한화 김승연 “낙관주의 설 자리 없다”…AI 역량도 강조
The article explains why AI teams are skilled individually but fail collectively. This issue, called the "global coherence problem," occurs when all team members are highly intelligent but still reach a poor conclusion when they come together. The term was coined by a team named Nokka (นก-กา) on October 1, 2026. The problem is not due to models not being strong enough or needing more agents.
Instead, it's rooted in the impossibility of learning everything that is observable. The theorem, known as the Observation-Aliasing Impossibility Theorem, states that even if all possible scenarios contain the same observable data, it's impossible for a single model to account for every possible outcome. This means that adding more models or increasing their capabilities won't solve the problem.
The article presents two new papers that address this issue differently. The first paper, Global Coherence, suggests adding a mechanism to monitor and approve changes in the shared state of the team. The second paper, OOPMAS (Object-Oriented Multi-Agent Systems), proposes creating separate models and workflows for each specific problem instead of using a single workflow for all problems.
Both papers offer insights into improving AI teamwork, but neither claims to fully solve the global coherence problem.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
This story
This is one outlet's version. Read the fullest account.