The 85% Rule for Learning and The Near Future of AI is Action-Driven: Connecting the Dots

The 85% Rule for Learning and The Near Future of AI is Action-Driven: Connecting the Dots

In both education and artificial intelligence (AI), there seems to be a common thread - the optimal success rate for learning and training is around 85%. Scott H Young, in his article "The 85% Rule for Learning," suggests that a success rate of 85% indicates that students are learning the material effectively while also being appropriately challenged. Similarly, in the realm of AI, it has been found that the optimal error rate for training is approximately 15.87%, which translates to a training accuracy of around 85%.

This 85% rule highlights the importance of finding the right balance between difficulty and support. Lev Vygotsky's zone of proximal development theory suggests that tasks slightly beyond our current capabilities, but achievable with assistance, maximize learning. Therefore, if we are getting more than one out of every five problems wrong, it may be beneficial to seek additional help. On the other hand, if we are consistently getting most of the problems right, it might be time to increase the difficulty level to foster further growth.

Interestingly, the concept of difficulty also plays a role in motivation. Robert Eisenberg's theory of learned industriousness suggests that success on challenging problems can be motivating, while failure can be demotivating. Therefore, finding the sweet spot of difficulty where we are likely to succeed but still face a significant challenge can encourage effort and persistence in the future.

Moving into the realm of AI, the concept of action-driven models becomes increasingly relevant. The ReAct model, as described in "The Near Future of AI is Action-Driven," takes a three-step iterative approach: Thought, Act, and Observation. By actively choosing actions and observing their outcomes, the model functions more like an agent, resembling the characteristics of artificial general intelligence (AGI). LLMs (large language models) have shown improved performance in question-answering tasks when prompted to think step by step, but their capabilities can be further enhanced by utilizing external cognitive assets.

OpenAI's 002-text-davinci model has achieved remarkable results by combining instruction tuning and reinforcement learning from human feedback. This combination allows the model to learn and improve based on the success ratings provided by humans. However, the true potential lies in actual reinforcement learning, where the system can be trained to produce better results based on specific metrics of interest.

Looking ahead, the future of AI lies in the development of powerful feedback loops. Startups that can solve customer pain points, collect data to enhance their solutions, train their models for consistency, and iterate on the process are likely to succeed. This iterative approach will serve as a moat in the AI landscape, expanding the possibilities and automation capabilities as the agents become more domain-general.

In conclusion, the 85% rule for learning and the emergence of action-driven AI models reveal fascinating connections. By finding the right balance between difficulty and support, both humans and machines can optimize their learning and training processes. Additionally, the incorporation of external cognitive assets and the utilization of reinforcement learning techniques hold immense potential for the future of AI. To make the most of these insights, here are three actionable pieces of advice:

1. Fine-tune the level of support or difficulty based on your success rate. If you're struggling, seek additional help, but if you're consistently succeeding, challenge yourself further.

2. Embrace an action-driven approach in your learning or problem-solving endeavors. Actively choose actions and observe their outcomes, allowing for a more agent-like experience.

3. Explore the possibilities of reinforcement learning and feedback loops. Continuously gather data, train models to improve, and iterate on the process to achieve better results.

By implementing these strategies, we can enhance our learning and problem-solving abilities while also contributing to the advancement of AI technologies.