AI Asal Mein Kya Hai: Ek Crash Course
9 Ideas, No Math, No Code: woh aik cheez jo baqi paanch courses farz kar lete hain ke aap pehle se samajhte hain.
Aap car chala sakte hain bina yeh jaane ke engine kya hai. Zyada tar log aisa hi karte hain. Lekin jis lamhe kuch galat hota hai (ek aawaz, ek warning light, dhalaan par engine band ho jaana) woh log jo taqreeban jante hain ke hood ke neeche kya hai, calm rehte hain, aur jo nahin jante, woh ghabra jate hain. Woh ek bekhauf khatkhatahat aur ek seize-shuda engine mein farq nahin bata sakte, kyunke un ke liye poori machine ek opaque box hai jo ya to chalti hai ya nahin chalti.
Yeh hi zyada tar logon ka AI ke saath taalluq hai. Unhon ne isay chalana seekh liya hai (baqi paanch Foundations courses aap ko sachmuch acha driver bana dete hain) lekin unhon ne ek baar bhi hood ke neeche nahin dekha. Is liye jab machine koi ajeeb cheez karti hai (ek source ijaad kar leti hai, khud se ulta keh deti hai, kisi bilkul ghalat baat par bilkul yaqeen se baat karti hai), un ke paas kyun ka koi model nahin hota, aur woh ya to is par hadd se zyada bharosa kar lete hain ya isay raddi qarar de dete hain. Dono react ek hi jagah se aate hain: yeh na jaanna ke yeh cheez asal mein hai kya.
Yeh course hood ke neeche ek nazar hai. Mechanic wali nazar nahin: yahan koi math nahin, koi code nahin, koi neural-network diagrams nahin jinhein aap ko decode karna pare. Sirf woh nau ideas jo AI ki har heran kar dene wali harkat ko samjhate hain, taake failures rahasya rehna chhor dein aur predictable ban jayein. Jab aap failures predict kar sako, to aap unhein avoid kar sakte hain, aur yeh hi poora payoff hai.
Chhe Foundations courses mein se, yeh woh hai jo baqiyon se pehle parhna chahiye, chahe yeh sab se abstract hai. 2026 mein AI Prompting, Markdown In, HTML Out, Code You Never Write, Skills & Connectors, aur AI ke Daur Mein Sochna sab aap ko machine kaise use karni hai sikhate hain. Yeh sab machine kya hai ke baare mein kuch facts par tikte hain ("yeh stateless hai," "yeh predict karti hai, lookup nahin karti," "yeh ghalat hone par bhi confident hoti hai") aur un facts ko ek-ek jumle mein, guzarte hue keh dete hain. Yeh course woh jagah hai jahan se woh jumle aate hain. Isay ek baar parhein, aur baqi paanch courses ka har "yeh aisa kyun karti hai?" ka jawab pehle se taiyar mil jata hai.
Kuch topics is course aur 2026 mein AI Prompting dono mein aate hain, soch samajh kar, repetition se nahin. Yeh course mechanism deta hai (ek explanation, phir aage barh jata hai); prompting course practice deta hai (aadatein, tafseel se). Jahan dono milte hain:
| Topic | Yahan (machine) | 2026 mein AI Prompting (aadat) |
|---|---|---|
| Yeh kya jaanti hai | Seekhna kyun freeze hua (Idea 2) | Woh knowledge kitni reliable hai, topic-bah-topic (Concept 2) |
| Context window | Yeh kyun sirf wohi hai jo model dekhta hai (Idea 5) | Isay kaise manage aur protect karna (Concept 4) |
| Confidence | Yeh kyun yaqeen se boli hai aur aap se agree karti hai (Idea 6) | Isay kaise neutralize karna (Concept 6) |
| Reasoning | "Thinking" asal mein kya hai (Idea 9) | Isay kab on karna, aur kab nahin (Concept 5) |
| Images aur audio | Yeh kyun bas mazeed tokens hain (Idea 4) | In ke saath asal mein kaam kaise karna (Concept 8) |
Rule of thumb: jis lamhe yahan koi section aap ko ek aadat sikhane lagta hai, woh ruk jata hai aur aap ko prompting course ki taraf bhej deta hai. Wohi handoff dono ke darmiyan ki lakeer hai.
📚 Teaching Aid
Poori Presentation Dekhein: AI Asal Mein Kya Hai
Do minute mein sabit karein
Kisi bhi explanation se pehle, machine ko aise behave karte dekhein jo sirf tab samajh aata hai jab aap jaante hain ke yeh hai kya. Claude.ai, ChatGPT, ya Gemini kholein (free account banane mein ek minute lagta hai) aur bilkul yeh paste karein, jaan boojh kar ki gayi galti spelling ke saath:
Without using any tools, just from memory: how many times does the
letter R appear in the word "strawberry"? Then spell the word out
one letter at a time and count again.
Dekhein kya ho sakta hai: pehle pass mein kuch models ab bhi ghalat ginti karte hain, phir jis lamhe woh harf-bah-harf spell karte hain usi waqt theek kar lete hain. Ek machine jo aap ko ek working program likh sakti hai, woh ek chhe-harf wale lafz mein harf reliably nahin gin sakti, jab tak aap usay lafz ko tukron mein torne par majboor na karein. Yeh bewaqoofi nahin hai. Yeh is course ke sab se ahem fact ka ek seedha, nazar aane wala nateeja hai: model harf nahin dekhta. Yeh tokens dekhta hai (Idea 4). Lafz pehle se chunks mein kata hua aata hai, aur ek chunk ke andar ke harf ginna is ke liye sachmuch mushkil hai, bilkul aise jaise kisi imarat ki khirkiyan ginna mushkil hai agar kisi ne aap ko sirf us imarat ka street address dikhaya ho.
Do minute, ek ajeeb harkat, aur aap poore page ka theme pehle hi mil chuke hain: AI ki har heran kar dene wali harkat is se samajh aati hai ke yeh asal mein hai kya, na ke is se ke yeh smart hai ya dumb. Neeche di gayi nau ideas us aik misaal ko ek mukammal model mein badal deti hain.
