First screen, captured 10 days ago. Open the video.

The Little Book of Generative AI Foundations: An Intuitive Mathematical Primer

What HN thinks

A math primer thread that turned into a book club, a reading-device support group, and briefly a fan club for a cat named Princess Donut.

Notation complaint mostly rejected
One commenter said the first formula's ê1 went unexplained, but most replied that it is standard unit basis vector notation and that the book does define it.
Swapping favorite ML books
Many used the thread to recommend other deep learning texts, especially ones with exercises, and resources on attention.
Readable but different scope
Readers found it surprisingly accessible from linear algebra basics, though some noted it leans toward diffusion-style generative models rather than transformers.

Consensus: The book is seen as a readable, well-built primer, and the ê1 notation is considered standard. People mostly wanted more recommendations to go with it.

Recommended in the comments

The Little Book of Deep Learning ↗ (book): A favorite concise primer, and it offers a phone-formatted PDF.
Understanding Deep Learning ↗ (book): Has lots of exercises, including many programming ones.
Deep Learning: Foundations and Concepts ↗ (book): Recommended as having the best-balanced theoretical exercises.
Dive into Deep Learning ↗ (book): A practical, hands-on option with mixed exercises.
Why Machines Learn (book): Praised for its pacing and for building intuition gently while keeping some rigor.
Formal Algorithms for Transformers ↗ (other): Called the most succinct introduction to attention.

Summarised by Claude from the comment thread 9 days ago.

Points: 351
Comments: 55
Best rank: #1
Posted: 2026-11-03 by ibobev
Category: AI & ML

generative-models diffusion-models variational-autoencoders gans normalizing-flows machine-learning-math arxiv

All captures:

Full page

Captured 10 days ago. Click for the whole screenshot.

Looks similar

Screenshot of arxiv.org: Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation

Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation (arxiv.org)

171 points by theanonymousone 4 days ago | 39 comments

arXiv paper introducing Next-Concept Modeling, a coarse-to-fine text generation method using a multiscale residual VQ-VAE and parallel transformer decoding, faster than diffusion LMs.

language-models diffusion-language-models vq-vae parallel-decoding

Screenshot of arxiv.org: Multi-Token Prediction Done Right

Multi-Token Prediction Done Right (arxiv.org)

3 points by matt_d 9 days ago | 2 comments

arXiv preprint introducing Soft-MTP, a one-line change to multi-token prediction training that keeps only a KL self-distillation gradient signal.

multi-token-prediction speculative-decoding language-models llm-pretraining

Screenshot of arxiv.org: Memorilla: Latent Semantic Memory for LLMs

Memorilla: Latent Semantic Memory for LLMs (arxiv.org)

101 points by handfuloflight 7 days ago | 10 comments

arXiv paper introducing Memorilla, a system that compresses document collections into query-dependent memory tokens in the input embedding space of frozen LLMs.

llm-memory memory-tokens long-context retrieval

Screenshot of arxiv.org: USDB: Scaling Storage for Next-Generation Recommender Systems at ByteDance

USDB: Scaling Storage for Next-Generation Recommender Systems at ByteDance (arxiv.org)

47 points by 1equalsequals1 10 days ago | 7 comments

arXiv paper presenting USDB, ByteDance's unified sequence storage system for recommender model streaming and batch workloads, using CascadeLSM compaction and hybrid columnar format.

usdb bytedance recommender-systems lsm-tree

Screenshot of arxiv.org: The alien space of science: Coherent but cognitively unavailable directions

The alien space of science: Coherent but cognitively unavailable directions (arxiv.org)

195 points by martincsweiss 22 days ago | 80 comments

arXiv abstract page for a NeurIPS 2026 paper proposing a method to sample scientifically coherent research directions that existing researcher communities are unlikely to propose.

research-ideation llm scientific-discovery novelty

Screenshot of arxiv.org: Proving and improving the Platonic Representation Hypothesis with JEPAs

Proving and improving the Platonic Representation Hypothesis with JEPAs (arxiv.org)

156 points by Anon84 12 days ago | 78 comments

arXiv paper by Balestriero and LeCun proving a Platonic Representation Theorem for JEPAs and showing good views align modalities more efficiently than next-token prediction.

jepa platonic-representation-hypothesis representation-learning multimodal-alignment

Full page

Looks similar

Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation (arxiv.org)

Multi-Token Prediction Done Right (arxiv.org)

Memorilla: Latent Semantic Memory for LLMs (arxiv.org)

USDB: Scaling Storage for Next-Generation Recommender Systems at ByteDance (arxiv.org)

The alien space of science: Coherent but cognitively unavailable directions (arxiv.org)

Proving and improving the Platonic Representation Hypothesis with JEPAs (arxiv.org)

Related stories

Generative AI for Beginners (microsoft.github.io)

The Little Book of Deep Learning [pdf](fleuret.org)

The Little Book of Linear Algebra (github.com)

Learn Your Way: Reimagining Textbooks with Generative AI (research.google)

Mathematical Introduction to Deep Learning: Methods, Implementations, and Theory (arxiv.org)

Sharing AI progress in mathematics (openai.com)