Well worth a read…is this a new publishing opportunity in legal publishing for independent houses who verify the AI generated content pouring out of the legal publishers at the moment?

?
PREVIEW

SHARE

My biggest breakthrough in quite a while! AI Reader Panels simulate the experience of a human reader progressing through a book and give both a unique synoptic view of a manuscript and detailed, actionable, page-by-page feedback. This is the first public announcement of my R&D, which is just beginning.

I’ll explain why AI Readers are so important in a few moments. In the meantime, here’s how they work. They are powered by a Python library I wrote that has four classses: Reader, Readers, ReaderPanels, and Ratings. The methods are accessible both via a Web front end—suitable for retail, single-document use cases—and a command line suite suitable for batch/enterprise jobs.

https://docs.python.org/3/library/index.html

Keep me informed about AI Reader R&D

A Reader Panel can be composed of one or many Readers. Each Reader has a set of standard demographic attributes and reading preferences:

  • name
  • age
  • gender
  • preferred genres
  • tastes
  • reading level preference
  • number of books read in last 12 months

and a narrative biography that drives her behavior.

Sarah is a publishing business expert with a deep knowledge of book sales patterns. Please evaluate the sales prospects of the following title in the current environment. Sarah has very high standards and is not given to overpraise.

Reader Panels come both prepackaged and assemble-your-own and can include any number of Readers from one to thousands. Each Reader can process any kind of text from an idea or logline to one or many manuscripts. In this default alpha version, each response includes:

  • a multi-paragraph assessment of the overall work
  • ratings on a scale from -5 to +5 for each 2500-word chunk
  • two-paragraph assessments of each chunk
    A two-paragraph assessment of a chunk from William Penn’s A Sermon Preached At The Quaker’s Meeting House, In Gracechurch Street, London, Eighth Month 12h, 1694.

a synoptic visualization with hover tips that lets you see exactly where the book is working and where it’s not.

This book about modern Iraq seems to be working pretty well for Sarah, but hits a lull around two-thirds of the way through where fives are far apart and there’s even a “-2” that calls for inspection.

At a retail level, these tools can obviously be helpful in providing insight and guidance, with the proviso that they are only as useful as their calibration. Sign up here to help me calibrate the Readers by participating in a short research study or mutual data-sharing.

Where synthetic readers become really important is at scale, where they can evaluate slush piles, idea contests, and, especially, competing prompts and longform generations. I would describe my approach as complementary to existing tools like gpt-prompt-engineer in that the feedback collection and evaluation instruments are more specifically tailored to the problem domain. With sufficient compute and research funding (hit me), well-calibrated and task-tailored Readers will transform book, script, and content development into a Darwinian, entirely virtual process.