I always find it interesting that we try to do what computers are good at (computations, storage) and vice versa (pattern recognition followed by inference and decision making).

There are many names for the logic, probabilistic/statistical and other numerical optimization appraoches ranging from AI, decision-making, signal processing, machine learning, statistics, econometrics, operation research etc.

The problem is that is just appears as AI though almost all the time you fix some input data format, run a numerical optimization procedure and do predictions either directly for discriminative approaches or after model parameter estimation. This can be seen as kind of "hard-coding".

Just look at the DARPA autonomous car competition. Here Stanford's Sebastian Thrun discusses his approach. These cars got a massive amount of sensors and measurement devices and still if it encounters unseen environments it can be quickly led to very sub-optimal decisions.

A similar thing holds for chat bots as Eliza. I think you just need to input a huge number of specific phrases to make it pass the Turing test ("being able to tell the difference between a machine and a human") for most humans.

All these scenarios share a common factor: widly heterogeneous "things", i.e. input data formats, environment types (different road types, shades, reflection etc.), different discussion topics. The obvious brute force solution is to use databases of examples and emulate the really unseen ones by similarity. It is similar to the situation with chess programmes right now. Even the top players can no longer consistently beat the top programmes. The programmes did not suddenly get "smart" but rather make use of a huge library of possible moves. People also make use of experience but also are more inclined to apply heuristics due to resource limitations.

Another problem is that model parameter estimates are very sensitive. For example doing microarray experiments on patients in hospital A and B may yield very different parameter estimates. Thus a very hot new branch of machine learning looks at transfer learning (multi-task learning) right now where you try to see how the estimates carry over across different datasets. In this way it may also be interpreted as online learning for non-stationary distributions. Here is a video presentation about this topic by Massi Pontil from University College London.