is an optimized communication language better for training and inference?

[ouvaa]

let's say we

1. reduce the letters to 13 instead of 26, that makes the requirement to process at least halved.
2. reduce base words for faster processing.
3. remove plurals, one word per form.

let's say we have this new language, will we have faster inference / training speed?
https://github.com/ouvaa/ouvaa.github.io

[BrickBee]

reduce the letters to 13 instead of 26, that makes the requirement to process at least halved.

That's not how it works. When you copy your full posting here then you'll see that most words get replaced by a single token, a single symbol. The number of characters per word does not matter, if you exclude misspelled words.

reduce base words for faster processing.

Not necessarily faster processing, but most importantly also faster training. While a more minimalist language could help a bit in that regard, it's not needed. Check out the TinyStories paper that followed a similar approach with plain English. The resulting model is so tiny that it fits into the CPU cache and is thus extremely fast.

The description of the language in your git repo is incomplete. Furthermore, you will not be able to train a model without sufficient text in that designed language. Once you have that, you could use train-text-from-scratch from llama.cpp to see what happens. As far as I remember dropout wasn't implemented yet, so you might want to use another framework if you do not have a lot of training data.

[ouvaa]

@BrickBee yes u are right, i understand what u mean.
i was thinking about doing text translation into the new language without typos (using llama.cpp).

there are other areas other than just pure training that can be assisted with a language whereby we control and limit the parameters of how it is used. which can be highly benefitial in being cross used in CPU or old school search engines.

can you give reference to "dropout"? what does that mean? the training data will depend upon translating existing english corpus to the new language.

[ouvaa]

looking for contributors to improve the new minimalist language. imagine the GBs we can save with a universal unambigiuos minimal language designed just for AI-human communication

[ouvaa]

@BrickBee i've updated https://ouvaa.com/, pls take a look into it.

The description of the language in your git repo is incomplete.

can you pls give me feedback on this rudimentary version of what i have done? what else do you think i need to improve upon other than having a large examples, dictionary words etc? (these i can get chatgpt to generate)

Anything else u think i will need to enhance for llama training etc?