Wednesday, October 17, 2018

Reading :: Mind over Machine

Mind Over Machine
By Hubert L. Dreyfus and Stewart E. Dreyfus

I've been rereading some of the mid-1980s work that introduced activity theory to the West, specifically paying attention to how it was positioned as an alternative to information-processing cognitive psychology (IPCP). Among others that were heavily cited at the time is this 1986 book, which asks the question: why hasn't artificial intelligence (AI) yielded the results that had been predicted in the 1950s and 1960s?

The 1950s and 1960s were a fertile time for beginning AI research, with Newell and Simon working on cognitive simulation at RAND (pp.6-7). Also at RAND was Stewart Dreyfus (henceforth SD), who was programming JOHNNIAC. Hubert Dreyfus (HD), a philosopher at MIT, expressed misgivings about AI to his brother and was in short order hired as a consultant for RAND in 1964 (p.5).

HD recognized that AI researchers were animated by the "continuum hypothesis": they believed that they were making the first steps, and if they continued, they would make steady progress. In contrast, HD saw a pattern in which AI researchers would solve a simple problem, consider it a first step to more complex problems, then encounter "failure when confronted with the more intuitive forms of intelligence" (p.7). His observations were not greeted with enthusiasm (p.8), but constituted the first detailed criticism of AI. As SD says later on, "Current claims and hopes for making progress in models for making computers intelligent are like the belief that someone climbing a tree is making progress toward reaching the moon" (qtd. on p.10).

To explore the contrast between AI approaches and human expertise, the authors distinguish between "know-how" and "know-that" knowledge—i.e., tacit and embodied knowledge vs. explicit knowledge (p.16). They propose a model with five steps to expertise:

  1. Novice
  2. Advanced beginner
  3. Competence
  4. Proficiency
  5. Expertise (Ch.1).
In the early stages, formalized or explicit knowledge is critical. These also represent the areas in which AI is most suited to assist, since AI excels at processing formal knowledge. But in the later stages, what is required is intuition: "Intuition or know-how, as we understand it, is neither wild guessing nor supernatural inspiration, but the sort of ability we all use all the time as we go about our everyday tasks" (p.29, their emphasis). See Klein for more on intuition in this vein. And like Klein, the authors state that "When things are proceeding normally, experts don't solve problems and don't make decisions; they just do what normally works" (pp.30-31, their emphasis). Put another way: "Competent performance is rational; experts act arationally" (p.36). In this sense, computers are "ideal beginners" (p.63) 

There's more to the book, but let's stop here, because this criticism and this model constitute the enduring legacy of the book. If you're interested in the historical development of AI and understandings of expertise, or in a model of expertise, definitely pick up this book.

No comments: