About
Most writing about machine learning research is optimized for speed. This site is optimized for the opposite: notes written slowly, by a person, with the paper in one window and its reference implementation in the other.
What this site is
Technical notes on LLMs, VLMs, multimodal models, quantization, training, inference, and AI systems. Each note is built from primary sources: the paper, the authors' own code, and the benchmark tables with their setups attached. The aim is that you can understand the method, judge the evidence yourself, and know what the implementation actually does before you attempt a reproduction.
How notes are structured
Every note follows the same anatomy:
- Core insights — what actually matters in the work, and what surprised the author.
- Method — the mechanism from first principles, with equations, including at least one the author re-derived or checked.
- Paper vs. code — the paper's algorithm matched line by line against the reference implementation, pinned to a commit. What the code does that the paper never says is usually the most useful part.
- Claims & evidence — each headline claim traced to the experiment behind it, with a verdict: verified, partial, unverified, or refuted.
- Benchmarks — results next to their baselines, with hardware, precision, and evaluation setup stated.
- Limitations & open questions — what the paper does not show, and what the author still doesn't understand.
- Reproduction notes — what was actually run or read, stated precisely. A code audit is labeled a code audit, not a reproduction.
Verification rubric
| Verdict | Meaning |
|---|---|
| verified | The claim is directly supported by the paper's own experiments and has survived independent replication or our own checks. |
| partial | Supported under the paper's specific setup, but weaker or unconfirmed outside it (different scale, hardware, or data). |
| unverified | Stated in the paper but not yet checked against evidence beyond it. |
| refuted | Contradicted by later work or failed replication. |
What this site is not
Not an AI news site, and not a paper-summary feed. Notes are published
when the analysis is done, not when the paper trends. Every note lists
its own status: draft notes are working
documents, and reviewed notes have
been checked end to end. The house style is documented in the repo
(WRITING.md) and enforced on every note, including the
rule that a claim is labeled by its source — the paper says, the code
does, or the author verified.
Contact
Corrections are the most valuable contribution a reader can make. Open an issue on GitHub if a claim, equation, or number is wrong.