🧩 🧱 Tokens (Sub-Word Chunks)
🧒 For Kids: Words chopped up into tiny Lego blocks. "Strawberry" is chopped into three Lego bricks: "Straw", "ber", "ry"!
⚡ For Pros: Sub-word byte-pair encoding (BPE) chunks. 1,000 tokens ≈ 750 English words.
🎒 🧠 Context Window (Working Memory)
🧒 For Kids: The robot's short-term memory backpack! If the backpack holds 100 pages, page 101 pushes page 1 out of its head!
⚡ For Pros: The maximum token capacity accommodated inside the Key-Value (KV) attention cache during a single inference call.
🦄 👻 Hallucination (Fact Drift)
🧒 For Kids: When the robot dreams and invents fairy tales by accident, but says them with total confidence like they are real!
⚡ For Pros: High-confidence probabilistic generation of factually ungrounded claims caused by distribution shifts or training gaps.
📦 🗜️ Quantization (Weight Compression)
🧒 For Kids: Squishing a giant toy so it fits inside your lunchbox, without losing any of its fun buttons!
⚡ For Pros: Compressing 16-bit floating-point weights (FP16) into 4-bit (INT4 / GGUF) integers to slash memory by ~70% with negligible perplexity loss.
🔍 📚 RAG (Retrieval-Augmented Generation)
🧒 For Kids: An open-book test! The robot gets to look at your diary or textbook before giving its answer.
⚡ For Pros: Retrieval-Augmented Generation. Vector database similarity search inserts live factual context into the prompt before generation.
🧠 Reasoning (CoT)
🧒 For Kids: Talking out loud in your head and double-checking your work before blurting out an answer!
⚡ For Pros: Test-time compute allocation via Chain-of-Thought reinforcement learning (RL) generating hidden tokens to explore decision trees.
🌡️ 🎲 Temperature & Top-P (Randomness Steering)
🧒 For Kids: The wildness dial! Zero makes the robot super strict and predictable; turning it up makes it tell wild jokes!
⚡ For Pros: Softmax probability scaling (Temperature) and cumulative probability threshold filtering (Top-P nucleus sampling).
👥 🏢 Mixture of Experts (MoE Routing)
🧒 For Kids: A council of mini-heroes! When a math question comes in, only the math wizard wakes up, letting the art wizard sleep!
⚡ For Pros: Sparse neural architecture where a gating router dynamically activates a subset of feed-forward layers per token (e.g. DeepSeek activates 37B out of 671B).
⚡ MLA (Latent Attention)
🧒 For Kids: A secret super-memory notebook that writes notes in tiny shorthand code so it never runs out of paper!
⚡ For Pros: Multi-Head Latent Attention. Projects Key and Value matrices into compressed low-rank latent vectors, slashing KV cache memory by over 80%.
💾 🧮 KV Cache Overhead (Context RAM)
🧒 For Kids: The table where the robot puts all the open books it is reading. If you give it 100 books at once, you need a giant table!
⚡ For Pros: Stored Key and Value attention matrices maintained in GPU VRAM across generation steps. Scales linearly with context length and batch size.
⚖️ RLHF vs DPO vs GRPO
🧒 For Kids: Gold star stickers from teachers! When the robot is polite, it gets a gold star. When it is rude, it gets sent back to practice.
⚡ For Pros: Post-training alignment methods. DPO simplifies RLHF with direct policy loss; GRPO (used in DeepSeek-R1) eliminates separate critic networks via group scoring.
📐 🌌 Vector Embeddings (Meaning Coordinates)
🧒 For Kids: Giving every word an address on a giant treasure map. Words with similar meanings live right next door to each other!
⚡ For Pros: High-dimensional floating-point vectors (e.g. 1536-dim) capturing latent semantic similarity calculated via cosine distance.
🔧 Tool / Function Calling
🧒 For Kids: Giving the robot hands to press calculator buttons or look outside the window!
⚡ For Pros: Emitting structured JSON payloads matching external API schemas, pausing generation for client execution, and ingesting results.
🎯 Zero-Shot vs Few-Shot
🧒 For Kids: Asking a puzzle right away (Zero-Shot) versus showing two completed example puzzle cards first (Few-Shot)!
⚡ For Pros: In-context learning prompting paradigm. Providing input-output demonstration pairs primes model attention to output specific formats.
🧭 System Prompts & Steering
🧒 For Kids: Whispering in the actor's ear: "In this play, you are Sherlock Holmes and you must speak in old English!"
⚡ For Pros: Hidden privileged instructions placed at the root of the context window that condition generation tone, boundaries, and safety constraints.
🪶 LoRA & QLoRA
🧒 For Kids: Sticking a mini sticky-note on your big textbook instead of rewriting the whole 500-page book!
⚡ For Pros: Low-Rank Adaptation. Freezes base model weights and trains small low-rank adapter matrices (rank 8-64), enabling fine-tuning on a single consumer GPU.
🔌 🌐 Model Context Protocol (MCP Standard)
🧒 For Kids: A universal USB cable that lets any AI robot plug into any database, browser, or game!
⚡ For Pros: Open standard created by Anthropic allowing LLMs to securely connect to external data sources, server tools, and local IDE environments via JSON-RPC.
👁️ Multimodal Tokenizers
🧒 For Kids: Slicing a picture or audio sound into puzzle pieces that the robot can eat and understand just like words!
⚡ For Pros: Vision transformers (ViT) chopping pixel grids into 14x14 patches projected into the same semantic vector space as text tokens.