The recent study titled “Cloning of Thoughts: Learning to Think in the Process of Action, Imitating Human Thinking” has opened up new perspectives in the field of artificial intelligence. Researchers from the University of British Columbia, Vector Institute, and Canada CIFAR AI Chair have presented a pioneering method for creating Intelligent Agents or Non-Human Workers that mimic human actions and generate and rationalize their actions.
Here’s the crux of the new method:
- The method promotes a high-quality increase in the speed and efficiency of AI training. It provides a mechanism to prevent AI crimes, ensuring the safe and effective implementation of AI.
- Intelligent Agents or Digital Employees learn from multiple information streams: observing actions, such as game moves, and a parallel stream of thoughts that explain these actions.
- For instance, in a real-time strategy game, the AI observes a player’s move and receives a text explanation like “Do not allow the enemy to cross the bridge”. This dual-stream process facilitates the creation of correct associations between behaviors and goals.
- Using thought cloning, these Non-Human Workers can produce thoughts in natural language at each interval and then base their actions on these generated thoughts.
- The methodology, aptly named “Interference before the crime”, also contributes to AI safety by allowing us to monitor an agent’s thoughts, diagnose potential issues, guide the agent by correcting their thinking, or prevent unsafe actions. It’s a system that echoes the pre-crime concept from the movie “Minority Report”.
Why is this important?
The approach enhances the training efficiency of large language models (LLMs), leading to more human-like digital employees.
It offers an unprecedented ability to generate and justify their actions, making these AI agents more reliable and predictable.
Proactive safety measures minimize the risk of AI malfunctions or harmful behaviors, fostering trust in AI systems and promoting their widespread adoption.
See here for the original article and more details about the ideal cloning method. The project’s results, including model weights, training code, and data generation code for training and testing, can be found on GitHub.
Ultimately, the innovation of thought-cloning is a significant leap towards creating Intelligent Agents that can function as Non-Human Workers in various sectors. The potential applications are vast, from customer support to personalized marketing, creating exciting prospects for businesses and consumers alike.
