Christie's first AI art auction has caused controversy, with a transaction volume of US$728,000
178
English
English
Microsoft's research team recently launched an artificial intelligence technology called "Large Action Model" (LAM), marking a new stage in AI technology. Different from traditional language models, LAM can not only conduct conversations, but also independently operate Windows programs and perform specific tasks.
Advantages of LAM
LAM can understand multiple input methods (such as text, voice, images) and convert these inputs into specific task steps. It can not only formulate plans, but also adjust action strategies according to actual situations. The construction process of LAM includes:
Task decomposition : Break down complex tasks into steps.
Translate into action : Convert plans into actual actions through AI systems such as GPT-4o.
Explore new solutions : Find new solutions on your own.
Fine-tuning training : Optimizing performance through rewards.
Test results
In the experiment, the LAM model based on Mistral-7B performed well in the Word test :
Success rate : LAM's probability of successfully completing the task is 71%, exceeding GPT-4o's 63%.
Execution speed : LAM only takes 30 seconds per task, GPT-4o takes 86 seconds.
Data expansion and training
The research team improved the training effect by expanding the data set from the original 29,000 pairs of mission plans to 76,000 pairs. Ultimately, approximately 2,000 successful task sequences were used for training.
Continuous challenge
Despite LAM's excellent performance, the research team still faced several challenges, including task errors , regulatory issues , and technology scalability . However, LAM represents a major advancement in AI technology and can help humans complete more practical tasks in the future.
Summary of key points
LAM can autonomously execute Windows programs, breaking through the limitations of traditional AI.
LAM performs tasks faster and with a higher success rate.
Data expansion and training optimization enhance the effectiveness and accuracy of LAM.
