00🧭
Meet AI
Build the big picture: AI history, the evolution from symbolic systems to deep learning, and why modern LLMs work.
Your personal AI learning platform · 2026
Understand AI.
Build what comes next.
Explore the complete journey behind modern AI — from the foundations and architecture that make a model possible, through training and inference, to complete AI systems and production engineering.
The curriculum
Eight connected learning tracks
One primary home for each concept. Learn the fundamentals, understand the model, then build real systems.
The dashboard and detailed roadmap share one numbering system: 00–07. Career growth and live labs are supporting resources across the tracks.
01∑
Foundations
Build the mathematical and programming foundations needed to understand learning, vectors, optimization and neural networks.
02⌘
Language & Model Architecture
Follow text into tokens, vectors and embeddings; then understand recurrent models, attention, QKV, Transformers and GPT.
03⚙️
Model Training Pipeline
Understand how raw data becomes a trained, aligned and deployable language model across 14 connected stages.
04✧
Model Inference Pipeline
Trace a user question through tokenization, embeddings, Transformer computation, next-token prediction and response delivery.
05🕸️
AI Systems Engineering
Connect models to external knowledge, context, tools and actions with RAG, vector databases, agents and MCP.
06🚀
Production & Optimization
Move from experiments to reliable AI services with evaluation, safety, efficient models, optimized inference and production operations.
07✦
Advanced AI & Frontier
Explore reasoning models, multimodal architectures, diffusion, Vision Transformers, Mixture of Experts and open research questions.
Your track progress
0 / 8Completion is saved in this browser.
Parts A, B and C
The complete GPT lifecycle
Training creates the model. Inference uses the trained model to produce an answer.
PART A · MODEL CREATION
Model Training
How the model learns patterns, acquires capabilities and has its weights updated.
Data preparationArchitecturePretrainingSFTPreference alignmentSafety & evaluationOptimization & deployment
Output: trained model, tokenizer and deployable artifacts
Explore 14 training stages →Read all 14 training stages ↗→
PART B · USER INTERACTION
Model Inference
How a deployed model processes a prompt and generates a response using fixed model weights.
User inputTokenizationEmbeddingsTransformerPredictionDecodingOutput
Output: response delivered to the user
Open interactive inference lab →Read all 11 inference stages ↗The connection: model training learns the parameters. Deployment packages and serves the trained model. Inference runs forward computation with those parameters. Tokenization is used in both training and inference.
Part D
Training vs inference
Two different operating modes of the same model lifecycle.
| Dimension | Model training | Model inference |
|---|---|---|
| Purpose | Learn patterns and adjust model parameters. | Use learned parameters to process an input and produce output. |
| Input | Prepared training examples and targets. | User prompt, instructions, history and other context. |
| Computation | Forward pass, loss, backpropagation and optimizer updates. | Forward computation and next-token probability calculation. |
| Weights | Updated during training. | Normally fixed for a deployed model version. |
| Output | Trained model checkpoint / artifacts. | Generated response. |
| Frequency | Development or post-training runs. | Every request / generation session. |
Recommended learning order · Part E
18-step study sequence, mapped to the 8 tracks
The 18 steps provide a chronological study order. The eight tracks remain the main curriculum structure; this sequence does not create a second numbering system.
Meet AI & Orientation
Build the mental model and understand the evolution of modern AI.
Before Step 01Explore track →Build the foundations
Python, math, linear algebra & probability, ML fundamentals, deep learning & backprop.
Steps 01–05Explore track →Understand language models
NLP sequences, tokenization, embeddings, Transformers, QKV attention and GPT architecture.
Steps 06–10Explore track →Teach the model
LLM pretraining, supervised fine-tuning, RLHF and DPO alignment.
Steps 11–13Explore 14-stage pipeline →Run the trained model
Inference, decoding, probability sampling and KV-cache acceleration.
Step 14Explore 11-stage pipeline →Connect knowledge & tools
RAG, embeddings for retrieval, vector databases, chunking, reranking and hybrid search.
Step 15Explore track →Make models production-ready
Evaluation & safety, quantization & optimization, deployment & serving.
