Why a Modality-Agnostic Representation Doesn't Inherit a Single-Vector Ceiling
In August 2025, Google DeepMind researchers proved that single fixed-length vector embeddings, the architecture behind most modern AI retrieval, have a hard mathematical ceiling on what they can represent, regardless of model size or training data. This paper traces exactly what that proof does and doesn't mean for a representation system that was never built on single-vector retrieval in the first place, and draws a careful line between what's patent-disclosed, what's current architecture doctrine, and what remains an open claim.
In August 2025, Google DeepMind researchers published a proof that single fixed-length vector embeddings, the dense-retrieval architecture underneath most modern AI search and memory systems, have a hard, provable ceiling: for any embedding dimension, there exists some combination of documents no query can retrieve, regardless of training data, model scale, or how the vectors are optimized. The authors built a stress-test dataset, LIMIT, specifically to surface the failure, and state-of-the-art embedding models failed on tasks simple enough that a human wouldn't think twice about them.
This paper argues that a discrete, compositional representation, MindAptiv's Meaning Coordinates, isn't subject to that specific proof, because it was never performing the operation the proof is about: comparing one fixed-length vector to another by dot product. That is a real, narrow, defensible architectural distinction. It is not the same as the further claim that the same representation spans the full electromagnetic spectrum and classical-to-quantum compute substrates, which rests on separate, current architecture documentation rather than on the patent family this paper also examines. Keeping those two claims attributed to their actual sources, rather than letting one borrow the other's evidence, is the paper's central discipline.
It's worth being exact about what DeepMind proved, because the temptation on every side of this conversation is to round it up or down. The result applies to a specific computational operation: a single, fixed-length, real-valued vector, compared to another such vector by dot product or cosine similarity. The number of distinct top-k document subsets that operation can express is bounded by the dimension of the vector space, provably, regardless of training data, regardless of model scale, regardless of how the vectors are optimized. The paper shows the ceiling holds even when embeddings are directly optimized against the test set itself.
That is a proof about an operation, not about meaning, retrieval, or intelligence in general. Systems that don't perform that operation aren't inside the proof's scope. That isn't a rhetorical dodge; it's what the mathematics actually covers. The paper's own authors frame it the same way, calling not for abandoning retrieval but for architectures that don't inherit the single-vector paradigm's specific constraint.
Meaning Coordinates, the semantic layer underneath both the Nebulo® data engine and the PowerAptiv execution substrate, take a different starting position than the single-vector paradigm does. They don't ask how to compress a signal into a fixed-length code. They ask what intent the signal expresses, regardless of which modality carried it. Relevance is worked out by following relational structure, is-a, has-a, as-a, stored as formula graphs, not by measuring distance between two points in a vector space.
That's a genuinely different computational object, and the difference is what does the actual work here, not the difference in scale or ambition.
| Property | Single-Vector Retrieval | Meaning Coordinates |
|---|---|---|
| Representation | Compresses meaning into one fixed-length vector | Discrete, compositional primitives, not one vector |
| Relevance operation | Dot product / cosine similarity between two points | Relational graph traversal (is-a / has-a / as-a) |
| Capacity ceiling | Bounded by embedding dimension, provably | No single-vector dot-product operation in the path |
| Relation to DeepMind's proof | Directly, provably bounded by it | Outside the proof's scope, a different question applies |
It would be a mistake to source the electromagnetic-spectrum and quantum-substrate claim specifically from MindAptiv's earlier signal-processing patents (US 10,037,592 and its continuations, all sharing a June 5, 2015 priority date). Those specific terms, electromagnetic spectrum, quantum units, don't appear in that filing. Reading them directly out of it would be an extrapolation the text doesn't support.
That said, the patents disclose more than images, video, and audio alone, and it would understate them to say otherwise. US 10,037,592's own title names its scope as “…for images and other data types,” and its specification, shared near-verbatim across all three patents since continuations can't introduce new subject matter, states the field of invention as digital signal processing of static images, moving images, and other data types; performance optimization of multi-processor systems; and natural language interfaces. All of that carries the 2015 priority date, including the two later continuations, whose 2020 and 2022 grant dates describe when they issued, not what they're entitled to claim priority to.
So there are two separate, correctly-sourced claims here, not one. "Data broadly, multi-processor systems, natural language interfaces" is patent-disclosed and 2015-dated. "Full electromagnetic spectrum, classical-to-quantum substrates" is current PowerAptivs architecture documentation, specifically Principle 01: intent must be representable as a semiotic system that maps any signal of human intent to its governing primitives, regardless of the modality through which that intent arrives, across the full electromagnetic spectrum. The second claim doesn't need to borrow authority from the first, and the first is stronger than treating it as merely an image-processing patent would suggest.
The pattern connecting DeepMind's result to this broader architecture claim is the same at every scale, and it's worth stating precisely so it doesn't get inflated in the retelling. A system built around one representational commitment inherits that commitment's specific limits. A system built to be agnostic to that commitment from the start does not. Single-vector retrieval commits to compressing everything into one fixed-length code; it inherits a proven ceiling on what that code can express. A modality-agnostic semantic primitive system commits to representing intent independent of the signal that carried it; it inherits a different question, how far that primitive system's own composability actually reaches, which is an open, ongoing engineering question, not a solved one, and not the same question DeepMind answered.
MindAptiv has demonstrated Meaning Coordinates operating across images, video, audio, and structured data through the Nebulo® and Essence® platforms. Radar, RF, spectroscopic, and other non-visible-spectrum applications would use the same primitive system by architectural design, that is a claim about the representation layer's scope, not a claim that every one of those applications has shipped. The distinction matters, and it's the difference between the two scenarios below.
A discrete, relational representation isn't subject to the specific sign-rank bound DeepMind proved, because it was never comparing one fixed-length vector to another. That's a real, defensible, architectural distinction, not a claim to have resolved an open research question. The further claim, that the same representation spans the full electromagnetic spectrum and classical-to-quantum compute, rests on separate, current architecture documentation, not on the 2015-priority patents examined here, even though those patents disclose more (data broadly, multi-processor systems, natural language interfaces) than a narrow reading would suggest. Keeping each claim attributed to the evidence that actually supports it is not a hedge. It's what makes the claim survive contact with a careful reader.
Request Platform Access → Full White Paper Series