no-js web accessible help

[kdmoyers]

The crashing weight of the llm revolution/tragedy has come to my company too, and it is a shame that the llms seem to have trouble accessing help info for winbatch. 

If you start at https://docs.winbatch.com/ you get the cool left side bar interface, which seems to baffle the llms.  I don't think they do the cool interface thing.

If you start at https://docs.winbatch.com/Contents.htm you can't actually penetrate to the content. there just don't seem to be any links inward.

I've seen https://docs.winbatch.com/mergedProjects/WindowsInterfaceLanguage/html/HTMLWIL_WIL001.htm
mentioned -- is this the best no-js entry point to recommend to the llms?

This icky future is here, and I hate to see us drift away from winbatch simply because the llms don't know it.  It doesn't matter to me, I already know it, but the junior programmers are lazy...

thanks in advance,
Kirby

[td]

Are you sure that JavaScript is the problem? It could just be the delay before the context sidebar is populated.

The big boys scan this Website and the Techsupport Website constantly, and they seem to grasp the WinBatch basics. I have been having some fun developing my own machine learning system, but have not tackled training the model with WinBatch scripts and documentation yet. Probability and information theory are enough to try to wrap my head around for now...

[td]

I didn't really answer your question so no guarantees but maybe give this link a try.

https://docs.winbatch.com/mergedProjects/WindowsInterfaceLanguage/html/HTMLWIL_M__003.htm

[bottomleypotts]

LLMs aren't coding because they are reading the manual. They're coding because they have examples. Until there is a tonne of well coded and commented on examples of winbatch, LLMs won't be able to help us.

[td]

Generally, training machine learning models involves both ingesting documentation and examples. It can be very tedious because coding examples should have an associated prompt. At the very least, the examples should contain quality code comments.

[bottomleypotts]

It's true for the instruction-tuning / supervised fine-tuning stage.

However, that paired data is not where most of an LLM's coding ability comes from. That's built during pre-training on the massive self-supervised phase where the model simply predicts the next token across trillions of tokens of raw internet sourced data from GitHub repos, Stack Overflow threads, notebooks, etc.

[td]

Perhaps read my previous post a little more carefully. Also, most programming language documentation contains examples. That is true of WinBatch documentation. And models can be adjusted once created with settings like top-t, top-p, tau, eta, mu, etc. These settings adjust the accuracy and performance. Fun math.

Anthropics Claude AI Chatbot will tell you that training for model updates comes from "publicly indexed places (forums, GitHub, official docs)."

[bottomleypotts]

I did read your post carefully.

The vast majority of an LLM's coding ability does not come from the handful of paired prompt+example snippets you see in official docs or during supervised fine-tuning. That stage is small.

What actually teaches the model how to code is the pre-training phase — next-token prediction on trillions of raw tokens scraped from the entire internet: GitHub repos, Stack Overflow, notebooks, forums, etc. That's where the model learns patterns, idioms, error handling, best practices, etc. for popular languages.

WinBatch is extremely obscure.

• It has almost zero presence on GitHub (the "winbatch" topic barely exists and has virtually no stars or activity).
• Stack Overflow has a handful of ancient questions.
• The official WinBatch site and tech support database have maybe a couple thousand examples total.

That is a rounding error compared to the scale of pre-training data. A few dozen (or even a few thousand) examples in the docs are nowhere near enough for the model to actually learn the language the way it learned Python, JavaScript, or even PowerShell.

The sampling settings you mentioned (top-p, temperature, top-k, etc.) are inference knobs — they only change how the already-trained model picks the next token. They don't add new knowledge.

And yes, Claude is correct that training data comes from "publicly indexed places." The problem is WinBatch simply isn't there in any meaningful quantity. That's exactly why LLMs suck at it.

So no — the official docs with their limited examples are not going to magically give the model WinBatch superpowers. We'd need orders of magnitude more well-commented, real-world WinBatch code in the wild for that to happen.

[spl]

Very interesting discussion. But, you know me... I go on tangents. Training LLM to wrote code {correctly} can be a holy grail, but I have become interested in Local LLM to digest schema information from a database [a post with code I made a couple of months ago for making schemas]... and then ask as a 'bot' to answer questions about the data in natural language. I am using "llama3.1" as the model, and the requests are from a local api - "http://localhost:11434/api/generate" - all done via HTTP request. The request could be made from WB via a CLR System.Net.HttpWebRequest, or possibly the COM WinHttp. The issue is streaming the result content so now using PS and the ConvertFrom-Json cmdlet. But as both Tony and BP noted - lot of behind the scenes work required. Designing and training a db to respond to natural questions like "Who had the least leads from our ads last week and why do you think that is?" without an analyst crunching data in Excel... scary, but not difficult.

[td]

I did read your post carefully.

