Autonomous AI Self-Improvement and the Next Generation of Agents | 29th Sep 2026

DailyNews (0:00)

Welcome to Colaberry AI Podcast, brought to you by Colaberry AI Research Labs and Carl Foundation.

SPEAKER_2 (0:05)

Glad to be here for this one. It is a really fascinating topic today.

DailyNews (0:10)

Yeah, it totally is. So imagine an AI from a company called WICO that just completely redesigned its own underlying codebase. Right.

SPEAKER_2 (0:19)

In exactly eight days, which is crazy.

DailyNews (0:22)

Exactly. Eight days and not a single human engineer typed a line of code or intervened in any way.

SPEAKER_2 (0:30)

Nothing. Zero human touch.

DailyNews (0:32)

And when this system was finished, it definitively beat a model that top-tier human researchers had spent two full years building by hand.

SPEAKER_2 (0:40)

Two years of human work beaten in eight days.

DailyNews (0:44)

Yeah. So today's deep dive is about the mechanics of how that actually happened.

SPEAKER_2 (0:48)

Right. And we are looking at a transcript today from the AI Revolution Channel titled The first real RSI is here and it's evolving fast.

DailyNews (0:57)

And just for those following the technical vocabulary of the industry, RSI stands for Recursive Self-Improvement.

SPEAKER_2 (1:04)

Right. It is the moment an artificial intelligence becomes capable of rewriting its own code to become smarter, right? Faster than human developers can upgrade it.

DailyNews (1:14)

Exactly. So if you are listening right now, whether you are a developer looking for things to try with new AI tooling or a security researcher mapping out vulnerabilities.

SPEAKER_2 (1:24)

Or honestly just someone curious about where the tech industry is heading.

DailyNews (1:27)

Yeah, totally. Understanding the precise mechanisms of this RSI experiment is critical for you to know what happens next.

SPEAKER_2 (1:34)

Because WIDCO didn't just like type a prompt saying, hey, get smarter.

DailyNews (1:38)

Right. They engineered a very specific dual model architecture. So let's look at how they set the board.

SPEAKER_2 (1:44)

Yeah. The architecture is the entire reason this succeeded.

WCO essentially built a system with two AIs working in a rigid tandem.

DailyNews (1:54)

A boss and a worker, right?

SPEAKER_2 (1:55)

Exactly. They created a boss model and a worker model. The boss was WCO's hand-built AI, which they called AID, and it was running on Claude Opus 4.7.

DailyNews (2:06)

Which is a heavy hitter.

SPEAKER_2 (2:07)

Oh, absolutely. And its sole job in this environment was evaluation. It just observes how the worker is processing tasks.

DailyNews (2:16)

It rewrites the instructions for how the worker thinks, tests the new version, and only keeps it if it actually beats the previous one.

SPEAKER_2 (2:23)

Right. Demonstrably beats it.

DailyNews (2:25)

I want to focus on the worker model for a second though, because this is where the resource allocation gets really fascinating.

SPEAKER_2 (2:31)

The worker choice is key here.

DailyNews (2:33)

The worker was not running on a massive flagship model like Opus. It was actually running on Gemini 3 Flash. And looking at the data, the choice to use a lightweight model wasn't an accident. They were operating on a strictly fixed compute budget.

SPEAKER_2 (2:47)

Which fundamentally changes the experiment, honestly.

DailyNews (2:50)

When you are on a fixed budget, your cost per token is like your most important metric.

SPEAKER_2 (2:54)

And Gemini 3 Flash is vastly cheaper to run.

DailyNews (2:58)

Exactly.

SPEAKER_2 (2:58)

By using the cheaper model as the worker, the system could afford exponentially more attempts, you know?

DailyNews (3:04)

More iterations, more trial and error runs for the exact same amount of money.

SPEAKER_2 (3:08)

Exactly.

DailyNews (3:08)

It reminds me of a corporate structure where the CEO is an incredibly expensive high level consultant.

SPEAKER_2 (3:13)

That's the Claude Opus model.

DailyNews (3:15)

Yeah. And they are tasked with optimizing a massive fleet of cheap rapid fire interns, which is the Gemini Flash model.

SPEAKER_2 (3:22)

I like that analogy.

DailyNews (3:23)

But because they have a strict corporate budget, they can't just, you know, buy their way to success by throwing money at cloud computing. Right.

SPEAKER_2 (3:30)

They have to win by figuring out how to work more intelligently.

DailyNews (3:33)

Exactly. Getting smarter, not just spending more money.

SPEAKER_2 (3:36)

And that analogy actually maps perfectly to the tasks those interns were given. WCO didn't give the worker easy jobs.

DailyNews (3:43)

No, not at all.

SPEAKER_2 (3:44)

It had to handle three highly complex domains, like training other AI models, solving difficult routing and scheduling puzzles.

DailyNews (3:52)

Like the classic traveling salesman problem, right?

SPEAKER_2 (3:54)

Yes, exactly that. And also debugging real functional software code.

DailyNews (3:58)

Wow. And to make sure the AI wasn't just faking its progress on those, WCO instituted this brutal grading methodology.

SPEAKER_2 (4:07)

Yeah, the dual score system.