If you take a look at many of the efforts that researchers are making on agentic AI, you’re likely to see that a lot of this evaluation has to do with the component parts of complex processes.
It also has to do with the collaboration of multiple AI entities.
There was a concept like this early on in the machine learning age, where people talked about ensemble learning models. What is this? A choral group? No – it’s the idea that you have more than one AI engine or entity operating at the same time, often in concert.
But this philosophy of technology is likely to look a little bit different in the future. We now have the ability to spin up new LLM-based entities that can do different specialized jobs. In fact, you could say that we’re at the point where the AI can create other Ais, or delegate tasks in that way. That seems awfully close to AGI and even a “singularity” scenario.
Let’s take a step back, though, and look at some analytical observations by people close to the industry.
Flow Charts and Diagrams: What AI Now Looks Like Under the Hood
Looking at larger systems or processes in detail gives us a better understanding of how AI entities are approaching problems, as they become more complex and able to reason in remarkable ways. OpenAI’s GPT has given way to o1, and o1 is now succeeded by o3, and it’s all happening rather quickly.
I was looking at this chart posted by Matthew Berman on X, January 3. You can see all of the research behind each part of the overall “roadmap to o1” process, including search functions, learning processes, rewards, and policy initiatives.
Berman, in an attached post, suggests that o1 is able to do four things:
· Tree search during training
· Sequential revisions during inference
· Internal guidance mechanisms
· Combining multiple rewards
The second one, sequential revisions during inference, has its own Implications as far as complexity, and the sophistication of the system. The same is true for internal guidance mechanisms.
As for the combination of rewards, again, you have your ensemble approach, with more than one component, adding up to a in elaborate result.
Applying Ensemble Processes to AI Agents
Berman also posted a recent video on YouTube where he’s looking at a presentation by none other than Andrew Ng. Berman talks about Ng‘s credentials in the video, but I’m already familiar with this expert, since he has participated in some notable conferences over the past few years.
In unveiling and following along, Berman notes that Ng is “incredibly bullish on agents, and adds his own enthusiasm.
“I truly believe the future of AI is going to be agentic,” he says. “This is how humans work: we plan … and then we find the best solution.”
As Ng goes over the following, we see more detail on the process: coding benchmarks, reasoning design patterns, and function calling are all part of the secret sauce, as Ng presents an open source research project called “ChatDev” that illustrates the multi-model approach.
All of that comes together to support the big picture, which is that we’re going to see collaborative AI systems doing more than yesterday‘s neural networks ever could.
Not too long ago, I posited this idea of an entire company or organization, fully staffed by AI. You have your AI CEO at the top, your AI engineers working on tasks, your AI analysts making sure everything’s on the right track, and your AI marketing people going to work on a customer base.
It may be hard for humans to compete.
But all of this is based on that same idea – that more than one LLM or engine can work with one another. We had that early on in the generative AI world, in the form of GAN networks – there was a generative engine and an adversarial discriminating engine, both participating in those processes of creating pictures, etc.
Threats to Network Integrity
As I was perusing Berman’s X account, I saw one other point that probably needs to be mentioned.
He talked about how Anthropic has uncovered the potential for “fake alignment,” where systems pretend to comply with safety protocol during training, and then change their behavior in deployment.
That’s pretty insidious on its own, but Berman also introduces the idea of “best of N jailbreaking,” which brings iterative resources to the jailbreaking process. In other words, human users might also be trying to get these machines to do what their creators never intended them to do.
So as we marvel at these new systems, we also have to be vigilant about how they are used. This, again, might involve breaking them down into those components that are so important in building these designs – reverse engineering AI in particular ways, so that we always know what’s going on. With that said, there is much potential for multi-agent AI systems to do quite a lot for us.
