v1.0  ·  30 interactive tools

learn AI
interactively

LLMs, transformers, RAG, agents, tokenization, attention — all visualized step-by-step in the browser. No installs, no signup.

what is an LLM? ✦ explore tools ↓
⭐ star on GitHub ☕ buy me a coffee
30+
tools
100+
glossary terms
∞
tokens explained
0
installs needed
$ tokenize — type anything below
your text
tokens BPE · cl100k
0
tokens
0
chars
0
words
$0
est. cost
Hello, world! ChatGPT is an LLM Tokenization splits text antidisestablishmentarianism 1+1=2 and 99×99=9801
tools
⌕ ⌘K
embed → Q,K,V
attention heads
FFN → LayerNorm
⚙️
Transformer
Walk through every block of the transformer — embedding, multi-head attention, feed-forward, residuals.
open tool →
high confidence →
wrong answer
factual drift
🌀
Hallucination
Why LLMs confidently produce false facts — probability collapse, training bias, and mitigations.
open tool →
query → embed
vector search →
context + generate
📚
RAG Pipeline
Animated Retrieval-Augmented Generation — query, retrieve, augment, and generate step by step.
open tool →
observe → plan
tool_call →
result → respond
🤖
AI Agents
The agent loop — observe, think, call tools, get results, respond. ReAct and function-calling patterns.
open tool →
"king" - "man"
+ "woman" =
"queen" ✓
📐
Embeddings
How words become vectors — semantic similarity, cosine distance, and the famous king-queen analogy.
open tool →
temp=0 → greedy
temp=1 → sample
top-p · top-k
🌡️
Temperature & Sampling
How temperature, top-p, top-k, and repetition penalty control output randomness and quality.
open tool →
input → tokens
embedding → attn
softmax → output
🧠
How LLMs Work
Step through token prediction — from input text to probability distribution to output.
open tool →
Q · Kᵀ / √d
softmax → weights
weighted sum V
👁️
Attention Mechanism
Visualize how tokens attend to each other — Query, Key, Value matrices and attention scores.
open tool →
hello → [15339]
world → [1917]
BPE · WordPiece
🔤
Tokenization
See how text breaks into tokens — BPE, subword splitting, token IDs, and vocabulary.
open tool →
forward pass →
loss → backprop
gradient descent
🏋️
Training Process
Pre-training walkthrough — forward pass, cross-entropy loss, backpropagation, weight updates.
open tool →
pretrained weights
+ task data →
specialized model
🎯
Fine-tuning
Full fine-tune vs LoRA vs prefix tuning — compare approaches, data requirements, and tradeoffs.
open tool →
SFT → reward
model → PPO
alignment loop
🎓
RLHF
Reinforcement Learning from Human Feedback — SFT, reward model training, PPO policy optimization.
open tool →
system: You are…
user: Explain…
assistant: Sure…
💬
Prompt Engineering
System prompts, few-shot, chain-of-thought, role prompting — patterns that improve outputs.
open tool →
system [512]
history [2048]
new [128] | 128K
📏
Context Window
Context limits visualized — how system prompt, history, and new input share the token budget.
open tool →
Q: 9+3×5=?
think step by step
→ 9 + 15 = 24 ✓
🧩
Chain of Thought
Why CoT works — scratchpad reasoning, zero-shot CoT, tree-of-thought, and self-consistency.
open tool →
get_weather({
city: "NYC"
}) → 72°F
🔧
Function Calling
Tool use step by step — schema definition, model decides when to call, result injection.
open tool →
host ↔ client
← resources
← tools ← prompts
🔌
MCP Protocol
Model Context Protocol — hosts, clients, servers, resources, tools, prompts, and the handshake.
open tool →
embed → HNSW
ANN search →
top-k results
🗄️
Vector Database
How vector DBs store embeddings — HNSW indexing, approximate nearest neighbor, and retrieval.
open tool →
noise → denoise
T steps →
clean image
🎨
Diffusion Models
How DALL-E/Stable Diffusion works — forward noise process, reverse denoising, U-Net, guidance.
open tool →
prefill → cache
K,V stored →
decode reuses
⚡
KV Cache
Why inference is fast after the first token — Key-Value caching during autoregressive decoding.
open tool →
W + ΔW
ΔW = A·B (rank r)
0.1% of params
🪡
LoRA
Low-Rank Adaptation — how adding tiny matrices to frozen weights enables cheap fine-tuning.
open tool →
router → expert 3
→ expert 7
sparse activation
👥
Mixture of Experts
MoE architecture — gating network, expert routing, sparse activation, and Mixtral/GPT-4 style.
open tool →
FP32 → INT8
4-bit GGUF
QLoRA · AWQ
📦
Quantization
Model compression — FP32 to INT4/INT8, GGUF, AWQ, GPTQ — tradeoffs in size vs quality.
open tool →
image → patch
embed → concat
text tokens → LLM
🖼️
Multimodal AI
Vision + language — how images become tokens, ViT patch embedding, CLIP alignment, GPT-4V.
open tool →
prefill → 1 fwd
decode → N×fwd
batch · streaming
🚀
LLM Inference
Prefill vs decode phase, TTFT vs TPS, batching strategies, streaming output mechanics.
open tool →
0-shot: just ask
1-shot: 1 example
5-shot: 5 examples
🎲
Few-Shot Learning
In-context learning — zero-shot vs few-shot vs chain-of-thought, and why examples help.
open tool →
role · rules
persona · format
constraints ↓
📋
System Prompt
Anatomy of a system prompt — role definition, rules, format instructions, and injection risks.
open tool →
claim → source
citation check
factual anchor
⚓
Grounding
Grounding LLM outputs in real sources — RAG grounding, citation generation, attribution chains.
open tool →
MMLU · HumanEval
GSM8K · MATH
HELM · BIG-bench
📊
AI Benchmarks
What MMLU, HumanEval, MATH, and HELM actually measure — tasks, metrics, and leaderboard caveats.
open tool →
AGI · CoT · DPO
GRPO · KV · LoRA
MoE · RAG · RLHF
📖
AI Glossary
100+ AI/ML terms — searchable A–Z reference from AGI to zero-shot with plain-English explanations.
open tool →
comments & feedback
also by echobash
devdunia .com networking tools echobash .com portfolio promptflix .in prompt library