About Above the Fog

Above the Fog is an AI hype radar for the tech, San Francisco, AI and LLM scene. Every day it reads 34 sources — developer communities, research feeds, newsletters, podcasts, news, search interest, package downloads and app rankings — and measures, for each of 96 tracked topics, how big the attention is, how long it lasts and which sources saw it first. The last complete UTC day is Sep 27, 2026.

What this is

The interactive map on the home page (it needs JavaScript) draws the topics and the stories that connect them. Every topic and every theme also has a plain page, and every page has a markdown twin at the same address plus .md. The site is rebuilt from the radar’s data after each daily run; nothing on it is written by hand except the topic descriptions.

How the hype score works

• Each source gives one daily number per topic: mentions, posts, articles, page views, downloads, search interest, a position in a trending list.
• The radar compares that number with the topic’s normal level on that source: the median of its previous 14 normal days. Days that were themselves unusual are left out, so a long hype does not become its own normal.
• A source is above normal when the number stands out from its usual day-to-day variation (z ≥ 2), is at least 30% over normal and, for counts, at least 3 more than normal.
• The score (0–100) adds two parts: up to 75 points for how strongly the sources point at hype (weighted by source and by the size of the topic on that source) and up to 25 points for breadth — how many independent source families are above normal at once (4 or more give the full 25).
• Sources that watch the same audience count once for breadth (Hacker News and time on the Hacker News front page, for example). The contributions of the sources add up exactly to the score, and every topic page breaks it down.
• A source takes part once it has at least 5 days of history for the topic.
• Every day is a UTC day. The current day is partial and is never scored: the site shows the last complete day.

The four stages

A day is hot when the score is at least 40, or at least 25 with at least 2 independent source families above normal. One source alone never makes a hot day. A run of hot days is an episode (one cold day inside it is tolerated); “in hype since” is the first day of the current episode.

“In hype” means rising or hot, everywhere on the site. “First seen on” names the monitored source that went above normal first for a topic — in its current hype episode (counting the 3 days before it) or, outside one, in the last 14 days: where the radar noticed it first, not where the conversation started.

Sources

34 of the 36 sources are active. Each watches one kind of audience: early — the people who build (developers, researchers, early adopters); amplification — the people who explain and repeat (newsletters, tech press, podcasts, events); mainstream — general news and search; adoption — the people who use it (downloads, app rankings, developer forums).

Themes and classification

Every story the sources bring is classified by the Jev model (TypeSafe AI) into themes — a story can carry several — the payment rails it names, the main company involved and the kind of story. A first pass keeps only the stories that are about the scene and have substance; the themes count the kept stories.

In the 7 days to Sep 27, 2026, 38,324 stories were classified and 18,782 kept (49%).

The themes form one tree: Models, Agents, Usage & adoption, Protocols & infra, Business & policy, SF scene. See the theme tree.

How to cite

Above the Fog, AI hype radar, https://abovethefog.app/, Sep 27, 2026.

For one topic: Above the Fog, “Jev — AI hype score and trend”, https://abovethefog.app/topics/jev, Sep 27, 2026.

Data

• Topic data (JSON): every topic with its stage, hype score, 30-day history and themes (open to any website)
• Theme data (JSON): every node of the theme tree with its numbers
• Atom feed: one entry per daily brief
• llms.txt and llms-full.txt: the site for language models
• Markdown: every page’s address plus .md.

Headlines and links belong to their publishers.

Privacy

The site counts its readers — people, search engines and AI agents — to learn who reads it. Here is what it collects.

• The site sets no cookies and stores nothing in your browser (no local or session storage). The visit script’s only identifier is a random page id kept in memory; it is gone when you leave the page.
• When JavaScript runs, the script sends the page address and its kind (a topic, a theme, the map…), the views you open on the map (the part of the address after #), the page that linked here and any utm_ tags in the address.
• It also sends coarse facts about the browser, used only to tell automated browsers from people: whether the browser reports being automated (navigator.webdriver) or names a headless browser in its user agent, the outer size of the window, the number of touch points, the number of browser plugins, the WebGL renderer name and, from the browser’s client hints, its brand, version and whether it is mobile.
• With them come a few facts that are stored with the page view and not used for anything today: the screen size, the inner size of the window, the pixel ratio, the language, the time zone, the number of languages, the platform the client hints name, whether a built-in PDF viewer is on, whether window.chrome exists and the site’s notification permission state (only read: the site never asks for it).
• The first time you interact with the page, it sends how long after the load that was and whether it was a pointer, a touch, a key or a scroll. When you leave, it sends a summary of how the page was used: the time it was visible, how far down you scrolled, the time to the first interaction, and counts of pointer moves (with the distance moved), clicks (and clicks with no pointer movement just before them), key presses, scrolls, touches and events generated by a script, with how straight and how even the pointer’s movement was. Counts only, never which keys or what you type. Search text typed in the item feed never leaves the page.
• A click on a link to another site records only that site’s host name (news.ycombinator.com, for example), not the full address.
• The hosting provider (Firebase Hosting) logs every request: the address, the time, the browser’s user agent, the referring page, the IP address and the country and city it maps to. The radar copies these logs into its own database (on Supabase), where the script’s events also arrive.
• There, the IP address is cut to its network — /24 for IPv4 (the first three numbers), /48 for IPv6 — and turned into a daily visitor code: a one-way hash of the address and the user agent with a random salt that changes every UTC day. Each day’s salt is deleted after 7 days.
• A full IP address is kept only for requests whose user agent names a known crawler that publishes its IP addresses (GPTBot, for example), to check that claim against the published list.
• No name, no e-mail address, no account: the site has no sign-in and asks for nothing.

Publisher

Above the Fog, an independent project. The pages are generated from the radar’s data once a day.