For those of you who remmber the days of monolithic monster programs,you also remember the process of throwing more hardware at the problem. Then someone came along and asked "What if we broke it into chunks and then only load what we are using, when we are using it?
There was a time when the entire OS and the program you were running where both loaded into RAM and you ran from there. Limiting capabilities because hey 1MB of RAM was expensive.
We are reaching a similar point with LLMs. As useful and wonderful as they may be. The idea that you can just through a larger context window at it and it will solve your problem is proving false.It's being proven that the more you have the LLM actively remember, the dumber it can get.
This White Paper marks my research into this issue and what I've found out about how to maximize what an SLM can remember, preserving data integrity, performance,and keeping your data local rather than feeding it to the all consuming.
Feel free to give me feedback and thanks for reading.
For those of you who remmber the days of monolithic monster programs,you also remember the process of throwing more hardware at the problem. Then someone came along and asked "What if we broke it into chunks and then only load what we are using, when we are using it?
There was a time when the entire OS and the program you were running where both loaded into RAM and you ran from there. Limiting capabilities because hey 1MB of RAM was expensive.
We are reaching a similar point with LLMs. As useful and wonderful as they may be. The idea that you can just through a larger context window at it and it will solve your problem is proving false.It's being proven that the more you have the LLM actively remember, the dumber it can get.
This White Paper marks my research into this issue and what I've found out about how to maximize what an SLM can remember, preserving data integrity, performance,and keeping your data local rather than feeding it to the all consuming.
Feel free to give me feedback and thanks for reading.