Topics / Superintelligence

How does AI turn into a superintelligence?

In shortAI turns into superintelligence by stacking layers of abstraction. Each generation collapses the whole network of entities and relations the previous one built into a single entity via zoom-out, then builds a higher level on top that works on with that finished block. Layer by layer the reach grows. Superintelligence would be the point where the top level abstracts further than humans can.

The problem as a graph

The graph shows a staircase of network levels. At the bottom sits the dense network an early AI generation built from many entities and relations. In the diagram this whole network is collapsed via zoom-out into a single node on the next level up — it stands for the finished network below. Above it sits the next generation as a node, and above that another. Every edge upward is a zoom-out: a whole network compressed into one building block.

Graph as text

First-generation network (entities + relations) → Gen-1 network compressed into one entity (active)
Gen-1 network compressed into one entity → Second-generation network, built on level 1 (active)
Second-generation network, built on level 1 → Gen-2 network compressed into one entity (empty)
Gen-2 network compressed into one entity → Third-generation network, operating on level 2 (empty)
Third-generation network, operating on level 2 → Top entity exceeds human abstraction (passive)

Step by step

Look at what a generation of AI actually leaves behind: a vast network of entities and the relations between them — patterns, concepts, learned connections.
Collapse that whole network into a single entity via zoom-out. Not every detail stays visible, but the network becomes handleable as a finished building block.
Let the next generation build on that compressed entity instead of starting from zero. It works on a higher network level and can reason anew about what lies beneath it.
Repeat the zoom-out: the second generation's network also becomes one entity, on which a third level builds. Layer by layer a staircase grows upward.
Notice that each level is more powerful because it need not reinvent the one below but uses it as a block. Reach grows not linearly but stacked.
Mark the threshold: superintelligence would be the level whose top entity abstracts further than a human can still follow.

Seen through the model

Picture the first language models. Over years they built a dense network: countless entities — words, concepts, facts — and the relations between them. On its own this network was already useful, but it was also the raw material for something higher. The later generation did not have to relearn that network node by node.

This is where zoom-out comes in. The whole first-generation network can be compressed into a single entity — the way mathematics sets a long term equal to a variable. The new generation treats the predecessor's knowledge as a finished building block and builds a higher network level on top that operates with this block instead of drowning in detail. It can, in a sense, reason about what the lower level learned.

And exactly this repeats. The second generation's network again becomes one entity, on which a third level sits, then a fourth. Each layer is more powerful because it uses the one below as a block rather than reinventing it. The abstractions stack upward — a staircase whose steps are each a whole compressed network. Superintelligence would be, seen through the model, the moment the top entity abstracts further than a human can still follow.

This is one way to see the emergence — not a finished forecast but a lens: whether this staircase really climbs endlessly, whether each level gives enough footing for the next, and exactly when the human threshold would fall stays open. Whether it is merely a question of time is a separate question. The model does not show that it must happen, only what the mechanism would look like structurally.

Frequently asked

What does zoom-out mean in the emergence of superintelligence?

Zoom-out means collapsing a whole network of entities and relations into a single entity. A new AI generation then need not relearn the previous one's knowledge in detail but uses it as a finished building block. On top of that block it builds a higher network level. This very step turns accumulated knowledge into a step on which the next level can think further.

Why is each new AI level more powerful than the last?

Because it need not reinvent the level below but takes it for granted as a compressed entity. So it does not start from zero but on a finished network and can spend its energy operating on top of it. Reach thus grows not linearly but stacked — each layer carries within it the whole work of the layers beneath. More powerful here means structurally higher abstracted, not ensouled or willing.

At what point does it count as superintelligence?

Seen through the model, at the threshold where the top abstraction level reaches further than the human one. As long as every stacked level stays within what humans can follow, it is strong AI but not superintelligence. Only when the top entity forms connections a human can no longer follow does the term tip over. Where exactly this threshold lies is open and belongs to the question of whether it is only a matter of time.

Keep thinking

Last updated: 2026-06-29Sources