From Newsgroup: comp.lang.forth


https://youtu.be/Dqa0ZxmXfE0

Hi, Terry the human here, no AI summary this time.

I had a couple of problems, one was my Deepseek bill which was averaging about $3.50 USD a day, dirt cheap for sure. but Im an Australian old age pensioner
so that's about $5.00 AUD a day, or $150 AUD a month. About 1.4x my food bill and well outside my budget.

Problem two was I was beginning to fight with Deepseek. It's incredibly powerful and wants to do everything, and frankly, it's an embedded Forth programming novice, it just does not have the training data as Forth is
very unpopular these days compared to C.

Sure, embedded Forth is ideal for AI, but the lack of training data cancels that out, well and truely.

So I needed to use my local Qwen3.8-27B-Q4_K_M.gguf agent to do a lot more:
a. Explain coding issues and methods to me because I'm not a programmer, I'm a electronics technician.
b. Write all my Wiki content in Fossil-scm markdown.
c. Use my logic analyser based on my description of what I want, and document the results.
d. Answer my questions with data from my databases, the source of my truth in all things STM32.

The big problem is this is a 28B-Q4 of about 19GB and after it's tuned to perfection by Deepseek, it only has about 16K of context available !

This is a total showstopper right ?

Actually I created a work around, the proof and the explanation is in the video. Let me know what you think in the comments ?

Cheers,
Terry

Disclaimer: Produced with the help of AI without which this video would not have been possible


--
Creator and maintainer of "The Unauthorised Mecrisp-Stellaris UserDoc (2014)" https://mecrisp-stellaris-folkdoc.sourceforge.io/
--- Synchronet 3.22a-Linux NewsLink 1.2