Beyond AGI: Google DeepMind’s Roadmap to Superintelligence | 16th June 2026

**SPEAKER_1** (0:00)
Welcome to Colaberry AI Podcast, brought to you by Colaberry AI Research Labs and Karl Foundation.
So I want you to imagine, just for a second, that you're tasked with writing a highly complex engineering manual, right? Like for a revolutionary new interstellar spaceship.

**SPEAKER_2** (0:16)
Okay, a spaceship manual. Sounds intense.

**SPEAKER_1** (0:18)
Right. But before you publish it, you just slap a Post-It note right on the title page. And this note explicitly says, hey, autopilot system, when the human pilot asks you to read this manual to them, make sure you don't compress the critical definitions and be sure to critically judge the conclusions.
You are leaving instructions not for your human peers, but for the machine that's eventually going to mediate the knowledge. And that is exactly what Google DeepMind just did.

**SPEAKER_2** (0:44)
It's wild. It really is.

**SPEAKER_1** (0:45)
We are getting into a really dense, just fascinating 57-page paper today, titled From AGI to ASI.
And literally, right there on page one, section one, there is a block of text labeled summary instructions. And it's written explicitly to prompt the future AI models that you and I will eventually use to summarize this very paper.

**SPEAKER_2** (1:05)
Yeah, it's a brilliant, albeit slightly unnerving, structural choice. But it tells you everything about the reality these researchers are operating in right now.

**SPEAKER_1** (1:15)
Unnerving is the perfect word for it.

**SPEAKER_2** (1:17)
Right, because they are formally acknowledging that the primary consumer of high-level scientific literature is transitioning from biological to digital. Like, the AI is the reader now.

**SPEAKER_1** (1:29)
Yeah, and we really need to look at the pedigree of the authors here because it demands attention.

**SPEAKER_2** (1:34)
Oh, absolutely. I mean, we're looking at a roadmap authored by 14 of the top minds in artificial intelligence. You've got Shane Legge, the co-founder and chief AGI scientist at DeepMind, and Marcus Hutter.

**SPEAKER_1** (1:45)
Hutter is a big deal in this space.

**SPEAKER_2** (1:47)
A massive deal. For those who track theoretical computer science, Hutter is the architect of the AIXI theory, which is essentially the mathematical proof for universal artificial intelligence.
So this document we're looking at, it's not some speculative sci-fi philosophy piece.

**SPEAKER_1** (2:03)
No, not at all.

**SPEAKER_2** (2:04)
It is a rigorously engineered, deeply technical projection of the post-AGI landscape.

**SPEAKER_1** (2:10)
And that title alone, from AGI to ASI, it totally shifts the goalposts.
I feel like we spend so much time debating the timeline to artificial general intelligence, right? Like, when will it happen? But this deep dive is entirely about what happens the morning after we cross that line.

**SPEAKER_2** (2:26)
Exactly. The day after.

**SPEAKER_1** (2:28)
So to ground this for anyone listening who tracks the space, how exactly is DeepMind defining these thresholds? Because AGI, in this specific context, it isn't some omniscient super being, is it?

**SPEAKER_2** (2:40)
Far from it. No, the paper anchors AGI as a system, performing at roughly the median human level across cognitive tasks.

**SPEAKER_1** (2:47)
Median human level, just average.

**SPEAKER_2** (2:48)
Right, just a completely average worker. It means the underlying architecture can reason, it can execute robust planning, navigate novel environments, and interface with external tools via APIs, all at the competency of a typical average human. It's basically the baseline of general utility.

**SPEAKER_1** (3:04)
So it's not the smartest quantum physicist in the room, it's just a highly reliable average employee. Right. But then they introduce ASI, Artificial Superintelligence, and the metric they use to define that leap is just mathematically staggering to me.

**SPEAKER_2** (3:19)
It really is. The bar for ASI is set at an output equivalent to tens of thousands of the absolute top human experts.

**SPEAKER_1** (3:28)
Tens of thousands?

**SPEAKER_2** (3:29)
Yes. Perfectly coordinated, operating without a single ounce of friction on a single complex problem for a decade.
So we're not talking about a model that just beats a grand master at chess anymore. We are talking about an entity that can match the entire aggregate research output of a global scientific discipline.

**SPEAKER_1** (3:46)
That is just wow. And hanging above both of those concepts is Marcus Hudders AIXI, right? Yeah. Which the paper uses as this anchor point.

**SPEAKER_2** (3:53)
Yeah. AIXI is vital to understand here, because it acts as the theoretical speed of light for intelligence.

**SPEAKER_1** (3:58)
It's like an absolute limit.

**SPEAKER_2** (3:59)
Right. It's a mathematical model of an agent that perfectly optimizes its reward in any computable environment. Basically combines reinforcement learning with Solomonov induction. But the catch, the big caveat here is that AIXI is formally uncomputable.

**SPEAKER_1** (4:13)
Because it would take infinite resources.

**SPEAKER_2** (4:15)
Exactly. It requires infinite compute to search through all possible programs.