Julia Turc

AI researcher and content creator specializing in efficient machine learning techniques, particularly around model compression and training optimization.

  • Video: How does 4bit quantisation work (2026-04-14) - Discusses evolution of training large-language-models with reduced precision, focusing on 4-bit floating-point (FP4) training challenges. Highlights cost of training LLMs (Gemini Ultra: ~78M in 2023) and shift toward Quantisation techniques to reduce computational demands.
  • Key insight: Reducing precision from 16-bit to 4-bit significantly lowers training costs while maintaining model performance through advanced Quantisation methods.

2026 04 14 How does 4bit quantisation work

Source Notes