A proposed capability stack for understanding and measuring progress toward truly versatile machine intelligence
In September 2026, Executive Order 14434 directed federal agencies to use "Super Intelligence" and "SI" in place of "Artificial Intelligence" and "AI" in official, non-statutory communications. The order states that Super Intelligence carries the same meaning as artificial intelligence under the existing statutory definition at 15 U.S.C. 9401(3). This model takes no position on the terminology. It exists to answer the question underneath it.
The stated rationale is that frontier systems now do more than imitate or automate discrete aspects of human intelligence, and that a stronger word better reflects their promise and pace.
Researchers have used "superintelligence" for decades to describe hypothetical systems exceeding human performance across virtually all domains, so the usages are not identical and the overlap is contested.
A label is not a capability claim until it names a domain. This stack names four, and only one of them is close to finished.
Whatever the technology is called, the spatial gap does not move. A system that writes a dissertation and cannot reliably put a cup on a shelf has an uneven profile, and the profile is the thing worth measuring.
| Risk Domain | Why Spatial Capability Changes Everything |
|---|---|
| Autonomy | Spatially capable systems can act without humans typing commands |
| Irreversibility | Physical actions cannot be undone like deleting text |
| Verification | Harder to audit actions in the physical world than text outputs |
| Containment | A robot can walk out of a lab; a language model cannot |
| Deception | Physical manipulation opens new deception surfaces |
| Dual-use | The same capability that builds can destroy |
| Attribution | When behavior is learned rather than coded, no document records who set the limit |
Author's judgment, not a benchmark. The marker on each bar shows the 2025 estimate, and the fill shows the 2026 revision.
Incremental. The layer that was nearly finished is still nearly finished.
The year's real movement. Tool use, long-horizon task execution, and multi-step orchestration all shipped.
Robot foundation models became purchasable rather than experimental. Deployment still trails production, and teleoperation still fills the gap.
Persistent memory improved. Persistent identity did not.
What gets called "AGI" today is advanced linguistic intelligence with a growing agentic layer on top. The spatial gap is the chasm between a conversational system and a versatile one, and renaming the field does not narrow it.
Six questions that separate a demonstration from a deployment, applicable to any announcement in the spatial layer.
A demonstration assisted by a remote operator measures human spatial intelligence, not machine. The percentage is the claim.
First pilots are marketing. A second site from a buyer who already ran the first one is evidence.
A more capable model that still needs months of per-task setup has not changed the economics of anything.
When a behavior emerges from training rather than from a rule, nobody can point to the decision. That is a governance problem before it is a safety one.
Fixtured cells, known geometry, and a stack built to a spec move the hard part onto the floor plan. Picking an unfamiliar item out of racking somebody else built is the spatial problem itself.
A demonstration ends when the camera stops. A deployment leaves three marks: somebody downstream who quit verifying its work, a machine from another vendor accepting its output, and at least one thing it reported that nobody asked it to look for. Expect them to fail in that order, because nothing gets handed another vendor’s work before it is trusted, and nothing notices the unasked thing from inside a lane built for it.