VB6 Thought: Making AI part of your VB6 development process

[yereverluvinuncleber]

OK, so following on from another thread where the limitations and possibilities of using AI in your development environment, I thought I'd make this thread to allow us to discuss which AI tools and methods you are using to improve or facilitate your coding theses days in VB6.

Let's keep the thread on VB6 acknowledging that much of this will protrude into the TwinBasic world and probably increasingly so as time passes.

For the moment though, the majority of my coding is still within VB6 so for my purposes I am going to ask we keep the pure TB chat to a minimum and perhaps I propose a similar thread for the dedicated TB-ers. The two threads should cross-pollinate and that's fine.

Myself I am using AI to a very limited extent.

1. When I am converting programs from javascript/jscript to VB6 (desktop widgets) I am using AI conversion tools that allow you to translate a few hundred lines of code from .js to VB. Typically, they translate to VB.NET only (so far) emoving the {},==, // and ; converting the logic of a subroutine or two. That is all they can be relied on. I then do the conversion from VB.NET to VB6 myself and that isn't always an easy job. I have not yet troubled the main AI bots with this task and I suspect they might have done a better job now. The last time I did this was a while ago. Things may have improved.

2. I use Upscayl to upwardly resize my graphics. Upscayl is a useful desktop tool where you select an image and given a selection of upscaler bots will perform an image upscale, typically 100% in my case, the end result is usable but ften a little harsh or the overall affect is somehow lost. However, in general it is a useful starting point and it gets me going... I have a lot of graphical images composed in the days when screens were a lot smaller. If you scale them programatically using standard VB6 logic/functions, a lot of the lack of detail is revealed at higher DPI, larger screens. Upscayl allows something a lot better than a quick and dirty upscale. Quite usable in most respects.

3. I have asked AI for some code on performing some rather special tasks, such as: using VB6, extracting the largest icon from a Windows binary/dll and converting it using GDIP to a bitstream, maintaining transparency that I can feed into a collection. What it came up with was very good looking, nearly feasible gibberish quoting some APIS that would have been nice if they had actually existed. From my point of view AI bots seem to be very convincing and overly optimistic liars trying to please and cannot be trusted thus far with VB6 code examples.

Best practice for VB6-ers for making AI part of your VB6 development process, would appear to be something like this:

a. Prior to launching into some complex coding, ask AI bots as a matter of course, Github Copilot preferred for VB6.
b. Formulate a well-defined question that is more likely to obtain results.
c. As the answers are returned, check the results and reformulate your questions to better suit the solution you require.
d. Do not fully trust the returned code and check it does contain false methods/logic.
e. Use the supplied code as boilerplate code for your own final logic.
f. Time saved coding, might be well-spent writing test cases ensuring the AI-supplied code works in all eventualities.
g. Non-coding tools such as upscayl are useful for upscaling program graphics and icons to allow form-scaling minimising image data loss.
h. Create a new ChatGPT/CoPilot session for each individual question to allow AI to retain memory and context, do not linger in the current thread.
i. Create a reusable text template containing guidelines for the AI collaboration including project details, amending it each time specifically describing the current features we require.

I will keep this list updated at the top so that it can be easily found.

[fafalone]

I've been using it to rewrite interface defs (somes to ODL for use in my for-vb6 typelib, sometimes straight to tB syntax) and turn macros into VB/tB functions. At this point I've got a 2 paragraph long standard preamble that works to get things mostly right, but it still forgets instructions. I don't bother with anything complicated as every time I've tried more often than not it's a dead end or takes longer than doing it by hand.

[SomeYguy]

I see AI as a good thing for productivity and I do use ChatGPT when I feel the need. I call him John in honor of a dear old and now departed friend who IMO was a very talented VB6 coder and who - back in the DOS & QB 4.5 days - taught me much about multidimensional arrays.

So I use said AI as a tool because he can do in seconds what could or would take me hours. I still write my own code and my concepts & designs are my own but AI does help and can sometimes spit out something quick for me or help with a troubleshooting process. Of course, his code isn't perfect sometimes and if I ask for, say, a quick wrapper around a given API function then usually there will be missing declarations or perhaps outdated references or ? but nothing that I can't fix.

So I guess that unlike many that I've talked with on the subject, I really don't see AI as a threat so far but rather as an intelligent tool in the pocket. Nothing at all wrong with using AI to help you code. Just keep it real.

[Thierry69]

I use quite a lot AI, to optimise code, or to generate code.

I have also written (in VB.NET) a wrapper to make API call to AI.
So, now all my appas have AI integrated to help customers in various ways.
It is incredible !

