David

I’m a 16 year old design engineer • currently design engineer at Paliverse App

[David @akhmedbayev]

All right, people, let's start at the beginning one last time.

My name is David Akhmedbayev, I'm a 16 designer engineer from the East Coast.

Looking to connect with more designers, engineers, and founders.

[David @akhmedbayev]

I don’t know whether I should feel good or bad about myself. @XBOX

[David @akhmedbayev]

praying for my mac studio rn. about to train a model on it overnight on the cpu and gpu concurrently. force setting fans to 100%

[David @akhmedbayev]

why tf was i using pytorch. mlx is 2x faster, 1.1x more accurate, and uses 4.8x less memory. great job @Apple

[David @akhmedbayev]

pretty excited ngl

#RingsOfPower

[David @akhmedbayev]

rewriting my training loop with MLX instead of PyTorch. didn’t know i could use MLX for training but i guess you can. if anyone has any recommendation or experience i’d love to hear about it

[David @akhmedbayev]

speed testing my first model at 16

[David @akhmedbayev]

going to have to train my model at float16 instead of 32.

[David @akhmedbayev]

anyone wanna fund my model training ✌️

[David @akhmedbayev]

@vinvan how i get to do that someday

[David @akhmedbayev]

beginning to write the train loop for my first model

[David @akhmedbayev]

takes 15 mins to tokenize my training data on my M2 Max Mac Studio wonder how much faster it would be on a M5

[David @akhmedbayev]

@atlou_ @mainframe can’t wait

[David @akhmedbayev]

@vinvan @mainframe love it

[David @akhmedbayev]

@jsngr @mainframe can’t wait to try!

[David @akhmedbayev]

compiling a dataset for my first model at 16

excuse the loud fan sound in the video my M2 Max Mac Studio is like a 100°C rn

[David @akhmedbayev]

generating 100K training tokens for my first ai model at 16. it’ll take over an hour but it’s worth it.

[David @akhmedbayev]

in a typical text to text model the problem and answer is the current word and the next one which the model predicts. but a music to note works differently. it’s problem and answer is the inputed music and the actual note played. which means i need to create the music and answers.

[David @akhmedbayev]

here’s why is building a music to note model is harder than a typical llm…

[David @akhmedbayev]

@davidpukhaa wassup, lets connect

[David @akhmedbayev]

@thechainalhaji @hkofdesigners damn