Any AI-integration with VAST on the roadmap?

[Matthew Boulter]

Hi, 

   I appreciate marrying the two together might be a tall order and perhaps it doesn't
meet ROI metrics, but it would be lovely to have a Github Copilot-esque way to discuss 
our enormous legacy code base with an LLM.

I have had encouraging results filing-out the two-dozen odd large applications, each into their own application-named folder and then each class as individual .ST files, that together we pack as our "Application" that we ship.

I put all of that into a VsCode project, I did install the most popular Smalltalk extension to at least get syntax highlights, etc.

When I challenged it with a range of requests, Copilot did a 70% job of knowing what it was doing.

The clincher was the other 30% ... which was it hallucinating classes that do not exist in the VASmalltalk Abt-world.

It got the intent right for the solution but missed the mark on the correct class to use.

I'm not familiar with any dialect other than VASmalltalk ... but it was so specific I assumed that it was trained on another dialects core classes, but I know LLMs are exceedingly good at hallucinating "convincingly".

It was so really close that I couldn't help but wonder if Instantiations may go down that route and train an LLM or perhaps VASmalltalk MCP server? 

Feels like it is not that far to reach to get there... just need some boffins smarter than me to do it :)

These are interesting times indeed.

Merry Christmas to all.

Regards Mattew.

[Joachim Tuchel]

Hi there,

I don't have an answer. But I have a comment ;-)

I find it interesting that most of the times when people ask about the combination of Smalltalk or VAST with AI they talk about a coding assistant. I'd be much more interested in marrying VAST with an AI model in a way resembling what Jupyter Notebooks do. Looking at a Jupyter Notebook it really feels like the 21st century version of a Smalltalk workspace. It is a list of snippets combined with a bit of documentation and the display of results all in one place. It can be saved and loaded back and thus be stored in a vcs. A complete AI workbench in a single window. It's a REÜPL for AI training and training data preparation.
A Jupyter Notebook is what a Workspace would look like if it were invented today. 

What if we had a Jupyter Notebook-like Workspace to prepare and massage training data, start training and keep an eye on the training results etc. right in our Smalltalk environment? And all that without the hassle of installing Anaconda and whatnot on our machines?

I guess this would be a real boost to our applications and projects. 

So yes, having more assistance in coding Smalltalk would be funny and sometimes even helpful. Maybe even in educating yound developers. But that is like stopping right in front of the curtain instead of moving it aside and make a step into "the real thing". 

Joachim

[Marcus Wagner]

I came from AI (heuristic search) in research and from a highly sensitive business area with strict regulations (Financial industry, somehow Fintech predecessors). In developing business software,  Smalltalk played an important role for me. That period ended about 20 years ago, my activities changed recently to consult govermental services. 

Given my personal bias, as long as there is no convincing concept to avoid unwanted transfer of information, any remote AI support is a potential risk. Insofar the absence of AI support up to now was a benefit - to pinpoint, even access to online help or to any public repository was regularily denied.

Besides of that, LLMs are based on propabilities (of matching solutions to a problem). Answers from there come from a distribution of gathered solutions (which came from training or previous questions) without any established (human) quality control.

If coding is seen similar as writing text, having something like high-lighting or correcting spelling orders, this is ok. 

However, if coding has to provide solutions to problems in the reality (beginning in the design, architecture and higher levels), any inclusion of remote AI support implies to establish professional quality control and acceptance of  information transfer to the public, that means loss of data confidentiality.

Everybody has to decide the boundary where this implied information transfer has to stop. Personally, with my limited experience and confrontations with AI generated codings, I stay sceptical to any public cloud based AI integration in software building tools.

[Greg Schultz]

Greetings All,

It's been a productive and very busy end-of-year for us, so I wanted to jump back into this thread and share an update on AI and VAST.

Long story short: Instantiations is already investing heavily in AI. We currently have multiple initiatives underway, including bringing Copilot-like assistance to VAST using popular models such as Gemini and ChatGPT. Beyond that, our broader roadmap includes a custom trained LLM for VAST, an MCP server, knowledge system capabilities, and several additional projects that we will talk about as they reach the right stage for public discussion. For those who attended Smalltalks 2025 in Argentina, some of this will sound familiar from our Company Update session.

As with most VAST features, our goal is flexibility. Some tools will integrate with widely used AI models and be more open and exploratory. Others are being designed with enterprise needs in mind, with a focus on privacy, data sovereignty, and the requirements that come with operating in highly regulated environments.

We're excited about the direction of this work and look forward to sharing more as features become ready for broader discussion and release.

Kind regards, and Happy Holidays to Everyone!

Greg Schultz Executive Vice President - Growth & Operations  gsch...@instantiations.com  /gjschultz  instantiations.com

--
You received this message because you are subscribed to the Google Groups "VAST Community Forum" group.
To unsubscribe from this group and stop receiving emails from it, send an email to va-smalltalk...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/va-smalltalk/76643e94-bf71-4539-9be9-8ed7c63549fan%40googlegroups.com.