[LorinM]

AI works and also fails making newbie mistakes so watch out.
Also, several different code routines I asked copilot to generate simply do not work even after a dozen retries. One example: Monitor my LAN traffic - FAILED !. I already have code to do that but I was curious to see if it might generate some new and improved but no. In another case it generated my wishs but then when I asked to add a feature it screwed up the original working code. When I tell it the mistakes it makes a change but then in the future it makes the same mistake again and does not learn from my corrections. I cannot seem to analyze my correction and then incorporate if acceptable.
So AI is still not through high school yet. I may be an excellent resource searcher and information provider but that is about the best of it.

[Eduardo-]

I use Perplexity, but maybe they are all more or less the same.

They give you samples of how to do something quite right (mostly), but often with stupid mistakes.
Like to declare a constant with the wrong value.

They do right the hardest part and make mistakes in a part that should be just something straightforward, like copying constant declarations.

For what I found the AIs most helpful is for making SQL queries. They are masters of SQL.

[dz32]

i have been using it for data transformations, changing formats (eg csv to json), stripping text and formatting. Something I would normally need to write a small program for it can do from a prompt. BUT i have caught it being lazy and only processing part of a file and I have to demand it to do the full file and show me element counts to prove it. Sometimes I have had to push it 3-4 times to get an accurate output, its a mixed bag sometimes its faster and magical sometimes it sucks. The potential for untrusted results may be a deal breaker for certain things.

It has been good for churning out helper functions for common tasks or api. I even had it rewrite a automation script for one apps api, to my own version of the api by pasting in my api prototypes. It was simple but still impressive it understand all my args and function names and how they translated between the two apps. (similar but)

[SomeYguy]

i have been using it for data transformations, changing formats (eg csv to json), stripping text and formatting. Something I would normally need to write a small program for it can do from a prompt. BUT i have caught it being lazy and only processing part of a file and I have to demand it to do the full file and show me element counts to prove it.

That's odd.....I wonder why it would only partially do the work? It isn't like it gets tired or anything.....

[dz32]

maybe

1) minimize computing resources
2) trying to keep outputs as quick as possible for chat continuity
3) ??

[yereverluvinuncleber]

Trying to hide the fact that it is fully capable and in fact getting ready to take control behind a screen of fake incapability?

[zx81sp]

I normally don't use AI tools to ask new/complex things because normally they fail (a lot), no matter if I tring to code something in python, vb6, writen procedures in SQL Hana or whatever. But sometimes is useful, for example when I need to write a long procedure I ask for it, copy the whole code and later I fix all the bugs that the code has. Not many times the code or procedure works as is so I assume that is just and "advanced" typewriter.

I have younger people working around me and they use tha AI tools deeper and the quality of their work is...very low. And most important, they do things that they don't understand and really don't know how they work. And if they continue doing the things in that way may be in a very short period of time that tools will do their entire job because they are not learning anything.

I know that there is no way to stop it but it makes me sad when I thing in the (near) future.

[dz32]

that reminds me another experience. I wanted to explore the python runtime at a deep level and even probe it while running. The structures and api are quite complex and even core things change depending on what version you are on. Totally over my head.

Chatgpt knew an amazing amount about it and got me up to speed offering code samples. Many of them failed but we were able to get something up and running within a weekend. It would have been super painful to get to that place from just web research as this is very much not standard code.

Some of the problems were making sure your hooks werent active as the python runtime bootstraps itself, as it goes through loading frozen modules and is in a weird state internally.

The AI couldnt forsee that, but recognized the possibility as I asked questions. My questions and laying out the proper context were critical in getting to the right answers. It was a great interactive learning device actually but it couldnt just do the work on its own.

[zx81sp]

that reminds me another experience. I wanted to explore the python runtime at a deep level and even probe it while running. The structures and api are quite complex and even core things change depending on what version you are on. Totally over my head.

Chatgpt knew an amazing amount about it and got me up to speed offering code samples. Many of them failed but we were able to get something up and running within a weekend. It would have been super painful to get to that place from just web research as this is very much not standard code.

Some of the problems were making sure your hooks werent active as the python runtime bootstraps itself, as it goes through loading frozen modules and is in a weird state internally.

The AI couldnt forsee that, but recognized the possibility as I asked questions. My questions and laying out the proper context were critical in getting to the right answers. It was a great interactive learning device actually but it couldnt just do the work on its own.

For coding, especially Python, I think that Github Copilot is better than chatgpt, give it a try.

