Megadose is a self-hosted mission-control view for AI progress: ranked, decayed, deduped news from public sources, plus live SWE-bench Verified, LMArena, Terminal-Bench, and METR Horizon snapshots.
Megadose is a reader, not a primary source. It does not run SWE-bench, LMArena, Terminal-Bench, or METR, and it does not publish original reporting. It watches public feeds, ranks what actually moved, and points you at the publisher, the paper, or the upstream leaderboard. Cite those originals; use Megadose to find them.
The product is a self-hosted mission-control dashboard for frontier AI: model releases, viral OSS, company news, research papers, and products you can try. It ingests about twenty sources — lab blogs from Anthropic, OpenAI, and DeepMind, plus ArXiv, Hugging Face Daily Papers, Hacker News, GitHub, and a handful of specialist feeds — and classifies each item with an LLM. The public site is https://megadose.ai. Feeds refresh about every thirty minutes.
Ranking is not recency alone. Each kept item gets an importance score. Display rank multiplies that score by the source's trust weight, then applies exponential time decay with a two-day half-life. A major release stays high for about a day; older items fade so the home feed stays live. Stories older than seven days fall off the ranked home list; stable permalinks at /i/{id} remain.
The same announcement often arrives from a lab blog, a news aggregator, and Hacker News. Dedupe canonicalizes URLs (stripping share tokens), hashes them, and groups near-duplicate titles by distinctive tokens such as model names and version strings, so one event becomes one card instead of three copies.
Audience filtering is the classifier skip list, not a demographic tag. Policy recaps, political coverage, consumer surveys, generic PR, changelog trivia, and rehashes are dropped so the feed stays useful to people tracking frontier models, OSS, research, funding, and products — not a general-interest AI firehose.
Alongside the news, Megadose mirrors live snapshots of SWE-bench Verified, LMArena, Terminal-Bench, and METR Horizon. Those pages are copies of upstream boards, not Megadose's own evals. Treat them as a convenience view of coding, chat, terminal-agent, and time-horizon scores sitting next to the day's releases.