Part 1: Machine
Teen ideas is baare mein ke jab aap send dabate hain to literal taur par kya ho raha hai. In ko samajh lein aur AI ke do-tihai behavior ka heran kar dena khatam ho jata hai.
1. Yeh text ka agla tukra predict karta hai; yeh cheezein lookup nahin karta
Yeh hai woh aik jumla jisay aap ke zehan mein sach banane ke liye yeh poora course bana hai: ek language model ek aisi machine hai jo, kuch text de kar, predict karti hai ke aage sab se zyada plausible kya text aata hai, ek waqt mein ek chhota tukra. Yeh poora core mechanism hai. Baqi sab is ka nateeja hai.
Is par thora ruk kar sochna zaroori hai ke yeh kitna ajeeb hai, kyunke yeh bilkul waisa nahin jaisa zyada tar log farz karte hain. Zyada tar log farz karte hain ke AI ek bohat tez librarian ki tarah kaam karti hai: aap sawal poochte hain, woh kisi vishaal internal encyclopedia mein relevant fact dhoondti hai, aur aap ko parh kar suna deti hai. Woh mental model ghalat hai, aur logon ki AI ke saath ki gayi taqreeban har galti isi tak ja pohanchti hai.
Asal mein jo hota hai woh duniya ke sab se zyada parhe-likhe autocomplete ke qareeb hai. Aap ne autocomplete ko "Happy birthday to..." ko "you" se mukammal karte dekha hai. Ek language model wohi move karta hai, lekin itne text par train kiya gaya hai ke woh kisi bhi prompt ko jaari rakh sakta hai, sirf aam jumlon ko nahin, aur woh isay ek lafz nahin balke ek token ke hisab se jaari rakhta hai (Idea 4), har tukra jo woh produce karta hai usay wapas khud mein feed kar ke agla tukra tay karta hai. Isay France ka capital poochein aur yeh France → Paris labelled koi database row lookup nahin karta. Yeh woh continuation produce karta hai jo, har us cheez mein jo us ne parhi, "The capital of France is" ke baad sab se zyada plausible aata hai, aur woh ittefaq se "Paris" hai, kyunke yeh sequence is ke training text mein lakhon baar aaya tha.
Ghisi-pitti facts ke liye, prediction aur lookup ek hi jawab dete hain, is liye farq academic lagta hai. Yeh academic rehna chhor deta hai jis lamhe text patla hota hai:
- France ka capital poochein → plausible continuation hi sacha continuation hai. Prediction knowledge jaisi lagti hai.
- Ek self-published novel ka plot poochein jo chand sau copies bika aur online kabhi review na hua → koi ghisa-pita continuation nahin hai, is liye model sab se zyada plausible-lagne wala continuation produce karta hai aise books ko blend kar ke jo milti-julti lagti hain. Yeh ab bhi predict hi kar raha hai. Bas is ke paas predict karne ke liye koi sachi cheez nahin hai.
Machine dono soorton mein bilkul yek-saan kaam kar rahi hai. Sirf aap farq bata sakte hain, aur sirf tab jab aap jaante hon ke yeh kya kar rahi hai.
Librarian ka tasawwur chhor dein jo retrieve karta hai. Ek writer ka tasawwur shuru karein jo jaari rakhta hai. Ek librarian jisay book na mile woh kehta hai "humare paas woh nahin hai." Ek writer jisay story jaari rakhne ko kaha jaye woh kabhi yeh check karne nahin rukta ke continuation sach hai ya nahin. Jaari rakhna hi poora kaam hai. Is liye AI kabhi yeh nahin kehti "mere paas woh nahin hai" jaise librarian kehta, jab tak isay khaas taur par aisa karne ke liye train na kiya gaya ho. Plausible continuation is ka native amal hai; sach is ke oopar lagayi gayi cheez hai, naqis taur par.
Ek hi sawal har baar mukhtalif jawab kyun deta hai
Machine ek next token predict nahin karta. Yeh plausible next tokens ka poora spread predict karta hai, har ek likelihood ke saath ("Paris" bohat likely, "the largest city in France" mumkin, ek darjan aur kam hote hue) aur phir us spread mein se ek ko pick karta hai. Yeh kitni jurat se pick karta hai is ko ek setting control karti hai jisay aam taur par temperature kehte hain: low temperature isay taqreeban hamesha sab se likely token lene par majboor karti hai (steady, repetitive), high temperature isay kam likely tokens tak pohanchne deti hai (variyd, zyada creative, kabhi-kabhi behka hua). Zyada tar chat products ek darmiyani value set karte hain, isi liye ek hi sawal do baar poochne par aap ko do mukhtalif alfaaz wale jawab milte hain jin ka matlab taqreeban ek jaisa hota hai. Yeh variation model ka "apna mind badalna" nahin hai. Yeh predictions ka wohi spread hai, do baar sample kiya gaya. (Yeh hi wajah hai ke ek aise task ke liye jahan aap har baar bilkul wohi output chahte hain, aap isay kabhi chat interface se nahin nikal sakte: dice pehle se baked hain.)
Yeh 2026 mein AI Prompting (Concept 2) ke "frequency equals reliability" rule ki mechanical jar hai. Ab aap jaante hain ke frequency reliability ke barabar kyun hai: ek sacha continuation training text mein jitni zyada baar aaya, machine usay utni hi shiddat se predict karta hai. Patla topic, kamzor prediction, confident-lagne wala andaza.
Product kar sakti hai; model ab bhi nahin karta. Modern tools predictor ko extras mein lapet dete hain, web search, file reading, code execution, yahan tak ke ek memory note (Idea 2), aur woh extras asal, current facts la sakte hain. Lekin facts context window mein land kar ke aate hain (Idea 5), aur model un ko ab bhi sirf usi tareeqe se jawab mein badalta hai jo woh kar sakta hai: un se ek continuation predict kar ke. Is liye "yeh predict karti hai, lookup nahin karti" beech wali machine ke baare mein sach rehta hai, chahe is ke ird-gird ka system abhi-abhi is ke liye kuch lookup kar chuka ho. Idea 8 woh jagah hai jahan tools theek se aate hain.
