Andrew Ng discusses enhancing LLMs through iterative, agentic workflows for better, collaborative outputs.
Last week, I checked out a recent presentation by noted A.I expert, Andrew Ng – during which he shared his insights on agentic workflows, a new way to interact with large language models (LLMs) that could significantly improve their capabilities.
In the past, LLMs were used in a non-agentic way, where a user would provide a prompt and the LLM would generate a response in a single shot. This is similar to asking a person to write an essay without allowing them to revise or edit their work.
Agentic workflows, on the other hand, are iterative. The user prompts the LLM to complete a task, then the LLM can perform actions like revising its output, searching for information, or calling other LLMs to help. This back-and-forth process allows the LLM to generate much better results.
Ng presented four design patterns for using agentic workflows:
• Reflection: Here, the LLM reflects on its own outputs and identifies areas for improvement. For instance, an LLM that generates code can be prompted to review the code for correctness and efficiency.
• Planning and Multi-agent Collaboration: This involves using multiple LLMs working together. For example, one LLM could be a coder, while another LLM could be a reviewer. These LLMs can communicate with each other to complete a task.
• Tools: Many LLMs can already access and use different tools. For instance, an LLM can use a code-generating tool or a web search tool to complete a task.
• Many-to-Many Use: This involves combining different LLMs, each specializing in a different task. An LLM can generate code, while another LLM can find information relevant to that code.
Ng believes that agentic workflows will significantly improve the capabilities of LLMs. He argues that we need to move away from the expectation of instant results from LLMs and instead be patient with the iterative process of agentic workflows.
The concept of A.I agents and Agentic Workflows is relatively new - Sounds like a promising approach to unlock the full potential of large language models.
So, what’s next? How can we start experimenting with agentic workflows?
How about incorporating iterative processes into our current use of LLMs? This could involve prompting the LLM to revise its outputs, search for additional information, or collaborate with another LLM to complete a task. This could be a good way to get models to enhance problem-solving capabilities and start learning more about the benefits of Agentic Workflows.
