About
    • AMMDI is an open-notebook hypertext writing experiment, authored by Mike Travers aka mtraven. It's a work in progress and some parts are more polished than others. Comments welcome! More.
Search
Page Contents
Full
Page Tree
Incoming links

Marvin Minsky/critic agents

11 Aug 2025 - 12 Aug 2025
Open in Logseq
    • Marvin always emphasized the importance of critics, that is, parts of the mind that would detect and deal with bugs or mistakes in other parts. "The Power of Negative Thinking" (SOM 14.8)

    • Suppressors (SOM 27.2)

      • It would be wonderful never to make mistakes. One way would be to always have such perfect thoughts that none of them is ever wrong. But such perfection can't be reached. Instead we try, a-s best we can, to recognize our bad ideas before they do much harm. We can thus imagine two poles of self-i-prou.-int. On one side we try to stretch the range of the ideas we generate: this leads to mor" id""r, but also to more mistakes. On the other side, we try to learn not to repeat mistakes we've made before. All communities evolve some prohibitions and taboos to teil their members what they shouldn't do. That, too, must happen in our minds: we accumulate memories to tell ourselves what we shouldn't think.
      • Suppressor-agents* wait until you get a certain "bad idea." Then they prevent your taiiig the cinesponding action, and make you wait until you think of some alternative. lf a suppressor could speak, it would say, "Stop thinking that!"
      • Censor-agents need not wait until a certain bad idea occurs; instead, they inter- cept the iates of mind that usually precede that thought. lf a censor could speak, it would sdy, "Don't even begin to think that!"
    • The roots of this idea are from Freud (acknowledged in the text), and first developed in Minsky's Jokes paper. Also probably influential, some of the work of Minsky's students like Sussman (Hacker) and Winograd (SHRDLU) which explored computational models that could detect and correct bugs or mistakes.
    • It's notable that present-day LLM AIs, despite having an architecture entirely different from Minsky's agent model, still have a censorship layer (based on reinforcement learning) which acts as a kind of superego.