A lightweight repository for running an AI Reading Club on foundational papers in modern language models.

Positioning:

From papers to executable understanding.

The club reads foundational AI papers, discusses what they really mean, and connects them to modern LLM systems. Some sessions now pair with executable workshop artifacts in hghalebi/rust-ml, especially for Rust, typed tiny ML, and category-theory-inspired reconstruction. An additional companion resource is hghalebi/category_theory_transformer_rs, where we implement a tiny ML model in Rust from scratch through a category-theory lens.

• Cadence: one paper every two weeks
• Format: 10-15 minute volunteer overview, followed by about 45 minutes of discussion
• Joining (Discord): https://discord.gg/5rAMsuVXXp
• Schedule: sessions/schedule-2026.md (started on 2026-03-11; confirmed history is tracked in docs/workshop-history.md; no sessions in August)

See:

• docs/workshop-history.md (confirmed session and workshop archive)
• docs/announcement-template.md (announcement template)
• docs/why-read.md (motivation)
• docs/organizer-tips.md (organiser tips)

1. Neural Machine Translation of Rare Words with Subword Units (2015)
2. Attention Is All You Need (2017)

3. What Does BERT Look At? An Analysis of BERT's Attention (2019)
4. Attention is not Explanation (2019)
5. Transformer Feed-Forward Layers Are Key-Value Memories (2020)

6. The Curious Case of Neural Text Degeneration (2019)

7. Datasheets for Datasets (2018)
8. Croissant: A Metadata Format for ML-Ready Datasets (2024)

9. FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness (2022)
10. LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale (2022)

11. LoRA: Low-Rank Adaptation of Large Language Models (2021)
12. QLoRA: Efficient Finetuning of Quantized LLMs (2023)
13. The Flan Collection: Designing Data and Methods for Effective Instruction Tuning (2023)
14. LIMA: Less Is More for Alignment (2023)

1. Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity (2021)

Detailed rationale and paper links are in curriculum/README.md.

• Create one GitHub issue per paper (use the "Paper Session" issue template).
• Assign a discussion lead for each session; they prepare a short slide deck or document.
• Add three guiding questions before the session so the discussion has a clear starting point.
• If the mathematics is dense, focus on the abstract, introduction, diagrams, and conclusion.

• curriculum/: the ordered reading list + paper links
• docs/: announcements and organiser guidance
• sessions/: session notes and templates
• docs/workshop-history.md: confirmed AI Reading Club and Rust/ML workshop history
• sections/: workshop and implementation assets grouped by module, including BPE materials under sections/tokenization/
• sections/tokenization/ch02/: BPE notebook walkthrough and assets
• sections/tokenization/rust_bpe_tokenizer/: Rust BPE implementation used in the same module
• sections/bert_attention_paper/: Rust walkthrough that reimplements the BERT attention-analysis paper with runnable step-by-step binaries
• .github/: issue templates and PR template