Why AI wants names, not translators

There is an engineering case against gluing systems together with translators: they multiply, they break, they cost more every year. But there is a deeper reason a world of translators fails the thing we are now building for, which is AI. It is not that translators are inefficient. It is that they are illegible to a reasoning system in a way a shared name is not.

Reasoning is about things, not observations

An intelligent system does not think in pixels or point clouds. It thinks in nouns: this door, that vehicle, the crack in this beam. Language models are the clearest example, they reason by referring to stable entities and asking questions about them. The whole power of the approach is that "the bridge" can carry everything known about the bridge, and a chain of reasoning can be built on top of that reference.

A translator gives an AI the opposite of a stable noun. It hands over a correspondence that is true only at this instant, between this pair of systems: "object 47 in system A is object 91 in system B, right now." Nothing about that survives to the next frame, the next system, or the next day. The AI cannot build a belief about a thing on a fact that expires the moment it is stated.

A translator gives an AI a coincidence. A name gives it a subject it can reason about.

Four things an AI can do with a name that it cannot do with a translator

Refer

Point at one specific thing, unambiguously

With a shared name, an agent can say "this window" and mean exactly the entity every other system means, no bounding box, no coordinate guess, no per-system id to reconcile. Reference is the atom of reasoning, and a name is what makes reference stable. Through a translator, "this window" is only ever "whatever my current mapping says it is," which is not something you can safely act on.

Retrieve

Pull just what it needs, not the whole scene

Ask a name for its context and you get a small, bounded description, class, pose, affordances, relations, provenance, that fits in a model's working memory. That is spatial retrieval built for reasoning. A translator layer offers no such handle; to learn about a thing an AI would have to ingest and re-derive whole scenes, which is both expensive and lossy. Names make the world queryable at the granularity of a thought.

Remember

Reason across time, because the subject persists

Because a name is the same across observations, an agent can ask "what changed here since last time," "what has this asset done," "is this the same thing we saw yesterday." Those questions require a persistent subject to attach memory to. Translators, being momentary, erase the thread the moment it forms; every encounter starts from zero. Names are what let an AI have a memory of the world rather than a series of unrelated glances at it.

Justify

Explain why it believes what it believes

A name can carry its own provenance: which sensors saw it, with what confidence, fused by what model, confirmed or not. So an agent can answer not just "what is here" but "why do we believe it, and how sure are we." A translator carries none of that; it converts a value and forgets. For AI that has to be trusted, and increasingly has to act, the ability to justify a belief is not a luxury, and it lives on the name.

Translators are illegible; names are a substrate

Consider what an AI agent actually faces in a world stitched together with translators. To reason about a single object across three systems, it must know which translators exist, trust that each is current, understand each one's quirks, and stitch the fragments itself, and redo all of it when any system changes. The reasoning is drowned in bookkeeping about the plumbing. Every translator is a place the AI can be silently wrong.

A shared name removes the bookkeeping. The agent refers to the thing, and every system that speaks the name agrees. The reasoning is about the world, not about the connectors between views of the world. This is the same reason the web became the substrate for software agents: not because HTTP was clever, but because a URL is a stable name an agent can hold, pass, and act on without negotiating with every server first.

The multi-agent stakes

It gets sharper when more than one agent is involved, which is where everything is heading. Two agents can only collaborate on a thing if they can name the same thing. Hand a task from one agent to another, an inspection to a scheduler, a target to a planner, a property to a financing agent, and the handoff is only meaningful if the referent survives it. Through translators, every agent-to-agent handoff needs its own reconciliation, and the errors compound with every hop. With a shared name, the handoff is just passing the name; both agents are provably talking about the same thing at the same version of the world.

Agents cannot cooperate on a world they cannot co-refer to. A shared name is the precondition for multi-agent anything.

The honest boundary

None of this makes the hard part disappear. Deciding that two observations are the same thing, the re-identification that lets an AI attach a new sighting to an existing name, is genuine work, and it is where the real research goes. A shared identity layer does not do that reasoning for the AI; it gives the reasoning a place to land. The alternative is not "no hard problem." It is the same hard problem plus a maze of translators on top, solved worse.

Why this is the moment

For most of computing's history, spatial data was consumed by rendering engines and analytics that did not need to reason, they needed to draw or to count, and translators were a tolerable tax. AI changes the requirement. A system that reasons and acts needs to refer, retrieve, remember, and justify, and all four of those run on stable identity and none of them survive a translator. The world is about to be operated on by agents, and agents want names.

4D-ID is that layer: one persistent name per thing, resolvable by every system and every agent, carrying its own history and provenance. Not a smarter translator between spatial systems, but the shared vocabulary that lets an AI think about the physical world the way it already thinks about text.

Watch an agent reason over named things →

← All writing