Four new cards from three operators. Two themes about how to produce writing worth reading when AI is everywhere. One structural finding about where AI search citations actually come from. One architecture analysis that changes the economics of personalized AI.
Theme 1, AI cannot ghostwrite, but writers can publish before they are ready
Tomasz Tunguz and Simon Willison reach the same place from different angles this week. Tunguz is writing about AI and authenticity. Willison is writing about technical blogging. The underlying claim is the same: what a reader receives is the only thing that matters.
Tunguz argues that AI fails as a ghostwriter because every writer using it produces the same output. A ghostwriter harmonizes their voice with the author's. An editor suppresses their own voice to serve the author's. AI, trained identically for every user, cannot harmonize with any individual. It can edit: it improves structure and fact-checks without erasing the author's idiosyncrasies. Readers have noticed. In a market flooded with AI-generated content, they test for authenticity markers. An ampersand, a neologism, a grammatical fragment. Tunguz deliberately preserves these signals in his own writing: "In the age of slop, readers test authenticity. I lace my posts with subliminal sincerity." The practical instruction is to use AI to edit drafts, not to rewrite them. AI homogenizes voice because all writers share the same ghostwriter, so use AI as an editor and preserve your own stylistic markers.
Willison's advice arrives from the opposite problem. Writers hold back publishable work because an imagined better version is in their heads. Willison's instruction: "My number one tip for blogging is to lower your standards! Aim to hit publish while you are still actively unhappy with what you have written." The mechanism is simple. Readers never see the imagined version. They evaluate only what is published. Every unpublished draft has zero reach. Publishing something imperfect immediately begins producing reach, feedback, and accumulated credibility. The flaws the author perceives are invisible to a reader who lacks the author's internal standard. The most effective blogging habit is publishing while still dissatisfied, because readers cannot see the imagined perfect version.
- Tomasz Tunguz, AI homogenizes voice because all writers share the same ghostwriter, so use AI as an editor and preserve your own stylistic markers. AI homogenizes voice because all writers share the same model; use it as an editor, preserve your own stylistic markers.
- Simon Willison, The most effective blogging habit is publishing while still dissatisfied, because readers cannot see the imagined perfect version. Publish while still unhappy with the draft; readers cannot see the imagined perfect version and only what is published generates any value.
Theme 2, External domains supply over 80% of AI citations, leaving the owned site in a corroboration role
Aleyda Solis analyzed citation sources across SaaS, ecommerce, and finance and found a consistent structural pattern: external domains account for 69 to 82 percent of AI search citations, while brand websites occupy only 10 to 11 percent of the top citation slots. The brand site establishes canonical facts. Third-party sources cite.
The vertical differences reveal the mechanism. In SaaS, social and community platforms (YouTube, Reddit) dominate at 45.7% because buyers trust peer experience over vendor claims. In ecommerce, competitor and marketplace domains dominate at 36.6% because buyers compare options on neutral ground. In finance, news and review sources lead at 19.4% because credentialed authority reduces perceived risk. AI engines weight independent validation over brand assertion. The practical implication is that optimizing only owned content reaches a small fraction of the citation surface. The majority of the surface lives on Reddit, YouTube, review platforms, and news sites where the brand cannot publish directly. Solis frames the model: "AI search is largely a 3rd-party citation problem, supported by an on-page corroboration base." This connects with Kevin Indig's UGC citation floor data from two days prior (UGC platforms are the most consistent AI citation source, holding a floor share across every buying stage) and with the cited-page versus traffic-page split (The pages AI engines cite and the pages that receive AI referral traffic are structurally different assets requiring separate optimization). The picture now has three layers: owned pages corroborate, external pages cite, and the pages that receive AI referral traffic are different from both. AI search is a 3rd-party citation problem: external domains supply over 80 percent of citations while owned pages corroborate.
- Aleyda Solis, AI search is a 3rd-party citation problem: external domains supply over 80 percent of citations while owned pages corroborate. External domains supply 69 to 82% of AI citations across three verticals; the brand site corroborates accurate facts while third-party sources do the citing.
Theme 3, Test-time training shifts AI provider constraints from memory to compute, making per-user models viable for specific use cases
Tomasz Tunguz analyzes what test-time training changes for AI providers and customers. Standard transformers carry conversation history in a growing attention cache. Memory and latency grow linearly as context lengthens. Test-time training instead takes gradient steps on each prompt, folding history into a fixed-size weight update. Memory stays flat. Stanford research shows 2.7x faster inference on small models with constant latency independent of context length.
The cost moves elsewhere. One frozen checkpoint serving millions becomes millions of slightly different models, each shaped by one user. The provider's constraint shifts from memory to per-user GPU capacity. Tunguz frames the decision rule: "Standard AI is limited by memory, test-time AI is limited by compute and chips." A per-user model earns its cost when the personalization it produces generates returns unavailable from a cheaper shared model. A coding agent that accumulates knowledge of a specific codebase and its conventions justifies the per-user overhead through developer retention and reduced context-loading time. A one-off customer support query does not. This connects with the pricing data from Enterprise AI workloads run overwhelmingly on non-frontier models because buyers optimize for price-over-performance, not benchmark scores, which showed that 84% of tokens on open routing platforms go to non-frontier shared models. Both findings point to the same underlying pattern: AI infrastructure pricing bifurcates around the specific value personalization creates. Test-time training makes per-user AI model variants viable only when personalization returns exceed per-user compute cost.
- Tomasz Tunguz, Test-time training makes per-user AI model variants viable only when personalization returns exceed per-user compute cost. Test-time training makes per-user model variants economically viable for high-value personalization use cases, but frozen shared models remain cheaper for one-off queries.