Lesson 18 — Beginner

Understanding the Model Landscape

Aaj har hafte naya model aata hai — aur har ek apne ko behtar batata hai. Is lesson mein hum landscape ko teen raston mein todenge (closed API, open weights, managed platforms), phir wo cheezein dekhenge jo model cards aur pricing pages se padhni chahiye — license, cutoff, versions. Aur ant mein — kyun koi universally best LLM nahi hota.

⏱ 10 min• Beginner• LLM Capabilities & Failure Modes

Teen raste — API, open weights, managed platforms

WHAT

Model landscape mein teen mukhya raste hain — closed models (provider ke servers par API), open weights (file download, apne hardware par) aur managed platforms (apne cloud account ke andar dono). Har raste ka apna trade-off hai, koi free lunch nahi.

WHEN

Jab bhi tum naya feature banao ya model switch karo — tabhi. Pehle requirement likho (privacy, latency, budget), phir rasta chuno — ulta kabhi nahi.

WHERE

API dashboards, cloud consoles (Azure AI Foundry, AWS Bedrock, Google Vertex AI jaise platforms) aur open-weight releases (model cards, licenses, hosting providers).

Sabse zaroori principle pehle: koi universally best LLM nahi hota. Best model workload par depend karta hai — tumhara task, latency budget aur privacy rules decide karte hain, leaderboard rank nahi. Isi vajah se poora khel comparison ka hai, ranking ka nahi.

Closed API vs open weights vs managed

Closed/API models: sabse frontier capability, sabse kam setup — par data provider ke paas jaata hai (privacy), vendor dependence (price badal sakta hai, version retire ho sakta hai), control kam.
Open weights: file download, apne GPU ya cloud par chalao — privacy poori, customization (fine-tuning, quantization) possible, par infrastructure aur MLOps ka poora kaam tumhara.
Managed platforms: apne cloud account ke andar closed aur open dono — compliance ka raasta aasan, billing aur identity unified.
Inference providers: open weights kisi aur ke hardware par, API jaisi suvidha — self-host ka faayda, ops ka jhanjhat nahi.

Decision rule: Privacy strict + ops team + budget → self-host/open weights. Speed to market + frontier capability → API. Enterprise compliance + apna cloud → managed platform. Aksar teams mixed rakhti hain — flagship API par, cost-sensitive kaam open weights par.

Model card padhne ke liye aur distinctions

MODEL CARD PADHO — ye distinctions samjho:

Small vs Large        : chhota = tez/sasta, bada = zyada capable (kabhi kabhi)
Dense vs MoE          : compute aur memory ka farq (M1.3)
General vs Specialized: chhat-chaudha banaam domain-specific (medical, code)
Reasoning vs Non      : test-time compute ka farq (M1.7)
Text-only vs Multimodal: image/audio/video input-output (L12)

Ye distinctions leaderboard rank se zyada bataati hain — "kya model chhota hai", "kya multimodal hai" jaise sawaal tumhare requirement se seedhe judte hain.
Small vs large ka farq sirf capability nahi — cost, latency aur hosting ka bhi. Chhota model aksar hi kaafi hota hai; bada tab chuniye jab task sach mein mushkil ho.

Knowledge cutoff, versions aur deprecation

Har model ka gyaan ek cutoff date par ruk jaata hai — uske baad ka duniya ka kuch bhi model ko nahi pata. Wo jankari sirf context, retrieval ya tools se aa sakti hai. Aur model ka naam hamesha dated snapshot pin karo — "latest" kabhi bhi badal sakta hai, deprecation ek din aayega hi.

Cutoff ke baad ki khabar chahiye? Context mein daalo, retrieval lagao, tool call karo — model ko guess mat karo (yahi L4 ka kaam hai).
Production mein pin karo: exact version/snapshot use karo, aur usse apne test suite par pass karke hi upgrade karo.
Deprecation notices padho: provider aksar mahine-bhar ka notice deta hai. Upgrade plan pehle se rakho, emergency kabhi nahi bano.

License padho — aur principle yaad rakho

License teen tarah ke hote hain: permissive (ship freely), restricted-use aur non-commercial. "Open" ka matlab license open nahi hota — weights open hote hain, shartein alag.
Open weights ka faayda tabhi hai jab control chahiye — warna API hi sasta padega. Hosting ka kharcha (GPU, ops, updates) mat bhoolo.
Koi bhi model naam se nahi chuno — hype, brand aur rank teeno dhoka de sakte hain. Apne task par test karke chuno.

Interview line: Comparison ka sahi formula — requirement pehle, model baad mein. Closed ka convenience, open ka control, managed ka middle path — aur best model wahi jo tumhare workload par empirically chala hai.

Exercise

Question: Privacy strict hai aur data bahar nahi ja sakta — kaunsa rasta chunoge? (self-host ya public API)

Question: Training ke baad ka duniya model ko nahi pata — us line ka naam kya hai? (do shabd)

Question: Production mein model pin karne ka matlab — version pin karo ya latest use karo?

Question: Open weights mein open hote hain — weights ya license?

Common mistakes

Leaderboard number 1 ko hamesha chunna: Rank kisi aur ke task ka hai. Requirement se shuru karo — capability, latency, cost, privacy — phir model chuno. Rank sirf shortlist ke kaam ki hai.
Open weights ko free samajhna: License padho (non-commercial aur restricted-use alag hote hain), aur hosting ka kharcha khud uthana padta hai — GPU, ops, updates sab. API se sasta tabhi jab volume ho.
Model name "latest" chhod dena: Floating model name ek din breaking change laata hai — output ka format badal jaaye toh app tootegi. Pin karo, test pass karke upgrade karo, deprecation notice padho.

Desi example: Teen sawari ke raste — landscape ka desi analogy

Closed API taxi jaisi hai — chadh gaye, pahunche, driver (provider) dekhega. Open weights khud ki gaadi — control poora, par maintenance bhi tumhara. Managed platform leasing — gaadi bhi, driver bhi. Aur chunna pehle safar (task) dekh kar, gaadi dekh kar nahi:

Closed API    : "bas chadh jao"  — infra nahi, vendor par depend
Open weights  : "khud chalao"   — control zyada, maintenance tumhara
Managed       : "leasing wala"   — apne cloud mein dono options
Pehle safar decide karo (task, privacy, budget), phir gaadi.

Interview angle — yahan se sawaal aate hain

Q: Closed aur open-weight model ka core farq? — Closed: convenience aur frontier capability, par vendor dependence aur privacy; open: control aur privacy, par infra ka kaam tumhara.
Q: Managed platform kya deta hai? — Apne cloud account ke andar closed aur open dono models — compliance, billing aur identity ka aasan raasta.
Q: Knowledge cutoff ka matlab? — Model ka gyaan training data ki cutoff date par ruk jaata hai; uske baad ki cheezein sirf context, retrieval ya tools se.
Q: Open-weight model ship karne se pehle kya padho? — License — permissive, restricted-use aur non-commercial ka farq, warna legal problem pakki.

Lesson complete?

Landscape clear. Agla lesson — Benchmarks & Leaderboards — mein dekhenge unhe kaise padhna hai bina dhoke mein aaye, aur unhe shortlist tak kaise istemaal karna hai.