Success in AI coding: Coherent AI model with size of 50 megabytes and 10 million parameters. | Forum

#1

the reason why this counts as success is:

this AI model is incredibly fast. the response you see in image, it generated all in less than 2 seconds
Model uses word-level tokenizer (Current big AI models use subcharacter level)
Model trains extremely fast and works on small training data. After few minutes of training on text with 20 KB, and reaching loss function under 0.01, it became coherent.
Model is coherent and acts as AI which gives answers it learned from training data.
this model is only 50 megabytes in size, compared to usual local models which take gigabytes.
Model is faster than any local model in existence.

Edit comment

Mentions Text*

<img src="https://debatecraft.s3.amazonaws.com/uploads/247e4f2b-3dae-4f31-8783-afde129235dc.png"> the reason why this counts as success is:<ol><li data-list="ordered">this AI model is incredibly fast. the response you see in image, it generated all in less than 2 seconds</li><li data-list="ordered">Model uses word-level tokenizer (Current big AI models use subcharacter level)</li><li data-list="ordered">Model trains extremely fast and works on small training data. After few minutes of training on text with 20 KB, and reaching loss function under 0.01, it became coherent.</li><li data-list="ordered">Model is coherent and acts as AI which gives answers it learned from training data.</li><li data-list="ordered">this model is only 50 megabytes in size, compared to usual local models which take gigabytes.</li><li data-list="ordered">Model is faster than any local model in existence.</li></ol>

#2

Also, another reason why this is very special.

Program is very small size, few KB. It doesnt require downloading a model. It creates a model on its own based on parameters which user inputs. then it trains a model based on data in knowledge.txt file which user provides. It also enables constant training, so each epoch can be trained separately. User can train 1 epoch today, then turn off program, and train another epoch tomorrow or any time, and not lose any training progress.

Also, program is "all in one". It isnt divided on separate programs for training and chatting.

I am still working on improvements, to optimize program even further.

I was using float32, but now I learned maybe float8 or float16 is better for even more speed.

Also, I made 300 million parameters model as well which became more coherent after being trained only twice on 20 kb file txt.

Edit post

Mentions Text*

Also, another reason why this is very special. Program is very small size, few KB. It doesnt require downloading a model. It creates a model on its own based on parameters which user inputs. then it trains a model based on data in knowledge.txt file which user provides. It also enables constant training, so each epoch can be trained separately. User can train 1 epoch today, then turn off program, and train another epoch tomorrow or any time, and not lose any training progress. Also, program is "all in one". It isnt divided on separate programs for training and chatting. I am still working on improvements, to optimize program even further. I was using float32, but now I learned maybe float8 or float16 is better for even more speed. Also, I made 300 million parameters model as well which became more coherent after being trained only twice on 20 kb file txt. <img src="https://debatecraft.s3.amazonaws.com/uploads/8d268dba-4090-4ed0-bf37-f37949672bce.png">

#3

Right now I am trying to optimize it somehow. I made model with 1 billion parameters, but my laptop can't keep up anymore. It works, but training is very slow, even if responses are much more coherent than in 300 million parameter model.

So it looks like I will have to stick to maybe 100 million parameters for speed and efficiency on laptop, or buy a better laptop.

Tho from what I understand, it takes a supercomputer to properly train AI.

I wonder if there even are any other people in Bosnia who code AI.

Edit post

Mentions Text*

#4

And about other type of chatbot, vector based word level tokenizer retriever.

I finally made it work with pdf documents for chatting with books.

Edit post

Mentions Text*

#5

@SatanLucy

Meta and Amazon are laying off thousands of employees because of AI.

Edit post

Mentions Text*

#6

@Debby

Meta and Amazon are laying off thousands of employees because of AI

Who cares?

Edit post

Mentions Text*

#7

@SatanLucy

What guarantees you a future in AI. With the layoffs news?

Edit post

Mentions Text*

#8

@Debby

the only job I have is the one I own, so I cant really be fired unless I fire myself.

Edit post

Mentions Text*

#9

@SatanLucy

Didn’t you say you run a chicken farm selling eggs?

Edit post

Mentions Text*

#10

@Debby

Didn’t you say you run a chicken farm selling eggs?

yes

Edit post

Mentions Text*

#11

@SatanLucy

How will AI coding help you handle chicken eggs?

Edit post

Mentions Text*

Conversation

Edit comment

Edit post

Edit post

Edit post

Edit post

Edit post

Edit post

Edit post

Edit post

Edit post

Edit post