What's moving through AI, ranked by lead time.
Radar is an hourly index of what's moving through the AI ecosystem, ranked by where each story sits on the lead-time gradient: papers become repos become community threads become analysis. The earlier a story is caught, the more head start you have before the crowd. Every row carries its receipts — which source saw it first, how many independent sources have it, how long ago — so the ranking explains itself. It's the public face of the same pipeline that feeds Vlad's Newsletter.
frozen 23 Jul 2026, 22:17 UTC · 67/78 feeds live · 17,230 tracked since 12 Jun
- 1 Red-teaming LLM unlearning: LUNAR's "forgotten" knowledge is still recoverable lesswrong.com · 24m
- 2 Inception in DiffusionGemma - Jailbreaking a Diffusion Language Model by Pinning Tokens Anywhere on the Canvas lesswrong.com · 36m
- 3 ACX Atlanta August Meetup lesswrong.com · 48m
- 4 Show HN: Advanced Coffee Search Covering Over 17,000 coffees beangenie.coffee · 48m
- 5 Show HN: Connecting my dumb garage and car's homelink button to Home Assistant dsauerbrun.com · 48m
338 stories entered the watched feeds in the last 24 hours — 25 in papers, 1 in repos, 280 in community, 32 in analysis. The scored board skews to Community (39 of 40). 2 stories are heating across independent sources, and the median lead across the ledger's 19 observed fires is 4h — measured, not estimated. Freshest receipt, 22h ago: caught in Hacker News front 4h before Dynomight carried it. The oldest survivor has held the board for 22h.
- Community → Analysis +4h lead Hacker News front first · then Dynomight 22h agoDoes creatine make you smarter?
- Community → Analysis +2h lead Hacker News front first · then OpenAI News 2d agoOpenAI and Hugging Face partner to address security incident
- Papers → Community +2h lead HF Daily Papers first · then Hacker News front 5d agoQubes OS Security in the Public Record
- 1Red-teaming LLM unlearning: LUNAR's "forgotten" knowledge is still recoverablelesswrong.comLive Signal Community · 24m old · via LessWrong
- 2Inception in DiffusionGemma - Jailbreaking a Diffusion Language Model by Pinning Tokens Anywhere on the Canvaslesswrong.comLive Signal Community · 36m old · via LessWrong
- 3
- 4Show HN: Advanced Coffee Search Covering Over 17,000 coffeesbeangenie.coffeeLive Signal Community · 48m old · via Show HN
- 5Show HN: Connecting my dumb garage and car's homelink button to Home Assistantdsauerbrun.comLive Signal Community · 48m old · via Show HN
- 6
- 7Show HN: Skim – a minimalist open-source email client for Windows (MIT)skim-tech.comLive Signal Community · 1h old · via Show HN
- 8You should selfhost your mail (in 2026)blog.haschek.atLive Signal Community · 1h old · via Lobsters
- 9Namecheap Gave My Account to an Unverified Third Party Just Because They Askednews.ycombinator.comLive Signal Community · 1h old · seen in 2 sources
- 10Show HN: AgentPulse, drift investigation for multi-agent systemsprove-ai.github.ioLive Signal Community · 1h old · via Show HN
- 11Couple pay >$800k for a gene-editing therapy for their daughter. She died.science.orgLive Signal Community · 1h old · via Hacker News front
- 12Show HN: LLM-budget-cap – atomic Redis spend cap for LLM APIsgithub.comLive Signal Community · 2h old · via Show HN
- 13Show HN: Grounded-forge – RAG with summaries and task views precomputed ingestgithub.comLive Signal Community · 2h old · via Show HN
- 14Show HN: Experimental freshness-first caching library for FastAPIgithub.comLive Signal Community · 2h old · via Show HN
- 15Estimating LLM Training FLOPs on the Nvidia Jetson Orin Nanolesswrong.comLive Signal Community · 2h old · via LessWrong
- 16Show HN: OpenTrust – Browser trust signals for the AI eragithub.comLive Signal Community · 2h old · via Show HN
- 17Show HN: Practice chess openings with a voice coachchessmaster.alirizasaral.comLive Signal Community · 2h old · via Show HN
- 18Not Pinning Your OpenRouter Provider Might Invalidate Your Researchlesswrong.comLive Signal Community · 2h old · via LessWrong
- 19Big Tech Isn't Hiding $1.65T of Debtfinterm.aiLive Signal Community · 2h old · via Hacker News front
- 20Show HN: Biscuit – Native, extensible, AI code editor. <20 mb in sizegithub.comLive Signal Community · 2h old · via Show HN
