K3D represents the architectural convergence that Tim Berners-Lee's Semantic Web promised and that centralized AI has failed to deliver—a standards-aligned, spatially-grounded knowledge representation that enables truly sovereign cognitive systems.
The web's original architect envisioned machines sharing meaning across a decentralized graph. That vision stalled not because it was wrong, but because it demanded explicit semantic markup from publishers who had no incentive to provide it. AI "solved" this problem through extraction rather than cooperation—and in doing so, created the very centralization, opacity, and data exploitation Berners-Lee now warns against. K3D proposes a third path: spatial knowledge representation that makes semantics implicit in geometry, enabling machine understanding without requiring centralized training or user data harvesting.
Tim Berners-Lee introduced the "Giant Global Graph" concept on November 21, 2007, describing a three-phase evolution: from interconnecting computers (the Internet), to interconnecting documents (the Web), to interconnecting the things documents are about (the GGG). His reflection was pointed: "The Semantic Web maybe should have been called the Giant Global Graph."
The distinction matters profoundly for AI. The original 2001 Scientific American vision—authored by Berners-Lee, James Hendler, and Ora Lassila—promised intelligent agents that could "carry out sophisticated tasks for users" without requiring "artificial intelligence on the scale of HAL or C-3PO." The mechanism was explicit: machine-readable meaning encoded in RDF triples and ontologies, enabling inference and reasoning across a decentralized web.
What we got instead was centralization masquerading as intelligence. As Berners-Lee himself acknowledged, AI companies achieved the machine-readable internet "through extraction rather than cooperation." The irony is precise: LLMs accomplish semantic understanding by scraping the entire web into centralized training sets, raising exactly the data sovereignty concerns Berners-Lee now addresses through the Solid Project.
K3D's spatial approach offers resolution. Rather than requiring publishers to annotate content with RDF (the adoption barrier that stalled the Semantic Web) or extracting meaning into centralized models (the privacy violation that concerns Berners-Lee), spatial knowledge representation makes semantics intrinsic to structure. Knowledge entities have positions, orientations, and relationships defined by their geometric configuration—no external annotation required, no centralized training necessary.
Berners-Lee's September 2025 Guardian article crystallizes the stakes: "We have learned from social media that power rests with the monopolies who control and harvest personal data. We can't let the same thing happen with AI." His solution—Solid's personal data pods with fine-grained access control—addresses data sovereignty but leaves a critical gap: how do AI systems reason over decentralized data without centralizing it for training?
Current approaches fail this test. Cloud-based LLMs require data transmission to external servers. On-device models like llama.cpp provide inference privacy but still depend on centralized pre-training. Federated learning distributes computation but aggregates gradients centrally. None enable genuine cognitive sovereignty—the ability to learn, reason, and adapt using only locally-controlled data and computation.
K3D's architecture addresses this gap through three mechanisms aligned with Solid's principles:
This isn't merely technical alignment—it's philosophical convergence. Berners-Lee's call for "personal AI that works for you like your doctor or your lawyer, bound by law, regulation and codes of conduct" requires an architecture where reasoning happens within the user's control boundary, not extracted to external systems. Spatial knowledge representation enables this by making cognition a function of local geometric structure rather than remote model weights.
The W3C AI Knowledge Representation Community Group, launched July 3, 2018, defines its mission as exploring "the requirements, best practices and implementation options for the conceptualization and specification of domain knowledge in AI." This scope explicitly includes spatial-temporal reasoning as a form of Meta KR—one of five knowledge representation categories the group has identified alongside heuristic, procedural, declarative, and structural approaches.
The group's stated deliverables for 2025 include publishing a "concept map of the domain" and "natural language vocabulary to represent various aspects of AI," with a long-term goal of developing "a web standard for Neuro-symbolic Integration." Their TPAC 2025 discussions centered on "explicit, shared knowledge representation standards" for explainable and trustworthy AI systems. K3D's spatial approach directly addresses these objectives:
| W3C AI KR Focus Area | K3D Alignment |
|---|---|
| Neuro-symbolic integration | Geometric primitives bridge continuous representations (vectors) with discrete structures (graphs) |
| Explainable AI | Spatial relationships provide interpretable reasoning traces |
| Knowledge exchange and reuse | glTF-based format enables interoperability with existing 3D ecosystems |
| Support for AI agents | Spatial grounding addresses embodied cognition requirements |
The W3C provides a clear pathway from Community Group incubation to formal standardization. JSON-LD—the semantic web technology most directly relevant to K3D—successfully transitioned from Community Group specification to Working Group recommendation. The AI KR CG's stated trajectory toward "eventually transition to a formal Working Group" creates exactly the standards track appropriate for novel knowledge representation architectures.
What the group considers valid contributions: Documents, specifications, test suites, tutorials, demos, code, concept maps, and vocabulary development. Participation requires only a W3C account (free) and signing the Community Contributor License Agreement. The group explicitly welcomes research notes exploring implementation options—precisely the category of contribution K3D represents.
