Recursive conversations

Where we are going (near future)

When you simultaneously see a similar idea publicly emerging from very independent directions you can be sure that there are a LOT more people privately working on the idea. Such is the case with "recursive conversations" in generative AI, which goes beyond langchain, which uses a series of actions or "orchestration" to achieve a specific LLM goal.

An early example of a recursive conversation, and helpful in understanding the general concept, researchers created a virtual town in which generative AI agents took on roles within a community and dynamically interacted with one another (Generative Agents: Interactive Simulcra of Human Behavior). For example, one of the agents is configured to run for mayor -- as "he" talks with other agents to explain why he is running, those agents take on the messaging which they then repeat to other agents. See the linked research paper for some in depth examples.

The two examples that I have seen more recently relate to software development and applying the same idea of recursive conversations between agents to the different roles in writing a software program: design, programming, testing, etc. In the first example, Communicative Agents for Software Development, researchers set up a virtual company called ChatDev and configured agents to play all of the different roles in the software development process. The researchers report of the interaction between different agents:

Each stage engages a team of agents, such as programmers, code reviewers, and test engineers, fostering collaborative dialogue and facilitating a seamless workflow. The chat chain acts as a facilitator, breaking down each stage into atomic subtasks.

A simple task was given to this "team" to write a software game (Gomoku) and they report that the virtual team of agents was able to complete the task in about 7 minutes... In the "discussion" section of the report the researchers do highlight some specific challenges and risks to be addressed but the concept seems quite compelling especially today for taking on simple development tasks.

And then independently a group of developers has created something called GPT-Pilot which they have posted on github so any developer can now play with the idea of a recursive conversation. As the author's explain:

The primary premise here is that AI has reached a point where it can autonomously generate a substantial portion of code for an application, potentially up to 95%

But the challenge is how to test and debug this code. This is where GPT-Pilot comes in, testing the code, fixing the code, retesting the code...

Software development is one of the first categories of work that generative AI has entirely reconfigured, with Github claiming that 92% of US-based developers already using coding assistants. So it makes sense that software development would be ground zero for this next enhancement in how generative AI is used. But the lessons from a virtual town and creating software programs should extend to many other categories of content. Imagine the virtual room of script writers where a set of agents are set up to write SNL skits -- each is configured to bring a different set of ideas to the interplay and the team as a group is given the task of writing a particular kind of script - they go back and forth rewriting for timing, humor, current events, etc. and produce in 7 minutes what a human writing team might take hours to produce... Will it be as good as the human writers? If it can get to 95% does it at least change the way this (and other tasks) are done in the future?