Note

Reading Log Index

Reading Log

A running log of papers, blog posts, and technical knowledge I read.

Date	Title	Type	Source
2026-02-12	GLM-5 - From Vibe Coding to Agentic Engineering	Blog Post	Z.ai (Zhipu AI)
2026-02-12	DeepSeek-V3.2 Paper	Paper	DeepSeek-AI

Date	Title	Type	Source
2026-02-12	Attention Is All You Need	Paper	Vaswani et al. (Google)
2026-02-12	DeepSeek-V3.2 Paper	Paper	DeepSeek-AI

Date	Title	Type	Source
2026-02-12	Recursive Language Models	Paper	Zhang, Kraska, Khattab
2026-02-12	DeepSeek-V3.2 Paper	Paper	DeepSeek-AI

Date	Title	Type	Source
2026-02-12	GLM-5 - From Vibe Coding to Agentic Engineering	Blog Post	Z.ai (Zhipu AI)
2026-02-12	DeepSeek-V3.2 Paper	Paper	DeepSeek-AI

2026-02-12 — Attention Is All You Need — The foundational 2017 paper introducing the Transformer architecture. Replaces RNNs with self-attention (Q/K/V dot-product mechanism), enabling O(1) sequential operations and massive parallelism. Every modern LLM descends from this.
2026-02-12 — DeepSeek-V3.2 Paper — Introduces DeepSeek Sparse Attention (DSA): a two-stage pipeline (Lightning Indexer + top-k selection) that reduces O(L²) attention to O(L*k), cutting long-context inference cost by 2x with negligible quality loss. Also features scaled RL and agentic data synthesis. Frontier-level performance.
2026-02-12 — GLM-5 - From Vibe Coding to Agentic Engineering — Z.ai launches GLM-5: 744B MoE (40B active), MIT-licensed, trained on Huawei Ascend. Best open-source on reasoning/coding/agentic benchmarks. Integrates DSA for 200K context. Introduces "slime" async RL infrastructure.
2026-02-12 — Recursive Language Models — RLMs treat prompts as external variables in a REPL, letting the model write code to recursively decompose and process inputs up to 100x beyond context windows. 8B model post-trained with 48 H100 hours approaches GPT-5 on long-context tasks. Solves long-context from the algorithm side (vs DSA from the architecture side).