Cyc: History's Forgotten AI Project

Original Article ↗ Hacker News Discussion ↗

Role and History of Cyc

  • Seen as a flagship GOFAI / expert system project aiming at encoded common sense and general reasoning via a massive hand-built knowledge base.
  • Some commenters report working on early Cyc implementations, porting from one Lisp environment to another, and designing multi-user KB editing, which was unusual at the time.
  • Cyc is described as still implemented in Common Lisp (notably Allegro CL), with delivery layers on other platforms. Unclear how much of the original AGI vision remains active.

Related and Successor Projects

  • OpenCyc (now unofficially mirrored on GitHub) is cited as the closest open-source relative, though only a subset of the full KB.
  • Other symbolic/semantic efforts mentioned: OWL/RDF/SPARQL reasoners, Neo4j and similar graph systems, upper ontologies like UMBEL and SUMO, biomedical ontologies (GO, Reactome, Uberon), and probabilistic KGs (e.g., PR-OWL).
  • Historical crowd-sourced common-sense efforts like Open Mind Common Sense and Mindpixel are recalled.

Symbolic AI vs Deep Nets

  • Critics call Cyc a “bad idea,” arguing knowledge graphs are too labor-intensive, brittle, and don’t scale, while transformers do.
  • Others counter that “it depends on the application”; symbolic systems still excel in explainability, causality, and well-defined domains.
  • There is debate over whether deep nets are universally applicable; some assert they can tackle any useful problem (though not always optimally), others highlight classes where algorithms or symbolic methods remain superior.
  • Handling inconsistency is flagged as a core weakness of classical logic-based systems; paraconsistent and probabilistic logics are mentioned as partial remedies.

Neuro‑symbolic Hybrids and LLM Integration

  • Multiple commenters propose “Cyc 2.0” with LLMs generating or extending symbolic KBs, and using KBs to reduce hallucinations or provide checks during training/inference.
  • Active work is reported on integrating frames/slots with RNNs/LSTMs, text-to-knowledge-graph pipelines, and more general “neuro-symbolic AI” (e.g., DeepProbLog, Logic Tensor Networks).
  • Retrieval-augmented LLMs over structured KGs and richer cognitive architectures with explicit world models are suggested, but concrete, mature systems are mostly “early days” or research prototypes.

Impact, Applications, and Open Questions

  • Cyc reportedly reduced medical query turnaround dramatically in at least one clinical setting, and some domain-specific expert systems in industry were technically successful but commercially limited.
  • Several see parallels between Cyc’s “scale up the rules” bet and today’s “scale up the model and data” LLM strategy, and question whether LLM progress will eventually plateau without deeper structural advances.
  • Overall sentiment mixes respect for Cyc’s ambition and historical importance with skepticism about manually curated, monolithic knowledge bases as a path to general AI.