Training the AI to write better LiveCode

[Mark Wieder]

On 1/13/23 07:30, Mike Kerner via use-livecode wrote:
> "AI" has always been a misleading term, for like 40 years.
> gawd, i'm old.
> There is pattern matching, tree traversals, bayesian probability, etc.
> it's definitely not anything like a wiki.
> it might just be helpful at slogging out a lot of the code that we have to
> write or include by hand, especially if you are trying to tackle something
> new, especially for code that isn't neatly organized and instantly
> available.
> but, it's also not amazing and mature, either. we can be both surprised at
> the program a ten-year-old writes, and recognize that it's...not amazing.

I think maybe it's time to stop using the ridiculous "AI" term and refer 
to this more properly as Machine Learning. From my experience using 
TensorFlow the quality of your results depends mainly on the corpus of 
data used in training. Not just the quantity of data used to train the 
system but also the quality of the data used. It doesn't matter whether 
you're training the system for facial recognition or grammar checking or 
code generation, the better quality data you can throw at the system the 
better your results will be. Any ML system will have a bias towards the 
data it was trained on, so it stands to reason that having better code 
to draw on will allow it to make better judgements about code creation.

-- 
  Mark Wieder
  ahsoftware at gmail.com