Hand-built meaning or learned meaning

1. 1959
   proposes writing ★ seminal

   Programs with Common Sense

   The 'advice taker': a program that decides what to do by deducing it from facts and rules stated in formal logic. [src]

2. 1984
   proposes project ★ seminal

   Cyc

   Begun by Douglas Lenat at MCC in 1984 to assemble a comprehensive knowledge base of common sense, written largely by hand as logical axioms: about 24.5 million of them by 2017. [src]

3. 1985
   proposes project ★ seminal

   WordNet

   A lexical database of English in which words are grouped into sets of synonyms linked by relations such as broader and part-of; it became a shared vocabulary for language processing. [src1][src2]

4. Sep 1998
   proposes writing ★ seminal

   Semantic Web Road Map

   Leaving aside the AI problem of training machines to behave like people, the Semantic Web will develop languages for expressing information in machine-processable form, toward a Web in which machine reasoning is ubiquitous. [src]

5. 7 Nov 2003
   critiques writing ★ seminal

   The Semantic Web, Syllogism, and Worldview

   ↳ critiques Semantic Web (proposed by Semantic Web Road Map)

   Calls the Semantic Web a machine for creating syllogisms, which do not work on slippery, partial or context-dependent data. Because metadata describes a worldview, any attempt at a global ontology is doomed. [src]

6. Mar 2009
   critiques writing ★ seminal

   The Unreasonable Effectiveness of Data

   ↳ critiques Semantic Web (proposed by Semantic Web Road Map)

   Simple models trained on very large amounts of web data beat elaborate hand-built ones. Hand annotation is slow, expensive and hard for experts to agree on, while useful semantic relationships can be learned from data such as search queries. [src1][src2]

7. 16 May 2012
   workaround writing ★ seminal

   Introducing the Knowledge Graph: things, not strings

   'Things, not strings': Google's Knowledge Graph launches with more than 500 million objects and 3.5 billion facts drawn from Freebase, Wikipedia and the CIA World Factbook, used inside a statistical search engine. [src]

8. 16 Jan 2013
   proposes writing ★ seminal

   Efficient Estimation of Word Representations in Vector Space

   Meaning learned from which words occur together: word vectors in which relations such as king to queen appear as directions, with no one writing the relations down. [src]

9. Aug 2014
   proposes writing

   Knowledge Vault: A Web-Scale Approach to Probabilistic Knowledge Fusion

   A knowledge base built by machine: facts extracted across the Web and combined with existing knowledge bases, each with a probability of being true, rather than curated by people. [src]

10. 12 Jun 2017
    proposes writing ★ seminal

    Attention Is All You Need

    The transformer, built on attention, becomes the architecture of the large language models that learn language, and much of what it says about the world, from text alone. [src]

11. 24 Apr 2024
    proposes writing

    From Local to Global: A Graph RAG Approach to Query-Focused Summarization

    A language model derives an entity knowledge graph from source documents, then summarizes groups of related entities to answer questions about the whole collection. The graph is learned, not hand-built. [src]

Dated events that show the problem moving, newest last.

1. 1 Oct 2025

   Wikimedia Deutschland opens a public vector database of Wikidata for AI applications, reachable through MCP. Wikidata[src]

Under each question, the answers people have put forward or built. The catalogue gives no odds on them.

• Can knowledge graphs built by language models be trusted without people checking them? [src1][src2]
  GraphRAG and similar methods have a model write the graph; the 2023 roadmap lists LLM-built graphs as one of three ways to combine the two.
  • Graphs extracted by language models are used without people checking them. GraphRAG
    GraphRAG builds a knowledge graph from source texts with a language model and answers questions from it directly.[src]
  • Models draft knowledge and people or curated graphs check it. Unifying Large Language Models and Knowledge Graphs: A Roadmap
    The 2023 roadmap proposes 'LLM-augmented knowledge graphs', in which models help build and complete graphs, alongside graphs that ground models.[src]
• Is hand-written knowledge still worth its cost when statistical methods keep improving? [src1][src2]
  Cyc's knowledge base took decades of hand axiom-writing; the 2009 essay argued that data at scale does better.
  • Curated knowledge becomes the grounding that models are checked against. Wikidata
    Wikidata's embedding project makes its hand-curated statements available to AI systems for fact-checking and source attribution.[src]
  • General statistical methods keep overtaking hand-built knowledge.
    Richard Sutton's 'bitter lesson' (2019): over 70 years of AI research, general methods that scale with computation have beaten methods that build in human knowledge.[src]

Pressures that push this problem one way or another, as their sources argue them. All forces

• General methods that scale keep winning

  Favours statistical learning over hand-written knowledge as computation grows cheaper.[src]

• Language models need knowledge to check against

  Brings knowledge graphs back as a partner to statistical models rather than a rival.[src1][src2]

• Conference on Neurosymbolic Learning and Reasoning (NeSy) conference

  Annual meeting on combining neural learning with symbolic knowledge and reasoning; the 20th ran in Lisbon in September 2026. [src1][src2]

  checked 21 Oct 2026
• GraphRAG repository · GraphRAG

  Open-source implementation of graph-based retrieval-augmented generation, in which a language model builds the knowledge graph it then retrieves from. [src1][src2]

  checked 21 Oct 2026