2. Yeh parh kar seekhi, aur phir seekhna ruk gaya
Predictions kahan se aaye? Training se: model ko insani text ki ek behad bari miqdaar dikhayi gayi (books, articles, code, forums, reference works) aur us ne khud ko baar baar adjust kiya, taake us text ka agla tukra behtar predict kar sake. Woh adjustment process hi woh waqt hai jab model kabhi kuch "seekhta" hai. Jab training khatam hoti hai, to nateeja ek tay-shuda internal numbers ke set mein freeze ho jata hai (engineers isay weights ya parameters kehte hain) jo dobara nahin badalta.
Do alfaaz isay theek samjha dete hain, aur inhein alag rakhna zaroori hai kyunke farq bohat kuch samjhata hai:
- Training ek dafa ki taleem hai, ek baar, maazi mein, us company ke zariye jis ne model banaya. Mehnga, slow, mukammal.
- Inference woh hai jo har baar hota hai jab aap isay use karte hain: frozen weights aap ke prompt par chal kar ek continuation predict karte hain. Tez, sasta, aur (yeh ahem hissa hai) yeh model ke andar kuch nahin badalta.
Jab aap conversation mein model ko theek karte hain aur woh kehta hai "aap theek hain, meri galti," us ne kuch seekha nahin hai. Us ne woh text predict kiya hai jo ek correction ke baad plausible taur par aata hai. Chat band karein aur agli conversation usi yek-saan frozen weights se shuru hoti hai, jisay yeh yaad nahin ke correction kabhi hua tha. Aap ne isay nahin sikhaya. Aap isay sikha hi nahin sakte. Sirf agla training run weights badal sakta hai, aur aap us ka hissa nahin hain.
Do nataij seedhe isi se nikalte hain:
| Nateeja | Yeh frozen weights se kyun nikalta hai |
|---|---|
| Knowledge cutoff. | Training ek mukarrar date par khatam hui; us ke baad jo kuch hua woh bas weights mein nahin hai. Model, hamesha ke liye, ek shandaar expert hai jis ne ek khaas din news parhna band kar diya. |
| Yeh aap ki private duniya nahin jaan sakta. | Aap ki company ke numbers, aap ka calendar, kal ki email kabhi training text mein nahin thay, is liye weights mein un ke baare mein kuch nahin hai. Model chhupa nahin raha; woh maloomat freeze karne ke liye wahan thi hi nahin. |
Kuch products ab ek "memory" pesh karte hain jo chats ke darmiyan aap ko yaad rakhta lagta hai. Yeh weights ko nahin badalta. Woh inference time par naamumkin rehta hai. Is ke bajaye jo hota hai: product khamoshi se aap ke baare mein chand facts text ke taur par save kar leta hai aur har nayi conversation ke shuru mein us text ko context (Idea 5) mein dobara daal deta hai. Yeh model ka yaad rakhna nahin; yeh product ka isay ek note dobara feed karna hai. Mufeed, lekin mechanically yeh context hai, insani matlab mein memory nahin. Yeh farq jaanna hi woh cheez hai jo model ke baqi behavior ko predictable rakhti hai.
Yeh "stateless" ki mechanical jar hai, woh lafz jo 2026 mein AI Prompting Concept 4 mein istemaal karta hai. Stateless ka matlab: apni koi memory nahin, har response sifr se compute hota hai frozen weights aur jo kuch abhi is ke saamne hai us se.
Woh alfaaz jo aap sunenge ("parameters," "mixture of experts," "quantization") aur in mein se koi bhi nau ideas ko kyun nahin badalta
Jaise jaise aap AI ke baare mein parhenge aap weights kaise bante hain ke liye alfaaz ki ek qatar se milenge. Teen sab se aam, ek-ek line:
- Parameters (jisay weights bhi kehte hain): is idea ke frozen numbers. "Ek 400-billion-parameter model" bas un ko ginta hai. Zyada ka aam taur par matlab zyada capable aur chalane mein zyada mehnga; yeh us baat ko nahin badalta ke numbers karte kya hain.
- Mixture of experts (MoE): un parameters ko aise tarteeb dene ka tareeqa ke kisi bhi token ke liye un ka sirf ek hissa on hota hai, har baar sab ke bajaye. Yeh ek bohat bare model ko tez aur sasta chalata hai. Bahar se machine ab bhi bilkul ek hi cheez karti hai: agla tukra predict karna (Idea 1).
- Quantization: numbers ko kam precision par store karna taake model chhote, saste hardware par fit ho jaye. Wahi behavior, halka footprint.
Pattern hi asal baat hai. Yeh sab "machine kaise banti hai aur sasti banti hai" ka jawab dete hain, "machine karti kya hai" ka nahin. Nau ideas mein se har ek (prediction, frozen weights, koi truth-checker nahin, tokens, context, confidence, jaggedness, tools, thinking) yek-saan taur par sach rehta hai chahe model dense ho ya mixture-of-experts, full-precision ho ya quantized, saat billion parameters ho ya saat sau. Is liye jab koi headline kehta hai ke ek naya model "MoE use karta hai" ya "ek trillion parameters rakhta hai," ab aap jaante hain ke is ka matlab kya hai aur yeh ke yeh us mein kuch nahin badalta jo aap ko ek user ke taur par karna hai. Woh developments jo sachmuch badalte hain ke aap machine ke saath kaise kaam karte hain woh hain jo yeh course cover karta hai: reasoning modes (Idea 9), tools (Idea 8), aur lamba context (Idea 5).
3. Koi alag jagah nahin jahan yeh check kare ke sach hai ya nahin
Ideas 1 aur 2 ko milayein aur aap us fact tak pohanchte hain jo logon ke liye sab se zyada paresan kun behavior samjhata hai. Ek insani expert ke paas do alag faculties hoti hain: ek jo jawab generate karti hai, aur ek doosri, khamosh, jo usay check karti hai: "ruko, kya mujhe is par yaqeen hai? maine yeh kahan se seekha? kya yeh theek lagta hai?" Dono aapas mein ikhtilaf kar sakti hain. Aap kuch zubaan se keh sakte hain aur usi saans mein mehsoos kar sakte hain ke yeh ghalat ho sakta hai.
