Co-founder · summer 2026 – now
The short version of a book
- To create a new reading medium
- Equal to the full book in its quality
- Status
- MVP, 47 readers
- Live
- diple.ai/mvp
A diple (>) is the mark ancient editors drew in the margin of Homer next to lines worth noticing. On this page it marks the lines I’d want you to notice.
Why
Summaries lose what you remember from a book.
There are books you don’t want to read in full. Part of the book doesn’t interest you, but you don’t want to lose what it’s saying either. Summary apps cut a book down to a page of takeaways and drop exactly what you remember from reading: the scenes, the people, the numbers.
I wanted the same book, only shorter: the same chapters in the same voice, at a fraction of the length.
The product
From search to the last chapter.
Diple is live as an MVP. Every screen below is a screenshot of the live product.
Find a book
Search covers almost any book, with covers, Goodreads ratings, author and publisher pages, and a link to buy the full book. You pick the language of the summary: English or Russian with the full interface, or Spanish, German, French, Italian, Portuguese, Chinese and Japanese for the summary itself.
Get the text
Diple looks for the full text in open libraries: Project Gutenberg, Wikisource, or the author’s own free edition. If the book isn’t there, you upload a copy you own: PDF, EPUB, TXT, DOCX or FB2, up to 60 MB.
You can also add one short instruction, like “focus on technical details” or “prioritize stories and examples”.
Diple builds it
It reads the whole book, maps the chapters and writes an author dossier first. The first chapter is ready in about twenty minutes. The rest are written while you read, and each one opens on its own when it’s done.
If one chapter fails, you rebuild just that chapter. If the whole build fails, the credit comes back. Every build has a spending cap, so a bug can’t quietly drain my card.
Read it
The reader is the part people actually use, so it got the most care: three themes, three text sizes, a contents page that shows which chapters are ready, and on every chapter the reading time against the original.
The book ends with five books to read next. Some go deeper into the subject, some come at it from another angle, and some move into a neighbouring field.
Language
Getting Diple to rewrite a book the way a person would is one of the hardest parts of the project. Any language model can retell a book, and it will do it in the manner people now recognize as AI. Few readers put up with that for the length of a book they care about.
So my team and I took Russian and English books, wrote our own short versions by hand and compared them with what the model produced. We studied what lets compressed text read easily in each language, and we catalogued the habits of the model itself. Here are a couple of examples:
English
“Not X but Y” only when someone actually claims X.
A model reaches for the contrast to sound significant: “Venture capital is not just about money. It’s about vision.” Nobody said it was only about money, so Diple states the second half directly.
No evaluation the author didn’t make.
Before a word like “elegant” or “strikingly simple” stays, it has to answer two questions: whose word is it, and what does it hold? If removing it loses only colour, it goes.
Russian
A word chosen through its English meaning is a calque.
The model drafts the thought in English and then translates it, so it writes «адресовать проблему» (to address a problem). The sentence is grammatical, and no native speaker would say it; Diple writes «заняться проблемой».
Prose without idioms reads as machine-made.
Russian leans on set phrases, and models almost never use them. Where the author would say «палка о двух концах» (a stick with two ends), Diple keeps it instead of the word-for-word «обоюдоострый меч».
Diple retells books in nine languages. For the seven besides Russian and English, we have started hiring people from China, Portugal, France and other countries to do the same work, so that a retelling in their language reads as if the best writer in that country had shortened the book personally.
Length
The second problem that took real work was length. Diple would not cut a book to a third of its size. It shortened books one and a half times, twice at most, so we came up with this:
- 1Code
Count the book
Diple splits the book into chapters and counts the words in each one.
- 2Model
Read it first
The model reads the book and notes how the author writes and what kind of text each chapter is: a story, an argument, a manual.
- 3Code
Set the length
Code decides how long each chapter will be and splits it into a few parts, each with its own word limit. A story gets about 175 words per part, dense analysis about 85. The model never chooses its own length.
- 4Model
Choose what stays
The model goes through the chapter paragraph by paragraph and marks each one: keep, shorten or cut. It starts from the chapter’s five main ideas, so those are never cut.
- 5Model
Write and count down
The model writes each part with a number in front of every word, counting down from the limit: 175, 174, 173… At zero the part ends. A number it has to write at every step is much harder to ignore than a target it read once.
- 6Code
Check the count
Code removes the numbers and checks that none were skipped or repeated. Then it measures the result: total length, paragraph length, and how many names and numbers from the original made it in.
- 7Model
Edit
A last model pass fixes what the checks flagged. If a chapter still runs long, it may cut a whole part, starting with the least important one.
Results
Where Diple stands in October 2026.
Steve Jurvetson wrote back. The co-founder of Future Ventures answered our email about Diple on 18 September with “what a cool idea”, sent a book he wanted compressed twenty times, and asked whether we could dial the ratio up from three.
Next