This post is by request from Joe Simone, who is active in the OpenMind project.

Joe asks:

Is Cyc destined for failure? Are we even close to bootstrapping AI such that it can learn from dialog?

I try to keep abreast of all the areas of artificial intelligence by virtue of accumulating books, by my membership in the AAAI and attending many AAAI conferences. However, given my seven years experience with Cyc, I will narrow my remarks to the niche surrounding Cycorp, whose knowledge base possesses these distinguishing features:

  • knowledge represented as predicate calculus assertions
  • assertion objects are symbolic and crisp, as opposed to fuzzy or probabilistic
  • single, consistent, ontology that attempts to cover the full range of human discourse
  • definitional assertions for each term, (i.e. its basic properties and a comment)
  • terms may be named concepts, or functional compositions of terms, or assertions
  • implication rules as well as ground facts
  • higher order expressiveness, e.g. properties are terms, types of terms are themselves terms, rules may contain property variables
  • powerful deductive inference engine
  • facilities to perform abductive and inductive inference
  • lexicon associated with many terms
  • KB content realized as a large number of ground facts, and a lesser number of rules, that support answering the simple sorts of queries that people can answer using their commonsense
  • hand-crafted KB (to-date)

Is Cyc destined for failure?
Let’s examine this question by drilling down on Cyc’s mission.

  • Will Cycorp remain a going concern as a business? I think so, for the foreseeable future. Cycorp has been a corporation for over 15 years and has a well-regarded body of research and track record, that will continue to draw the support of US government research sponsors. Cycorp has retained its key developers and ontological engineers despite its fluctuating income. To date however, Cycorp has no significant commercial products.
  • Will some other company adopt Cyc’s approach and subsequently surpass it? I think not. Cycorp has used its funding to build a hand-crafted KB containing several million assertions. Likewise they have a multi-million line program code base. This high barrier to entry, and Cycorp’s competitiveness within its niche, in my opinion preclude anyone else achieving AI with the same approach.
  • Will Cycorp eventually create artificial intelligence? I think yes, if not beaten by some other organization (having a different approach). Doug Lenat’s vision is that having commonsense knowledge and reasoning ability is the key milestone on the path to AI. So that is what the Cyc project has been doing since the early 1980’s. But because the KB does not yet achieve Lenat’s goals, Cycorp continues to put its resources into incrementally improving the KB content and its inferential and lexical behaviors. Eventually, I expect that Cycorp will meet the KB goals and move on to implementing the remaining behaviors for AI. If another commercial organization creates AI, then it is likely that Cycorp would be partner or close collaborator of that entity.
  • Will some other organization create artificial intelligence before Cyc does? I think yes, or I would not be working on Texai. Here are the factors that motivated me to attempt my own approach:
    • The Cyc KB is mathematically and philosophically organized in a manner agnostic to concepts from the viewpoint of a particular natural language. Thus, connecting it to an English dialog system is harder than it otherwise needs to be.
    • Cycorp is a commercial organization, and is closed source with respect to the majority of Cyc content and for the entirety of its program source code. This prevents Cyc from interacting closely with the open source communities for programming (e.g. GNU libraries) and open source content (e.g. Wikipedia). It also precludes mass collaboration from volunteers. John DeOliveira founded the Cyc Foundation after he left Cycorp to partially address these issues.
    • The focus of Cycorp is on perfecting the Cyc KB, improving its deductive inference engine, and improving natural language interfaces. Cycorp will not, in my opinion, focus on other required mechanisms for AI. Sponsors have chosen other organizations when funding AI mechanisms apart from Cycorp’s area of excellence.

Are we even close to bootstrapping AI such that it can learn from dialog?

I think so, and hope to demonstrate that with Texai sometime this year, in the limited domain of language acquisition. No other large-scale (i.e. ambitious) project is learning from English dialog, in the deep understanding sense. Incremental Fluid Construction Grammar is big step in this direction. Various chatbots modify their behavior based on experience but that will not achieve bootstrapping AI.