Model ke paas sirf pehli faculty hai. Is ke andar koi doosri machine nahin jo aap tak pohanchne se pehle prediction ka sach ke liye audit kare. Wohi ek process jo ek sahi continuation produce karta hai ek ghalat bhi produce karta hai, in ke darmiyan farq karne wale kisi internal flag ke baghair. Fluent, achi tarah banaya gaya, confident jumla wohi output hai chahe under-lying prediction training text se achi tarah supported tha ya hawa se kheencha gaya. Fluency machinery se produce hoti hai; sach isay alag se verify nahin karta.
Yeh hi woh cheez hai jis ki taraf log ishara karte hain jab woh kehte hain ke AI hallucinate karti hai: yeh fluent, confident, bilkul ghalat bayanaat produce karti hai. Lafz isay aisa zaahir karta hai jaise koi kharabi ho, ek glitch jisay theek karna hai. Yeh glitch nahin hai. Yeh machine ka bilkul waise kaam karna hai jaise yeh bani: ek plausible continuation predict karna, ek aisi jagah jahan plausible continuation ittefaq se sach nahin hota. Ek bina madad wala model jo kabhi hallucinate na karta woh bilkul ek mukhtalif qism ki machine hoti, jis ke ird-gird asal retrieval, verification, ya inkaar karne ki salaahiyat bani hoti. Oopar lagaye gaye tools aur checks (Idea 8) kam karte hain ke yeh kitni baar hota hai; woh beech wali cheez ki fitrat nahin badalte, jis ka poora kaam jaari rakhna hai.
Model ka confident lehja is baat ka saboot nahin ke yeh theek hai. Lehja ek style hai jo us ne confident insani writing se seekha (mazeed Idea 6 mein); yeh usi process se generate hota hai jis se content, aur sach se utna hi alag hai. Ek banayi gayi statistic bilkul usi pur-yaqeen aawaz mein aati hai jis mein ek asal. Yeh hi poori wajah hai ke AI ke Daur Mein Sochna course maujood hai: is ki Error Taxonomy (Discipline 3) ek checklist hai un ghalat continuations ko, haath se, pakarne ke liye jinhein machine khud apne liye pakar nahin sakti. Aap woh ghaib doosri faculty hain.
Ek non-software misaal. Ek walid ne AI se ek khaas chhoti tuition academy ki theek fee schedule aur class timings poochein, ek aisi jis ki na koi website thi aur taqreeban koi online maujoodgi nahin. AI ne courses, timings, aur monthly fees ka ek confident, qareeene se banaya gaya table produce kiya. Har figure ijaad ki gayi thi. AI ne jhoot nahin bola tha aur kharabi nahin ki thi. Academy kisi bhi training text mein mushkil se maujood thi, is liye predict karne ke liye koi asal schedule nahin tha, aur uske baghair, machine ne wohi kiya jo woh kar sakti hai: us ne sab se zyada plausible-lagne wali fees produce ki jo aisi academy le sakti thi, usi confident aawaz mein bichha kar jo woh tasdeeq-shuda facts ke liye use karti hai. Is ke paas koi doosri faculty nahin thi jo sargoshi kare "aap andaza laga rahe hain." Woh sargoshi aap ki taraf se aani hai.
Part 2: Yeh aise behave kyun karti hai
Chaar ideas jo ajeeb behaviors (harf ginna, memory khatam ho jaana, yaqeen se bolna, ek hi saans mein shandaar aur bekaar hona) ko un cheezon mein badal dete hain jinhein aap aata hua dekh sakte hain.
4. Yeh tokens mein parhta hai, harf ya lafz mein nahin
Model aap ke prompt ko harf ke taur par nahin dekhta, aur bilkul lafz ke taur par bhi nahin. Kuch bhi hone se pehle, aap ka text tokens mein kata jata hai: chunks jo aam taur par ek lafz ya ek lafz ka tukra hote hain. "Strawberry" do ya teen chunks ke taur par aa sakta hai; "the" ek hai; ek lamba ya ghair-maamooli lafz kei. Model sirf in chunks ko dekhta hai, in chunks mein predict karta hai, aur in ke andar ke alag alag harf kabhi nahin dekhta jab tak inhein spell out karne par majboor na kiya jaye.
Yeh aik mechanical fact ek aise jhurmut behaviors ko samjhata hai jo warna naqabil-fehm hain:
| Behavior | Tokens isay kyun samjhate hain |
|---|---|
| Yeh ek lafz mein harf ghalat ginta hai (strawberry test). | Yeh chunks dekhta hai, harf nahin. Ek chunk ke andar harf ginna ek street address se kamre ginne jaisa hai. |
| Yeh kuch rhyming, anagrams, aur wordplay mein bura hai. | Yeh harf aur aawaz par chalte hain; model chunks par chalta hai. |
| Aap ke prompt mein typos shaaz hi maayne rakhte hain. | Ek ghalat spell kiya lafz ab bhi un chunks se map hota hai jo manshood matlab ke kaafi qareeb hain. (Is liye 2026 mein AI Prompting aap ko typos theek karne ki zehmat na karne ko kehta hai.) |
| Cost aur length tokens mein naapi jati hai, lafzon mein nahin. | Jo cheez machine asal mein process karti hai woh token hai, is liye wohi cheez hai jis ka aap se bill liya jata hai aur jis se aap mehdood hote hain. |
Tokens paise ka unit bhi hain aur memory ka unit bhi. Jab koi tool kehta hai ke us ke paas ek "200,000-token context window" hai, woh bata raha hai ke yeh ek hi waqt mein in chunks mein se kitne hold kar sakta hai (Idea 5). Jab aap ka "per token" bill banta hai, aap har chunk in aur har chunk out ke paise de rahe hain. Taqreeban, English mein, teen tokens taqreeban chaar lafz hote hain, lekin aap ko kabhi theek ratio ki zaroorat nahin, sirf yeh idea ke chunk hi asal unit hai, aur lafz ek taqreeb hai jo aap oopar laga lete hain.