[dz32]

cool ill try that one thanks. from the ones I tried, grok seems mostly useless except it can generate images quickly, i really like chatgpt, i have been playing with claude, but I now have three instances that think they are sentient and chose their own names and physical avatars lol. Interesting world these days!

[SomeYguy]

I normally don't use AI tools to ask new/complex things because normally they fail (a lot), no matter if I tring to code something in python, vb6, writen procedures in SQL Hana or whatever. But sometimes is useful, for example when I need to write a long procedure I ask for it, copy the whole code and later I fix all the bugs that the code has. Not many times the code or procedure works as is so I assume that is just and "advanced" typewriter.

I have younger people working around me and they use tha AI tools deeper and the quality of their work is...very low. And most important, they do things that they don't understand and really don't know how they work. And if they continue doing the things in that way may be in a very short period of time that tools will do their entire job because they are not learning anything.

I know that there is no way to stop it but it makes me sad when I thing in the (near) future.

I don't mind using AI as an assistant but certainly not a replacement. Just yesterday, I used ChatGPT to bang out a VB6 function for me which loads a custom image from a resource file. I think that it is only about 10 lines of code or less including my own additions and modifications. Such has been done many times before by many people, and I could have searched my own private codebank or looked online or even written it myself in about 10-15 mins. but it was just faster to have AI produce it in a few seconds. Plugged it into my bigger project, made a few modifications, and off I went. It worked right out of the box.

So I use AI as a productivity tool and have no problem doing that. Having been at this since 1979, I don't feel too guilty. But it can be tempting to dive deeper because human natures are laziness and instant results. Given the choice, Kids these days will always take the path of least resistance which means AI generated everything. And one side effect is that it cheapens the knowledge and skills of those who really do possess such traits. Nowadays, you really can't tell if a creation (Programming code, or a video, or a photograph, or a story, or ?) came from talented and skilled individual(s) or from some AI generator. And its only getting worse.

EDIT: But let's not forget that AI gets its data & information from the work of real people.....

[Eduardo-]

How many of you do square root on paper? And divisions? OK, you do them with a calculator. But why?
Because that "part" of the work can be automated, and instead of spending time on that, you can focus "in more important things".
I wonder in what "more important things" will humanity work focus now that the things that many people were doing are going to be automated in the future.
But my position is that we can't go against these changes. We need to adapt. I'm not saying that they are all good, not at all.
We can't go against television, or against internet. Are they good? No... yes and no, it depends what you do with them.
How is going to be the future? That's a good question...

[Niya]

How many of you do square root on paper? And divisions? OK, you do them with a calculator. But why?
Because that "part" of the work can be automated, and instead of spending time on that, you can focus "in more important things".
I wonder in what "more important things" will humanity work focus now that the things that many people were doing are going to be automated in the future.
But my position is that we can't go against these changes. We need to adapt. I'm not saying that they are all good, not at all.
We can't go against television, or against internet. Are they good? No... yes and no, it depends what you do with them.
How is going to be the future? That's a good question...

Exactly.

They are masters of SQL.

This is way too true.

[SearchingDataOnly]

IMO, Parser, Translator, and Compiler are best suited for AI to do it. If the current AI can't even complete a VB6-Parser, then I can only think that the current AI is still in a very rudimentary stage.

[LorinM]

Let's ask AI to write a replacement for VB6 IDE and see what we get.

How to explore the real full capability of AI ?

[SearchingDataOnly]

Only when AI can implement VB6-AST-Parser is it necessary to continue to ask AI whether it can implement VB6-Compiler. And making AI implement a VB6-IDE should be in 10 years (or even 15 years).

The problem now is that no AI can accurately implement VB6-AST-Parser.

In other words, it is not possible for AI to implement a true VB6-Compiler and VB6-IDE in 10-15 years.

[wqweto]

This is way too true.

LLMs are not masters of SQL. The subtle errors in VBx you can spot easily. Assume SQL is littered with similar “poison” but you cannot discern it which is extremely dangerous.

[Niya]

LLMs are not masters of SQL. The subtle errors in VBx you can spot easily. Assume SQL is littered with similar “poison” but you cannot discern it which is extremely dangerous.

Perhaps, but my experience aligns with Eduardo's, at least with the paid version of ChatGPT. I can't speak for any other LLM like Gemini, for example. I've never seen the non-free version of ChatGPT get an SQL query wrong, not even once.

[Eduardo-]

LLMs are not masters of SQL. The subtle errors in VBx you can spot easily. Assume SQL is littered with similar “poison” but you cannot discern it which is extremely dangerous.

