Method

About 4o4 / AI

The AI press is broken. Same story, thirty outlets, no signal to weight. We rebuilt the stack from scratch.

4o4 / AI is an AI news publication that does what news used to do: it tells you which stories are real, which are echoes, and which are quiet but matter. We track 376 sources — frontier labs, academic institutes, safety researchers, trade press, newsletters, podcasts, even Discord channels — and weight them by what they actually are.

Most AI coverage today is pasted press releases reordered into clickbait. We don't reorder. We cluster, weight, and synthesize.

How we weight sources

Every source sits in a tier. The tier reflects what kind of authority the source carries, not how popular it is.

T0 ×10

Canonical / Primary. The lab itself. OpenAI's own blog, Anthropic, DeepMind, Meta AI, NVIDIA, Mistral. When they post, that is the story.

T1 ×5

Authoritative analysts. SemiAnalysis, Import AI, Simon Willison, Stratechery, Apollo, METR, Latent Space. Their take is itself news.

T2 ×3

Quality trade press. The Information, MIT Tech Review, Platformer, FT, Bloomberg, Reuters, Semafor.

T3 ×2

General trade press. TechCrunch, The Verge, Wired, VentureBeat, Ars Technica.

T4 ×1

Community / aggregators. Hacker News, Reddit, arXiv, Hugging Face Daily Papers.

T5 ×0.5

Ambient. LinkedIn posts, generic YouTube creators, podcasts re-summarising elsewhere.

Five OpenAI re-blogs ≠ five distinct outlets covering it. Five distinct outlets, three of them T1, with the lab itself absent? That's a real scoop.

The signals we extract

Every story gets tagged. Each tag tells you something different about the shape of the coverage.

Official: The lab posted it. You're reading it directly from source.
Scoop: T1 + T2 outlets independently confirm a story the lab hasn't acknowledged. Strong signal.
Analysis: Synthesis or commentary on a known development. The take, not the news.
Underrated: T0/T1 sources have it; the trade press hasn't picked it up yet. Read this first.
Breaking: Velocity spike — cluster doubled in two hours. Story is unfolding.
Context: Background on why a current story matters. Useful when the news cycle assumes you've been paying attention.

What we don't do

No sponsored content. No "sponsored by" sections. No affiliate links.
No engagement bait. No 'Top 10 AI tools you NEED in 2026' lists.
No re-treads of yesterday's news with a different headline.
No claims without source links. Every fact is auditable.

How the pipeline works

RSS, Atom, and a long tail of per-source adapters (sitemaps, YouTube, Reddit JSON, arXiv, Hugging Face Papers) feed a normalised store. New items are embedded and clustered against a 72-hour rolling window. Clusters score by weighted source count, canonical-source presence, velocity, and recency. A pass with Claude Haiku triages; surviving clusters go to Sonnet for synthesis. Every published piece links back to every source it draws from.

It is, as much as we can make it, a machine for distinguishing noise from signal.