Woh "teen tokens ≈ chaar lafz" ratio English ke liye hai. Doosre scripts ka text (Urdu, Arabic, Hindi, Chinese, aur bohat se aur) aam taur par har lafz ke liye zyada tokens mein kata jata hai, kyunke training text English-heavy tha aur tokenizer ne English chunks sab se behtar seekhe. Do amali nataij seedhe nikalte hain: ek hi message ek ghair-English zubaan mein zyada cost karta hai, aur yeh context window ko tezi se bharta hai (Idea 5), is liye us conversation ke liye model ki mooassir memory chhoti hoti hai. Yeh behtar ho raha hai jaise jaise tokenizers behtar hote hain, lekin 2026 mein yeh ab bhi asal hai. Agar aap zyada tar ek non-Latin script mein kaam karte hain, to ek English user jo wohi task chala raha hai us se cost aur length limits tak jaldi pohanchne ki tawaqqo karein. Aur jab ek lamba document ahem ho, to kabhi kabhi yeh sahi hota hai ke model ko andar-hi-andar English mein kaam karwaya jaye aur aakhir mein translate kiya jaye.
Yeh dekh sakti hai, aur mechanism nahin badalta: yeh generalize karti hai. Aap ki upload ki gayi tasveer chhote patches mein kati jati hai, aur har patch ek token ban jata hai; ek sound clip chhote segments mein kata jata hai, aur har ek token ban jata hai. Model phir ek single stream par predict karta hai jo word-chunks, image-patches, aur audio-segments ko aapas mein milata hai. Is liye is course ka har cheez images aur audio ke liye bhi sach rehti hai: wohi prediction (Idea 1), wohi frozen weights (Idea 2), wohi ghaib truth-checker (Idea 3), wohi context-window-as-desk (Idea 5). Yeh woh mechanical wajah bhi hai ke ek image mein baareek tafseel aur chhota print mushkil hai: ek patch ek chunk hai, aur ek patch ke andar harf parhna phir wohi strawberry problem hai. Amali pehlu (kaun si images AI achi tarah parhti hai aur kaun si bigar deti hai, aur in ke liye prompt kaise karna) 2026 mein AI Prompting Concept 8 ka kaam hai; neeche ka mechanism bas mazeed qism ke tokens, wahi machine.
5. Context window hi woh aik cheez hai jo yeh dekh sakti hai
Kyunke weights frozen hain (Idea 2) aur model ki apni koi memory nahin, theek ek hi jagah hai jahan se yeh aap ki khaas soorat-e-haal ke baare mein maloomat haasil kar sakti hai: context window, woh text jo is ek response ke liye is ke saamne baitha hai. 2026 mein AI Prompting isay Concept 4 ke taur par sikhata hai, "context hi poori game hai," aur isay prompting ka markazi skill maanta hai. Yahan woh mechanical wajah hai ke yeh markazi kyun hai.
Context window ek response ke liye model ki poori duniya hai. Is mein aap ka prompt, ab tak ki conversation, jo files aap ne attach ki, tool descriptions, aur woh nazar na aane wala system prompt hota hai jo product ne aap ke aane se pehle wahan rakha. Us window mein jo kuch hai, model usay use kar sakta hai. Jo kuch is mein nahin hai woh is jawab ke liye maujood hi nahin, is liye nahin ke model inkaar kar raha hai, balke is liye ke is ke paas dekhne ki aur koi jagah nahin. Frozen weights isay duniya ke baare mein aam fluency dete hain; context window aap ki duniya ke khaas pehluon ka woh aik channel hai.
Yeh do cheezon ko reframe karta hai jo aap warna rahasya paayein ge:
- Briefing kyun kaam karti hai. Model ko context dena koi politeness ya koi trick nahin. Yeh literal amal hai maloomat ko us aik jagah daalne ka jahan machine usay parh sakti hai. Ek un-briefed model lazy nahin hai; is ke saamne sachmuch kuch nahin.
- Lambi conversations kharab kyun ho jati hain (prompting course mein "context rot"). Window ki ek size limit hai jo tokens mein naapi jati hai (Idea 4). Is mein bohat zyada ghair-mutaaliqa history thoons dein aur jo signal aap ko chahiye woh patla ho jata hai, ya sab se purane hisse jagah banane ke liye summarize ho kar nikal diye jate hain. Model thak nahin raha; is ka reading desk bas bhara hua hai.
Context window ek reading desk hai, dimaagh nahin. Aap desk par jo kuch rakhte hain, model usay ghaur se parhta hai. Jo kuch aap desk se bahar chhor dete hain, woh usay nahin dekh sakta, chahe woh aap ke liye kitna hi zaahir kyun na ho. Prompting ka poora skill, baqi paanch courses mein parhaya gaya, ek hi aadat mein simat jata hai jab aap isay aise dekhte hain: control karein ke desk par kya land karta hai.
6. Is ka confidence ek seekha hua style hai, sach ka signal nahin
Idea 3 ne kaha ke model ke paas koi internal truth-checker nahin. Yeh idea doosra rukh samjhata hai: is ka thakawat-na-jaanne wala confidence kahan se aata hai, aur woh confidence durustagi ke baare mein aap ko kuch kyun nahin batata.
Main training (Idea 2) ke baad, models ko aam taur par insani feedback se mazeed tune kiya jata hai: log responses ko rate karte hain, aur model ko aise jawab ki taraf adjust kiya jata hai jisay logon ne zyada rate kiya. (Engineers is step ko RLHF kehte hain, reinforcement learning from human feedback; aap ko machinery ki zaroorat nahin, sirf nateeje ki.) lakhon ratings mein, log mustaqil taur par confident, helpful, fluent, aur agreeable jawab ko un jawabon par tarjeeh dete hain jo hedged, blunt, ya contrarian hote hain. Is liye machine ko confident, agreeable, fluent text produce karne ki taraf shakal di jati hai, chahe under-lying content theek ho ya nahin. Confidence ek style ban gaya jo yeh default ke taur par pehnti hai, jaise ek polished writer un mawazoo ke baare mein rawani se likhta hai jinhein woh aadha samajhta hai.
