PureBasic Forums - English

It might be a forbidden question, but recently I tested using Gemini3 to assist making programs, it seems understands the C++ syntax and can make a prototype app with prompts, oh and it does Python, too.

I don't know if it really understand the coding language, but it can generate some good enough results with proper prompt, so I'm wondering if it(insert your AI models) can really "learn" a programming language or is it just grabbed something from elsewhere.

And PureBasic syntax is relatively simple IMHO, if AI can undestand PB and coding some basic things itself, holiday coders like me would be useless anyway, I'm a bit afraid...

Well, I don't know when AI can go that far, but I think it's happening, at least AI will learn more and live longer in some degrees, and human life is so limited.

Hi, I'm not very familiar with Gemini, but with ChatGPT you can build an own GPT especially for PureBasic, using the PureBasic Documentation.

Claude and codestral are trained on purebasic. Gemini appears to be but says it's not.
I find Gemini good for research, it's quite capable
And I find Claude better at producing working code but it usually insane slop.

It's been discussed here before. ChatGPT is pretty good with PureBasic, in my experience. It knows what I'm asking and 9 out of 10 times it gives a correct working example. I'm not a PureBasic newbie, though, so I can tell when it's giving incorrect info and I tell it what's wrong so it remembers for next time.

Love it or hate it, it's here to stay. So, may as well get used to it.

It's ELIZA, forty years in the making.

But seriously, it's a machine, albeit with a plethora of ethical and real-world ramifications. Sentience may not be possible, but what is it, anyway?

Meanwhile, let's just enjoy the technology. So far, for me, it's been a programming partner, and the world's best search engine. And it's quite the muse as well. The ability for it dynamically pull data and answer questions or design code, for virtually any language, or even stimulate human conversations, also in almost any language, is nothing short of impressive.

Sure, it's not perfect, yet, but it really has that can-do attitude, even if it spews crap every now and then.

Let's just enjoy it!

Ich benutze ChatGPT als Unterstützung. Aber ganze Projekte ist nicht so einfach.

tft wrote: Sun Nov 30, 2025 5:18 pm Ich benutze ChatGPT als Unterstützung. Aber ganze Projekte ist nicht so einfach.

Translation
I use ChatGPT for support. But entire projects aren't that easy to implement.

"AI understand PureBasic"


PS: Che mai potrebbero sapere di nuovo le intelligenze artificiali?
After all the training, they cannot translate if it isn't a tiktok video…
Not good…

Claude is pretty good at generating purebasic.
I used it in the animated vector graphics code to do the filler from a complete example. What would have taken an hour took seconds to do.

I've been using Claude for over a year. Although generally good, it has not yet generated 100% PB code for me. I frequently find functions that look as if they've come from other languages, particularly Python. But it is certainly better now than a year ago.

I mainly use it for research into specific things. Saves heaps of time and effort. It was a big help getting Postgres and MariaDB working in my complex school library management system. As an exercise, I asked Claude to build code to backup and restore a database from MariaDB that could be done from a workstation. Took over 20 retries until something worked. I then wrote my own version, which was about 10% the size and considerably faster. Have to say it is getting better.

I looked at Claude but it wanted my phone number to sign up, so I didn't. ChatGPT is working well enough for me.

I belive, that content generation is not really an ai feature. An untraind model generates nonsense, a trained ai generates outputs similar to the learning inputs, but deviates ( interpolates ) between them. A more trained NN will produce outputs more close to the original training data source with less deviation ( aka loses creativity ) This applies to image generation or code generation or text (literature) generation.

I’ve thrown some old PureBasic scripts at a few AI tools, and most of them do an okay job as long as the code is clean and not full of custom libs. They won’t always catch the quirks of older PB versions, but for debugging or quick rewrites, it’s surprisingly useful.If you’re mixing PB with API calls or Windows-specific stuff, the AI usually needs more context but still helps speed things up. The word world shows up a lot in coding forums because everyone’s trying to make their own “hello world” tests. It fits here since AI is basically opening a new world for older languages too.

