One-Third of Web Pages Are Now Written With AI, Says Pew Research

Slashdot, 2026-08-24. A single-number quantification of a trend that has been anecdotal for a year. If the number holds up under scrutiny, "detect AI text" loses meaning as a task — the base rate is now a third.

We never use AI. For anything

HN → corkmac, 2026-08-24. A macOS developer publishing an explicit "no AI" stance on the product page. Notable not for the position but for the marketing calculation: manual work has become a differentiator worth advertising above the fold.

Why is Anthropic's public writing style so unlike Claude's?

HN → cmart, 2026-08-24. Meta-question on model-vendor prose. The observation — vendors don't ship their own models' voice — is the mirror image of the corkmac stance and the Pew number: everyone is now legible on which side of the AI-authorship line they sit.

Executable Is a SQLite Database

HN → Farid Zakaria, 2026-08-24. A short systems piece pushing SQLite into the packaging layer. Reads as one more entry in the "SQLite for X" catalogue that has been quietly accumulating all year (bench, WAL-2, executables).

Kids outlearn AI — and we still don't know why

MIT Technology Review, 2026-08-24. Language acquisition still favors small humans over large models. Useful counter-data on days when scaling arguments dominate the feeds.

Anthropic's best AI model struggles to attract users as cheaper tools thrive

Simon Willison, 2026-08-23. Willison quoting the FT piece. Sits directly upstream of Monday's Anthropic API outage — cost pressure and reliability landing together is the harder story than either alone.

Why your biggest hidden AI cost isn't GPUs

InfoWorld, 2026-08-24. Companion piece: Can neoclouds corner AI compute?. Two InfoWorld posts framing the compute-cost story as an operations problem, not a chip problem. Watch the "neoclouds" framing — new bucket for CoreWeave / Lambda / Together tier.

AI authorship becoming legible

Slashdot: Pew — one-third of pages AI-written (Mon), corkmac: we never use AI. For anything (Mon), cmart: why Anthropic's writing ≠ Claude's (Mon), MIT TR: kids outlearn AI (Mon).

Anthropic economics + reliability

SW: Anthropic top model struggles vs cheaper tools (Sun), status.claude.com: elevated errors for multiple models (Mon), InfoWorld: hidden AI cost isn't GPUs (Mon), InfoWorld: neoclouds cornering AI compute? (Mon).

Harness > model (carryover)

Latent Space: Evolution of the Agent Harness (Sat), Latent Space: simulation is taking over (Sat), TC: Nvidia harness is the hero (Fri), TNS: Opus 5 30% → 100% wrapped in AVO (Fri).

Performance revival (carryover)

Ronacher: Fast and Hard Code (Sun), Malisper: JIT in 5μs (Sun), Dan Luu: no reason for software to be slow (Sat), counter: reasons for software to be slow (Sat).

AI governance at the distro / policy layer (unresolved)

Slashdot: Debian voting on AI contributions (Sun), TNS: Debian proposed banning AI code (Thu), InfoWorld: crates.io backdoor (Fri).

