Class 3 · 3.1

LLMs aslaan kaise kaam karte hain

Course ka woh aik wazan-daar theory hissa — aur faida bhi bara. Woh chaar tricks samajh lein jo “autocomplete on steroids” ko agent bana deti hain, to aage aane wali har technique samajh aa jayegi.

🧠 Maqsad: machine asal mein kya kar rahi hai yeh jaanein⏱ Parhein: 12 min👋 Experts: 2× welcome

Do seedhi seedhi warnings

Aaj zyada tar parhai hai, karna kam — aur yeh bunyadi cheez hai, to kuch log yeh pehle se jaante honge. Koi baat nahi: jaani pehchaani cheezein jaldi guzar lo. Lekin baaki sab kuch isi bunyad par khara hai, isliye agar yeh naya hai to yahan dheere chalein.

01LLM aslaan kya hai

GPT jaisa ek large language model ek kaam karne ke liye bana hai: aage kya aayega yeh andaza lagana. Aap usse input sequence dete hain, aur woh batata hai ke aage kya aane ka imkan hai. Yeh autocomplete on steroids hai — ek zabardast statistical pattern-matcher jo beshmar text par train kiya gaya hai.

Do precise baatein jo log aksar nazar andaz kar dete hain:

Input words nahi hote, tokens hote hain — kuch haroof ke tukde, kabhi poora lafz, kabhi uska kuch hissa.
Output “agla lafz” nahi — balke har mumkin aglay token ki probability hai. “Two plus two is” dene par “four” token ko bahut zyada probability milti hai; “bananas” ko na ke barabar.

LLMs ek token at a time generate karte hain, agla token predict karne ke liye ab tak sab kuch dobara parh kar.

LLM aur AI application — inhe alag rakhein

Ek jaisi lagti hain, par aham farq hai:

	LLM	AI Application
Misal	GPT	ChatGPT
Kya hai	Ek model jo agla token predict karta hai. Bas yahi.	Software jo LLM calls ke irda girda bana hai — memory, web search, tools, UI ke sath.
Isi tarah ke	—	Cursor agent · Duolingo Max · Atlassian Rovo

Jab se ChatGPT 2022 mein aaya, hum hairaan hain ke predictive text kitna samajhdar lag sakta hai. Yeh zahiri samajhdari software mein likhi char tricks ki wajah se hai jo raw next-token engine ke irda girda hain.

02Char tricks

Trick 1 — Yadasht ka dhoka

LLM ko har call bilkul stateless hoti hai. GPT ko “I’m Ed” kaho to woh kehta hai “Hi, Ed.” Dobara — bilkul nayi call — “Who am I?” kaho to woh kehta hai “I don’t know.” Pichli call ki koi yaad nahi.

To ChatGPT aapka naam kaise yaad rakhta hai? Ek chaalaki: har baar woh poori guzri baat-cheet wapas bhejta hai. “I’m Ed / Hi Ed / Who am I?” sab ek saath jaate hain — to aglay tokens ki zyada imkan wali baat ban jaati hai “You’re Ed, you just told me.” Yadasht ek dhoka hai jo har baar sab kuch dobara bhejna se banta hai.

Trick 2 — Reasoning (sochna)

Yeh ek daryaft se shuru hua: prompt mein “think step by step” lagane se behtar jawab aate the. Phir models ko jawab se pehle apni soch likhne par train kiya gaya. Ajeeb lagta hai, lekin approach describe karne wale tokens sach mein natija behtar karte hain.

Coin wali tricky sawal

“Aap do coin uchhalte hain; ek heads hai. Doosre ke tails hone ka kya chance hai?” Reasoning ke baghair model jhat se “aadha” bolta hai. Sochne ko kaho to woh trick pakad leta hai aur do-tihai jawab deta hai — kyunki aap ne nahi bataya ke kaun sa coin heads hai. Reasoning tokens hi wajah hain ke aaj ke models yeh sahi karte hain.

Trick 3 — Tools

Samjhein ke LLM ke tokens sirf text jawab nahi ho sakte — yeh koi kaam karne ki request bhi ho sakte hain. Prompt mein aap kehte hain: “Aap web search, calculator, ya Python chalane ke liye special tokens se jawab de sakte hain. Agar dein, to main chalaonga aur nataija lekar wapas aaonga.”

ChatGPT · ek chhota tool misal

# Aap question se pehle ek rule add karte hain:
To use Python to answer the next question, just reply
Python: and then a Python expression.

# Aapka sawal:
What is the square root of pi?

# Woh number nahi, yeh kehta hai:
Python: __import__('math').sqrt(math.pi)

Haqeeqat se naata na torein

LLM aslaan mein kabhi web search nahi karta ya code nahi chalata. Woh sirf tokens generate karta hai ke woh karna chahta hai. Har baar aapka software un tokens ko samjhta hai aur calculator chalata hai, phir nataija wapas bhejta hai. Isi poori dance ko “calling tools.” kehte hain.

Trick 4 — Loop

LLM ko ek baar bulane se behtar kya hai? Usse loop mein bulana: chalao, pocho “ho gaya?”, agar nahi to dobara bulao… aur dobara… jab tak maqsad na mila. Yahi seedha sa khayal agents ko ek single request se bahut badi cheezein hasil karne deta hai.

Char tricks

Autocomplete se agent tak

1 Yadasht ka dhoka · 2 Reasoning · 3 Tools · 4 Loops. Mil kar yeh 2022 ke ChatGPT se aaj ke Cursor aur Claude Code tak le jaate hain.

Behtareen tarif

Agent kya hai

Ek LLM jo maqsad hasil karne ke liye loop mein tools chalata hai. Simon Willison ne late 2025 mein maqbool kiya — yeh trick 3 aur trick 4 ko jodta hai.

03“Agent” ki tarif

“AI agent” kai cheezein matlab rakhta raha. Tarif badlti rahi:

Daur	Tarif	Source
Shuruwaat	“AI systems jo aap ke liye mustaqil kaam kar sakein” — yeh sirf bolta nahi, kaam karta hai.	OpenAI & doosre (misal: Operator / GPT Agent)
Early 2025	“Systems jahaan LLM workflow control karta hai” — uske output tokens tay karte hain aage kya hoga.	Hugging Face; Anthropic’s Building Effective Agents
Late 2025 →	“Ek LLM jo maqsad hasil karne ke liye loop mein tools chalata hai.”	Simon Willison (raaij tarif)

Aap ek dekh chuke hain

Kal ka Cursor game ek classic agent tha: aap ne usse maqsad diya (“web page mein first-person shooter”), woh clearly ek loop mein chala (files ek ke baad ek aati rahi), aur usne tools use kiye (files likhna, code chalana). Maqsad + tools + loop = agent.

✓ Aham baatein

LLM aglay token ki probability predict karta hai — aur kuch nahi. Yeh stateless hai.
LLM (GPT) ko application (ChatGPT) se alag samjhein jo uske irda girda bana hai.
Char tricks usse samjhdar mahsoos karati hain: yadasht ka dhoka, reasoning, tools, loops.
Agent ek LLM hai jo maqsad hasil karne ke liye loop mein tools chalata hai.