Waa Maxay LLMs | Large Language Models | AI #3

Kani waa muuqaalka saddexaad ee taxanaha AI, waxaanan ku sharraxayaa sida LLMs-ku u shaqeeyaan. Waxaan ugu talagalay developer-ka iyo qof kasta oo maqlay Large Language Models oo doonaya faham sax ah.

LLM waa system AI ah oo lagu tababaray data qoraal ah oo aad u badan, kaas oo fahmaya oo soo saari kara qoraal bini'aadam u eg. Waxaan sharraxayaa saddexda dimension ee uu large ku noqday: cabbirka training data-da oo trillions of words ah, tirada parameters-ka laga bilaabo GPT-1 ilaa GPT-4, iyo computational scale-ka oo GPU-yo aad u badan iyo bilo shaqo ah qaadanaya. Kadib waxaan qaadayaa foundation models: hal tool oo shaqooyin badan qabta, halkii hore mid walba shaqo gooni ah u qaban jiray. Waxaan sharraxayaa Transformer-ka iyo attention mechanism-ka anigoo tusaale u soo qaadanaya miis toban qof ku fadhiyaan.

Waxaan kala saarayaa laba marxaladood oo training ah: pre-training oo internet-ka oo dhan la siiyo, iyo fine-tuning oo lagu daro aqoon ama dabeecad gaar ah. Waxaan sharraxayaa auto-regressive generation: model-ku hal eray ayuu mar walba saadaaliyaa, taasina waa sababta streaming-ku u jiro. Ugu dambayn waxaan isbarbardhigayaa open source iyo closed source: cost, performance, privacy iyo maintenance, iyo goorta mid walba loo doorto.

The third video in Dugsiiye's Somali-language (Af-Soomaali) AI series, an 18-minute explanation of how large language models work. It defines an LLM, then breaks down the three dimensions that make it 'large': training data measured in trillions of words, parameter counts from GPT-1 through GPT-4, and the computational scale of thousands of GPUs over months. It explains foundation models as one tool that handles many tasks, the Transformer and attention mechanism through a dinner-table analogy, and the two training phases of pre-training and fine-tuning, using Claude's coding strength and ChatGPT's conversational strength as examples. It then covers auto-regressive generation one word at a time and why responses stream, context limits and response variability, and closes by comparing open-source and closed-source models on cost, performance, privacy and maintenance. Free on YouTube.