I was hesitant to call this project smallest, but the fact that there is 
already a project called smallGPT that does not do justice at all or even is 
guaranteed to be GPT because he made it on a weekend, encourages me to the 
depth of hell to call my project something much much smaller and simpler. It's 
not my fault they made it so jumbled up to work on GPT.

I was hesitant to ask for money yet, because I should be able to do some of the 
simpler things, but after some thinking, all the things I need done go into the 
same bucket. It's a lot of work, and it's offshorable work. The bucket is an 
easy bucket that has a low pay. I need maybe 500USD to do the following jobs, 
so I can focus on my main meat of understanding it and building on it (and my 
original project at the same time). Usually I always get someone that can do 
the job, I simply am not a [pro] in this "area":



1) Find the GPT code that produces GPT-1 results, some implementations may use 
less layers or be kid-i-fied and may not give GPT-1 results. I want the GPT 
code in Python, there cannot be ANY C++ or C etc code it is calling or using 
like in tensorflow and pytorch etc [by the end of this project].

2) Get all needed code, Pytorch, etc, or the C++ if there is no complete python 
implementation in existence as of yet, and put it all into 1 folder, and put 
all of the code into 1 file, and remove all imports that call code like torch 
and tensorflow because we want to have /that/ larger amount of (usually C++) 
code in our single file and remove the code that calls other files and the code 
that calls and builds using tensorflow as well (those ex. 300 lines in 
smallGPT). Calling other code only makes code size larger.

3) Remove all comments in the code. A description can be made in notepad if 
need. Separate the functions in the code with 2 lines, this should only happen 
ex. 5-10 times in the code like where GPU code is, BPE, embeds, tokenization, 
ADAM, GELU, etc.

4) Make code smaller by removing self classes, etc. Put things into 1 large 
loop, so there is less def functions to achieve less code. GPT by the end of 
this project should be about 400 - 600 lines of Python code, currently the 
smaller implementation is still ~6000 lines of code, definitely not 300 lines 
like they say it is.

5) Make the project a 1 click press, so it runs and eats the dataset in the 
folder, there should not be any cmd or work involved, it should just run on GPU 
and run in my Pycharm IDE (which I love).
------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T74cd23135c964c0e-Me53efaf257ff471cbfd65b2c
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to