On 19 February 2015 at 12:05, Vic <lug@???> wrote:
>
>> 10 years from now
>> we will simply ask a machine to write software for us
>
> I first heard that argument some 40 years ago. It wasn't true then, and it
> isn't true now. The reason for this is simple - code generation is a
> purely mechanical process, but defining the solution to the problem space
> requires semantic understanding, and that is the stuff of sci-fi.
>
> Many years ago, I was involved in writing some tools for Z. The idea was
> that a non-programmer could specify what he wanted, and then the Z
> compiler would generate the code. The project was a spectacular failure,
> because it turns out that getting that spec to be complete and accurate is
> exactly the same job as writing the code - the spec and the implementation
> can be considered synonymous. And this situation will persist until an AI
> is created that can properly *understand* a requirement. I don't know if
> we will ever get to that point - it's not just AI, it's Artificial
> Consciousness - but it won't happen in my lifetime.
>
I think writing software could be improved and made easier to do.
For example, at work, I am a Technical Architect / Project Manager at
the moment.
So, I might create a spreadsheet describing how one system should talk
to another, lets call it an interface spec.
I then pass it to developers and they write he code.
Occasionally they come back and ask for a clarification on this or
that, but on the whole they have enough to complete the program.
The developers were able to interpret my documents and write software from them.
I did not have to specify the threading model to be used, the message
priority model, they just selected what they thought might be the most
appropriate one based on what I told them the system had to do.
Once they had written the first working prototype, I reviewed the code
and made some changes, but I did not have to change much.
I believe that an AI could eventually do this, so long as a user also
understands the resulting code, and can review it and fix it where
needed.
I know that to get it right first time, the spec and implementation
are essentially the same, but I would prefer there to be a mis-match
between the spec and implementation with the AI making some
assumptions along the way, and then a review cycle to correct any
wrong assumptions.
What is missing is the AI to read and understand my spreadsheet or
document like a developer would.
I.e. Read a document that is not written in source code, and be able
to understand it as well as if they were reading source code, even in
the semantics are less strict.
James
--
Please post to: Hampshire@???
Web Interface:
https://mailman.lug.org.uk/mailman/listinfo/hampshire
LUG URL:
http://www.hantslug.org.uk
--------------------------------------------------------------
