RightSignalLearn · track
How LLMs work
From a single neuron to the chatbots you use: how large language models learn, read text, write answers and run fast.
0 of 14 doneStart: Gradient descent
- Neural networksLayers of weighted sums that learn a task from examples · Coming soon
- Gradient descentHow a model nudges its weights to get a little less wrong each step · 6 min
- BackpropagationHow a network works out which weights caused its error · Coming soon
- TokenizationWhy models read text as chunks of words, not letters · Coming soon
- Word embeddingsHow words become lists of numbers that capture meaning · Coming soon
- AttentionHow each word decides which other words matter to it · Coming soon
- TransformersThe layer design behind GPT, Claude and Gemini · Coming soon
- Next-token predictionHow a model writes one token at a time, and what temperature does · Coming soon
- KV cacheWhy a model doesn't reread the whole chat for every new word · Coming soon
- QuantizationShrinking a model's numbers so it fits on smaller hardware · Coming soon
- Speculative decodingA small model drafts, the big model checks several words at once · 4 min
- Mixture of expertsHuge models that only use a small part of themselves per word · 5 min
- Retrieval-augmented generationLetting a model look things up before it answers · Coming soon
- RLHFHow human preferences turn a text predictor into a helpful assistant · Coming soon