We’re building an AI editor called Maxwell Litura, and the one thing it will never do is write your book.

Ask it to rewrite a sentence, and it tells you what the sentence needs to accomplish instead. Not one word of your manuscript will be AI-generated. That constraint is the product, and enforcing it turns out to be a far harder engineering feat than most of the other technical challenges we’ve had.

The name Maxwell loosely comes from a prominent fiction editor who famously kept his own voice out of his famous authors’ books. He was a fine novelist himself. He simply didn’t want anyone to see him in other people’s work. That is our standard.

This newsletter discusses, in very candid terms, what I’m learning as we build an application that honors this standard. The newsletter goes out most Wednesdays. Every post is about one real thing that went wrong or one hard decision we struggled with, told with the actual dates and numbers rather than a case study written after the fact. Some weeks this means admitting we lost someone’s work. Some weeks it means explaining why an AI model that agrees with you is worse than one that’s wrong.

It’s written for two kinds of people who turn out to want the same thing: writers and editors deciding whether any AI-centered technology can be trusted with their manuscript, and engineers building with language models and finding that the hard part isn’t AI or their coding abilities.

If you write, or you build, or you’re just plain suspicious of AI—you’re in the right place. And by the way, I put that em dash there.

Jeb