Noothi Akshith

Explains Tinker (a training API) and how techniques like LoRA enable parameter-efficient fine-tuning, allowing researchers to run large-model experiments from local workflows while retaining control over training loops and reproducibility.

Analyzes sources of nondeterminism in LLM inference (floating-point non-associativity and lack of batch invariance) and proposes engineering batch-invariant kernels (RMSNorm, matmul, attention) to achieve reproducible, deterministic outputs.

Discusses research showing that tiny subsets of parameters (sometimes a single weight) can disproportionately influence model behavior, and explores implications for interpretability, fine-tuning, robustness, and model compression.