The path from "unknown architecture" to "industry standard" is well-documented. Georgi Gerganov's llama.cpp—now at 85,000+ GitHub stars with 700+ contributors—began as an independent developer's side project in Bulgaria. The credibility pattern is instructive:
Stage 1: Solve an unmet need. llama.cpp enabled LLM inference on consumer hardware without GPUs when no alternative existed. K3D addresses an equally unmet need: sovereign cognitive systems that reason over decentralized data without centralized training.
Stage 2: Open development. llama.cpp's MIT license, pure C/C++ implementation, and zero dependencies enabled explosive organic adoption. George Hotz's tinygrad similarly gained credibility through live-streamed development and radical simplicity (under 10,000 lines of code). Transparency is non-negotiable for novel architectures.
Stage 3: Enable ecosystem integration. llama.cpp became infrastructure for Ollama, LM Studio, GPT4All, and jan. GGUF format succeeded by being "opinionated about one thing (efficient local inference) while remaining flexible about everything else." K3D's glTF integration follows this pattern—leveraging an established 3D format ecosystem rather than requiring new infrastructure.
Stage 4: Benchmark validation. The ARC-AGI benchmark—created by Keras author François Chollet—has become what its creators call "the most important unsolved AI benchmark in the world" because it measures novel problem-solving rather than pattern matching. For cognitive architectures specifically, the credibility path runs through theoretical grounding (as ACT-R's 50+ years of development at CMU demonstrate) and practical application (as SOAR's military training systems validate).
The OpenCog cautionary tale illuminates what doesn't work: predictions without delivery (Ben Goertzel's unfulfilled 2011 prediction of AGI by 2021), PR stunts without substance (Sophia robot criticized as "complete bullshit" by Yann LeCun), and symbolic approaches positioned against dominant paradigms without empirical validation.
Current AI sovereignty initiatives—India's Project Indust, Denmark's Gefion supercomputer, Singapore's SEA-LION—represent national infrastructure investments, not architectural alternatives. They reduce geopolitical dependency on US providers while preserving technical dependency on the same centralized training paradigm. The sovereign cloud market may reach $169 billion by 2028, but sovereign clouds running replicated architectures don't produce sovereign cognition.
The edge AI market (projected at $66.47 billion by 2030) demonstrates both capability and limitation. On-device inference works: smartphones hold 80.5% market share in edge AI hardware, and NPUs enable complex model inference on mobile platforms. But as MIT Media Lab research identifies, five technical challenges block truly decentralized AI: privacy, verifiability, incentives, orchestration, and user experience in distributed contexts.
Current LLMs face a fundamental architectural barrier to sovereignty. As SAGE Journals analysis notes, LLMs demonstrate "behavior discrepancies between LLM inference and human reasoning, insufficient grounding, and hallucination." The root cause is architectural: pattern matching over statistical distributions doesn't produce genuine reasoning, world models, or metacognition. Local inference provides privacy; it doesn't provide cognitive capability independent of centralized pre-training.
K3D proposes an architectural alternative. Spatial knowledge representation grounds cognition in geometric structure rather than statistical distributions. Knowledge acquisition happens through spatial placement rather than gradient descent. Reasoning traces are interpretable paths through geometric space rather than attention weight matrices. This isn't an incremental improvement to existing architectures—it's a different computational substrate for cognition.
Andrew Tanenbaum's January 29, 1992 dismissal of Linux as "a giant step back into the 1970s" and "too closely tied to the x86 line of processors to be of any use in the future" exemplifies how established experts evaluate innovations using criteria from existing paradigms. Jamie Dimon's September 2017 declaration that Bitcoin was "a fraud" (followed by his January 2018 acknowledgment that "the blockchain is real") illustrates how even sophisticated critics can reverse positions when paradigm shift evidence accumulates.
Clifford Stoll's infamous February 1995 Newsweek article "The Internet? Bah!" dismissed online databases, telecommuting, electronic commerce, and interactive libraries—every prediction wrong. His later reflection deserves quotation: "Of my many mistakes, flubs, and howlers, few have been as public as my 1995 howler. Wrong? Yep... Now, whenever I think I know what's happening, I temper my thoughts: Might be wrong, Cliff…"
The pattern recognition is robust across domains:
The Semantic Web itself faced this gatekeeping. Cory Doctorow's 2001 "Metacrap" essay called it "a pipe-dream, founded on self-delusion, nerd hubris, and hysterically inflated market opportunities." Aaron Swartz blamed "the formalizing mindset of mathematics and the institutional structure of academics." Yet JSON-LD, Schema.org, and Google's Knowledge Graph—all Semantic Web descendants—now structure how billions of web pages communicate meaning to machines.
"This doesn't make sense" is the predictable initial response to paradigm-shifting architectures. Novel approaches require building new mental models rather than extending existing ones. Spatial knowledge representation violates the implicit assumption that cognition must be either symbolic (logic-based) or connectionist (neural network-based). The concept that geometric structure itself can encode semantic relationships and support reasoning requires cognitive reframing—exactly as the concept that packets could replace circuits required reframing for telecommunications engineers encountering the internet.
