The new LAM model can perform tasks in Word.

Microsoft researchers have developed the LAM (Large Action Model), an AI capable of controlling Windows applications, including Microsoft Office. Unlike traditional language models like GPT-4, which primarily generate text, LAM can transform user requests into real actions, creating step-by-step plans. It processes input from text, voice, and images, adjusting its actions in real-time.

In a test with Word, LAM successfully completed tasks 71% of the time, outperforming GPT-4, which succeeded 63% of the time when visual information wasn't included. LAM was also faster, taking just 30 seconds per task, compared to GPT-4's 86 seconds. However, when visual data was used, GPT-4 achieved a higher accuracy rate of 75.5%.

Training involved 29,000 "task-plan" pairs, later expanded to 76,000 examples with the help of GPT-4. Despite challenges like AI errors and technical limitations, researchers believe LAM marks a significant step toward Artificial General Intelligence (AGI).

The creation of LAM follows four main stages: breaking tasks into logical steps, using advanced AI like GPT-4 to convert plans into actions, developing new solutions independently, and refining the system through reinforcement learning.