The new LAM model can perform some tasks in Word.

Microsoft researchers have developed the LAM (Large Action Model), an AI designed to control Windows applications, including Microsoft Office. Unlike traditional language models like GPT-4, which focus on generating text, LAM can turn user requests into real actions by creating step-by-step plans. It processes input from text, voice, and images, adjusting its actions in real-time.

In tests with Word, LAM completed tasks successfully 71% of the time, outperforming GPT-4, which achieved 63% when visual data wasn't included. LAM was also faster, completing tasks in 30 seconds compared to GPT-4's 86 seconds. However, when visual information was provided, GPT-4 had a higher accuracy rate of 75.5%.

For training, researchers used 29,000 "task-plan" pairs, later expanding this to 76,000 examples with GPT-4's help. Despite challenges such as AI errors and technical limitations, researchers believe LAM represents an important step toward Artificial General Intelligence (AGI).

LAM's development follows four key stages: breaking tasks into logical steps, using advanced AI like GPT-4 to convert plans into actions, finding new solutions independently, and improving the system through reinforcement learning.