#74 | AGI by 2030 — What if, actually?

TL;DR: Demis Hassabis confirmed at Google I/O that the AGI timeline is still on track: 2030, plus or minus a year, confidence interval narrowed. Isomorphic Labs already has drug compounds in the preclinical stage.

👋 Hello,

Fifteen years is long enough to know whether a plan is holding.

Shane Legg, DeepMind’s co-founder, put the date in writing in 2010: AGI around 2028 to 2030. A twenty-year mission with the destination already marked from the first day.

At Google I/O, Demis Hassabis, DeepMind’s CEO, sat down for ​this interview​ and was asked how those timelines had moved. His answer was briefer than the question deserved.

“I think we’re on track. And I would say my confidence interval has narrowed. These days I’m thinking it’s 2030, plus or minus a year.”

Two years ago, Demis was still saying five to ten years.

Alongside the AGI timeline, the drug discovery timelines hardened too. Both are moving in the same direction, on the same schedule.

In this edition: You’ll know what Demis Hassabis actually said about AGI timelines and the specific gaps still separating us from it, where Isomorphic Labs stands in drug discovery right now, which human skills hold their value as the tools get stronger, and what any of this means for how you work, whether 2030 holds or not.

What a jagged intelligence still can’t do

Demis had a name for current AI systems: “jagged intelligences.” Remarkably capable in certain areas, genuinely unreliable in others, sometimes within the same task.

• Physics modeling is still imperfect, and memory is brute force in his description.
• Consistency and continual learning are research problems that haven’t been cracked yet.
• Long-term reasoning and planning sit on that list, too.

Each gap he named has active research behind it at DeepMind.

And those gaps have daily implications you will have become aware of.

• Every session starts from zero.
• Delegation still requires oversight.
• The parts of your work that require memory and judgment over time remain yours.

On the agentic side, systems are genuinely useful now. Reliable enough for specific tasks, but not yet reliable enough to fully hand off.

Drugs in months, not a decade

​Isomorphic Labs​, the DeepMind spinout working on AI drug discovery, already has test compounds in the preclinical stage. Oncology and immunology are first, because the clinical pathways for serious diseases move faster.

Compressing drug discovery from ten years to months is the ambition. Not the clinical trial stage, which still runs for five to ten years. Just the part that goes from a disease profile to a candidate compound worth testing.

“Think of it as AlphaFold… maybe half a dozen to a dozen other AlphaFold-level breakthroughs, all integrated into one continuous platform that goes from the target of interest to a lead compound.”

​AlphaFold​ predicted the 3D structure of essentially every known protein. Half a dozen breakthroughs at that scale, integrated into one continuous pipeline.

“This is the beginning of them becoming real, real cures.”

Hassabis said the first successful compound would be the biggest watershed moment in AI. Bigger than the AGI milestone itself. That is unusual for a founder to say about their flagship goal.

What you can’t hand to the tool

Somewhere near the end of the interview, the question finally shifted to human skills: which ones get more valuable as AI gets more capable?

“What separates the great scientist from the good scientist is not their technical capability. It’s that creative capability, that sort of taste and judgment.”

I’m sure he will be right, I hope so :). Taste is the one thing you cannot outsource without noticing immediately.

Taste is the one thing you cannot outsource without noticing immediately.

Taste, design sensibility, original thinking, and the ability to connect emotionally with an audience.

He was specific about one combination: creative and technical together. Bring both, he said, and you’re in an “amazing position.”

But we could also ask this: whether humans will actively build taste or quietly let the tools fill that space instead. Because taste is a matter of opinion, isn’t it?

Whether 2030 holds or not

Four years from now is not a long time.

Assume the timeline holds. AGI arrives around 2030. Delegation becomes reliable. What matters to our, your, or my careers shifts entirely to what you bring:

• taste, judgment, emotional resonance, and the combination of creative and technical thinking.

Call it the other scenario in which the timeline slips. Thus, memory stays brute force. Consistency remains unsolved. And human judgment still fills every gap the tools can’t cross.

Sounds better?

Well, both scenarios lead to the same preparation.

Waymo, the self-driving car service, came up as his example of how fast normalization happens. First autonomous ride: unsettling, gripping the door. Second ride: completely normal. “It’s weird that it’s normal,” he said. Each new capability arrives disorienting and leaves routine.

Whether you adapt ahead of that curve or behind it is the only question that matters. His point: the deeper you understand the technology, the better you (can) take advantage of it.

The part nobody followed up on

Near the end of the talk, with maybe ninety seconds left, Demis also mentioned cybersecurity. Something careful.

“I do worry about… the cyber worries about some of the models. And I think that’s just the beginning of some of the issues we need to guard against, as well as getting excited about all the amazing opportunities.

Maybe this is the time to try and push some standards and maybe international cooperation.”

Standards and international cooperation. Said once, at the end of an interview, about AGI arriving in four years.

He narrowed his confidence interval to one year either side of 2030. He’s also saying the conversation about standards hasn’t started yet.

That combination is a cause for concern, in my opinion. It means it won’t happen. International cooperation where it would matter, haha.

Cheers,

Mark
The AI Learning Guy​
👋⚡😎

Interesting Sources

• ​Demis Hassabis interview at Google I/O​ — YouTube
• ​Isomorphic Labs​ — Isomorphic Labs
• ​AlphaFold​ — Google DeepMind
• ​Google DeepMind​ — Google DeepMind
• ​Shane Legg on long-term AGI probability​ — LessWrong

Note: No single website has all the answers. This list serves as a starting point for those who want to explore or satisfy their curiosity about AI.
Links: Links with * are affiliate links. See disclosure below.