Today: What are LLMs? from the Hugging Face Agents Course.
The part that stuck most was tokens — LLMs don't read text as words or characters, they read it as tokens, chunks of text (often sub-words) mapped to numbers via a tokenizer. Everything downstream — context length, cost, even weird model behavior on rare words — traces back to this.
This actually pulled up an old memory: a tokenization module from an NLP class back in my master's, well before LLMs were the default lens for "NLP." Same core idea, different era. Nice to see the thread connect — it means I already had the vocabulary, just not the framing for how central tokens are to how modern LLMs actually operate.