Three new cards, two operators. Willison on a structural risk AI training creates for researchers who publish incrementally. Indig and Johnson on the economics of marketing automation. Ronacher on what AI detection tools actually score.
AI training creates a new competitive risk for researchers who publish as they go
Simon Willison's September 8 essay on the Navier-Stokes equations opens a risk case that has no clean prior analog.
AI training pipelines can turn partial public research into a competitive priority leak names the mechanism. A researcher who blogs about partial progress toward an unsolved problem adds that signal to the public web. Future training runs ingest it. A competitor working with a model trained after that post may find the model primed to complete the researcher's line of reasoning faster than the researcher herself. No bad faith required. Willison observes that even knowing a solution exists is enough to focus substantial compute on a solvable target.
"Just knowing that there is an unpublished solution to a problem might trigger millions of dollars in LLM spending to get there first."
The mechanism is speculative but structurally sound. Training pipelines that continuously absorb public web content create a new asymmetry for researchers whose practice is to share work openly as it develops. The norms of open science (pre-registration, incremental disclosure, public preprints) were designed for a world where competing researchers worked at human speed. The question Willison raises is whether those norms need rethinking.
- Simon Willison, AI training pipelines can turn partial public research into a competitive priority leak. Partial public research disclosures may enter future training corpora, enabling a model user to reach the same solution before the original researcher completes it.
AI automation moves marketing work upstream, it does not delete it
Kevin Indig and Amanda Johnson's September 14 post in Growth Memo is one of the cleaner explanations of why most organizations get no return on AI investments.
AI automation redistributes where marketing effort goes rather than reducing total effort names the mechanism. MIT research found 95% of organizations get zero return on GenAI investments. The cause is consistent: teams automate the execution step they already perform, not the step that most needs changing. The work doesn't disappear; it shifts. The highest-value 30 minutes in a marketing team's week is now the pre-automation design phase, where a marketer decides what to automate and whether the current step is worth keeping at all.
"The most valuable 30 minutes in marketing right now are the 30 minutes you spend thinking _before_ you start building any AI automation."
The practical heuristic Indig and Johnson offer is: swap one step, not one job. Function-level replacement with no role elimination. Automation scoped to a single defined step produces measurable efficiency. Automation scoped to a full role produces unchanged or worse output, because the judgment that made the role valuable didn't go anywhere.
"AI does not delete marketing work, it moves it"
This connects directly to AI tool maintenance in marketing teams creates invisible overhead that the official hours count does not capture, making net productivity gains structurally smaller than they appear from September 7, where Indig observed that AI labor hours go uncounted in marketing teams. The same invisible effort now concentrates in the pre-automation design phase.
- Kevin Indig, AI automation redistributes where marketing effort goes rather than reducing total effort. AI automation redistributes where marketing effort goes rather than reducing total effort; the 30-minute pre-automation thinking phase is now where competitive advantage concentrates.
AI detectors score the LLM's structural fingerprint, not the human text written over it
Armin Ronacher's September 14 post on Pangram is a precise account of a measurement artifact that affects a large number of writers who use AI as a writing tool.
AI detectors score LLM structural artifacts that persist through human editing, not the human-authored words themselves names the mechanism. AI detectors like Pangram score statistical properties of text: word entropy, sentence complexity variance, hedge density, clause parallelism. LLMs produce characteristic patterns in these dimensions. When a writer uses an LLM to generate structural scaffolding (headings, bullet outlines, paragraph templates) and then writes their own content into those slots, the structural pattern persists in the final text. The detector fires on that inherited scaffold, not on the human-authored words.
"if you rely on an LLM to give your text structure, it will score badly on Pangram even if you do plenty of edits over it."
The practical implication is that the score measures LLM involvement at any stage of production, not whether the words themselves were written by a human. A writer who uses an LLM for outline generation and then edits extensively will appear as high AI. A writer who uses an LLM for synonym suggestions with no structural output will appear as lower AI. Rewriting the structure reduces the score more than rewriting individual sentences.
- Armin Ronacher, AI detectors score LLM structural artifacts that persist through human editing, not the human-authored words themselves. AI detection tools score the structural artifacts LLMs leave in scaffolding; a human who adopts that structure and edits the content will still receive a high AI score because the structural signal persists through editing.