AI was huge in the 1980s. Expert systems (rule-based progamming), Logic programming (Prolog), and Case-based reasoning were spawning startups at a furious pace. There were great breakthru systems such as Emycin. Large companies were developing advanced systems (IBM did a Financial and Marketing Expert). AI was taking over the world. AI based on reasoning and understanding was the true path. Eventually the limits became apparent and the hype died. Money fled the market. Winter arrived.<p>Today we see great strides in perception and manipulation (P-and-M). We can recognize faces. We can "compile action" into networks that can perform tasks. There is a belief that this is "all". "Reasoning and understanding" is now seen as "Good old-fashioned AI" (GOFAI). The limits of the P-and-M approach are obvious if you look at them with a clear eye. Unsupervised learning is one glaring example. Eventually the limits will become apparent and the hype will die. Money will flee the market. Winter will arrive.<p>This will be dismissed as reasoning by false analogy. But consider trying to use P-and-M in manufacturing. A robot would be excellent at putting something together using deeply-learned perception of parts and deeply-learned actions on parts. But the technology is useless for attacking related problems. For example, given a CAD drawing of a new toaster, develop a plan to assemble the toaster from parts and create the robot motions to complete the assembly. How do you create P-and-M actions on unique parts? How do you reason about plans? How do you handle one-of-a-kind problems that don't have big-data sources?<p>The limits are obvious. The primary market will saturate. Hard problems will arrive that P-and-M can't address. Winning technology areas will become mundane. Fast money will move elsewhere. Winter will arrive.<p>The "next generation" of AI will arrive when we start combining things like knowledge representation (e.g. the concept of a wrench) with perception (is this a wrench?) and manipulation (how to use it). The concept of a wrench will be grounded by attaching deep-learning P-and-M routines. Knowing the concept of a wrench will also mean knowing how to recognize one and knowing how to use one. Self-modifying systems that learn by changing their structure will arrive. New experience will add new knowledge and modify existing knowledge structures. Planning based on these structures will lead to experience which will lead to modified knowledge. The system will learn to ride a bicycle, "compile the knowledge into manipulation", attach it to the "bicycle riding concept" as grounding, and be permanently modified. Facts learned through conversation will be used to modify the knowledge base, leading to new behavior, just as we teach new students now. Self-modifying systems will arrive. But first we have to let the deep-learning hype burn out. Sigh.