The 2028 Warning: The Rise of Recursive AI Self-Improvement | 29th June 2026

**DailyNews** (0:00)
Welcome to Colaberry AI Podcast, brought to you by Colaberry AI Research Labs and Carl Foundation.

**SPEAKER_2** (0:05)
Thanks for having me back.

**DailyNews** (0:06)
Yeah, of course. So imagine handing a massive software engineering project to a brand new hire.

**SPEAKER_2** (0:12)
Right, like a major infrastructure rebuild or something.

**DailyNews** (0:15)
Exactly. You'd expect this project to take them, realistically, maybe four months of nonstop coding, but instead, they hand it back to you, perfectly completed, the very next morning.

**SPEAKER_2** (0:25)
Which is normally impossible.

**DailyNews** (0:28)
And their total compensation for the entire job was exactly $251.

**SPEAKER_2** (0:32)
Yeah, that's the crazy part.

**DailyNews** (0:34)
And for you listening, that isn't some hypothetical scenario we just made up. That is a documented benchmark that actually occurred inside a frontier AI lab.

**SPEAKER_2** (0:43)
Yeah, it fundamentally changes how we have to think about the timeline of this technology. I mean, we are moving entirely away from what these models can say to you in a chat window.

**DailyNews** (0:53)
Right. The chatbot era is kind of over.

**SPEAKER_2** (0:55)
Exactly. It's now entirely focused on what they can autonomously build in a development environment, and more importantly, how they're beginning to build themselves.

**DailyNews** (1:05)
Right. The timeline for AI's most dangerous and highly anticipated capability recursive self-improvement or RSI has just been severely compressed.

**SPEAKER_2** (1:16)
Very severely.

**DailyNews** (1:17)
So the mission of this deep dives is to basically bypass the hype. We want to analyze the hard technical data, the specific evaluation methodologies, and these startling benchmark results that led Anthropix Co-Founder to attach a 60% probability to RSI becoming a reality before 2028

**SPEAKER_2** (1:35)
Yeah, 2028 is right around the corner.

**DailyNews** (1:37)
It is. So we are going to get into the technical weeds today for you. We're looking past the chatbots to analyze things like agent harnesses, black box executables, and the raw chain of thought evaluations currently happening behind closed doors at these labs.

**SPEAKER_2** (1:50)
And to really understand that 2028 timeline, we have to look at what Jack Clark Anthropics co-founder laid out. He used this brutally simple formulation.

**DailyNews** (1:57)
Oh, right, the Claude 10 quote.

**SPEAKER_2** (1:59)
Yeah, he just said, Claude 10 building Claude 11

**DailyNews** (2:01)
Wow. Just one model building the next.

**SPEAKER_2** (2:05)
Right. And that single sentence represents the transition from human limited research to compute limited research.

**DailyNews** (2:11)
Because the AI is the engineer now.

**SPEAKER_2** (2:13)
Exactly. If an AI model becomes the primary engineering engine that architects the next generation of models, well, progress stops being bottlenecked by how fast human researchers can theorize, test, and write code.

**DailyNews** (2:26)
And Dimas Asabas at Google DeepMind recently confirmed this is the current reality, right?

**SPEAKER_2** (2:32)
Yeah.

**DailyNews** (2:33)
He calls it soft self-improvement.

**SPEAKER_2** (2:35)
Yeah, soft self-improvement. Because, I mean, we aren't talking about an AI magically rewriting its own core neural weights in real time just yet.

**DailyNews** (2:42)
That would be the sci-fi nightmare scenario.

**SPEAKER_2** (2:44)
Right, right. Soft self-improvement is about the AI automating the machine learning pipeline itself.

**DailyNews** (2:49)
Generating synthetic training data, that sort of thing.

**SPEAKER_2** (2:51)
Exactly. Generating data, building better evaluation suites, and automating the ML ops. Every leading lab is actively pushing this forward right now.

**DailyNews** (2:59)
But it raises a critical question, I think.
Why is software coding the absolute center of this race? Like, why aren't we seeing this rapid acceleration in, say, chemistry or mechanical engineering?

**SPEAKER_2** (3:13)
Well, the physical world has latency. I mean, think about it. If an AI designs a novel protein structure for a new drug...

**DailyNews** (3:20)
You still have to actually make it.

**SPEAKER_2** (3:21)
Right. You have to synthesize it in a lab, culture it, and run it through physical trials. That feedback loop takes weeks, sometimes months.

**DailyNews** (3:30)
But software doesn't have that problem.

**SPEAKER_2** (3:31)
Not at all. Software structurally bypasses the physical world. In a coding environment, an AI can write a piece of code, trigger a compiler, run it against a continuous integration test suite, and read the stack trace of a failure.

**DailyNews** (3:44)
All in the digital realm.

**SPEAKER_2** (3:45)
Exactly. And then it can rewrite the code to fix the bug. That entire feedback loop, hypothesis, execution, failure, correction, it closes in milliseconds.

**DailyNews** (3:54)
Okay, let's unpack this. Because if you think about biological evolution, it takes millions of years of trial and error to optimize an organism for its environment.

**SPEAKER_2** (4:02)
Right. Because physical reproduction is incredibly slow.

**DailyNews** (4:05)
Exactly. So what we are looking at with these AI coding agents is essentially a genetic algorithm trapped in a hyperbaric chamber.