Skip to content
View YZ-LLM's full-sized avatar
🌍
🌍

Block or report YZ-LLM

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
YZ-LLM/README.md

From backend developer to AI researcher. Building analytical alternatives to backpropagation.

Original Contributions

🧮 BackLoss — Single-formula weight computation without gradient calculation. 22x faster than backprop in classification, 99.77% on MNIST (world #7).

🔗 PropLoss — Multi-layer gradient-free learning system. Outperforms backprop by 50.9 percentage points at 6 layers. pip install proploss

⚡ LazySmart Optimizer — Combines Shampoo-lite + Lookahead + P3. Beats AdamW by 14.4%. Trains Shakespeare language models on CPU in 120 seconds.

📐 Backloss Gradient Scaling — Analytical gradient correction for transformer linear layers. Derived from linear regression closed-form solution: grad_new = grad / (n × |x|)

🚀 ChunkBLAS Linear Attention — BLAS-friendly alternative to softmax attention. 42% faster on CPU at ctx=512 with O(n) complexity.

Research Areas

Transformer optimization · Gradient-free learning · CPU-first AI training · Small language models (SLM)

Repositories

"Intelligence is sometimes simplicity. My formula proves it."

Pinned Loading

  1. backloss-for-classifier backloss-for-classifier Public

    Python

  2. proploss-classifier proploss-classifier Public

    Python

  3. nanoGPT_more_acc nanoGPT_more_acc Public

    Forked from karpathy/nanoGPT

    The simplest, fastest repository for training/finetuning medium-sized GPTs.

    Python

  4. GPT_at_CPU GPT_at_CPU Public

    Python

  5. MyCL-Studio MyCL-Studio Public

    TypeScript 6