[Yomiuri Shimbun] Sakana AI's Proprietary System Visualizes "Cognitive Warfare" on SNS

Sakana AI just showed off a tool built for a very specific, very modern job: reading the internet's mind during a diplomatic spat. Working with The Yomiuri Shimbun, the Tokyo-based startup ran its proprietary system over 1.1 million social media posts tied to Chinese criticism of Japan, and the results went well beyond a simple sentiment count.

The interesting part isn't the volume of data, it's what the system pulled out of it. Take a post demanding that a Japanese official retract a supposedly erroneous statement. A keyword search for "Taiwan" would find nothing there. But Sakana's system flagged it as part of a broader narrative about interference in the Taiwan issue, connecting dots that live in implication rather than explicit text. That's the whole pitch: context and nuance over brute-force pattern matching.

Under the hood, three different large language models work in tandem, cross-checking each other to surface granular narratives before clustering them into bigger thematic groups. That layered structure lets analysts zoom from thousands of individual posts up to a handful of dominant storylines, which is presumably the only way to make sense of over a million posts without drowning in them. From there, the AI generates hypotheses with supporting evidence attached, and human analysts pick through them, asking for more digging where something looks promising and discarding what doesn't hold up.

One hypothesis to come out of this run: that Chinese criticism of Japan escalated only after officials debated a coordinated response following remarks by Prime Minister Takaichi in the Diet. Yomiuri reporters didn't just take the AI's word for it. They went and interviewed government sources on both sides and consulted outside experts to confirm the timeline actually held up, which is a good reminder that even sophisticated AI output still needs old-fashioned shoe-leather verification before anyone treats it as fact.

Sakana frames this as a template for a new kind of intelligence work, where AI surfaces patterns humans would never catch by hand, and humans stay in the loop to stress-test and act on them. The company is now naming defense and intelligence as a core business line alongside finance, language that signals it wants a piece of Japan's growing appetite for AI-driven information warfare tools, not just chatbots and coding assistants.