Turbovec and the trick of not training a quantizer
How TurboQuant uses a random rotation to precompute its quantizer, and why skipping the training step changes the operational story.
Posts with the tag aiml
All TagsPost List
- $Hadamard and the Random RotationHadamard and the Random RotationWhy a matrix of plus and minus ones does the work of a dense random rotation, in O(d log d) instead of O(d squared).
- $RAG vs MCP: Complementary AI ApproachesRAG vs MCP: Complementary AI ApproachesUnderstanding the differences between RAG and MCP, when to use each, and how they work together
- $Which Loss Function Do LLMs use?Which Loss Function Do LLMs use?Exploring Cross-Entropy Loss in Large Language Models.
- $What is Matryoshka Representation Learning (MRL)?What is Matryoshka Representation Learning (MRL)?Nesting Power and Flexibility into ML Embeddings
- $UE8M0 FP8 Number FormatUE8M0 FP8 Number FormatTraining LLMs without H100 using UE8M0 FP8 number format.
- $Understanding ML Numerical FormatsUnderstanding ML Numerical FormatsUnderstanding INT4, INT8, FP16, BF16, and TF32 formats in machine learning - their precision, speed, and memory trade-offs for training and inference.
- $What do GPT-OSS and Gemma 3 really offer?What do GPT-OSS and Gemma 3 really offer?GPT-OSS and Gemma 3: two new small-but-powerful language models pushing the boundaries.
- $What are Positional Embeddings?What are Positional Embeddings?The mathematical technique that teaches AI models where each word sits in a sequence.
- $Words, Tokens and EmbeddingsWords, Tokens and EmbeddingsHow language models convert token IDs into meaningful vector representations that capture semantic relationships.
