When the AI hype train was just getting rolling last year I thought I
would try this out to write a basic but non-trivial algorithm. I wanted
to implement a method to get the nodes and weights for Gauss-Laguerre
quadrature integration for the Gamma function for any number of nodes,
rather than just using a hard-coded series of arrays for the numbers. I
didn't need to do this (the hard-coded array approach works) but I
thought this was an interesting test because it's not an everyday garden
topic that the AI will just be able to cut and paste from Stack
Overflow. On the other hand it does have a known solution so it can be
shown whether the AI solution is any good or not.
The AI was able to generate a plausible looking solution very quickly
and confidently. Which was, at first glance, impressive. Wow. It
compiled first time and ran. Syntactically it was great. And that is
where the positives end. The code wasn't just wrong in terms of being
close, but inaccurate. It was just gibberish. The results weren't even
vaguely in the ballpark. People talk about AI being able to pass
university exam level questions... what a load of rubbish. This was
clearly plagarised so badly that the 'student' submitting the answer was
clueless that their answer was even wrong.
The AI had no idea -- and why would it. It is nothing more than a large
neural net that has produced a sequence of tokens based on another input
sequence of tokens. At its heart it is nothing more than a sophisticated
regression algorithm based on a series of data points which were used to
train it. What we are doing is the equivalent of deriving a line of best
fit through a cloud of data points and then asking the algorithm to
extrapolate based on that regression equation. It works -- of course it
does, it's just an equation -- but fundamentally it is grossly
misleading. Any decently educated person knows that extrapolation is
fraught with error. And that is all AI is (with the techniques used
today). A sophisticated extrapolation engine. Anyone who says otherwise
is deluding themselves.
Andreas is, in my view, completely right. If you rely on AI for coding
you will spend more time correcting the code than you would writing it
yourself. More than that, the code appears very neat, so the errors are
not obvious at first. Writing the code yourself will also bring you
closer to it, and that familiarity will breed insight that AI never had
to start with and certainly cannot impart to its users. Why would you
want to give up that side benefit from writing your own code?
AI is, in my view, one of the more dangerous fads to come out recently.
Not necessarily because of the harm it might do itself, but because it
drives a false sense of security amongst those that do not understand it
-- which is to say most of the world, as there are very few people I've
met who are able to explain to me how it works, even on a basic level.
Perversely this could lead to a loss of knowledge and intelligence
amongst the non-machine population.
Maybe I'm wrong and I'm just a modern-day Luddite. I'd be interested to
know if there are any genuine real-world cases where AI code is useful.
I was thinking that maybe boilerplate code, like when writing GUI etc.
but NetBeans Matisse already has that covered just fine. How would AI
improve this? It wouldn't have a chance improving on many of the
algorithms I write for solving engineering problems for the simple
reason that many of those algorithms never existed before I wrote them
and I'm proud to admit that I had to fall back on real intelligence, and
not artificial intelligence, to write them :-)
Anyway, an interesting topic!
Peter
On 2024/02/08 16:17:51 Andreas Reichel wrote:
> On Thu, 2024-02-08 at 11:16 -0500, Alonso Del Arte wrote:
> > There isn't, but probably soon there will be. If it's built-in, it
> > better come with an off switch. But if BlueJ adds an AI assistant,
> > then we're really in trouble.
>
> I have tried them all on IntelliJ and I can tell you: they are rubbish!
> OpenAI is great for looking up JAVA or C API efficiently and for this
> case you have a 50% chance to get something useful (although 50% chance
> of getting completely wrong information, like Java Libraries that don't
> exist.)
>
> But in the IDE it is just annoying. You will spend more time correcting
> the suggestions than writing code.
>
> Example: ask for a Java Example of Prophet Time Series Forecast! Or ask
> for a sample of RGBA to ARGB Byte Swapping using SSE or AVX.
> The given information will be TOTALLY wrong and fantasy!
>
> Cheers
> Andreas
>
>
>
-
To unsubscribe, e-mail: users-unsubscr...@netbeans.apache.org
For additional commands, e-mail: users-h...@netbeans.apache.org
For further information about the NetBeans mailing lists, visit:
https://cwiki.apache.org/confluence/display/NETBEANS/Mailing+lists
