> extensively documenting and testing their software if three sentences are enough to encode it
Was just hyperbole for "from plain English specs/requirements".
I'll admit to being uninformed about GPL, but your understanding of large language models is also limited. They actually learn to interpolate between data points meaning they can compose sequences not found in the training data. Further, GitHub added a feature that checks existing code for a match and rejects predictions if any match occurs.
Nobody disputes their ability to interpolate, I think (at least not me), but the problem is the starting points for these interpolations contains GPL licensed code, hence it derives GPL licensed code.
This derivation brings GPL in, and the model doesn't understand this. As a result, every time a GPL training data is mixed into the interpolation, you're converting the code GPL, or if you're not converting your code to GPL, you're violating GPL.
It's plain and simple.
On the other hand, I'm hearing "we'll write the specs, and computer will just auto-generate it" gospel since 2002. This time it won't be different. Human brain, intuition and creativity is beyond algorithmic modeling.
So, no, computer will not autogenerate the code from specs. It might link boilerplate together, which can be already done today.
Was just hyperbole for "from plain English specs/requirements".
I'll admit to being uninformed about GPL, but your understanding of large language models is also limited. They actually learn to interpolate between data points meaning they can compose sequences not found in the training data. Further, GitHub added a feature that checks existing code for a match and rejects predictions if any match occurs.