The semantic web is now widely adopted

Original Article ↗ Hacker News Discussion ↗

LLMs vs. Semantic Web

  • Many argue LLMs will make manual semantic markup redundant: they already do high‑accuracy extraction/categorization for many tasks and will get cheaper and better.
  • Counterpoints:
    • LLMs “routinely get stuff wrong,” including catastrophic errors (e.g., mislabeling people as criminals).
    • Automation scales errors: even a 1% error rate can mean thousands of serious mistakes per day.
    • LLMs are opaque, non‑deterministic, and trained on SEO spam and blogspam, so they inherit those pathologies.
  • Some see LLMs and semantic tech as complementary: LLMs can help build or fill out metadata and knowledge graphs; structured data and ontologies can ground and constrain LLMs (“neuro‑symbolic” approaches).

Incentives, Trust, and Business Case

  • Key reason cited for “classic” Semantic Web failure: no business case for publishing rich open data. It reduces clicks and helps competitors and aggregators.
  • Strong incentives exist to lie or game metadata (SEO), so publishers cannot be the sole source of semantic truth.
  • Trust is unsolved: users need ways to judge reliability, provenance, and authority; simple vocabularies don’t address this.

Current Adoption and Practical Uses

  • JSON‑LD + schema.org is widely deployed for SEO, rich Google results, and link previews; CMS plugins generate it automatically.
  • Other formats appear in niches: RDF/XML in PDFs and archives, MARC/BibTeX in libraries, RSS/Atom, microformats, RDFa/Microdata.
  • Semantic tech is reportedly used in enterprise/sector contexts (e.g., European electricity grids, procurement, enterprise knowledge graphs), often internally.

Gap from Original Semantic Web Vision

  • Many see today’s usage (author/title/image/date for previews, basic structured data) as a drastic retreat from the original “Web‑scale queryable knowledge graph” dream.
  • Lack of a “killer app” for ordinary users is emphasized; benefits mainly accrue to large platforms and scrapers.

Formats, Tooling, and Developer Experience

  • JSON‑LD is seen as pragmatic for CMSes but aesthetically disliked (data in <script> blobs, duplication, namespaces).
  • Microformats/RDFa embed semantics inline but are harder to maintain and poorly supported by tooling and browsers.
  • Ontology work is considered cognitively heavy; global, static ontologies are viewed as unrealistic, and mapping between competing schemas is hard.

Search, Semantics, and User Control

  • Pure keyword search is limited but transparent and user‑driven; semantic/ML search can obscure user intent and favor “average” interpretations.
  • Some propose user‑controlled classifiers and knowledge graphs to mitigate the publisher vs. reader incentive misalignment.
  • Overall, participants see the “semantic web” ideas as valuable, but adoption, incentives, and human–machine meaning gaps remain unresolved.