The talk is mostly pessimistic. He warns about a work emergency, not enough people working and we will need AI to replace human work in the next years. Now, robots work in small domains, for example in the construction of standardised objects. But there is no automation of maintenance. Little or big maintenance is not accomplish by robots because this work usually has unexpected elements. Minsky has a gloomy view of the last 20 years of AI. He calls to recall old fields of AI like Sematic Information Processing. He also wants to emphasize the lecture of other books from the 60s and 70s. An example of restricted domain where AI was successful is symbolic integrals. Also algebra problems represented with natural languages have very good old solutions, hardly improved in the last years. But no commonsense knowledge is embedded in AI in these days. Human intelligence has multiple domains of knowledge in parallel: physical, social, emotional, spatial, mental, etcetera. He recommends his book: The Emotion Machine. Also warns about fad techniques and research areas, some of these will go away: genetic programming, insect robots, artificial neural networks, etcetera. They work with a well defined problem, not very general. For example with genetic programming, the problem is that only remembers what succeeded. No common mistakes are learn. On the other side, culture teaches common mistakes. Memes, propagated beliefs, not genes, include positive and negative information. Also the representation of knowledge is diverse. Minsky proposes a "Critic-Selector" model of the brain. It seems a very abstract model of the brain, including many levels, at least 6 of them. He emphatizes that you need theories of the mind before doing mental experiments with the brain. He observes that there are many Ways To Think: analogy, planning, simplify, reformulate, simulate, etcetera. Also that there are more parts than you need in the mind, there is no Occam's Razor in psychology! Many levels and structures. Minsky talks about different types of goals, seems to static to me if you have static categories. Finally he confess he professes isolationism, not connectionism, the existence of isolated levels and structures, that only interact when they need to.
I liked the emphasis on:
- Commonsense is mostly social.
- If you can solve a problem you cry for help, social intelligence comes to rescue.
- Any machine with Minsky's Model it's "... got to learn from people.".
Happy AI Hacking!