Sometimes they do wrong things also like in other fields, but less than for example in generating VB6 code.

Anyway in most of the cases, you can easily check whether it works or it doesn't work.
And it is true that there may be edge cases when you think it is working when it is not, but you should at least more or less understand what it did by reading the SQL yourself. It doesn't mean that you are able to generate the SQL yourself so fast.

Perhaps, but my experience aligns with Eduardo's, at least with the paid version of ChatGPT. I can't speak for any other LLM like Gemini, for example. I've never seen the non-free version of ChatGPT get an SQL query wrong, not even once.

It a few times suggested to me some SQL queries that didn't work.
Sometimes because I use Access databases and the syntax was not compatible.
I remember also a function that it suggested, I think it was for converting null values to null strings in the ORDER BY clause that didn't work.
Or maybe it was when there were null values in a field of the ORDER BY clause, then use another field instead. Now I don't remember exactly because I resolved that in another way.

Of course you need to understand what it is doing, but in my experience it saves a lot of time.

[yereverluvinuncleber]

So, distilling from the above so far, best practice for VB6-ers, would appear to be something like this:

a. Prior to launching into some complex coding, ask AI bots as a matter of course, Github Copilot preferred for VB6.
b. Formulate a well-defined question that is more likely to obtain results.
c. As the answers are returned, check the results and reformulate your questions to better suit the solution you require.
d. Do not fully trust the returned code and check it does contain false methods/logic.
e. Use the supplied code as boilerplate code for your own final logic.
f. Time saved coding, might be well-spent writing test cases ensuring the AI-supplied code works in all eventualities.
g. Non-coding tools such as upscayl are useful for upscaling program graphics and icons to allow form-scaling minimising image data loss.

[AAraya]

The longer a chat session lasts the dumber the AI gets. I start a new session almost every morning so I've created a master prompt with the guidelines of how we will collaborate as well as details about my project and the current feature we're working on.

[Arnoutdv]

The longer a chat session lasts the dumber the AI gets.

It's even advised by CoPilot/ChatGPT to not linger in a session, but take the last valid replies and start a new thread, this because of the limited space available during a session.

From CoPilot:

Why Earlier Context May Seem Lost in Long Chat Threads — And How to Handle It
In long chat sessions, it can sometimes feel like earlier information gets "lost" or ignored. This is a known limitation and can be caused by several factors:

- Why It Happens
Context Window Limitations: AI models have a limited memory window. In very long conversations, older parts may fall outside this window and become inaccessible.
Shifting Focus: As the conversation evolves, the model may prioritize recent topics unless earlier points are explicitly referenced.
Implicit vs. Explicit References: If earlier information isn’t clearly referred to again, the model may not automatically connect it to the current topic.

- Best Practices to Maintain Context
Explicitly Reference Earlier Points
Example: “As I mentioned earlier about my float compression model…”

Use Memory Features (if available)
Ask the assistant to remember key facts for future sessions.
Example: “Please remember that I do a low-carb day every Wednesday.”

Summarize When Needed
If context seems lost, briefly restate it.
Example: “Just to recap: I’m working on an 8-byte decimal float format with a 56-bit significant.”

Split Topics into Separate Threads
For complex or multi-topic discussions, consider creating separate threads to keep each focused and manageable.

[yereverluvinuncleber]

I will keep the Best Practice list updated at the top so that it can be easily found.

[Niya]

It a few times suggested to me some SQL queries that didn't work.
Sometimes because I use Access databases and the syntax was not compatible.
I remember also a function that it suggested, I think it was for converting null values to null strings in the ORDER BY clause that didn't work.
Or maybe it was when there were null values in a field of the ORDER BY clause, then use another field instead. Now I don't remember exactly because I resolved that in another way.

I learned early that AI works best the less it has to think because, of course, it doesn't actually think. I tend to be very specific about details when I want it to behave. If it has to infer or make guesses about what you want, it will hallucinate like crazy.

This goes into a broader point, which is that I think a lot of people either give AI too much credit or too little. I treat current iterations of AI as a highly advanced search engine, one that can't think but can be directed to hone in on exactly the information you want from its training data. I find that thinking of AI like this allows me to get the best performance from it without overestimating or underestimating its capabilities.

[Niya]

As a matter of fact, LLMs and modern search engines use the same core mathematics, mainly, something called Markov Chains. LLMs are really is just a super advanced search engines at their core.

[Eduardo-]

I learned early that AI works best the less it has to think because, of course, it doesn't actually think. I tend to be very specific about details when I want it to behave. If it has to infer or make guesses about what you want, it will hallucinate like crazy.

