On Wednesday, 4 September 2024 at 18:55:07 UTC, user1234 wrote:
On Wednesday, 4 September 2024 at 17:02:55 UTC, Vladimir
Marchevsky wrote:
In case that really needs some arguing, I would say
translation is not a programming.
There are semantical differences between two languages. Things
like implicit conversions, integral promotion, order of
evaluation, etc. still need a good old human behing the screen.
Yeah, the nice thing about translation is that the things that
LLMs have the most trouble with, ie. high-level semantic design,
class naming, abstraction, are already handled. But it's also not
like the LLM isn't making its own choices. Recently, I got a LLM
to *automatically* translate a program from asyncio-mqtt to
paho-mqtt.
That's not a syntax rewrite, it's a full-on migration, async to
threads. It still handled it fine. The boundary at the top of
abstraction work in programming where LLMs have trouble is really
not all that big.
The problem is that because current AIs are congenitally
incapable of backspacing, high-level abstraction basically
requires the AI to guess what it's gonna need from zero, and be
right, because it can't change course. Defining an abstract class
requires guessing what methods that class needs, because you
can't backtrack. It isn't - can't be - good at that, and because
of that, it isn't really trained in it either. But you can work
around that with careful prompting and some iteration.
For instance, you can see that we turned the Python segment
classes into functions. That's a thing that I suggested, but my
effort literally amounted to saying:
Honestly the more I think about it, the more I feel the
segments should really just be functions, since they don't
interact.
And later:
Can you please refactor the segment logic so that the functions
always return empty arrays for "don't include segments" and we
never look at `content.empty`? Also remove unneeded parameters
while we're at it.
The LLM works a lot better when it's going "from file to file"
making small to moderate incremental changes. (Like us, really.)
