David Ha became the deciding factor—the company's emphasis on open-endedness culture, clear business progress, and distinct technical advantages revealed enormous growth potential.

Open-endedness is an important concept in AI research, referring to a system's ability to continuously generate novelty and complexity without converging to a fixed solution. This concept originates from observations of biological evolution—nature's evolutionary process has no predetermined endpoint, yet continuously produces increasingly complex and diverse life forms. At the organizational culture level, this means encouraging researchers to explore unknown directions rather than merely optimizing known objectives. David Ha himself is a well-known advocate of open-endedness research, having published multiple papers on the topic, and this philosophy deeply shapes Sakana AI's research culture and organizational operations.

The deeper motivation was his pursuit of social contribution "beyond the framework of private internet services." The defense sector provided exactly such an outlet.

Two Core Directions in Defense AI Development

SNS Disinformation Countermeasures: AI Applications in Intelligence

The team has developed proprietary technology targeting disinformation in social media spaces. This technology has been deployed in collaboration with the Yomiuri Shimbun for visualization analysis of "cognitive warfare" on social networks.

Cognitive Warfare has been identified by NATO and other organizations as the "sixth domain of warfare" after land, sea, air, space, and cyber. Its core is manipulating information to influence the cognition, judgment, and behavior of target populations. Unlike traditional information warfare, cognitive warfare emphasizes using social media algorithms, deepfake technology, and Coordinated Inauthentic Behavior to systematically distort public perception. In recent years, from information operations in the Russia-Ukraine conflict to foreign interference in various national elections, cognitive warfare has become a central tool in modern geopolitical competition. Given that information warfare is increasingly central to modern conflict, the strategic value of this direction is self-evident.

Command and Control Systems (C2): Supporting Rapid Situational Awareness and Decision-Making

This is the more challenging direction—building command and control systems that support rapid situational awareness and decision-making for military operations. C2 (Command and Control) systems are the technical core of military command structures, responsible for integrating battlefield situational awareness, communication networks, and decision support into a unified operational command platform. Modern C2 systems have evolved into C4ISR (Command, Control, Communications, Computers, Intelligence, Surveillance, Reconnaissance) architectures that need to process multi-source heterogeneous data from satellites, drones, ground sensors, and more in real time.

The introduction of AI primarily manifests in automated situational assessment, threat identification prioritization, and assisted generation of courses of action, with the goal of compressing the OODA loop (Observe-Orient-Decide-Act) from hours to minutes or even seconds. The system needs to integrate massive data from various devices including drones, perform comprehensive analysis, and provide support for command decisions.

Tech Stack and the Special Challenges of DDIL Environments

From a tech stack perspective, the team uses a relatively standard combination: Python backend (highly compatible with ML), TypeScript/Next.js for Web UI, and Kotlin for Android applications. But the real technical challenge lies in DDIL environments (Denied, Disrupted, Intermittent, Limited)—extreme environments where communications are jammed, interrupted, intermittently connected, or severely bandwidth-constrained.

DDIL is a specialized term in military communications describing the communication environment on modern battlefields where the electromagnetic spectrum is actively jammed by adversaries or physical infrastructure has been destroyed. Specifically, "Denied" means communications are completely denied, "Disrupted" means communications are actively degraded through jamming, "Intermittent" means connectivity is on-and-off, and "Limited" means available bandwidth is extremely constrained. In such environments, conventional cloud architectures are completely unsuitable—systems must possess edge autonomous computing capabilities, checkpoint-and-resume mechanisms, intelligent data priority sorting, and efficient information compression and transmission under extremely low bandwidth. This requires technologies such as Edge Computing, federated learning, and CRDTs (Conflict-free Replicated Data Types) to achieve decentralized data consistency and system resilience. The U.S. military's JADC2 (Joint All-Domain Command and Control) program similarly identifies DDIL environment adaptability as a core technical challenge.

Entirely new distributed system architectures must be designed, placing extremely high demands on engineers' system design capabilities—and this is precisely where Itoh's decade of large-scale distributed systems experience comes into play.

He mentioned that defense development has required him to "fully mobilize all the technical accumulation from his entire career"—from edge inference and multi-device input integration to distributed system design in DDIL environments, every subsystem spans multiple technical domains.

Cautious Application of Generative AI in Military Scenarios

The Critical Boundary of Human-Machine Collaboration

One of the most thought-provoking points in the interview concerns the principles for applying generative AI in mission-critical scenarios. Itoh explicitly states:

"Considering that every one of our outputs relates to the lives of Self-Defense Force personnel, there is no higher level of tension and responsibility. Using generative AI in this mission-critical domain requires particularly deep attention—ensuring that humans are substantively involved and guaranteeing the quality of implementation code and system outputs is absolutely essential."

