About and method
The Neural Net covers AI releases and research for people who build things. This page explains exactly how the coverage is produced, including what is automated.
What is automated
Most stories on this site are written by a language model, without a human reading them before they publish. Every story says so at the bottom, and names the model that wrote it. We think you should be able to tell, so we tell you.
What is not automated is the judgement about what this site is for. The source list, the selection rules, the editorial constraints and the thresholds below are set deliberately and reviewed by a person.
We write from primary sources only
A story here is written from the announcement itself — the lab’s post, the release notes, the model card, the paper. We do not rewrite other publications’ reporting. When something is only covered elsewhere, it appears under “On the radar” as a link to the original, and we do not reproduce it.
Other outlets are still useful to us as a signal: if several cover the same announcement quickly, that tells us the story is moving, and it raises the priority of the primary item. They are an input to selection, never a source of text.
The rules the writer works under
- Every fact, number and quote must appear in the source. Nothing is inferred into existence.
- A company’s own benchmark is reported as that company’s claim, not as a finding.
- Where the source is silent, the story says so rather than guessing.
- No padding. A two-paragraph story stays two paragraphs.
- Vendor marketing, conference notices, hiring posts and listicles are rejected before writing.
Sources we watch
Primary (12) — written from: OpenAI, Google AI, Google DeepMind, Hugging Face, NVIDIA, AWS Machine Learning, Microsoft AI, Hugging Face Hub, vLLM releases, Transformers releases, llama.cpp releases, PyTorch releases.
Research (4) — filtered heavily; a preprint is only covered when it concerns a named system or has already been picked up: arXiv cs.AI, arXiv cs.CL, arXiv cs.LG, Berkeley AI Research.
Signal only (4) — never a source of text: TechCrunch AI, The Verge AI, The Decoder, Simon Willison.
The register
The model register holds a different standard from the reporting. The news side is selective: most of what we see is rejected, and a quiet day is a short one. The register is the opposite. It aims to be complete within its stated boundary, because a record with unexplained gaps cannot be cited.
Every field on a register page names the endpoint it came from and the date it was last checked. Parameter counts are exact and never rounded. Context length is max_position_embeddings from the model’s own configuration, which is the architectural maximum rather than whatever a provider chooses to serve, and the pages say so. Where we cannot source a field, it is blank. We do not fill gaps with estimates.
The whole register is available at /api/models.json and /data/models.csv under CC BY 4.0, and as a sortable table at /models/compare. Use it, cite it, and tell us when it is wrong.
Corrections
Every story links to the source it was written from. If a story and its source disagree, the source is right and we want to know. Corrections are made on the story itself, with the modified date updated.