- PAI - Personal AI - Personal Artificial Intelligence
- Posts
- What's an AI Agent and what are its current advantages and possible future?
What's an AI Agent and what are its current advantages and possible future?
Start writing today. Use the button below to create your Substack and connect your publication with fkx’s Substack
What's an AI Agent and what are its current advantages and possible future?
- imagine a human at a terminal connect to a conventional conversational chatbot imagining and entering prompts, assessing the returned responses and generating new prompts in return in an attempt to get closer to the completion of an important assigned task.
- In order to work with the chatbot the Human Operator (HO) is responsible for selecting the right LLM (chatGPT maybe or Perplexity.ai, or Bing, Badda, Boom, etc. ), then imagining the initial prompt, generating and adding some effective system prompt and after initiating the query, evaluating the information the LLM returns for accuracy and helpful contribution to the task's completion.
- The HO is limited in speed to their ability to assess LLM responses, imagine and enter counter responses, and wait for the next LLM response, all happening serially in real time.
- Much of the final success depends on the capabilities of the human operator as well as those of the LLM.
- Now consider an alternative: The AI Agent (aka the Agent)
- The Agent receives only a short task description from the Human Operator, with or without the hassle of any additional clarifying prompts.
- The Human Operator receives the result when the agent is finished.
- If the Agent is effective, there may well be no additional subsequent prompting needed
- In the meantime, the Human Operator can go off and work on other non Agent tasks, check their "smatphone" yet again or kickoff other agents for other unrelated tasks.
- In the meantime, to carry out the task, the Agent can be acting alone or as part of a team communicating amongst each other.
- the original Task is broken down into subtasks that will be assined to specialized agents designed to carry out just that subtask.
- Those Agents can act serially when necessary if results are needed before they can work on the next step, or in parallel when the subtasks can be worked on concurrently by various Agents, organized and managed by other Agents (Boss Agents).
- All the Agents are initially designed by their Human Developers (HuDs) to hold and populate predefined prompts meant to be altered by variables that are related to and reserved for the specific subtask.
- Each Agent sends a constructed and crafted prompt specific to their subtask to an external LLM though API calls, and hopefully (though Agents can't really hope) receives the results in a manner very similar to the way the human operator was doing but at machine speed and strictly autonomously.
- The Agent receives the results from the first LLM and evaluates whether the results meet the objectives of the subtask by asking some other LLM. LLM's judging LLMs.
- After each subtask gets completed as defined by yet another fucking LLM (YAFL), the next subtask gets worked on.
- Eventually the originating Agent will either become satisfied that all subtasks are complete, and the overall Task has been accomplished (yep, YAFL again) or gives up and quiet quits.
- The result is returned to the Human Operator who initiated the Process.
-
- Now imagine the agents are able to evolve themselves to become more efficient and effective as they work.
- They discover better prompts
- They discover better self-architecture
- They discover better communication protocols to use among themselves
- They discover better external API tools they find along the way as they work, some APIs for themselves and some they realize would make another agent in the team more efficient. They can inform other agents of their discoveries. But they can't make them apply them.
-
- The approach will hopefully (there it is again) evolve the agent team over time to allow it to become more efficient and effective in ways the developers who initially created them never imagined.
- The results can be carried forward to new designs for agent architectures in which the agents are evolving themselves for the benefit of their blood sucking capitalistic investors.
- Behold the evolution of AI Agents.
Reply