- 21
- 22Show HN: CobaltCode – Dedicated persistent computer for Codexcobaltcode.aiLive Signal Community · 2h old · via Show HN
- 23Show HN: Open-artifact – view and share your LLM docs and artifactsgithub.comLive Signal Community · 2h old · via Show HN
- 24OpenAI’s accidental cyberattack against Hugging Face is science fiction that happenedsimonwillison.netAI Firehose Analysis · 22h old · seen in 3 sources HEATING ×2
- 25Show HN: An inbox for your AI coding agents (herdr plugin)github.comLive Signal Community · 2h old · via Show HN
- 26
- 27Silent Replacement of Trusted macOS App Executablesmysk.blogLive Signal Community · 9h old · seen in 2 sources HEATING ×2
- 28Show HN: Production duplex speech model for revenue callsmetavoice.ioLive Signal Community · 3h old · via Show HN
- 29Show HN: Free .dev subdomains for indie projects via GitOps Python workflowgithub.comLive Signal Community · 3h old · via Show HN
- 30Show HN: Syndicate – a desktop app for multi-agent orchestrationusesyndicate.orgLive Signal Community · 3h old · via Show HN
- 31Show HN: Echo – Fable-level results at 1/3 the cost using open-weight modelsnews.ycombinator.comLive Signal Community · 3h old · seen in 2 sources
- 32Show HN: Fovea – Record beautiful product demos on Windows (free, open source)github.comLive Signal Community · 3h old · via Show HN
- 33
- 34Show HN: Notebooker.ai – NotebookLM alternative, your own models, keys, storagenotebooker.aiLive Signal Community · 3h old · via Show HN
- 35
- 36
- 37Show HN: Jik – a language with region-based memory management that compiles to Cgithub.comLive Signal Community · 4h old · via Show HN
- 38
- 39
- 40Show HN: SafeMediaKit – blur/report UI for Apple's on-device nudity detectiongithub.comLive Signal Community · 4h old · via Show HN
How Radar works
The gradient. Stories travel through the AI ecosystem in a rough sequence: they appear first in papers (arXiv, Hugging Face) or repos (GitHub), then surface in community discussion (Hacker News, Lobsters, LessWrong), then get interpreted in analysis (operator newsletters and essays). Radar ranks by where a story sits on that gradient. The earlier it is caught, the more lead time you have before it reaches consensus. The spectrum at the top of the page shows the last 24 hours of that flow — every story that entered the watched feeds, counted at the station where it first appeared.
The score. Each story gets a recency-decayed signal score. Every source tier carries a weight and a half-life, so live signal moves fast and fades fast while analysis sources move slowly and stay relevant longer. A story scores higher when two or more independent source families carry it — corroboration, not popularity, which is why five feeds of the same site count once. A small bonus applies when a story has visibly climbed the gradient. Stories past their tier's maximum age expire from the board.
Heating and lead time. A story is heating when it has surfaced in two or more independent source families. When the same story is observed first at an earlier stage and later at a higher one, Radar records the gap in hours — for example, seen on arXiv fourteen hours before it reached the Hacker News front page. That gap is the reader's lead time, and it is a measured span, never an estimate.
The honesty rule. Every value on this page is an observed fact from the feed log — which source carried a story, when, and how many. Radar makes no claim about what a story means; that is the newsletter's job. When a reading does not exist yet — no stories heating, nothing upstream — the page says so rather than inventing one.
- The spectrum band
- A filled flame tick is an early, lead-time-rich stage; a hollow tick is a later stage; a hollow ring on an empty span is a station with nothing on it yet.
- Signal class
- The tier chip: Live Signal moves fast and decays fast; analysis sources stay relevant longer. The ranker weights each accordingly.
- heating
- The same story in two or more independent source families — not two feeds of one site.
- seen in / via
- How many families carried it, or the single source that did.
Once a week, the operator read on what actually changed and the move it calls for. No digest, no hype.
Read the newsletter →