An Index — kept by Alek
A section of this.guide 4 reports  ·  established 2026

AI.

Close readings of machine learning papers

One paper at a time, worked from the primary source rather than the summaries. Where the authors release code, the mechanism is re-run locally and the page carries the numbers that came back — not the ones that sounded right.

N° 01Published

The reports

Ref Title Subject Source Published
A·04 Physical AI
The hardware the research is waiting on		 Embodiment · instrumentation · real-time systems Latency · tactile · sync · buses 2026-08-13
A·03 RLVR
Reinforcement learning with verifiable rewards		 Post-training · reasoning models GRPO · DAPO · Dr. GRPO · R1-Zero 2026-08-13
A·02 KV cache
The tensor that decides what your cluster can serve		 Transformer inference · memory systems GQA · MLA · PagedAttention · KIVI 2026-08-12
A·01 Engram
Conditional memory via scalable lookup		 DeepSeek-AI · Peking University github.com/deepseek-ai/Engram 2026-08-11
N° 02Method

How these are written

Three rules, meant to be enforceable rather than decorative.

01 — Sources

Primary only

The paper, the repository, the released weights or tokenizer. Secondary coverage is cited for context and never for a fact.

02 — Verification

Reproduce what is reproducible

If the arithmetic can be re-run, it is, and the output is checked against the reference implementation before anything is published.

03 — Honesty

Label every estimate

A figure redrawn from a plot rather than a table says so in its caption. A derived quantity shows the derivation.

The section publishes when a piece is finished rather than on a schedule. Next up: attention-variant teardowns, and a look at what the hyper-connection literature actually changes.