I have coded a working AI similar to current AI, but here are the problems which come after coding

Started by SatanLucy

Replies
12
Posts
13
Page
1 / 1

Conversation

#1 •••

Coding an AI is relatively easy.


I used AI to code AI, and eventually ended up with working AI which uses its own model file, saves weights in model file, learns based on training data in txt file, with training process too shown clearly so I know how it progresses.


By using trial and error coding method, I simply tell AI to give me a working code for AI model, with correct instructions for vectors, weights, training data...ect.


then I test program. Identify errors (saving logs makes this easy). tell AI to fix error (one at a time). Until eventually I get a working program.


And eventually, I ended up with coding my own working AI model.


However, then I faced problem which cannot be solved by coding.


I make AI use .txt file for training, because .txt files are most lightweight.


the files in folder are basically:


AI.py

Model file (here weights are saved)

Knowledge.txt file

Save.txt (for logs for errors)

tokenizer file


the only file which is really big is model file where weights are saved.


the problem is, this AI entirely depends on knowledge.txt file (its training data).



If I make knowledge.txt file be filled with bunch of knowledge, then training process becomes extremely slow (we are talking about hours, to days, to weeks depending on size of txt file).


If I make knowledge file smaller, training is completed sooner, but AI becomes much dumber smaller model which cant answer most of questions coherently, and can only answer small knowledge in its training data, cant reason much either.


I assume all current AI companies face very much same issue. Larger training data takes much more time, creates a slower smarter model, while small training data creates faster, smaller, but dumber model.


Quality of training data is crucial, btw.

Edit comment

#2 •••

My AI seems to work good.


the size of AI is very small.


the code itself right now (Coded by chatgpt) is only 13 KB.


When it runs, it creates a model file.


Model file size is about 5 megabytes right now (very small model, works well, program code creates model automatically by given code. Increasing values creates larger model).

Edit post

#3 •••

When I increase number of model parameters to 200 million, model file size becomes about 400 megabytes.


the problem, however, is, when I increase number of parameters, the training becomes even slower. So while it becomes smarter, it also becomes a slower learner.

Edit post

#4 •••

this AI still uses weights. Later I will try to code AI which doesnt use weights but uses some other system which is hopefully nearly as good. the disadvantage of AI which uses weights is size and training time. Size greatly increases with increase in number of parameters. And training time too increases when parameters increase.

Edit post

#5 •••

I have added some optimizations so AI runs faster. Mainly changing model format type. the goal is to make it as fast as regular AI models. Right now it answers almost fast enough, however problem is this: training time is too slow. So I need improved model, for answers but especially for training time.

Edit post

#6 •••

It takes me about 8 hours to train AI with 20 epochs on a txt file of 800 KB (bunch of text).


I optimized training by using word level tokenizer instead of subcharacter level.


Also, my GPU cant support AI, so I made a CPU version which can. GPU version I made too, but it still uses CPU because my GPU too weak.

Edit post

#7 •••

tho I will try to program next one to only do 3 epochs, so I can train it on even larger text.

Edit post

#8 •••
@SatanLucy

What do you hope to achieve. AI will succeed with or without you.

Edit post

#9 •••
@Debby

I have coded a small 10 million parameters model which is not like AI today. It uses word level tokenizer, which improves response speed.


Right now I optimized it so 800 KB training size takes 5 minutes per 1 epoch (It used to take 15 to 20).


Here is it at work and chat:


Edit post

#10 •••
@Debby

the image shows model trained on 20 KB txt file size. For more coherence, more training text is needed.

Edit post

#11 •••
@SatanLucy

Looks very complicated compared to ChatGPT.

Edit post

#12 •••
@Debby

Well, it still needs proper user interface, but right now I am working on improving it. User interface can come later.

Edit post

#13 •••
@SatanLucy

Will wait for the final result. Very curious!

Edit post