"This doesn't fit our scope" reflects categorical thinking that novel approaches intentionally transgress. The W3C AI KR Community Group scope explicitly includes "Meta KR: types of knowledge and logical reasoning" with spatial-temporal reasoning listed as an example. glTF's extension mechanism exists precisely to accommodate novel capabilities—KHR_xmp_json_ld brings JSON-LD semantic web integration into 3D formats, demonstrating that "unexpected" combinations are how standards evolve.
"Where's the working demo?" identifies the legitimate bootstrap challenge facing all novel architectures. llama.cpp's credibility required whisper.cpp's prior success. The demo-before-recognition pattern creates a chicken-and-egg problem that independent innovators resolve through focused proof-of-concept implementations rather than comprehensive systems. The appropriate response isn't "build everything first"—it's targeted demonstrations that validate core architectural claims.
"No major institution backs this" applies to every paradigm shift at inception. Linux was Torvalds' spare-time project. Bitcoin emerged pseudonymously. The World Wide Web was, in Berners-Lee's boss's words, "vague but exciting"—never an official CERN project. Independent researchers from Katalin Karikó (mRNA vaccines, facing "rejection after rejection, the scorn of colleagues, and even the threat of deportation") to Barbara McClintock (jumping genes, waiting 30 years for recognition) demonstrate that institutional validation follows demonstration, not precedes it.
The Khronos Group's glTF extension process offers a clear integration path. Extensions progress through three tiers: vendor extensions (any company can request a prefix via GitHub issue), multi-vendor extensions (EXT_ prefix when multiple implementations exist), and Khronos-ratified extensions (KHR_ prefix, voted by Board of Promoters). The OMI Group pathway provides an alternative route—extensions developed through the W3C Metaverse Interoperability Community Group can graduate to Khronos submission, as KHR_audio_emitter successfully demonstrated.
Existing semantic extensions establish precedent for K3D integration:
KHR_xmp_json_ld (provisional) adds JSON-LD compliance to glTF for product metadata, directly leveraging Semantic Web standards within the 3D format ecosystem. This extension demonstrates that Linked Data integration into 3D standards is not merely theoretical but actively implemented.
EXT_structural_metadata defines schema-based structured metadata with property tables, attributes, and textures—enabling semantic identifiers for interpretation. This extension, developed for Cesium's 3D Tiles, proves that complex metadata schemas integrate naturally with glTF's architecture.
NNEF (Neural Network Exchange Format) represents Khronos's existing AI/ML standard—a "PDF for neural networks" that encapsulates complete network descriptions independent of training tools. This precedent demonstrates Khronos's willingness to standardize AI-related formats.
WebGPU's ML capabilities (compute shaders, FP16 support, direct GPU access) enable in-browser neural network inference at near-native performance. WebLLM, ONNX Runtime Web, and TensorFlow.js all leverage WebGPU for client-side AI. K3D's spatial reasoning can utilize this same acceleration pathway.
WebXR's spatial primitives (XRSpace, XRReferenceSpace, XRPose) provide the coordinate system abstractions that anchor knowledge entities in physical or virtual space. The technical foundation for spatial knowledge representation already exists in W3C specifications.
K3D emerges at the intersection of multiple mature standards and urgent industry needs:
Berners-Lee's vision alignment: The Giant Global Graph concept (2007) described exactly what spatial knowledge representation provides—interconnecting the things documents are about rather than just documents. Solid's data sovereignty principles (2016-present) require cognitive architectures that reason locally without centralized training. K3D delivers both.
W3C standards integration: JSON-LD (W3C Recommendation), WebXR (W3C specification), WebGPU (W3C standard), and the AI KR Community Group's focus on "knowledge representation for AI" create a standards ecosystem ready for spatial knowledge representation. The pathway from Community Group incubation to formal recommendation is documented and precedented.
Khronos ecosystem leverage: glTF's extension mechanism, existing semantic extensions (KHR_xmp_json_ld, EXT_structural_metadata), and the OMI Group's community-driven development process provide technical and procedural pathways for K3D integration. NNEF demonstrates Khronos's AI standardization precedent.
Industry need: The sovereign AI market ($169B projected by 2028), edge AI expansion ($66B by 2030), and growing critiques of centralized AI dependency create demand for architectural alternatives. Enterprise concerns about data exposure (69% cite AI-powered leaks as top security concern), regulatory conflicts (US CLOUD Act vs. GDPR), and service discontinuation risks validate the need for sovereign cognitive systems.
Credibility pathway: The llama.cpp pattern—solve unmet need, open development, ecosystem integration, benchmark validation—provides a tested route from novel architecture to industry adoption. K3D's alignment with existing standards accelerates this path by reducing integration friction.
The web began as one physicist's "vague but exciting" proposal at CERN. The Semantic Web emerged from that same physicist's recognition that documents weren't enough—we needed to interconnect what documents meant. Now, as AI threatens to centralize exactly the knowledge flows the web was designed to distribute, the architectural answer may be what Berners-Lee intuited but couldn't implement: a Giant Global Graph where meaning is spatial, sovereignty is architectural, and cognition happens at the edge.
K3D proposes to build it.
W3C Specifications
Khronos Standards
Key Sources