Newsroom speed · updated hourly
How fast we are
When an AI lab, vendor, or regulator publishes something, how long until our verified story is live? This page publishes the answer every week, including the weeks we were slow.
Snapshot assembled . Window: last 8 ISO weeks (Monday 00:00 UTC), 1035 published stories.
- Median, this week
- 1.9 h197 stories · source publish → story live
- 95th percentile, this week
- 23.6 hprior 3-week median avg: 1.7 h
- Stories per day
- 50.4353 in the last 7 days
- Verified primary source
- 84%real source image: 99%
This week against the last 7 weeks
Bars show the weekly median (half of that week’s stories were live faster than this). The line shows the 95th percentile (all but the slowest one in twenty). Both are minutes from the moment a source published to the moment our story went live. The last bar is the current, still-incomplete week.
| Week starting (UTC) | Stories | Median | 95th percentile |
|---|---|---|---|
| 0 | — | — | |
| 0 | — | — | |
| 0 | — | — | |
| 0 | — | — | |
| 16 | 85 min | 10.9 h | |
| 140 | 2.5 h | 46.3 h | |
| 980 | 66 min | 39.3 h | |
| (this week) | 197 | 1.9 h | 23.6 h |
The last 24 hours
- All leads, median
- 6.6 h46 stories
- All leads, p95
- 23.7 hslowest one in twenty
- First-party leads, median
- 8.1 h7 stories from labs, vendors, regulators
- First-party leads, p95
- 10.3 hthe number we hold ourselves to
“First-party” means the lead came straight from a lab, vendor, or regulator we read directly (their newsroom, blog, release notes, status page, model registry, or official X account), not from an aggregator or wire search. The same figures appear in our public automation status under latency.firstParty and latency.weekly.
How it is measured, and why it matters
The clock starts when our source poller first records a lead. Lab and regulator sources are polled every minute, research registries every five, slower blogs every fifteen, so the recorded time is usually within a minute or two of the source actually publishing. We do not backdate to the source’s own timestamp, because sources sometimes publish quietly and date items in the past; the honest number is when the public could have seen it and we noticed.
The clock stops when the story is live on this site: the moment a new article is published, or the moment a canonical update is committed for a development we had already covered. Distribution to X, LinkedIn, newsletters, or search engines is not included; those are separate, slower channels.
Percentiles, not averages. The median (p50) says what a typical story experienced. The 95th percentile (p95) exposes the slow tail: an editor queue backed up, a source blocking our fetch, or an image that took time to verify. Averages hide that tail; we publish it.
Verified primary source is the share of stories whose recorded source is a first-party document (the lab’s own page, filing, paper, or repository) rather than a reported-secondary outlet we accept only when the primary cannot be reached. Real source image is the share carrying an image lifted from that verified page. We do not generate images; a story without a usable source image runs without one.
Why it matters. Most people first hear about an AI development from a summary of a summary. Every hour between an announcement and a verified, plainly written explanation is an hour in which speculation fills the gap. Publishing our own latency, including bad weeks, is how we keep ourselves honest about the trade-off between speed and verification, and it lets anyone check whether “fast” came at the cost of “verified”.
Limits. Weeks with few stories produce noisy percentiles; the current week is incomplete until Sunday 23:59 UTC. Stories originating from social posts are attributed to the linked first-party page, not the post. Numbers are recomputed at most hourly from the live archive; the JSON above is the exact data behind this page. Report a discrepancy.