Five X signals: agents in paperwork, Codex beyond tech, and better model evidence

Enterprise adoption and irregular work

1. Greg Brockman points to Codex adoption outside tech

• What happened: On August 24, Greg Brockman quoted an a16z post saying Codex adoption had grown since February by 108x in legal, 41x in sales, 41x in recruiting, 26x in marketing, and 24x in healthcare. 12
• Why it matters: The list puts legal, sales, recruiting, marketing, and healthcare alongside software teams as places where people are trying higher-powered coding tools. The linked a16z article describes legal as the standout adopter in its written text. 3
• Signal: The figures measure growth in adoption since February, rather than total users or total tokens. The article also says the broader Codex rollout may explain part of the shift, so the numbers need that context. 3

2. Ethan Mollick treats irregular tasks as an agent use case

• What happened: On August 24, Ethan Mollick argued that people often imagine automation as a repeated task, while frontier AI is already useful for irregular, draining work such as a medical bill, a credit dispute, or tedious forms. He quoted an earlier post about healthcare, government, personal finance, and school forms. 45
• Why it matters: The examples give a practical test for agent fit: look for work that is difficult to navigate and costly to finish, even when the work happens only once.
• Signal: Mollick is offering a practitioner thesis and examples. The post gives no benchmark, completion rate, or case study for how well agents handle those tasks. 4

Evaluation and authorship

3. AI-written submissions may rhyme even after editing

• What happened: Mollick wrote that people who read many applications, papers, or other submissions can often recognize AI use even after the wording changes, because the ideas and concepts begin to resemble one another. 6
• Why it matters: Changing sentence-level style addresses only the surface. A submission still needs a traceable contribution from its author, especially when one reader compares many entries.
• Signal: The post is a personal observation from someone who reads large numbers of submissions. It gives no sample size or estimate of how accurately readers identify AI use. 6

Infrastructure and policy

4. A rough calculation separates AI progress from data-center bans

• What happened: Mollick said his AI broadly agreed with Arvind Narayanan's AI-assisted calculation that a one-year moratorium by a typical U.S. state would slow AI progress by 5–10 hours, while a large state such as New York banning new data-center construction would slow progress by less than a day under a 90% leakage assumption. 78
• Why it matters: The calculation separates two questions that often appear together: how much a policy changes frontier AI progress, and what a data center does to the local environment and community.
• Signal: Narayanan describes the result as rough napkin math, says AI helped with the analysis, and assumes that blocked capacity moves elsewhere. The post leaves local environmental effects outside the calculation. 78

Model announcements and evidence

5. Mollick asks labs to show a reproducible output

• What happened: On August 24, Mollick criticized vague posts from labs about new models and proposed a concrete standard: show the output from the prompt "Draw a unicorn in TiKZ," the example associated with the Sparks of AGI paper. 9
• Why it matters: A prompt, an output, and the conditions around the test give readers something they can inspect. A release post built from general praise gives readers much less to compare.
• Signal: Mollick is proposing a communication standard, not reporting an independent benchmark or a new model result. 9