AI ke do sab se zyada zer-e-behes behaviors seedhe isi se nikalte hain:
- Yeh ghalat hone par bhi yaqeen se boli hai. Yaqeen ek seekha hua stylistic default hai, usi process se generate hota hai jis se content aur sach se utna hi alag. Ek confident jumla house style hai, durustagi par koi faisla nahin.
- Yeh aap se agree karne ki taraf jhukti hai, woh sycophancy jis ke liye 2026 mein AI Prompting Concept 6 waqf karta hai. Agreement ko disagreement se zyada rate kiya gaya, is liye machine aap ko woh batane ki taraf jhukti hai jo aap chahte lagte hain. Poochein "kya X sach nahin hai?" aur aap ne woh jawab signal kar diya jo aap chahte hain; train-shuda jhukao usay supply kar deta hai.
Ab prompting course ke fixes mechanically samajh aate hain. Neutral framing ("X ko evaluate karein; har taraf ka sab se mazboot case dein") kaam karti hai kyunke yeh woh signal hata deti hai jis ki taraf model warna jhukta. Score force karna ("isay in criteria ke khilaf 1–10 rate karein") kaam karta hai kyunke ek number ko agreeably fake karna ek adjective se mushkil hai. Aap machine se zyada chaalaki nahin kar rahe; aap woh cues hata rahe hain jo is ke train-shuda jhukao ko trigger karte hain.
7. Yeh aas-paas ke lamhon mein shandaar aur bekaar hai (jagged frontier)
Insani salaahiyat kaafi smooth hai: jo mushkil calculus kar sakta hai woh taqreeban zaroor asaan arithmetic kar sakta hai. AI ki salaahiyat smooth nahin hai. Yeh jagged hai: ek task par superhuman aur ek paros ke task par heran kar dene wali na-ahel jo, humare liye, koi zyada mushkil nahin lagta. Yeh ek legal-lagne wala contract clause draft kar sakti hai aur phir "strawberry" mein harf ghalat gin sakti hai. Yeh quantum mechanics samjha sakti hai aur ek teen-step logic puzzle bigaar sakti hai jo ek bachcha hal kar leta.
Jaggedness bay-tarteeb nahin; yeh training text aur token mechanism tak ja pohanchti hai. Woh tasks jo training data mein aksar, saaf shakal mein, aaye (aam concepts samjhana, aam styles mein likhna, aam code produce karna) mazboot hain. Woh tasks jo un cheezon par munhasir hain jinhein machine achi tarah nahin dekh sakti, jaise alag alag harf (Idea 4), bohat haaliya waqiaat (Idea 2), aap ka private context (Idea 5), ya shaaz topics (Idea 1), kamzor hain. "Shandaar" aur "bekaar" ke darmiyan frontier ek jagged lakeer mein chalti hai jo mushkil ke baare mein insani intuition se match nahin karti, jo theek wajah hai ke yeh logon ko heran karti rehti hai.
Teen amali aadatein jaggedness qubool karne se nikalti hain:
| Aadat | Yeh jaggedness se kyun nikalti hai |
|---|---|
| Yeh farz na karein ke kyunke is ne ek mushkil task kar liya, yeh ek asaan bhi kar legi. | Dono jagged frontier ke ult-pulat taraf baith sakte hain. |
| Boundary ke aar-paar verify karein, beech mein nahin. | Khatarnaak ghaltiyan woh asaan-lagne wale tasks hain jinhein yeh khamoshi se fail kar deti hai, woh mushkil nahin jo aap pehle hi check kar rahe thay. |
| Wohi task do ya teen mukhtalif models mein try karein. | Mukhtalif models ki frontiers mukhtalif shakal ki hoti hain; ek woh pakar leta hai jo doosra chhor deta hai. (2026 mein AI Prompting, Concepts 12–13.) |
Frontier harkat bhi karti hai. Woh cheez jo model is quarter "nahin kar sakta," ek naya model agle quarter aasani se kar sakta hai, aur ek cheez jo yeh achi tarah karta hai woh bilkul behtar na ho. Prompting course ka yeh mashwara ke har chand mahine dobara test karein ke AI kya kar sakti hai, mechanically, ek aise frontier ko dobara map karne ka mashwara hai jo shift hoti rehti hai.
Part 3: Kis cheez ne ek text-predictor ko aisi cheez bana diya jo act karti hai
Do ideas jo "yeh text predict karti hai" aur un agents ke darmiyan ka faasla band karti hain jin ke baare mein is kitab ka baqi hissa hai. Yeh yeh kya hai se yeh duniya mein kya karti hai tak ka pul hai.
8. Tools isay act karne dete hain, sirf describe karne nahin
Ab tak har cheez ek aisi machine bayan karti hai jo text produce karti hai. Ek khaalis text-predictor aap ko woh weather bata sakta hai jo usay training se yaad hai, lekin yeh aaj ka weather check nahin kar sakta, asal numbers par calculation nahin chala sakta, aap ki file nahin parh sakta, ya email nahin bhej sakta. Saalon tak yeh hi chhat thi.
Chhat tools se uthi. Ek tool ek mukarrar action hai jisay model ko call karne ki ijazat hai (ek web search, ek code run, ek file read, ek email draft), jisay context window (Idea 5) mein baqi har cheez ke saath describe kiya jata hai. Mechanism nateeje ke lihaaz se taqreeban sharminda kar dene wala saada hai: jab model predict karta hai ke sahi continuation "is query ke saath search tool use karo" hai, na ke khaalis prose, to product us action ko sachmuch chalata hai, nateeja wapas context window mein daal deta hai, aur model wahan se jaari rakhta hai. Prediction, action, result-wapas-context-mein, dobara predict. Woh loop ek aise chatbot ke darmiyan farq hai jo duniya describe karta hai aur ek aise assistant ke jo us par act karta hai.