This position reflects the current core consensus in military AI applications: AI is a tool for augmenting human decision-making capabilities, not for replacing decision-makers. In life-or-death scenarios, human judgment and oversight are indispensable. This principle is widely endorsed at the international level—the U.S. Department of Defense's 2023 Autonomous Weapons Systems Directive explicitly requires maintaining "Meaningful Human Control" in decisions involving the use of lethal force, and discussions under the UN framework on Lethal Autonomous Weapons Systems (LAWS) similarly emphasize the irreplaceability of human oversight.

AI Empowerment Strategy in the Development Process

At the same time, the team extensively uses generative AI within the development process itself—not only for code implementation but also for goal setting and problem extraction. Under strict information governance, AI tools help the team produce higher-quality outputs with fewer people in less time.

This dual-track strategy of "external caution, internal efficiency" may represent the best practice model for AI applications in highly sensitive domains. Externally (facing operational system outputs), the Human-in-the-Loop principle is strictly enforced, ensuring every AI-generated recommendation is reviewed by human experts. Internally (facing development workflows), AI is fully leveraged to accelerate coding, documentation generation, test case design, and other processes, maximizing efficiency gains in areas that don't involve sensitive decisions.

Team Culture and Collaboration Model in Defense AI Development

Breaking Down Traditional Role Boundaries

In the early product stage, there is no strict division between Applied Research Engineers and Software Engineers. The team adopts a tight collaboration model of "everyone does what they can together." This flat working style facilitates rapid iteration and helps new members integrate quickly.

Cross-Domain Expert Support System

Interestingly, the team includes not only technical personnel but also domain experts from Japan's Ministry of Defense and Ministry of Foreign Affairs. This cross-domain talent configuration effectively compensates for software engineers' lack of military domain knowledge. Itoh himself had virtually no expertise in land, sea, or air defense when he joined, and it was precisely through these experts' support that he was able to get up to speed quickly.

This "technology + domain" hybrid team model is not new in defense technology, but it is relatively rare in AI startups. Traditional defense contractors like Lockheed Martin and Mitsubishi Heavy Industries typically possess decades of accumulated military domain knowledge, while AI startups' advantages lie in technical agility and cutting-edge algorithmic capabilities. By bringing in domain experts from retired military officers or government agencies, Sakana AI effectively bridges this knowledge gap.

In terms of development style, the team eschews formalized processes and meetings in favor of truly agile methods: freely extracting tasks from concrete requirements research, implementing development after reaching consensus on priorities, then gathering feedback to enter the next iteration cycle.

Industry Implications: A New Paradigm for AI in Defense

Japan's Choice in the Global Trend

Sakana AI's entry into defense is not an isolated event. Globally, companies like Palantir and Anduril have already demonstrated the enormous commercial and strategic value of AI technology in defense. Palantir Technologies was founded in 2003 by Peter Thiel and others, initially providing big data analytics platforms for U.S. intelligence agencies. Its Gotham platform is widely used for counterterrorism intelligence integration and battlefield situational awareness, and its AIP (Artificial Intelligence Platform) launched in 2022 integrates large language models into military decision-making processes. Anduril Industries was founded in 2017 by Palmer Luckey (founder of Oculus VR), focusing on autonomous systems and computer vision applications in defense, with its Lattice platform providing AI-driven battlefield awareness networks. Both companies have achieved valuations exceeding tens of billions of dollars, demonstrating the enormous commercial potential of AI defense applications.

As a nation facing complex geosecurity challenges—including North Korean missile threats, Taiwan Strait tensions, and territorial disputes with China and Russia over the East China Sea and Northern Territories—Japan's application of cutting-edge AI technology to defense is an inevitable trend. Japan's government explicitly called for substantially strengthening defense capabilities in its revised National Security Strategy at the end of 2022, listing military applications of advanced technology as a priority, with defense budgets planned to double to 2% of GDP over five years. This policy backdrop provides institutional support for technology companies like Sakana AI to enter the defense sector.

Career Implications for Engineers

This interview sends an important signal: defense AI development does not require engineers to have a military background. The core requirements remain solid software engineering capabilities—distributed system design, full-stack development, ML system integration, and more. Domain knowledge can be supplemented through team collaboration and continuous learning.

As Itoh puts it, now is "the most interesting period where one can broadly and deeply influence decision-making." For engineers seeking both technical depth and social impact, defense AI may be a direction worth serious consideration.

Key Takeaways