The vast majority of an LLM's coding ability does not come from the handful of paired prompt+example snippets you see in official docs or during supervised fine-tuning. That stage is small.

What actually teaches the model how to code is the pre-training phase — next-token prediction on trillions of raw tokens scraped from the entire internet: GitHub repos, Stack Overflow, notebooks, forums, etc. That's where the model learns patterns, idioms, error handling, best practices, etc. for popular languages.

WinBatch is extremely obscure.

• It has almost zero presence on GitHub (the "winbatch" topic barely exists and has virtually no stars or activity).
• Stack Overflow has a handful of ancient questions.
• The official WinBatch site and tech support database have maybe a couple thousand examples total.

That is a rounding error compared to the scale of pre-training data. A few dozen (or even a few thousand) examples in the docs are nowhere near enough for the model to actually learn the language the way it learned Python, JavaScript, or even PowerShell.

The sampling settings you mentioned (top-p, temperature, top-k, etc.) are inference knobs — they only change how the already-trained model picks the next token. They don't add new knowledge.

And yes, Claude is correct that training data comes from "publicly indexed places." The problem is WinBatch simply isn't there in any meaningful quantity. That's exactly why LLMs suck at it.

So no — the official docs with their limited examples are not going to magically give the model WinBatch superpowers. We'd need orders of magnitude more well-commented, real-world WinBatch code in the wild for that to happen.

Me thinks you claim too much and too little at the same time.

[kdmoyers]

Thanks Tony!

[kdmoyers]

One more question: is there a link I could point the llms to that exposes the tech database in an llm friendly format?   How would this do?

https://techsupt.winbatch.com/webcgi/webbatch.exe?techsupt/nftechsupt.web+WinBatch

But maybe I should not suggest the link, because so much of the code in there is kinda old? 
Not sure.  What do you think?

[td]

Since you ask, I don't think the age of the examples matters much as long as the syntax is correct. I am not sure what the best approach to the tech database is. I would need to get back to you on that. I do know that several bots regularly navigate the tech database on their own. For example, Yandex, OpenAI, and Amazon bots are navigating as I write this. How deeply the bots are traversing the site is a question I haven't checked into yet.

Since WIL has a Lua-like syntax, it shouldn't take too many examples to train a model on basic syntax. For example, Anthropic's chatbot suggests that 500-1000 good examples are sufficient and better than 100,000 random ones. Reportedly, synthetic data generation is commonly used to train models on programming languages. Basically, AI trains AI. This would need techbro by/in to accomplish, and what could possibly go wrong? <grin>

My views on machine learning have changed a bit. I am still very sceptical, but my youngest has shown me what task specific models can do in computational biology, genetics, immunology, and several other fields. 

I learn by doing, so I have an AI-related side project made up of many smaller tasks. The results have almost no probability of seeing the light of day, but I might learn something in the doing.

And I apologize for wandering off topic again.

[bottomleypotts]

Me thinks you claim too much and too little at the same time.

I'm not sure what you're saying here. But good chat!

Basically, AI trains AI.

The big one everyone talks about. But when models train recursively on their own (or prior gen) outputs, they lose diversity, amplify subtle errors, and get narrower/confident-but-wrong. This has been documented since the 2023-2024 papers and is still a live concern in 2026 discussions. Good in theory.

My views on machine learning have changed a bit. I am still very sceptical, but my youngest has shown me what task specific models can do in computational biology, genetics, immunology, and several other fields.

Biology has clearer "ground truth" + faster, cheaper validation loops. That's why were making progress in this field.

Coding is open-ended engineering, not prediction. WinBatch is the poster child for the "long tail" problem in AI coding. Almost zero public training data. There is no verification loop payoff!

[td]

One more question: is there a link I could point the llms to that exposes the tech database in an llm friendly format?   How would this do?

https://techsupt.winbatch.com/webcgi/webbatch.exe?techsupt/nftechsupt.web+WinBatch

But maybe I should not suggest the link, because so much of the code in there is kinda old? 
Not sure.  What do you think?

The ChatGPT bot is crawling articles this morning. The bot is harvesting each topic without issue. I also need to mention that the Website throttles bots a bit because of CPU usage spikes. The big boys know better, but the Python based "kiddy scrappers" not so much.

[kdmoyers]

Thanks Tony!!  We're all new at this stuff.

[td]

I'm not sure what you're saying here. But good chat!

Maybe an AI model will explain it.

The big one everyone talks about. But when models train recursively on their own (or prior gen) outputs, they lose diversity, amplify subtle errors, and get narrower/confident-but-wrong. This has been documented since the 2023-2024 papers and is still a live concern in 2026 discussions. Good in theory.

Like I said, "what could possibly go wrong?" We are familiar with the chatter around the topic.

