Yatharth Anand | ML Engineer
Hello, I’m Yatharth. 🚀
I am an ML Engineer specializing in Large Language Model (LLM) internals, mechanistic interpretability, and low-level optimization. My work focuses on deconstructing the “black box” of neural networks to understand how they reason and where they break.
Currently, I’m deep-diving into Llama 3 on Apple Silicon using the MLX framework.
🔬 Featured Projects
Llama 3 MLX Research Lab
A dedicated lab for deconstructing Llama 3. Key discoveries include the “Tipping Point” of quantized weights and manual “brain surgery” on token embeddings.
- Tech: MLX, Python, Llama 3, Metal.
LLM Decoding Strategies
Exploring how different sampling methods (Contrastive Search, Top-P, Min-P) can “rescue” models with physical weight corruption.
- Tech: MLX, NumPy, Logit Manipulation.
✍️ Latest Blog Posts
The Tipping Point: Identifying the Threshold of Quantized Weight Corruption
🚀 Objective To determine the precise mathematical threshold at which uniform bit-level perturbations to 4-bit quantized LLM weights cause catastrophic model collapse. We aim to understand how robust these compressed models are to “bit-level noise.”...
🛠️ Toolkit
Frameworks: MLX, PyTorch, Transformers, JAX
Hardware: Apple M-Series (Metal Performance Shaders)
Interests: Mechanistic Interpretability, Quantization, KV-Cache Optimization, DPO.