TensorShin

What’s new

A running log of what we’ve shipped. Newest first.

Three new Paper Labs and a papers index

Sep 26, 2026

New Paper Labs for Adam (bias correction and why its step ignores the loss scale), Batch Normalization (forward and backward, inference with running statistics, training at a much larger learning rate) and word2vec (negative sampling, the unigram noise distribution, and an analogy probe against untrained vectors).
A papers index for signed-in learners lists every paper cited in the curriculum, with filters for domain, year and topic, and links to the canonical source and any lab built around it.
Topic pages now give each paper a clear Read paper link that names its source, plus Open lab and Open mission where a hands-on activity exists.
Curriculum paper references were checked against arXiv, Crossref and publishers: 456 papers added, incorrect identifiers fixed, and unofficial PDF copies replaced with canonical links.
Fixed pages across the app falling back to a serif system font.

Sep 20, 2026

Explore Attention Is All You Need through a small NumPy experiment: change future tokens, compare causal and open masks, and inspect saved attention weights by head.
Implement stable masked attention and head splitting, compare scaling on identical inputs, and test your code on a longer sequence.
Keep an account notebook with code versions, controlled comparisons, optional notes and an exportable decision record. The mission tests the attention mechanism without training a full Transformer.

Sep 16, 2026

YouTube players load only when you choose to load them. The viewer shows the original publisher, course source, and a link to watch on YouTube.
In-app playback is limited to reviewed video-to-topic selections. Other resource links open their original websites.
Updated Privacy & Terms to explain YouTube data collection and clarify that linked videos are not automatically licensed for copying or redistribution.

Sep 15, 2026

Paper Labs now covers five foundational papers: Attention Is All You Need, ResNet, BERT, LoRA, and Denoising Diffusion Probabilistic Models (DDPM). Each study includes original-paper references and explains what its experiments can establish.
Practice through 15 connected challenges: build working components, repair realistic bugs, and run controlled experiments. Explore residual gradients, masked-token learning, adapter training, denoising, and future-token leakage.
Run your Python implementations in the sandbox and inspect their plots and numerical results. Completion requires passing implementation checks, captured measurements, and a conclusion; reflections are no longer scored by text length.
Save notebooks to your account, recover browser drafts, and export your code, predictions, observations, and notes. Editing code clears its accepted experiment evidence until the updated implementation is verified.
The original attention-scaling experiment remains available as an introduction, with a link to the new causal-attention challenges. Related paper references in the curriculum link directly to their labs.
Fixed missing-module errors that could interrupt local development after a production build.

Jun 23, 2026

New Visual Calculus lab — the calculus that powers AI/ML, taught through its ML role: derivatives as loss gradients, partials and the Jacobian, the Hessian and curvature, and backpropagation as the chain rule through a network. Each concept pairs an interactive graph with ML-framed practice.
The knowledge graph has a new look

Jun 6, 2026

Coding problems can now draw a live graph from your solution — slide a parameter and watch the curve, decision boundary, or loss change.
Topic pages show intuition graphs next to the research papers (starting with the bias–variance tradeoff).
Every graph supports a short “what to notice” caption.

Jun 6, 2026

The landing page now features Tensor Code — 90+ from-scratch problems across guided tracks, run in an isolated sandbox.
The topic-page focus timer is now a calm Pomodoro with an optional, dismissible break nudge.