Google DeepMind published a long essay this week. It asks what comes after AGI. AGI means AI that matches typical human skill at most tasks.
ASI is the term the essay reaches for next. ASI means AI stronger than large groups of human experts combined. DeepMind wants readers to picture that gap now.
Researcher Tim Genewein wrote the essay with eleven co-authors. Shane Legg helped write it too. Legg co-founded DeepMind years ago.
Legg shared the essay on X himself. His post landed 2,442 likes. It drew 76 replies too.
The essay names four paths past AGI. One path is simply scaling up today's models. Another path moves past the design most models share now.
A third path groups many AI copies into one team. The fourth lets a model improve its own design over time. DeepMind calls that last path recursive self-improvement.
That fourth path worries safety researchers the most. A model that redesigns itself could move faster than any team can check. DeepMind names that risk directly in the essay.
The essay runs 57 pages long. It also makes a bold cost claim. Work that costs $1 billion in compute today could cost $1,000 in ten years.
The piece reached Hacker News too. It earned 74 points there. Commenters left 134 replies digging into each path.
Why did this catch fire? A top lab rarely plots its own path past today's models this plainly. Readers want to know which path a lab like DeepMind expects to win.
The reach says something about where AI leaders now look. Engineers once argued mainly about this year's benchmark scores. Some of them now argue about what happens once benchmarks stop mattering.
Watch which path DeepMind backs with real spending. An essay costs little to publish. The next model Google builds will show which of the four paths it picked first.