forums.silverfrost.com

DanRRight · Posted: Wed Jun 25, 2025 9:42 pm Post subject:

Ok, let's teach AI a bit of Clearwin+ syntax and also possibly learn from it.

Grok learned from our previous communications that I like to see visualization in FTN95 and offered me without even me asking to plot in Clearwin some functions I was discussing with it. I was pleasantly surprised. Of course responded Yes. Here is the source code it generated.

I do not know where Grok learned all these Clearwin controls and how to use them given that it clearly has no idea about the major thing with Clearwin that all its variables have to be REAL*8 not just REAL.

PaulLaidler · Posted: Thu Jun 26, 2025 6:32 am Post subject:

There is no %pn in ClearWin+. It looks like it is intended to define a graphics region with given attributes and dimensions. Maybe there is some graphics interface other than ClearWin+ that has 'canvas' etc..

DanRRight · Posted: Thu Jun 26, 2025 6:11 pm Post subject:

Paul,
Thanks for info about %pn hallucination. And how about this line

DanRRight · Posted: Fri Jun 27, 2025 5:54 pm Post subject:

What i just realized is that most probably this Clearwin code was using older Simpleplot plotting %pl utility I completely forgot about. I rewrote all my codes to new so called native %pl long ago. There indeed potentially might existed these "canvas" and different orders of calling conventions. But as I remember still the variables X and Y there were of double precision, not just REAL like Grok supplied to me, am I right?

I need all that to "discuss" with AI how it must learn all modern Clearwin things and unlearn older %pl and some older Simpleplot commands. Better of course if Paul would do that otherwise there will be a holly mess of DBOS, Salford graphics library, Simpleplot, Clearwin and native Clearwin codes all conflicting with each other

Even better would be if we prepare kind of new Standard how we all want to see AI will supply graphics solutions to us in the future. It must be ultimately simple and professionally looking at the same time . If one curve on the plot will be needed - there will be one kind of simple FTN95 codes, if two curves on the plot need to be presented - another code, with 3D plots - third way. But definitely the gibberish like today with older Clearwin the AI must to unlearn

(I'd say that the graphics Grok, DeepSeek and ChatGPT provide right now with Python is looking really good and is very very simple so it will be difficult to Clearwin to compete with that)

PaulLaidler · Posted: Fri Jun 27, 2025 6:31 pm Post subject:

Dan

If what you are asking for is feasible then it is outside of my skill set.

I have looked for clues on the Internet but made no progress.

Kenneth_Smith · Posted: Sat Jun 28, 2025 3:31 pm Post subject:

This may provide a way of quickly generating an AI tool with domain specific information e.g. Clearwin+.

https://www.youtube.com/watch?v=BK2rcC6zEaU

https://pickaxe.co/

I converted a chm file for an application I have to pdf, and used the "free" model to create a chat for this software. When I then started asking questions I found modelling capabilities which I had never found in the last 30 years!

Only took me about an hour to work out the necessary steps to train the model.

I am sure there are other such tools.

DanRRight · Posted: Sun Jun 29, 2025 7:21 pm Post subject:

Please send the chm file also to all other AIs and tell them to learn from it and use only this latest knowledge for Silverfrost Fortran and Clearwin. That is exactly what i would do literally the first. Do AI actually need conversion of CHM to PDF for that?

I afraid though that the major AIs will not re-train from this information, the training is long and expensive process done under supervision and is also probably intended to be not free to the end user. Probably it will simply keep this information from the CHM file just for specific user as a so called "context training" and will immediately forget it for any other users. Pichaxe probably also learning under context training or using simple and free AI for that, the're progressing fast and becoming very capable.

Will look at Pichaxe and try it, good to have such hints, thanks. Can it be installed locally?

PaulLaidler · Posted: Mon Jun 30, 2025 7:03 am Post subject:

Online (html) documentation for ClearWin+ is available from https://www.silverfrost.com/ftn95-help/clearwinp/clearwin.aspx.

On the same page you will find legacy documentation for ClearWin+ Fortran in pdf form.

I guess that the installed text file, normally located at "C:\Program Files (x86)\Silverfrost\FTN95\doc\cwplus.enh", could also be uploaded to an AI tool.

Kenneth_Smith · Posted: Mon Jun 30, 2025 9:23 am Post subject:

Summarising my investigations over the past few days:

When you are using ChatGPT, on the sidebar menu select GTPs. This brings up available domain specific or �custom GTPs�, and under the programming tab you will find several such trained custom GTPs.

Of these, both Code Copilot and Code know about Clearwin+, and I have the impression that Code Copilot has more knowledge of Clearwin+ than Code. These custom GPTs have been developed by third parties, who have made the resource available by this publishing route.

Such custom GTPs don�t represent global training � just a custom layer on top of the base model.

It requires a subscription to have the capability to create these in ChatGPT. The basic process is:
1. ChatGPT Plus subscription (a recurring cost)
2. Create a custom GPT, define purpose, upload data, refine instructions, test and refine, and publish, making it available for wider use. This process is like that in the PICKAXE video. I have not tried to do this.

Using PICKAXE, an individual can create a domain specific chat for their own use via for free (limited number of queries per month, limited training data uploads). They could share the link to the PICKAXE chat for use by others, but that consumes their available query budget. The main use of such a personal GTP is that it can be trained on a modest data set and it gives you a very quick way to query that data.

Alternatively, an individual ChatGTP user using the �free� service could upload training material during a chat, and if their setting for Memory is �ON�, they can tell the GTP to remember this information. Each time the response to a query is wrong, you can correct it and specifically tell the GTP to remember not to make that mistake again. This seems to work as it is possible to look at a text description of the information in memory, and I can see mistakes I have told it to remember and not make again! Such saved memories are not forgotten over time, and carry forward to the next chat session. This is available on the �free� service. Memory is �OFF� by default, meaning that everything is forgotten over time � even in the same chat.

Saved memory is only accessible to the user, i.e. it does not become widely available. ChatGPT does not learn globally from your individual chats.

So if an individual or organisation has a ChatGTP subscription, they can create a custom GPT for any subject, and if feeling philanthropic could also select to make the custom GTP publicly available.

For lightweight collaboration between users on ChatGTP:
1. One person creates the Custom GPT and sets up a shared workspace (e.g. using Google Docs or similar).
2. Team members contribute feedback, prompt examples, and intended behaviours.
3. The owner manually updates the GPT's instructions or files with these contributions.
It�s a manual loop, but doable for lightweight collaboration.