After 24 years in AI: Some lessons

Abstract

After 24 years of teaching and working in AI, here are ten lessons: 1. It is difficult to make those stupid machines intelligent. 2. Scaling up does not work well in AI. Techniques to solve simple, easy problems do not scale up to solve bigger problems. 3. Divide-and-conquer does not work well in AI. In hard problems, progress in subpart interferes with previous gains elsewhere. 4. To educate a good AIer: start early with AI, train her thoroughly in software engineering, and teach her to work in a team. 5. Training a good AIer requires getting her hands dirty, by having her study and subsequently develop, large and complex systems. 6. It is difficult to find large, complex systems which can be built in a few manmonths. Team-work and add-ons are only partial solutions. 7. Practical applications are a good source of training projects, although it is difficult to supervise them correctly, in particular to avoid much busy work of educationally doubtful value. 8. Practical applications only rarely advance basic AI. 9. Vagaries in AI funding, and the attached changes in fashion, prevent any deep and long term attack on hard AI problems. AI prefers to spread its activities, instead of deepening them. 10. Pseudo-experts and -expertise are a great danger to AI.