Cognita Works

Cognita Works is a research spin-off from a major Seoul university — five researchers, three with active academic appointments, all working on the boundary between fundamental and applied AI research. The office is two floors above a 24-hour study café in Gangnam. The café owner waves at them in the elevator. They have, over the past two years, hired three of his customers.

The firm’s work is reasoning-architecture research — small models specifically designed to demonstrate or refute claims about how reasoning emerges from scale. They publish in respected venues. They sell the trained models to enterprise customers as “interpretation benchmarks.” Per-model invoicing is academic-modest: low five-figure Cu for off-the-shelf benchmarks, low six-figure for custom architectures. Both audiences are happy. Both are sometimes the same person.

The lead researcher’s office has a whiteboard that is wiped clean precisely once a year, on December 31st. She insists this is because the year’s annotations should fit on the year’s whiteboard. She also insists this is a methodological choice. Her postdocs do not interrupt her.

Cognita Works’s models occasionally exhibit reasoning patterns that the team finds “interesting” in the field’s specific technical sense. They publish about half of these observations. The other half they retain internally, citing “ongoing analysis.” The half they publish has built the firm’s reputation. The half they retain has built nothing they will discuss.