Biology has clearer "ground truth" + faster, cheaper validation loops. That's why were making progress in this field.

Coding is open-ended engineering, not prediction. WinBatch is the poster child for the "long tail" problem in AI coding. Almost zero public training data. There is no verification loop payoff!

As my youngest, who works in both fields, might tell you, coding is more deterministic than the biological sciences.

I have a hypothesis concerning the coding part that I will be testing when time permits.

[td]

Thanks Tony!!  We're all new at this stuff.

That's the fun part. An opportunity to experience the awe and aha moments of learning.  I have been doing a lot of learning myself since I started down the machine learning rabbit hole.

[bottomleypotts]

Looking forward to seeing how your hypothesis turns out.

Your youngest is right — coding execution is indeed more deterministic than biological systems.

The real bottleneck has never been that part.

It's the open-ended engineering judgment, the extreme long-tail data scarcity (WinBatch is the textbook example), and the complete lack of a clean, objective verification loop like biology gets from physics and wet labs.

That's exactly why AI is shipping real, usable results in computational biology but still produces absolute slop when asked to code.

Keep us all informed when you finally run your tests.

[td]

I don't entirely agree with your assertions about AI success in science. I think there is more nuance than that. Based on personal experience, I have not found AI code generation to be as good as many claim. I am not referring to WinBatch here, but I do see a pattern of bias. That said, I do find at least one chatbot useful for conducting code reviews, and that is a very useful thing. 

[spl]

Our daughter teaches advanced biology but is aware of students using AI for their tests. So (this is true) for one student a test question was to describe a test to determine if a restaurant was serving top shelf fish or disguised Tilapia - but write it as a Pirate - sure enough the student, without reading just uploaded the whole question... and the results came back starting with "Aaarghh, matey... ya have to take the fish and..." and she submitted it verbatim.

[td]

That anecdote made my day.

[bottomleypotts]

I don't entirely agree with your assertions about AI success in science.

Chai-2 designs novel antibodies and miniproteins from scratch, achieving a 16% hit rate (over 100× better than prior methods) and picomolar-affinity binders that succeed in wet-lab assays within two weeks.

Evo 2 accurately predicts the functional impact of genetic variants across the genome.

AlphaFold 3 models complex protein interactions and is now used by major pharma companies for drug optimization.

Several AI-designed molecules have reached Phase II/III trials with higher early-stage success rates. Insilico Medicine's Rentosertib came up with the first fully AI-designed and AI-discovered drug.

AI-designed molecules are hitting 80–90% success rates in Phase I (safety and tolerability in healthy volunteers).
Historical industry average is only ~40–65%. Rentosertib??

The key advantage is biology's fast, objective wet-lab validation loop — something coding currently lacks.

You cannot argue with the math.

Based on personal experience, I have not found AI code generation to be as good as many claim.

I agree. I am hearing about AI coded crap all the time. Code clearup is discovering the same function coded differently in hundreds of places in the same project. AI isn't going to code a complex app like SAP anytime soon.

Not saying it doesn't have usecases either. AI is now discovering serious zero-day vulnerabilities in critical codebases that had evaded human experts for years.

AISLE identified all 12 new OpenSSL zero-days patched in January 2026 and was responsible for 13 of the 14 recent OpenSSL CVEs.

Mythos uncovered thousands more, including a 27-year-old flaw in OpenBSD, a 17-year-old remote code execution bug in FreeBSD, and a 16-year-old vulnerability in FFmpeg.

Although I'd hate to know how many vulnerabilities are being introduced now because of sloppy AI use.

[spl]

That anecdote made my day.

Christ. I laughed for 5 minutes....

[td]

Chai-2 designs novel antibodies and miniproteins from scratch, achieving a 16% hit rate (over 100× better than prior methods) and picomolar-affinity binders that succeed in wet-lab assays within two weeks.

Evo 2 accurately predicts the functional impact of genetic variants across the genome.

AlphaFold 3 models complex protein interactions and is now used by major pharma companies for drug optimization.

Several AI-designed molecules have reached Phase II/III trials with higher early-stage success rates. Insilico Medicine's Rentosertib came up with the first fully AI-designed and AI-discovered drug.

AI-designed molecules are hitting 80–90% success rates in Phase I (safety and tolerability in healthy volunteers).
Historical industry average is only ~40–65%. Rentosertib??

The key advantage is biology's fast, objective wet-lab validation loop — something coding currently lacks.

You cannot argue with the math.

Funny you should mention Alphafold. The original pre-Alphabet layered transformer protein fold model was developed about 15 miles from where I am sitting.

I am well aware of the successes, but they come with some caveats. When base models are used, they may or may not be publicly available. But either way, they almost always have target specific training for the task. Also, not all AI generated outcomes are as easily verified as your examples might suggest.