I find the “codic cortex” a pretty useless metaphor. We have no idea what such a thing would look like so positing it adds nothing toward producing a competent AI which can program, much less a recursively self-improving AI. One could add all sorts of tools to a human programmer and make her more productive also. To call it an external codic cortex seems more poetic than useful. There are economic more than technical reasons why there is a paucity of good programming tools. Those of us that have been around a few decades doing this work know there has not been a lot of improvement in day to day tools.

The main problem with human programmers today is that we are stuck writing programs that are actually more or less predictable and debuggable by us with the limited cognitive architecture we possess. Better tools can only help part of the problem as our predictive and thus understanding capacity is obviously architecturally limited.

Tim Tyler bring up an interesting point. Most human programming is about algorithms, more or less Turing machines. But increasingly the really important aspects of computer software is more about interacting ecologies of “programs”, humans, sensors, actuators. It is more about inter-relationships. But such “interactive [or interaction] software” is only beginning to surface as a concept in software engineering.

Computers already do most computer programming anyway – and they are much better and faster at it than humans. Humans just give a relatively high-level description to compliers these days – and the machines do most of the refactoring, conversion to bytecode and machine-code programming.