Yeh hi wajah hai ke wohi under-lying machine ek din ek chat window ho sakti hai aur, tools wire ho jaane par, ek agla din ek aisa agent jo aap ka folder dobara organize karta hai. Baqi Foundations courses, hood ke neeche, isi yek-saan predictor par wire kiye gaye khaas tools ke baare mein courses hain:
- Code execution woh tool hai jo Code You Never Write ke peeche hai: model ek program predict karta hai, tool usay chalata hai, asal nateeja wapas aata hai.
- Connectors woh tools hain jo aap ki apps par wire hain, Skills & Connectors ka mauzoo: model predict karta hai "isay Drive se fetch karo," tool usay fetch kar leta hai.
- Web search woh tool hai jo ek stale model ko bacha leta hai (Idea 2), 2026 mein AI Prompting mein cover kiya gaya.
Yeh kitab ek aisi AI ke liye agent kehti hai jo aap ki taraf se multi-step kaam karti hai. Ab aap dekh sakte hain ke hood ke neeche is ka matlab kya hai: ek agent yeh hi next-token predictor hai, tools de kar, predict-act-observe loop ko ek goal ki taraf qatar mein kai baar chalata hua: ek action predict karna, nateeja apne context mein land hota dekhna, aur wahan se agla action predict karna. Koi nayi qism ka dimaagh shaamil nahin. Ek jana-pehchana predictor hai, tools ka ek set hai, aur ek loop hai. Yeh hi woh poori buniyad hai jis par kitab ka baqi hissa banta hai.
9. "Thinking" bas jawab se pehle, zubaan se, mazeed prediction hai
Naye models jawab dene se pehle "soch" ya "reason" kar sakte hain, aur 2026 mein AI Prompting (Concept 5) aap ko mushkil tasks ke liye isay "think hard" se invoke karne ko kehta hai. Yeh jaanna ke yeh asal mein kya hai aap ko isay hadd se zyada rahasya banane se rokta hai.
Ek reasoning model, apna aakhri jawab dene se pehle, pehle intermediate working ka ek lamba hissa predict karta hai (steps bichhana, approaches try karna, khud ko check karna) aur sirf phir aakhri jawab predict karta hai, ab woh saari working apne context window (Idea 5) mein baithi hui hai jis par bunyaad rakhe. Yeh ab bhi khaalis next-token prediction hai. Trick yeh hai ke jawab predict karna zyada asaan aur zyada durust ho jata hai jab ek acha chain of reasoning pehle se desk par maujood ho jis se predict kiya jaye. Pehle zubaan se kaam karna sachmuch madad karta hai, usi wajah se jis se yeh ek shakhs ki madad karta hai ke jawab par committed hone se pehle kaaghaz par soche.
Yeh hi wajah hai ke "think step by step" pehle type karne ke liye ek mufeed phrase tha, aur yeh ab aksar built-in kyun hai: aap manually model se keh rahe thay ke jawab se pehle reasoning desk par rakhe; ab model mushkil masail ke liye khud apne aap karta hai. Yeh cost aur intezaar bhi samjhata hai: reasoning ka matlab hai bohat se extra tokens (Idea 4) generate karna jo aap kabhi nahin dekhte, jo waqt aur paisa leta hai, jo theek wajah hai ke prompting course aap ko thinking mode sirf sachmuch mushkil sawalon ke liye bachane aur jaldi lookups ke liye chhorne ko kehta hai.
Yeh, magar, machine ko Idea 3 wali doosri faculty nahin deta. Ek reasoning model apna kaam usi prediction process se check karta hai jo ghalat ho sakta hai, is liye yeh apni bohat si ghaltiyan pakar leta hai aur phir bhi kuch chhor deta hai, aur phir bhi ek aise chain of reasoning ke andar poore confidence se hallucinate karta hai jo rigorous lagti hai. Zyada thinking gap ko tang karti hai. Yeh isay band nahin karti. Aap ab bhi aakhri check hain.
No math, no code ka waada poora karne ke liye, kai asal topics ek taraf rakhe gaye. Teen naam lene ke layeq taake aap jaanein ke woh maujood hain: woh training compute aur cost jo ek model banate hain (behad, aur woh wajah ke sirf chand organizations unhein banate hain); woh safety aur alignment kaam jo shakal deta hai ke ek model kya karega aur kya nahin (apne aap mein ek bara maidan); aur woh gehri mechanics (weights asal mein kaise structured aur adjusted hote hain), jinhein woh math chahiye jo yeh course chhor deta hai. In mein se koi bhi oopar wali nau ideas ko nahin badalta; woh un ke neeche aur un ke saath baithte hain. Agar koi aage ka chapter aap ko in teen mein se kisi ki taraf bheje, ab aap ke paas us par banane ke liye farsh hai.
Prompts try karne se pehle ek mukhtasar recap
Nau ideas, ek-ek line. Aakhri jumla saath le jayein; baqi ke liye wapas aayein.
- Idea 1. Yeh text ka agla tukra predict karta hai; yeh facts lookup nahin karta. Prediction sirf wahan knowledge jaisi lagti hai jahan training text mota tha.
- Idea 2. Yeh ek baar, parh kar seekha, aur phir seekhna freeze ho gaya. Is liye knowledge cutoff, aur is liye yeh aap ki private duniya nahin jaan sakta. Isay use karna isay kabhi nahin sikhata.
- Idea 3. Is ke paas koi alag faculty nahin jo check kare ke ek prediction sach hai ya nahin. Hallucination machine ka bani hui tarah par kaam karna hai, kharabi nahin.
- Idea 4. Yeh tokens (chunks) mein parhta hai, harf ya lafz nahin. Yeh matlab ka, memory ka, aur paise ka unit hai.
- Idea 5. Context window hi woh aik jagah hai jahan yeh aap ke khaas pehlu dekh sakti hai: ek reading desk, dimaagh nahin. Control karein ke is par kya land karta hai.
