http://www.incompleteideas.net/IncIdeas/BitterLesson.html
>> The biggest lesson to be drawn from 70 years of AI research is that general methods using computation are ultimately the most effective, and by a wide margin. The ultimate reason for this is Moore’s Law, or rather his generalization of the ongoing exponential decline in the cost of a unit of computation. Most AI research has been conducted as if the computations available to the agent were permanent (in this case, the use of human knowledge would be one of the only ways to increase productivity), but over time a little more than a normal research project, much more computation inevitably becomes available. In an effort to make a difference in the short term, researchers are trying to harness human knowledge in the field, but the only thing that matters in the long run is the use of computation. These two tasks do not necessarily have to contradict each other, but in practice they do. Time spent on one is time not spent on the other. There are psychological attachments to investing in a particular approach. And the human-knowledge approach tends to complicate methods in ways that make them less suitable for taking advantage of general methods that use computation.
>> That's an important lesson. As an area, we still haven’t mastered it as we keep making the same mistakes. To see this and effectively confront it, we must understand the power of the attraction of these mistakes. We have to learn the bitter lesson that building the way we think into algorithms doesn’t work in the long run. This bitter lesson is based on historical observations: 1) AI researchers have often tried to incorporate knowledge into their agents, 2) it always helps in the short term and brings personal satisfaction to the researcher, but 3) in the long run it plateaus and even inhibits further progress, and 4) breakthrough progress is ultimately achieved by the opposite approach based on scaling computation through search and learning. Ultimate success has a bitter taste and is often not fully digested because it is a success over a favored person-centered approach.
>> One thing to learn from this bitter lesson is the enormous power of general-purpose methods, methods that continue to scale with increasing computation even as the computation available becomes very large. Two techniques that seem to scale arbitrarily in this way are search and learning.
The second general point to take from this bitter lesson is that the actual content of minds is extremely, incompressibly complex; we must stop trying to find simple ways to think about the content of minds, such as simple ways to reason about space, objects, multiple agents, or symmetry. All this is part of an arbitrary, inherently complex, external world. They should not be embedded because their complexity is infinite; instead, we should only embed meta-methods that can find and fix this arbitrary complexity. Essential to these methods is that they can find good approximations, but they should be sought by our methods, not by us. We need AI agents that can discover like we do, not ones that contain what we have discovered. Embedding our discoveries only makes it harder to understand how the discovery process can be performed. [/spoiler]