Dossier
Philippe Laban
Coverage of Philippe Laban in the Nexus archive.
- Microsoft researchers find AI models and agents can't handle long-running tasks
Microsoft researchers found that AI models introduce errors in long workflows, contradicting their potential use for automated tasks. The study tested large language models on multistep tasks and found substantial errors, with an average degradation of 50% across all models. This raises concerns about the reliability of AI agents for complex tasks.