Summary

  • Simular AI has created an AI agent called S2, which combines frontier models with models that are designed to use computers, allowing it to achieve state-of-the-art performance when it comes to tasks like using apps and manipulating files.
  • It uses a powerful general-purpose AI model to reason about how best to complete a task, while smaller open-source models will carry out other tasks, such as interpreting web pages.
  • It learns from experience and can record user feedback and actions to improve future actions.
  • However, even the most advanced AI agents are still troubled by edge cases and can sometimes exhibit odd behaviour.
  • One possible solution is to add human intelligence to the mix, with systems like CowPilot, a Chrome plugin that allows a human to intervene if the AI gets stuck on a task.
  • The combo of a human and AI agent can perform more tasks than either party alone, with the human-agent combo completing 95% of given tasks, compared to 95%/15% when each works alone.

By Will Knight

Original Article