Lesson 9 — Beginner

What an LLM Actually Is

LLM ka matlab Large Language Model — ek bada model jo text padh kar agla token predict karta hai aur isi predict, choose, append ke loop se poora jawab bana deta hai. Is lesson mein hum samjhenge ye hota kya hai, ye kya nahi hai, aur ek hi model se writing se coding tak sab kaise ho jaata hai.

⏱ 10 min• Beginner• Basic computer usage

LLM kya hai

WHAT

Large Language Model ek bada neural network hai jo text ke pattern seekhta hai aur phir har step mein yeh predict karta hai ki agla token (shabd ka tukda) kya hona chahiye. Poora jawab isi ek simple kaam ke baar-baar hone se banta hai.

WHEN

Jab input bhi text ho aur chahiye bhi text hi — summary likhwana, email draft, translation, classification, code likhna, sawaal ka jawaab. LLM kaam tab aata hai jab andar-undar likhna, samajhna ya badalna ho.

WHERE

Chat apps, coding assistants, support-ticket bots, search ke chhote summaries, document drafting tools — aaj kal har AI feature ke andar koi na koi LLM chhupa hota hai. Tum jis bhi app mein type karke jawaab pate ho, peeche wahi model hai.

Sabse pehle ek saaf idea bana lo: LLM ek text-janane wali machine hai, yaad-rakhne wali machine nahi. Ye apne andar laakho sawaal-jawaab store karke nahi baitha — ye pattern se naya text banata hai, har baar, fresh.

Yaad rakho: LLM = Large Language Model. Large ka matlab hai model ke andar laakho-crore parameters (sankhya ke knobs) hain, aur Language ka matlab hai uska poora khel shabdon ke saath chalta hai — input bhi token, output bhi token.

Next-token prediction — poore model ka dil

Prompt: "Bharat ki rajdhani"

Step 1  PREDICT  → agla token ki probability list banao
          New Delhi : 0.35   Bombay : 0.10   Kolkata : 0.08 ...
Step 2  CHOOSE   → list mein se ek token chuno (sample)
          → "New"
Step 3  APPEND   → "Bharat ki rajdhani New"
Step 4  REPEAT   → predict → choose → append … jab tak stop na ho

Final output: "Bharat ki rajdhani New Delhi hai."

Predict: model agla token ki poori probability distribution nikalta hai — har vocabulary token ke liye ek score.
Choose: us distribution se ek token chunta hai (kaise chunta hai — temperature, top-p — agle lesson mein).
Append: chuna hua token already-generated text mein jud jaata hai, aur model ka agla prediction isi badhe hue text par chalta hai.
Repeat: yahi loop tab tak chalta hai jab tak ek stop-token na aa jaaye ya limit na poore ho jaaye. Ek chhota sa ek-step machine — crore baar dohraya gaya.

Interview line: Ek line mein poochha jaaye toh bolo — an LLM generates text by repeatedly predicting the most appropriate next token given the tokens it has already seen. Generation ka poora khel isi ek sentence mein chhupa hai.

LLM kya NAHI hai

Database nahi — ismein jawaab store hoke nahi baithe. Tumhara sawaal naya ho toh bhi model pattern se jawaab bana dega.
Search engine nahi — ye web pages crawl karke result list nahi deta. Google links deta hai; LLM ek single generated jawaab deta hai.
Deterministic program nahi — C jaisa program hamesha same input pe same output deta hai. LLM ko thoda randomness ke saath chalaya jaata hai, isliye same prompt ke alag-alag runs mein alag wording aa sakti hai.
Store of saved answers nahi — isliye isse poochho "tumhe yaad hai?" matlab sense nahi; jo context mein diya gaya wahi uski bas ki baat hai.

Ye confusion isliye hoti hai kyunki LLM bohot confident dhang se bolta hai. Confidence aur accuracy alag cheezein hain — model apne tone se aisa lagata hai jaise sab kuch jaanta ho, aur yahi baat aage failure-modes lesson mein hamari problem banegi.

Ek hi model, kai tasks

Ek hi LLM summarization bhi kar leta hai, translation bhi, code bhi — kyunki task ka naam usse prompt mein hi mil jaata hai. Model ne training mein dekha hai ki "summarize" wale instructions ke baad chhota jawab aata hai, "translate" ke baad doosri bhasha aati hai. Instruction hi uska switch hai.