Steps 16–18Explore track →Explore the frontier
Reasoning, multimodal models, diffusion, MoE and research directions.
Advanced extension beyond the 18-step sequenceExplore track →How to read this: Track 00 is the orientation before the numbered study sequence. Steps 01–18 then progress through Tracks 01–06. Track 07 extends the curriculum into advanced and frontier topics beyond those 18 steps.
Learn by building
Labs, experiments & projects
Explore working visualizations, run hands-on experiments, and turn concepts into small implementations.
FEATURED INTERACTIVE JOURNEY
Inside the Answer — From Question to Answer
Follow an input prompt through an end-to-end visual playground, from system framing and token IDs to attention, sampling and the final synthesized response.
Stage 1 · The Prompt & System FramingStage 2 · Subword Tokenizer & IDsStage 3 · 1,536D Embedding Lookup MatrixStage 4 · Multi-Head Attention & QKV ScoresStage 5 · Feed-Forward (FFN) & Layer StacksStage 6 · Logits & Temperature SamplingStage 7 · Autoregressive Generation LoopStage 8 · Final Synthesized Output
🚀 Launch Question to Answer Flow →
Explore the mechanics
Interactive live labs
Focused playgrounds for understanding how language models process inputs and generate outputs.
INTERACTIVE LAB
Sentence → Response Chain
19 stages · full transformer walkthrough
›Full 19-stage walkthrough›Step 2 · Tokenization›Step 6 · Self-attention›Step 14 · Softmax + sampling›Step 19 · Final answer
Open Lab →
INTERACTIVE LAB
Attention Lab
Q·K scores · scaling · softmax · animated stages
›Animated stage-by-stage walkthrough›QKV vector visualiser›Softmax attention matrix›Live Ollama inference
Open Lab →
INTERACTIVE LAB
Attention Deep Dive
8-step math · QKV projection · scores · softmax
›Step 1 · Token inputs›Step 2 · Q, K, V projections›Step 3 · Dot product scores›Step 4 · Scaled softmax & weights›Step 8 · Complete multi-head block
Open Lab →
INTERACTIVE LAB
LLM Journey
Temperature · logits · softmax · sampling
Open Lab →INTERACTIVE LAB
Tokenization Live
Live input · token counts · BPE intuition
Open Lab →INTERACTIVE LAB
Vectors Live
Dot product · Euclidean · cosine similarity
Open Lab →INTERACTIVE LAB
Embeddings Live
Context changes meaning · cosine comparison
›Context comparison lab›“bank” · river vs finance›Static vs contextual embeddings›Live Ollama inference
Open Lab →
INTERACTIVE LAB
Transformer Architecture
8 stages · multi-head attention · FFN · residuals
›Stage 1 · Embeddings & positions›Stage 2 · Multi-head attention›Stage 3 · Add & LayerNorm›Stage 4 · Feed-Forward MLP›Stage 8 · Complete stacked block
Open Lab →
INTERACTIVE LAB
Inside Transformer · Signal Flow
Interactive sentence routes · QKV vectors · 6-layer stack
›Animated token-to-token flow›Interactive Q·K·V vector inspectors›Dynamic attention matrix heatmap›6-layer repeated block execution
Open Lab →
INTERACTIVE LAB
Generation & Sampling
8 steps · logits · temperature · top-k · top-p
›Step 1 · Unnormalized logits›Step 2 · Temperature scaling›Step 3 · Softmax probabilities›Step 4 · Top-K and Top-P filtering›Step 8 · Autoregressive text loop
Open Lab →
INTERACTIVE LAB
Agent Loop Live
Think · Act · Observe · Repeat
Open Lab →Build beyond the walkthroughs
Projects & learning resources
Continue into implementation projects, connected learning spaces and progress tools.
CURRICULUM RULEOne concept, one primary home
Shared concepts such as tokenization are taught once, then referenced from training and inference.
BUILD IN PARALLELLearn → visualize → code → test
Every important topic should lead to a tangible experiment, not just a reading checklist.
PRESERVE EXISTING WORKKeep your current labs
The inference walkthrough and existing topic lessons remain intact while the curriculum is reorganized.