Using Artificial General Intelligence With Autonomous AI Agents

James Dooley (0:00)

Artificial General Intelligence, also known as AGI.

Today I'm joined with Kasra Dash, and we're gonna be talking about what Artificial General Intelligence actually is. So Kasra, for anyone that doesn't know, can you briefly explain what AGI is?

Kasra Dash (0:19)

Yeah, so I think the easiest way to explain this is general intelligence of the entire world, and then you've got artificial general intelligence. So that's when an AI robot matches the general intelligence of the world.

James Dooley (0:37)

Yeah, for sure. And obviously nowadays we're using a lot of different AI harnesses that's integrating a lot of autumnus agents and agentic workflows.

So obviously there's a few tools out there now saying, like Codus, anyone who's watching this that hasn't used it, check out the link in the description where we review Codus as being the best AI harness. But with regards to using it as being AGI, they say Ferri, who's the founder, says that that's a code editor for AGI. So he's basically saying that the current agents now are performing at Artificial General Intelligence level. Do you agree with that? Or is there any reason why you might say it's not always the case?

Kasra Dash (1:24)

I think AI does a brilliant task of being able to code things and answer questions and stuff like that. But when you're doing longer tasks, and many people might not even experience this, so it might be like 5 percent of people watching this.

When you task it to code something massive, like, I don't know, build me a Stripe alternative, for example, what you will actually end up realizing is that when you start getting granular into certain features, like I want Apple Pay, I want Android Pay, there's certain times where you ask it to fix something and it will break something else, and that's solely down to the context window. But that's probably the only thing that I think from a coding point of view, using AI, it lacks, and I'm sure in maybe one or two months' time, they'll probably figure out a way to fix that as well.

James Dooley (2:20)

Yeah, for sure. So with regards to Artificial General Intelligence, do you have something set up where you're able to have a team or a swarm of agents being able to speak to each other? And if, let's say, one did go off and make mistakes with regards to the context window, have you got certain quality control measures where another AI agent can be checking and fixing it from a quality control measure?

Kasra Dash (2:48)

Yeah, so again, like what we were saying before, I use CODIS, the actual coding AI harness.

And basically what I'm able to do with that is I can speak to one orchestrator and that will spin up a load of different multi-agents underneath it. So one could be for design, one could be for backend, one could be for the server, one could be for quality control. And when we're shipping anything from an app point of view, it needs to pass all of those different AI agents. And then the orchestrator, the brain basically comes back to me and says, yeah, we're ready to ship this. And then we will actually ship it live to our audience.

James Dooley (3:33)

Yeah, for sure. And anyone who's watching this, make sure you check out the link where we're talking about ASI, Artificial Superintelligence. And I want to just talk a little bit now about the difference between AGI and ASI.

We've got a whole dedicated podcast about ASI, where artificial superintelligence is where it vastly surpasses humans. And I don't feel we're at ASI yet. But can you explain why when you start to deploy an agentic system and you have the quality control measures and you're able to get different AI agents to be able to speak to each other, how close that is now in creating that whole swarm of autonomous agents? Yes, they're only working at AGI levels, but how close that now is becoming to almost being an ASI level where they can corroborate with each other and you can see certain different agents arguing to get to the end result of what you actually require. How far off ASI do you think we are after deploying now the swarm of AGI agents?

Kasra Dash (4:46)

I think we're technically at ASI.

It depends how you use AI. So again, going back to what I was saying before, if you're just using ChatGVT inside your browser, inside chat, that's not really that advanced. Same goes with the people that are using Clod Code or Codex or whatever. Whereas if you're using a harness, right, you can have Clod Code and Codex or Astra 6, for example. So Fable 5.1 and Astra 6, both arguing and coming with an outcome. Now, to me, that's like two super intelligent people going back and forth and trying to work on the same task and finalise the actual task at hand. So to me, I think that's a little bit of ASI. I'm sure that over the next couple of years and stuff like that, we're going to get proper, true ASI. But currently, you can get the two strongest models out there, arguing against each other and getting new outcomes. So I think we're almost at ASI.