GM
I'm hearing that people are having Claude talk like a caveman to reduce token spend. Seems to make sense -- make the output less verbose, more efficient. It's not a human after all.
We could also take this further, and perhaps have it use some sort of standardized shorthand that the computer uses. Maybe build in some additional looping so that when it's doing the same thing but slightly differently, it doesn't write out the whole thing each time. I'd call this something descriptive. Like, say "function."
I bet you could also save even more tokens by using the same type of shorthand for input, instead of natural language. Using this standardized shorthand (we could call it something like "code") to pass instructions to Claude. It may save more than tokens too -- I assume that less bandwidth would be used for uploading the instructions, and likely, less memory and cpu cycles too thanks to the smaller size. We might even be able to shave resources even further, using this standardization, by creating a new model -- one that doesn't use the inefficiencies of natural language, but instead, only parses this "code." This new model, being much smaller, wouldn't be appropriately called an LLM, due to not being large. It'd deserve a name more fitting. Something like "interpreter."
I bet if it were made small enough we could even run it locally on user machines -- even cell phones -- instead of in a datacenter somewhere. Allowing anyone to use it basically for free on their own hardware, privately. To run even faster for "code" we want to use often, we could further slim it down by having an "interpreter" convert it, or "compile" it, into an even shorter version, that only the computer itself will understand, but will completely dispense with the extra punctuation, so it runs even faster.
Man that'd be a really cool advancement in computing.