# Wahi model, alag-alag prompts — alag tasks
prompt1 = "Is email ko 2 line mein summarize karo: ..."
prompt2 = "Is paragraph ko English se Hindi mein translate karo: ..."
prompt3 = "Neeche diya function likho jo list ka average de: ..."
prompt4 = "Is review ka sentiment batao: POSITIVE ya NEGATIVE"

# Har task ke liye alag model ki zarurat nahi —
# instruction badlo, kaam badal jaata hai.

Classification: label set do, ek label chuno — jaise sentiment, ticket category.
Transformation: rewriting, tone change, grammar fix — input ka hi ek naya roop.
Generation: email, code, poem — naya content jo training data mein bilkul nahi tha.
Adaptation: prompt mein 2-3 example de do (few-shot) toh model tumhara custom format bhi seekh jaata hai — bina kisi code change ke.

Capability kahan se aati hai

Data: jitna bada, saaf aur diverse text corpus, utne zyada patterns — grammar, facts, code syntax, reasoning ke tarike sab usi se nikalte hain.
Compute: laakho-crore examples par bade batches mein train karna possible hi tab hai jab GPUs ki barish ho.
Scale: zyada parameters matlab zyada capacity ki patterns ko store aur combine kiya ja sake (lekin sirf size kaap nahi — M1.3 mein dekhenge).
Training improvements: sirf pretraining nahi — SFT, preference tuning jaise steps model ko usable banate hain (M1.4 mein poora pipeline).

Pro Tip: Naya model aaye toh model-card padho — kab train hua, kis data pe, kya post-training kiya gaya. Yehi teen cheezein bataati hain ki uski capability kahan se aayi hai aur kahan tootegi.

Exercise

Question: LLM har step mein exactly kya predict karta hai? (do shabd)

Question: Same prompt do baar dene par output bilkul same aayega ya variation aa sakta hai? (ek shabd)

Question: LLM ko database bol sakte hain — haan ya na?

Common mistakes

LLM ko database ya search engine samajhna: Ismein saved answers nahi hain — search engine links deta hai, LLM ek naya generated jawaab deta hai. Poochhne ka sahi sawaal hai "kya ye is conditions mein reliably karta hai", "kya iske paas ye jawab hai" nahi.
Deterministic program ki tarah expect karna: Hamesha same output ki guarantee nahi — sampling ke chalte wording badalti rehti hai. Testing mein exact-string match mat karo; structure aur rubric se check karo.
Har task ke liye alag model dhoondna: Ek hi model instruction se task badal deta hai. Alag model tabhi chuniye jab capability, cost ya latency ka farq ho — kewal task-name ke liye nahi.

Desi example: Shayar ka khel — generation ka desi analogy

LLM ko kitaab mat samjho, ek shayar samjho — jo har shabd ke baad agla shabd apni feel se jodta hai. Shabd judte jaate hain, aur ek poori ghazal ban jaati hai. Usne ratta nahi maara — usne laakho shabdon ka pattern (andaaz) seekha hai:

Bharat ki rajdhani → [New: 0.35, Bombay: 0.10, …] → "New"
"Bharat ki rajdhani New" → [Delhi: 0.90, …]     → "Delhi"
"Bharat ki rajdhani New Delhi" → [hai: 0.80, …]  → "hai"

Shayar ne agla shabd chuna, jod diya, dohraya — ghazal tayyar.
LLM bhi yahi karta hai — bas andaz thoda mathematical hai.

Interview angle — yahan se sawaal aate hain

Q: LLM kaam kaise karta hai ek line mein? — Wo repeatedly predict karta hai ki agla token kya hona chahiye, uske paas jo tokens pehle se hain unhe dekh kar.
Q: LLM aur search engine ka farq kya hai? — Search engine stored pages se links deta hai; LLM ek naya generated jawaab banata hai, store se nahi, pattern se.
Q: Ek hi model summarization aur coding kaise kar leta hai? — Task ka naam aur tarika usse prompt mein hi mil jaata hai; instruction hi switch hai, alag-alag task-specific models ki zarurat nahi.
Q: LLM ko database kyun nahi bolte? — Database mein exact stored records hote hain; LLM har run mein text generate karta hai, isi liye variation aur hallucination dono possible hain.

Lesson complete?

LLM ka basic idea clear — ek next-token prediction machine jo crore baar dohra kar poora jawab banati hai. Agla lesson — Tokenization — mein dekhenge model text ko kaise todkar numbers banata hai aur usse cost kaise judti hai.