I'm new to AI and I want to learn more about its limits

[Marc Weber]

openai.com for instance has a 'unsolved' type of task: reading and understanding english descriptions of programs and turn them into code.

Looking at some of the descriptions reveals that it requires a lot of imagination and applying wisdom to contexts.

I'm personally interested in how well a machine might perform planning advertising for target audience or estimating the value of a brand for instance by looking at Facebook likes, how people react to the brand name and similar.

The more I think about how I solve problems the more I conclude its a mix of 1) creativity 2) wisdom (remembered details) 3) applying and following patterns such like a stack machine 4) trying different strategies such as how to reach the target or knowing the starting point what could I do ... 

Do you have any hints about where to start reading about which solutions exist right today to understand when computers will do my job?

Companies like sourced.tech are trying to understand huge amounts of code and links like http://news.mit.edu/2016/faster-automatic-bug-repair-code-errors-0129 indicate that AI might be useful to apply to real world problems fixing bugs today ..

I personally have spend huge amounts of time *reinventing the wheel* in the past - for instance when implementing a XDEBUG protocol plugin for Vim requiring some creativity - Vim strings cannot represent 0 bytes - but the protocol is using them as well as stupid work (fixing regex based XML parser library) ... But in the end Vim is flawed because VimL code can be interrupted when on resize autocommands without you as programming having a chance to setup locking.

So the perfect AI would return a reply such as:

1) a quick fix would be fixing this VIML that way

2) a real fix would be rewriting Vim in a different language (rust/go) and rewrite all the event handling because its broken by design...

There are a lot of fun tasks such as "write a completion engine" for all editors out there or turn an editor into something usable within a browser (I know about llvm -> JS) -> which I know can be done triviialy once you spend enough time on the tasks...

Once new forks arrive (neovim) (same happened to emacs in the past for legal/license reasons) ... the funnier it gets and the more AI could assist.

I don't fear more intelligent AI - neither do I fear more intelligent people :) Given the same environment things get more and more similar - same will apply to humans/AI.

chess/go -> soon StarCraft (Blizzard) -> See deepmind people and machines have been showing similar trates (over training) for quite a while as well.

Marc Weber

[Mark Nuzz]

On Wed, Feb 22, 2017 at 4:05 PM, Marc Weber <wanaho...@gmail.com> wrote:
> openai.com for instance has a 'unsolved' type of task: reading and
> understanding english descriptions of programs and turn them into code.

I think that this problem should be approached incrementally. Start
with simple programs, and a simple (home grown) coding language, and
increase the complexity of the language and/or problem description
gradually until you hit a roadblock. Sure, you may be able to solve it
at first with some shortcuts, which won't apply to real world use
cases, but at the very minimum you'll have some proofs of concept and
inspire interest on the problem by publishing your findings.