Hello! Got some updates on my AI in TurboWarp!
Here's a little live demo;
Yep, it's working real well and real fast!
I plan on optimizing even more, making it faster and letting me cram more power into this thing!
Alright, see you soon!
Yep, my TurboWarp/Scratch AI is finally done! Here's a little demo trained at 5k steps with around 2.0 loss. (if i remember correctly)
That's about 100 tokens in 7.23 seconds! This is a full transformer inference architecture, built on TurboWarp!
Yes, this might not be the first (Markov engines, RNN's LTSM's, etc) built in TurboWarp or Scratch, but it's literally the architecture inside most GPT's in the world! (ChatGPT, Claude, Gemini, etc.)
There are still some bugs (current responses are kinda quirky for some reason compared to old versions that spat out stuff like "persona a: i love to spend time with my family and friends too. persona b: that's a great way to spend a day in our family and friends" which is still kinda quirky, but that took like a good minute to generate.)
I'm not saying that this AI is better than any other AI in the website, this is just my shot on making an AI inside of Scratch/TurboWarp!
Now Ima' go on, fix bugs, and train in a chatbot dataset. Peace!
If you have any questions, drop em' below in the comments! I'll try my best to answer all of them.
This AI comes out September 15th!
Yep, my TurboWarp/Scratch AI is finally done! Here's a little demo trained at 5k steps with around 2.0 loss. (if i remember correctly)
That's about 100 tokens in 7.23 seconds! This is a full transformer inference architecture, built on TurboWarp!
Yes, this might not be the first (Markov engines, RNN's LTSM's, etc) built in TurboWarp or Scratch, but it's literally the architecture inside most GPT's in the world! (ChatGPT, Claude, Gemini, etc.)
There are still some bugs (current responses are kinda quirky for some reason compared to old versions that spat out stuff like "persona a: i love to spend time with my family and friends too. persona b: that's a great way to spend a day in our family and friends" which is still kinda quirky, but that took like a good minute to generate.)
I'm not saying that this AI is better than any other AI in the website, this is just my shot on making an AI inside of Scratch/TurboWarp!
Now Ima' go on, fix bugs, and train in a chatbot dataset. Peace!
If you got questions, drop em' in the comments! I'm all up for responding.
This AI is launching on September 15th.
My Scratch transformer (that uses an actual transformer architecture) is making good english words, but not good architecture. Took around 29.1 seconds for 500 tokens (or around 58 ms per token). Not bad!
My Scratch transformer (that uses an actual transformer architecture) is making good english words, but not good architecture. Took around 29.1 seconds for 500 tokens (or around 58 ms per token). Not bad!
Shakespeare-ish, also funny "Ay, that brocker I per?"
