About AI Trend Notifier

What this is

Most AI news is written once and never touched again, so the same background gets re-explained every week and nothing accumulates. This site inverts that. Each day’s reading is folded into permanent pages — one per lab, model, concept, paper and person — and those pages are linked to each other. The daily brief is the by-product; the wiki is the artefact.

It currently holds 108 pages and 373 cross-links, built from 57 daily briefs since 2026-05-16. The home page is that link graph, drawn from the real pages — not an illustration.

Where the material comes from

Everything starts from a source that actually published something — feeds are polled on a schedule, and nothing is discovered by asking a model what it remembers. The standing set:

• Papers — arXiv (cs.AI, cs.CL, cs.LG) and Hugging Face Daily Papers.
• Labs, first-party only — official blogs and release notes from OpenAI, Google DeepMind, Anthropic, Meta AI, xAI, Mistral and others.
• Chinese labs, on a rotation — DeepSeek, Qwen, Moonshot (Kimi), Z.ai/Zhipu and MiniMax, two per day. These are the releases English-language coverage reports latest: MiniMax M3 was picked up 46 days after the fact and GLM-5.2 after 27, both only once a Western outlet wrote about them. None of these labs publishes a usable feed, so this set is found by running searches rather than by polling — slower and less certain than the rest, which is why it is called out separately here.
• Policy — the US Federal Register and comparable primary sources, rather than reporting about them.
• Practitioner signal — a small set of researcher accounts and aggregators, used to notice things, never as the citation of record.

Announcements are cited to the announcing organisation wherever one exists. When a third party is the only available source, the page says so.

What the agent decides, and what it does not

Running on a schedule: polling, deciding what clears the threshold, writing and updating pages, maintaining the index and the log, producing the daily brief, and — weekly — a synthesis and a consistency lint over the whole wiki.

Reserved for a human: which sources are tracked at all, how interests are weighted, reviewing newly created entity and concept pages, and resolving contradictions between sources. The machine does the tedium; the editorial perimeter is set by a person.

Rules every page is held to

• No claim without a citation. Every factual statement points at a source — an original snapshot kept on this site, or an external URL.
• Sources are immutable. Snapshots are captured once and never edited, so a citation still resolves after the original page changes or disappears. The wiki on top of them is expected to change.
• Unknown is written down, not omitted. On a model page, a spec row with no reliable value reads unknown rather than being dropped — an omitted row is invisible, an explicit one is countable, and the weekly lint counts them.
• Conflicts are kept, not smoothed over. Where sources disagree, the body follows the most authoritative one and the others are recorded beneath it with their attribution.
• No fact is written twice. If it is already on a page, the brief links to it instead of restating it.

What it is not

It is not a newswire, and it does not try to be first. Items are picked up when a primary source publishes them, which for quieter labs can be a few days after the fact. It is not editorial commentary or investment advice. And because the pages are machine-written, errors are possible — treat the citation, not the sentence, as the authority.

Found something wrong? Corrections are welcome, and the page and its source will both be checked.

Who runs it

Built and maintained by undefined studio, an independent one-person studio. The system that produces this site is its own case study — see undefined-studio.dev.