Part 1: How LLMs work - tokens, vectors, attention

Module 10 · Sat 19 Sep

LLMs magic chat box illa. Token-based systems - context, prompting, product design ah poruthu behave pannum. Adhanaala PM ku basics theriyanum.

• Tokens dhaan real ah cost, processing unit. Token count ≠ word count - oru word pala tokens ah udaiyalaam.
• Vectors meaning ah represent pannura numbers; dimensions adhigam = meaning rich. Attention munnadi vandha mukkiyamana tokens ku weight kudutthu, plain vector ah contextual vector ah maathum.
• Model next-token prediction la, oru oru token ah ezhudhum.

Product levers: temperature (support, enterprise, factual assistant ku low; creative output ku high), max tokens (task ku etha maadhiri set pannu, default illa), input limits, model choice (perusu = capable, aana cost adhigam). Cost, latency, user need moonaiyum serthu yosi.

Adhiga context eppavum better illa - relevant context help pannum, noisy context kedukkum. Perusa documents ah paste pannaadha; relevant parts ah mattum retrieve pannu (class idha RAG, retrieval systems oda connect pannuchu). Business AI useful aagradhu domain context naala dhaan. Prototype easy - validation dhaan kashtam.