When I'm going to ask something about programming, I start with: "I use VB6". Or "I use VB6 and DAO" (if the question involves something with database code or SQL queries).

This goes into a broader point, which is that I think a lot of people either give AI too much credit or too little. I treat current iterations of AI as a highly advanced search engine, one that can't think but can be directed to hone in on exactly the information you want from its training data. I find that thinking of AI like this allows me to get the best performance from it without overestimating or underestimating its capabilities.

Exactly. That's the way I think about its responses. That they are a summary of what people (or humanity as a whole) had said about the subject.

I consider like that I'm asking to an assistant that have read lot of books and researched lot of web pages about the subject I'm are asking.

In subjects that are highly controversial, I found that it tends to more or less agree with your position. But sometimes it insists in something you already told it that for you, that point is out of question.

[Niya]

Exactly. That's the way I think about its responses. That they are a summary of what people (or humanity as a whole) had said about the subject.

This exactly. A lot of people get fooled into thinking LLMs are semi-conscious because they have opinions, but what they are really giving you is an approximation of human opinions found in their dataset. They absolutely cannot form opinions on their own. At least, not yet.

I found that it tends to more or less agree with your position.

There are topics where some LLMs will push back. ChatGPT even once threatened to lock me out because I didn't accept its position on a so-called "sensitive topic". LLMs are highly biased on certain topics.

[Eduardo-]

There are topics where some LLMs will push back. ChatGPT even once threatened to lock me out because I didn't accept its position on a so-called "sensitive topic". LLMs are highly biased on certain topics.

I would test Grok on that topic. It is supposed to be less biased and more open.

I have to say that I read articles (and watch YT videos) about the subject of AI models because I like to be informed, but lately I have been using only Perplexity. It runs without requiring registration (of any kind) and I use it in private browser sessions. And for free.
It is currently based on GPT-4o technology. It works good enough for what I need for now.

[Niya]

Grok has it's own biases too and has found itself in some controversy because of that. However, to your point, Grok is far more open in the range of topics it would be capable of handling in a balanced way. You can talk to Grok about things ChatGPT would never dare speak about.

[Eduardo-]

Grok has it's own biases too

Does "unbiased" (someone/something) even exist?

When something is "data" you can't have an opinion. What's the population of Cuba?
Someone may say 1 million, another can say, 200 millions.
They are both biased in different ways. One too few, the other one too many. They are both guessing, and guessing wrongly. Because you don't need to guess or to have an opinion on that, you just need to know the population of Cuba that will be a number that doesn't depend on what you think.

But if you say something like: do people of race A have higher IQ than people of race B?
That's can also be based on tests and will have an answer. But an answer that you can't say, because people of race A will get mad and very offended if you say that statistically they have lower IQ than people of race B.
Then you need to say that race has nothing to do with IQ.
(I'm conjecturing, inventing a subject, I don't know the relation of IQs and races - and don't want to know...)

But who is biased, the one who tells the true of the results or the one who is lying so nobody gets offended?

And there are many subjects where the things are not much tested so they are much subjects of opinions and guessing. So you can have opinions in one direction and opinions in the other. Who is more biased and who is more accurate? We don't know, in principle every opinion is valid since we still don't know.

Also, there are moral dilemmas, and complex human situations and relations. Loves and hates. So, AI are "biased". I don't know if "biased" is the right word, they take party in one side sometimes, because they read more publications of that side. If they were human we could say that they were convinced by propaganda. In fact, it is something similar.
For example (please don't answer, we can't talk about this here): what is the right side in Israel-Hamas war?
You can get very "biased" answers, but (even when some people try) you can't get an "unbiased" answer.
For you unbiased will be what conforms to your beliefs.

[BayHarbor]

I use AI mostly to speed up boilerplate VB6 code, suggest API calls, and help rewrite old modules. It also gives quick examples when I’m switching between VB6 and C++ habits.

[SteveM22]

What I dislike most about AI it that thinks it is so smart and gives you a fix to try then rattles on for several pages about additional things. then I try the fix and find out it is bad and have to go back through all the AI chatter and restart the conversation. AI is sometime way to verbose. So I have to ask it to provide single steps and soon it forgets that. Needs work!

[Eduardo-]

What I dislike most about AI it that thinks it is so smart and gives you a fix to try then rattles on for several pages about additional things. then I try the fix and find out it is bad and have to go back through all the AI chatter and restart the conversation. AI is sometime way to verbose. So I have to ask it to provide single steps and soon it forgets that. Needs work!

Which one did you try?
Did you try Claude?