@nataliaevelynsant: #foryou #fy #fypage #viral #reels

Natalia Evelyn Santo
Natalia Evelyn Santo
Open In TikTok:
Region: BR
Saturday 04 April 2026 18:33:28 GMT
10969
246
3
0

Music

Download

Comments

adelmafvasconcelo
adelmafvasconcelos :
🥰🥰🥰🥰
2026-04-04 18:59:51
1
user452397279
Jose Sousa :
🥰🥰🥰
2026-04-04 18:49:17
0
roberto.miranda100
Roberto Miranda :
🍀👏🙏
2026-04-06 13:53:05
0
To see more videos from user @nataliaevelynsant, please go to the Tikwm homepage.

Other Videos

Everyone can explain what an LLM is. Almost nobody can explain what happens in the 400ms after you hit enter. Wrong question:
Everyone can explain what an LLM is. Almost nobody can explain what happens in the 400ms after you hit enter. Wrong question: "which model is best?" Right question: "where does the time and the money actually go?" 12 concepts, one line each 👇 🟢 WHAT IT HOLDS 01 Tokenization — text becomes IDs; "strawberry" is 3 tokens, not 10 letters 02 Context window — how much it sees at once; when full, the oldest falls out 03 KV cache — it stops recomputing the past, so memory grows with the chat 🔵 HOW IT SERVES 04 Batching — many requests in one pass; a slot frees, the next joins 05 Streaming — tokens ship as they're made, so TTFT is what users feel 06 Speculative decoding — a small model drafts ahead, the big one verifies 🟡 HOW IT PICKS 07 Temperature — flatten or sharpen the odds before choosing 08 Top-p — keep the smallest set of tokens that reaches p 09 Top-k — keep a fixed count, whatever the odds look like 🟣 WHAT IT COSTS 10 Quantization — FP16 → INT8 → INT4; 16GB becomes 4GB 11 Latency — TTFT is the wait, tokens/sec is the flow 12 Cost — you pay both directions, and output runs ~5x input 👀 Watch cards 07-09. Same nine-bar distribution, three times. Top-p adapts to its shape — a confident prediction keeps 2 tokens, an uncertain one keeps 8. Top-k keeps the same count either way. That's the whole answer to "top-p or top-k", and why p is the safer default. And card 06 isn't a trick: speculative decoding provably returns the exact same output distribution as the big model alone. 1.5-2.5x faster, zero quality cost. The rule: latency and bill are decisions, not properties of the model you chose. Most teams tune the prompt for weeks and never touch batching, caching or quantization. Send this to whoever owns the inference bill. 📸 Screenshot it — 12 concepts, one card. Follow @hackproduct — scary AI concepts, made shippable. ⚡ . . #LLM #AIengineering #inference #MLops #GenAI

About