Agentic AI, the Next Phase in the Journey Toward General Artificial Intelligence
Yesterday I sat through a presentation by a data science professor at UT. He was covering OpenAI-01, which was officially released just a few days ago. 01 represents the next step up the chain toward General Artificial Intelligence. Here are OpenAI's stages in that chain:
Level 1: Conversational AI - AI systems have natural sounding conversations with humans. <<<<This is the level where most language models are today.
Level 2: Reasoning AI - Solve basic problems at a level comparable to a person with a PhD. <<<<This is v01. "basic" = "PhD" made me laugh a bit.
Level 3: Autonomous AI - AI agents can act autonomously, performing analyses and then executing tasks based on those analyses on behalf of users.
Level 4: Innovating AI - AI systems can develop innovative ideas and solutions.
Level 5: Organizational AI - This is the ultimate level where AI systems can perform the tasks of an entire organization
While not yet level 3, v01 adds support for some "agentic" or agent-based reasoning. The programmer assigns different tasks or points of view to different agents and then lets the agents battle it out to come up with an answer. For instance, in a customer service scenario, you might create 3 agents: 1 problem identifier agent whose job it is precisely determine what the customer is complaining about, 1 corporate optimizer that comes up with a solution that most benefits the company and 1 customer sat agent that comes up with the thing that will make the customer happiest (free replacement, a credit, etc). These agents would listen in on the human support rep's (or a chat bot's) interaction with the customer then battle it out behind the scenes to come up with the best solution for each complaint.
Agents don't always have to be diametrically opposed though. In a situation where you want GPT to look for correlations between data, say a company's stock performance and their pricing or inflation rate, 1 agent could be in charge of tracking stock performance, another tracks inflation rates, a 3rd pulls the analysis together and a 4th identifies and incorporates tools/libraries needed execute the analysis workflow, all without human intervention, leaving the user to review the results and tweak the agents, OR if things are taking too long/too expensive, to help the agents shortcut things by pointing them at particular data sources, etc.
This is fascinating stuff and the agent-based model takes us closer to the kinds of decision making and distributed tasking that we take for granted every day as humans.