Ontology: The Semantic Operating System Powering Large‑Model AI

Re‑interpreting Sutton’s “Bitter Lesson” for knowledge engineering

Richard Sutton’s The Bitter Lesson argues that lasting progress comes from methods that scale with compute rather than from manually encoded expert knowledge. The discussion warns against reading this as “knowledge is irrelevant”; instead, it stresses moving from costly hand‑crafted encoding toward computable, evolvable semantic structures.

Ontology as a semantic operating system for AI

Participants likened modern ontology to a “semantic operating system” or “semantic compiler”. It provides a shared language that aligns business concepts (e.g., “customer”, “contract”, “risk”) with data and technology, turning them into machine‑understandable units that can be tracked and acted upon.

Core insights (six points)

Resolving internal ambiguity – Enterprises often have inconsistent definitions across departments; a stable semantic layer is required to align cognition.

Semantic foundation for models and agents – Without a shared ontology, large models cannot reliably operate in business contexts, leading to hallucinations and mismatches.

Static schemas are insufficient – Traditional data‑warehouse schemas are rigid; approaches such as LPG/SPG (Semantic‑enhanced Programmable Graph) can absorb dynamic business changes.

From description to action – Future ontologies must be “actionable”, connecting data, APIs, models, and agents so AI can understand, locate the correct interface, and execute tasks.

Ontologies as private competitive barriers – As generic models improve, companies need proprietary semantic structures to retain knowledge and capability.

Timing of semantic investment – Delaying the build of a semantic layer risks missing the ecosystem momentum window.

OpenKG practice and emerging knowledge infrastructure

OpenKG demonstrates a modern knowledge stack:

SPG (Semantic‑enhanced Programmable Graph) – Unifies data, semantics, rules, and actions into an evolvable graph, turning static ontologies into programmable semantic layers.

KAG (Knowledge‑Augmented Generation) – Couples structured knowledge with large‑model reasoning, using logical forms to guide retrieval, decomposition, and solving, thereby improving controllability and reducing RAG‑related hallucinations.

SkillNet – Transforms knowledge into reusable skill units, moving from “knowing what” to “knowing how” for complex task execution.

MemOS – Provides a memory operating system for long‑term learning, enabling agents to manage persistent memory and self‑update.

These components illustrate a shift from static ontologies to a dynamic, programmable semantic layer.

Organizational reality as the primary bottleneck

Three practical obstacles were identified:

Insufficient business drive – departments prioritize immediate data needs over long‑term semantic alignment.

High cross‑department coordination cost – data, responsibility, and profit boundaries hinder unified modeling.

Unclear ROI – semantic benefits are long‑term and hard to quantify, making budget approval difficult.

Additional factors include scarcity of talent who understand both business and AI, and the legacy paradigm of building monolithic ontologies.

Two paths for enterprise‑level agents

Current implementations fall into two camps:

Ontology‑first (exemplified by Palantir) – Builds a stable, auditable semantic skeleton before applying models.

Model‑first (exemplified by Claude‑related work) – Centers on a base model and attaches external knowledge via a Context Graph for faster deployment.

The most competitive agents are expected to fuse both: using ontologies for deterministic constraints and models for creative reasoning.

Why open‑source is the viable way forward

Because a semantic foundation must be shared to be effective, open‑source collaboration is essential. The following repositories illustrate the public layers being co‑developed:

https://github.com/OpenSPG/KAG-Thinker

https://github.com/MemTensor/MemOS

https://skillnet.openkg.cn

Open‑source enables a common semantic base, interchangeable engines, composable capabilities, and continuous evolution.

Technical stack references

Semantic engine: https://github.com/OpenSPG/KAG-Thinker Memory core: https://github.com/MemTensor/MemOS Skill repository:

https://skillnet.openkg.cn

Conclusion

Large models will continue to grow, but enterprises need more than conversational ability—they require a stable, executable semantic infrastructure. Ontology is re‑emerging as the AI‑era semantic operating system, to be built as code, data, standards, and industry‑specific contexts, ideally through open‑source collaboration.