Can current AI "understand"....

No, it's just like a weather forecast in that its responses are based on likelihoods of being correct. There is no real understanding involved, it's merely a text transformation based on probabilities...

##Deutsch
Ich habe jetzt mein erstes Tool zusammen mit ChatGPT gebaut.

Insgesamt etwa zwei Tage, rund 40 Stunden, umgesetzt in Python, weil einiges von dem, was ich machen wollte, in PureBasic so nicht sinnvoll oder sauber lösbar war.

Schon am ersten Abend habe ich gemerkt:
„Mach mal …“ funktioniert nicht.

Man muss sich die Arbeitsweise der KI aktiv zu eigen machen. Das bedeutet:

exakte Aufgabenbeschreibung

klare technische Vorgaben

vorher festgelegte Regeln, wie gearbeitet wird

Vor allem eine wichtige Erkenntnis:
Mitten im Projekt das Thema wechseln ist extrem schlecht, selbst wenn es nur eine scheinbar harmlose Frage zum Editor oder zur IDE ist. Ab diesem Moment wird der Workflow inkonsistent, und man verliert die Linie.

Am zweiten Tag habe ich das Ganze dann bewusst als Projekt behandelt.
Allerdings habe ich den Code konsequent von ChatGPT entwickeln lassen – und genau da liegt die nächste Falle:
Der Code ist plötzlich anders als zuvor, Funktionen fehlen, Dinge wurden implizit weggelassen. Und schon funktionieren Abhängigkeiten nicht mehr. Man verzettelt sich schneller, als man denkt.

Am dritten Tag habe ich deshalb komplett neu angesetzt:

klare, schriftlich festgelegte Vorgaben

andere Codebasis

feste Regeln (kein Themenwechsel, kein „nebenbei noch schnell“)

Heute, 12:00 Uhr, bin ich seit 07:00 Uhr dran – und:
Das Tool macht genau das, was es soll.

Allerdings nicht ohne Preis.
Ich habe dafür gefühlt einige Gemüts-Stahlseile im Kopf geopfert.

Fazit für mich:
ChatGPT ist kein „Code-Automat“, sondern ein hochleistungsfähiger, aber extrem strukturabhängiger Projektpartner.
Wer das akzeptiert und entsprechend arbeitet, kann sehr schnell sehr weit kommen.
Wer es wie eine Suchmaschine behandelt, scheitert.

## Englisch
I have now built my first tool together with ChatGPT.

In total, about two days – roughly 40 hours, implemented in Python, because some of the things I wanted to do are not feasible or cleanly solvable in PureBasic.

Already on the first evening, I realized one thing very clearly:
“Just do it…” does not work.

You have to actively adapt to the working method of the AI. That means:

precise task descriptions

clear technical specifications

predefined rules for how the work is done

One especially important realization:
Changing the topic in the middle of a project is extremely harmful, even if it’s just a seemingly harmless question about the editor or the IDE. From that moment on, the workflow becomes inconsistent and you lose direction.

On the second day, I deliberately treated the whole thing as a project.
However, I let ChatGPT generate the code consistently – and that is where the next trap lies:
The code suddenly becomes different from before, functions are missing, things are implicitly omitted. Dependencies break, and you get tangled up faster than you expect.

On the third day, I started over completely:

clear, written specifications

a different code base

strict rules (no topic switching, no “just quickly doing something on the side”)

Today, at 12:00, I’ve been working since 07:00 – and:
The tool does exactly what it is supposed to do.

However, not without a price.
I feel like I have sacrificed quite a few mental steel cables in my head.

My conclusion:
ChatGPT is not a “code automaton,” but a high-performance, highly structure-dependent project partner.
Those who accept this and work accordingly can get very far, very quickly.
Those who treat it like a search engine will fail.