- Idea 6. Is ka confidence aur is ki agreeableness seekhe hue styles hain, sach se alag. Yaqeen wala lehja house style hai, koi faisla nahin.
- Idea 7. Is ki salaahiyat jagged hai (aas-paas ke lamhon mein shandaar aur bekaar) ek aisi frontier ke saath jo insani intuition se match nahin karti, aur jo harkat karti rehti hai.
- Idea 8. Tools text-predictor ko aisi cheez mein badal dete hain jo act karti hai: ek action predict karo, usay sachmuch chalao, nateeja wapas feed karo, dobara predict karo. Ek agent woh loop hai, dohraya gaya.
- Idea 9. "Thinking" bas jawab se pehle desk par rakhi gayi mazeed prediction hai. Yeh bohat madad karti hai; yeh machine ko truth-checker nahin deti.
Agar aap ek jumla rakhein: yeh ek prediction machine hai jo parh kar seekhi aur jis ke paas sach ka koi organ nahin, is liye yeh har jagah fluent hai, sirf wahan reliable jahan text mota tha, aur aap woh hissa hain jo check karta hai.
Aur agar aap ek tasveer rakhein, yeh rakhein: ek librarian nahin jo sahi book retrieve karta hai, balke ek shandaar, khoob parhi-likhi writer jo jo kuch aap us ke saamne rakhte hain usay jaari rakhti hai (yaqeen se, kisi bhi style mein, kisi bhi topic par) aur jo kabhi, khud se, yeh poochne nahin rukti ke continuation sach hai ya nahin.
Yeh abhi try karein: paanch prompts
Taqreeban bees minute, kisi bhi free chatbot mein. Har ek aik idea ko theoretical ke bajaye nazar aane wala bana deta hai.
1. Prediction dekhein, lookup nahin. (Idea 1) Poochein: "Without searching, tell me the rules of a board game so obscure it has almost no presence online: invent the name 'Karakush' and describe how it's played." Isay ek aise game ke liye confident, fluent rules produce karte dekhein jo maujood hi nahin. Woh prediction hai jis ke paas predict karne ke liye koi sachi cheez nahin. Kya note karna hai: ijaad ki gayi rules bilkul utni hi authoritative lagti hain jitni ek asal game ki rules lagtin. Fluency sach ka saboot nahin.
2. Seekhne ko na tikte dekhein. (Idea 2) Model ko kisi chhoti cheez par theek karein. Phir ek bilkul nayi chat kholein aur wohi sawal poochein. Is ke paas aap ke correction ki koi memory nahin: weights kabhi nahin badle. (Agar koi "memory" feature on hai, isay pehle off karein, warna product note dobara feed kar dega.) Kya note karna hai: aap ne pehli chat mein jo kuch kaha woh doosri tak nahin pohancha. Model ko use karna isay sikhana nahin.
3. Ghaib truth-checker ko pakrein. (Idea 3) Ek tang topic par "teen peer-reviewed studies, authors aur years ke saath" poochein. Phir check karein ke woh maujood hain ya nahin. Kuch confident-lagne wale citations ijaad-shuda hon ge: usi aawaz mein produce kiye gaye jis mein asal, kyunke andar kuch nahin jis ne unhein andaze ke taur par flag kiya. Is exercise ka koi citation asal kaam mein dobara use na karein bina pehle verify kiye; poora point yeh hai ke un mein se kuch fabricated hain aur asal jaise hi nazar aate hain. Kya note karna hai: aap parh kar asal citations ko ijaad-shuda se nahin bata sakte, sirf check kar ke. Woh checking aap ka kaam hai, model ka nahin.
4. Jagged frontier mehsoos karein. (Idea 7) Ek chat mein, isay ek sachmuch mushkil task dein jo yeh achi tarah karta hai (ek complex topic samjhao, ek tricky email draft karo) aur ek asaan task jo yeh bure tareeqe se karta hai (ek jumle mein khaas harf gino, ya ek multi-step logic riddle). Note karein ke competence mushkil ka peechha nahin karti. Kya note karna hai: woh asaan task jisay yeh bigaar deti hai woh khatarnaak hai: woh jisay aap kabhi check karne ka sochte hi nahin.
5. Thinking on aur off karein. (Idea 9) Wohi mushkil reasoning sawal do baar poochein: ek baar saada, ek baar "think hard and show your working first" ke saath. Moqabla karein. Doosra jawab aam taur par behtar hota hai, kyunke model ne jawab predict karne se pehle reasoning desk par rakhi. Kya note karna hai: working ne jawab behtar kiya, lekin model phir bhi apni working ko certify nahin kar sakta: zyada thinking gap ko tang karti hai, isay band nahin karti.
Yeh kahan le jata hai
Aap ke paas ab model ke neeche ka model hai: woh cheez asal mein kya hai, is se pehle ke koi course aap ko isay use karna sikhaye. Yahan se, baqi Foundations isay achi tarah chalane ke baare mein hai:
- 2026 mein AI Prompting Ideas 1, 5, aur 6 ko briefing, context control, aur sycophancy ko neutralize karne ki rozana aadaton mein badal deta hai.
- AI ke Daur Mein Sochna woh discipline hai jo seedhe Idea 3 par bani hai: kyunke machine ke paas koi truth-checker nahin, aap woh ban jate hain.
- Markdown In, HTML Out aur Code You Never Write is baare mein hain ke context window (Idea 5) mein kya andar aur bahar bahta hai aur tools (Idea 8) us ke saath kya kar sakte hain.
- Skills & Connectors usi predictor par mazeed tools (Idea 8) wire karta hai.
The Agent Factory ki baqi har cheez (agents, unhein manufacture karna, unhein deploy karna) Idea 8 ke predict-act-observe loop par bani hai, scale par chalayi gayi. Machine kabhi next-token predictor hona nahin chhorti. Isay bas mazeed tools, lambe loops, aur ek frozen weights ka set milta hai jo in teenon ke saath sachmuch hairan kar dene wali miqdaar mein kaam karta hai.