Drag to rotate · scroll to zoom · hover a node for its title · click to open the doc
Pre-ML Era: Classical NLP
Rule-based systems and statistical NLP — what came before learning, and why it hit a wall.
Supervised Learning → Neural Nets
ML as function search, loss and gradient descent, neural network anatomy — with an animated forward pass.
Transformers → Inference
Attention, training, MoE, FlashAttention — plus an animated prefill/decode loop showing why the KV cache exists.
Mathematics of AI
The linear algebra, calculus, and probability that make the previous two docs precise.
Prompt & Context Engineering
Tokens as currency: harnessing AI, saving tokens, fighting hallucination.
Embeddings & RAG
Meaning as geometry: vector search, HNSW, the animated RAG pipeline, and where retrieval fails.
Agents End to End
The tool-use loop animated, MCP, and the design patterns that survive production.
Evaluation
Why benchmarks mislead, the animated eval-building loop, and the LLM-judge problem.
Sampling & Decoding
Temperature, top-p, grammar-constrained JSON, speculative decoding — the animated sampling funnel.
Multimodal & Diffusion
How models see (VLMs) vs how they paint (diffusion) — animated denoising, and why image models can't spell.
Safety & Alignment
The alignment stack, jailbreaks vs prompt injection (animated attack walkthrough), and defense in depth.
Post-Training: RLHF, DPO & LoRA
How weights get their behavior — SFT, preference optimization with the KL leash, DPO's shortcut, and the LoRA math.
Test-Time Compute & Reasoning
Thinking tokens billed as output, RLVR, best-of-n math, and when extra thinking isn't worth it.
Long Context & Memory
1M-token windows: positional math, ring attention, RAG vs agentic memory, and what each approach costs.
Small Models & On-Device AI
The bandwidth wall on a phone: 4-bit math, NPUs, tokens-per-second from LPDDR, hybrid edge→cloud routing.
SOTA LLMs & MoE Architectures
DeepSeek-V3, Mixtral, Qwen3, GLM — the gate math, expert capacity, and why total vs active params is the new spec.
KV Cache Types
MHA → MQA → GQA → MLA — the design space of attention memory, with the unified sizing formula.
Agent Harness Architectures
Orchestrator-workers, memory banks, skills-as-progressive-disclosure — each mapped to the bottleneck it relieves.
Prefill & Decoding Strategies
Chunked prefill, disaggregated serving, speculative decoding with the exact acceptance-rate speedup math.
New Architectures Landscape
Mamba & SSM hybrids, RWKV, BitNet b1.58, byte-level & diffusion LMs, JEPA world models, JEV decision models.
Latest AI Enhancements
The moving target, dated: decision models, MoE everywhere, KV innovation wave, agent standards — linked into deep dives.
AI Security Engineering
Direct vs indirect prompt injection, threat models, the guardrail stack as AND-gates with correlation, agent hardening, red teaming as a regression loop.
Standardization: MCP → Governance
MCP & A2A protocols, interface convergence, NIST AI RMF & EU AI Act — and why standards are context engineering at scale (N×M → N+M).
The Developer Perspective
How the paradigm of coding is shifting for working engineers.
The Developer Harness
Skills, rules, workflows, memory banks, subagents — software answers to the context problem.
CPU vs GPU
Serial genius vs parallel army; memory bandwidth as the real bottleneck — with live data-traffic animation.
AI as an Observability Stack
Agents and models as runtime dependencies: telemetry ↔ quality ↔ semantics ↔ governance planes, SLI/SLO math, evals as production monitors.
AI & Human Evolution
Knowledge transfer, genes to GPUs — the long arc.
Inference Anatomy
Prefill vs decode, KV cache math, batching — animated request lifecycle from Send to streamed tokens.
Context Caching & Cost
Tokens as currency: animated cache hit vs miss, cache-friendly prompt anatomy, the three token price classes.
GPU Memory Hierarchy
The bandwidth wall, arithmetic intensity, and an animated naive-vs-FlashAttention walkthrough.
Software Context Solutions
Skills, rules, workflows, memory banks, subagents — each mapped to the hardware bottleneck it relieves.
The Inference Optimization Stack
Model → Memory → Runtime → Cluster: MoE, SSMs, GQA, quantization, PagedAttention, kernel fusion, tensor parallelism, Splitwise — one animated mental model.