What the wire reads, and what it does with it. Every figure is generated by
the pipeline on each publish; nothing is hand-entered, and the pipeline that
generates it is open source, so the method is
checkable too, not just the result. The full roll is on the
sources page.
The wirerunning totals
3,211verified sources
261,519items ingested36,858 this week
225,481stories clustered28,245 this week
1,855curated picks4 this week
61digests publishedlatest 2026-08-10
6.0avg relevance, 7d
Ninety days on the wireitems per day
busiest Jul 18 51,838quietest Jun 16 0mean 3,736 per dayitem tracking since Jun 9orange strip: picks curated that day
The funnelall time
items ingested261,519
stories clustered225,481 (86.2%)
articles read36,100 (16%)
picks curated1,855 (5.1%)
stories published793 (42.7%)
Read everything, keep almost nothing: of 261,519 items
ingested, 793 stories were published,
0.3% of the
firehose. Track widths are log scale; on a linear track every stage
after the first would be invisible. Percentages are conversion from
the previous stage.
Who feeds the wire3,211 sources
Expert blogs1,076
Newsletters350
Dev tools347
Research285
Tech news261
Security260
Science195
Hardware & science174
Physics72
News71
AI companies60
Aggregators60
tier 1 flagship 169
tier 2 core 1,305
tier 3 niche 1,737
paywalled 3% of the roll
Loudest feeds, 30 daysby items ingested
Google News Science4,523
World news | The Guardian4,269
dev.to top articles3,274
Hacker News3,263
Google News Business2,992
The Guardian2,605
Google News Technology2,580
Google News query: "artificial intelligence"2,376
arXiv quant-ph2,004
arXiv cs.LG1,900
arXiv cond-mat1,899
arXiv astro-ph1,887
arXiv cs.CV1,707
Stack Overflow hot questions1,697
arXiv cs.CL1,673
Where stories echoby clusters, 7 days
Google News Science1,245
World news | The Guardian1,018
Google News Business950
dev.to top articles851
Google News Technology755
The Guardian714
Google News query: "artificial intelligence"605
Hacker News572
Stack Overflow hot questions457
Al Jazeera English452
surfaces once 203,410
twice 18,290
3 to 5 times 3,706
6+ times 75
A surface is a place a story shows up: Hacker News, an arXiv section,
a subreddit. Most stories surface once and stay there. The ones that
echo across six or more surfaces are the news, and the pipeline
clusters every echo into one pick so nothing is counted twice. The
echo distribution counts every story ever clustered; the bars above
cover the last seven days.
The modelsconfigured, then observed
claude-sonnet-5Triage
claude-sonnet-5Deep read
claude-opus-4-8Digest
claude-sonnet-5 on subscription 1,403 calls, avg relevance 3.5
claude-opus-4-8 on subscription 587 calls, avg relevance 6.0
claude-opus-4-8 9 digests
The cards above are the configured routing; the strips are what
actually ran in the last 30 days, straight from the curation log.
The pipeline runs on a MacBook; judgment is rented by the token.
The barrelevance 0 to 10, 30 days
012345678910
Dark columns were kept, pale columns were judged and skipped. You are
looking at the cut line: where the pale mass ends and the dark mass
begins is the bar a story has to clear.
Method, briefly: the pipeline reads every source on a cadence, clusters
the same story across surfaces, fetches and reads the strongest link,
scores relevance on a 0 to 10 rubric, and keeps almost nothing. Every
number on this page regenerates on each publish. The full feed roll is
on the sources page.