AI Neural
An intuitive, interactive introduction to how AI learns.
Six hands-on courses take you from digit recognition all the way to multimodal LLMs, powered by a self-built on-device tensor engine that trains and runs for real — with matrix-level visualizations that make every step visible.
- 6 hands-on courses
- On-device tensor engine
- Matrix-level visuals
- First 3 courses free
What it can do
No walls of formulas: turn knobs yourself and watch matrices flow — go from “heard of it” to actually understanding it.
Hands-On Courses
MNIST → CNN → YOLO → Transformer → MoE → multimodal: six courses, step by step.
On-Device Engine
A self-built tensor engine accelerated by Accelerate, with automatic differentiation doing backprop — all on your device.
Matrix Visualizations
QKV, attention and weight heatmaps unfold layer by layer — abstract math becomes visible dots.
Real-World Play
A live YOLO camera detects objects around you — learn it, then play with it.
Hands-On Courses
Every course breaks into bite-sized levels: plain-language concept cards first, then hands-on tweaking with instant feedback.
Six Progressive Courses
Start with MNIST digit recognition, move through CNN, YOLO and Transformer, and arrive at MoE and multimodal LLMs.
Gamified Levels
Each course is a chain of small levels — finish an exercise to unlock the next, with real momentum.
Metaphor Cards
Weights as mixing-desk faders, attention as a spotlight — intuition first, math later.
The On-Device Engine
This isn’t a pre-recorded animation — training and inference happen on your device, and the parameter changes are real.
A Self-Built Tensor Engine
Matrix ops are accelerated by Apple’s Accelerate framework (cblas_sgemm) — small footprint, fast runs.
Automatic Differentiation
Backprop is handled by automatic differentiation — the forward pass defines the gradient, and the code is the textbook.
No Network Needed
Every course and model runs locally — learn just fine in airplane mode.
Matrix Visualizations
The “black box” unfolds into visible matrix flows — see exactly where attention is looking.
QKV, Step by Step
Watch queries, keys and values multiply, scale and normalize into attention — animated step by step.
Attention Dots
Attention weights render as dot grids, making it obvious which token the model attends to.
Weight Heatmaps
Watch weights shift live during training and literally see the network learn.
Gallery
From interactive levels to matrix views — a look at the real interface.
FAQ
How much does it cost?
The first 3 courses are completely free; a one-time VIP purchase unlocks all 6 courses plus future updates — no subscriptions.
Do I need math or coding skills?
Not at all. The courses assume zero background, using metaphors and visuals instead of derivations — hands-on play leads to understanding.
Where do the models run? Do I need internet?
Everything runs on your device. The app ships with its own tensor engine — training and inference never leave your phone, and all courses work offline.
Which devices are supported?
iPhone and iPad, requiring iOS 17 or later.
Train your first neural network today
Abstract AI starts with your first visible matrix.