• Monday feeds
  • Pew — one-third of web pages AI-written, Slashdot, 08-24
  • We never use AI. For anything, HN → corkmac, 08-24
  • Why is Anthropic's public writing style so unlike Claude's?, HN → cmart, 08-24
  • Executable Is a SQLite Database, HN → Zakaria, 08-24
  • Kids outlearn AI — and we still don't know why, MIT TR, 08-24
  • Elevated Errors for Multiple Models, status.claude.com, 08-24
  • Why your biggest hidden AI cost isn't GPUs, InfoWorld, 08-24
  • Can neoclouds corner AI compute?, InfoWorld, 08-24
  • Andreessen Horowitz is investing billions into a bleak future, HN → Model Republic, 08-24
  • Fast drilldown dashboards from a single Parquet file, HN → Ulmer, 08-24
  • OCR It — pull text out of un-copyable documents for your LLM, HN, 08-24
  • SCO Successor Xinuos Asks Court to Rehear Its Claims Against IBM/Red Hat, Slashdot, 08-24
  • Mecha Comet Linux PC-Phone hybrid, Slashdot, 08-24
  • Building certgrep.sh — free CT search engine, Lobsters, 08-24
  • 3 constraints before I build anything, Lobsters → Jordan Lord, 08-24
  • Tiny, Untyped Monads (2024), Lobsters, 08-24
  • Migrating a Synology NAS to a UniFi UNAS Pro 8, HN → Hanselman, 08-24
• Sunday carryover — quiet-day threads
  • Debian is Voting on Whether to Allow AI-Assisted Contributions, Slashdot, 08-23
  • Fast and Hard Code, HN → Ronacher, 08-23
  • JIT Compiling Code in 5μs, HN → Malisper, 08-23
  • Wi-Fi 8 isn't chasing speed, HN, 08-23
  • Foundational Verification of Running-Time Bounds, Lobsters → Chlipala, 08-23
  • I Dream of Quieter Computing, HN → Henry Codes, 08-23
  • Anthropic's best AI model struggles to attract users as cheaper tools thrive, Simon Willison, 08-23
  • One pull to wipe them all, TNS, 08-23
  • Why real-time AI at scale is so hard, TNS, 08-23
  • claude-code v2.1.241, releases, 08-23
• Saturday carryover — harness, performance, governance
  • The Evolution of the Agent Harness, Latent Space, 08-22
  • 10% worse, 100x cheaper, 10000x faster: Simulation is taking over, Latent Space, 08-22
  • There's no reason for software to be slow anymore, HN → Dan Luu, 08-22
  • There continue to be reasons for software to be slow, Lobsters, 08-22
  • Mathematical Discourse, Terence Tao, 08-22
  • More than just code review, Simon Willison, 08-22
  • llm 0.33, Simon Willison, 08-22
  • Just a rumour of a bug is enough to find a security exploit, OCaml.org, 08-22

Authorship legibility is the load-bearing frame today

Three items — Pew's one-third quant, corkmac's stake, and cmart's meta on Anthropic prose — all land the same Monday. Read together they mark a phase change: the question stops being "can we detect AI writing" (Pew says: the base rate makes detection meaningless) and becomes "which side of the line are you signaling from." The corkmac position is only interesting because "no AI" is now disclosure-worthy on a product page.

The Anthropic bind

Willison's quote of the FT piece (cheaper tools thrive) plus Monday's outage cluster plus InfoWorld's cost reframing form a coherent picture: the frontier lab is priced for a market that isn't buying frontier, while carrying inference volatility that cheaper providers don't. Not a crisis, but the first Monday where the three sides of the bind — pricing, reliability, cost structure — all showed up in the same day's feeds.

The neoclouds tag

InfoWorld's second post uses "neoclouds" as if it's a settled category. It isn't yet — nobody agrees which providers count — but the coinage is worth tracking. When analysts need a new bucket for CoreWeave / Lambda / Together / Crusoe tier, someone will land on the naming. Whether this is the naming is the question.

Sources that publish frequently but have gone quiet:

• Interconnects (7 days) — last item 2026-08-17.
• anthropic-generated (10 days) — last item 2026-08-14.
• Kenneth Payne (10 days) — last item 2026-08-14.
• METR (10 days) — last item 2026-08-14.
• Vicki Boykis (12 days) — last item 2026-08-12.
• GitHub Engineering (14 days) — last item 2026-08-10.
• Tailscale (18 days) — last item 2026-08-06.
• AI Snake Oil (19 days) — last item 2026-08-05.
• Jane Street (19 days) — last item 2026-08-05.
• All Things Distributed (20 days) — last item 2026-08-04.
• deepmind-blog (23 days) — last item 2026-08-01.
• Grafana Labs (24 days) — last item 2026-07-31.
• Bunnie Studios (25 days) — last item 2026-07-30.
• Klara Systems (26 days) — last item 2026-07-29.
• Marc Brooker (26 days) — last item 2026-07-29.
• The Markup (27 days) — last item 2026-07-28.
• Fly.io (31 days) — last item 2026-07-24.

build: 2026-08-24 | crawler-sha: afd8192 (Walsh-Research/1.2, compliance v1.4) | feeds: 65 active | items-considered: 4825 (14d, incl. 2913 arxiv-cs-ai) | warehouse: 38132 items | published: 21