Inchworm Translation Improvement Loop
Pay for sentence-level improvements only after readers prefer them
- Difficulty
- Advanced
- Time to result
- ~months to results
- Steps
- 7
- Confidence
- 96%
Sivers proposes a sentence-level marketplace for translations that improve incrementally. A finished source text is split into database entries, and a machine supplies a deliberately rough starting translation. Translators improve individual sentences rather than owning one entire edition. As accepted versions become stronger, the proposed payment rises, rewarding the harder marginal improvement. A separate group of readers sees the current and proposed sentence in random order and votes for the better version. The translator is paid only when the majority prefers the revision, preventing payment for unvalidated output. The loop repeats until the work is good enough to release. Sivers presents this as an active experiment rather than a solved system and acknowledges unresolved incentive questions, including why a skilled translator should contribute early when later improvements may pay more.
Origin
Extracted from The Tim Ferriss Show, where Derek Sivers described the experimental translation marketplace at inchword.com.
Core principles
- 01A translation can improve through multiple independent passes
- 02Sentence-level work lowers the unit of contribution
- 03Higher rewards can make later marginal improvements worthwhile
- 04Reader preference should gate payment rather than mere submission
How to run it
- 1
Atomize the source
Put each sentence of the completed work into its own database entry.
- 2
Create a baseline
Use a computer-generated translation to establish a rough version that is easy to improve.
Watch out Do not mistake the baseline for publishable quality.
- 3
Solicit one improvement
Let a translator propose a better version of an individual sentence.
Pro tip Keep the contribution small enough for specialists to target clear opportunities.
- 4
Blind the comparison
Show readers the incumbent and proposed versions in random order so position does not signal which is new.
- 5
Gate acceptance
Replace the incumbent and pay the translator only when a majority of readers judges the proposal better.
Watch out Reader quality and sufficient voting remain essential dependencies.
- 6
Raise the marginal reward
Offer more for later improvements as surpassing the accepted sentence becomes harder.
Watch out Higher later rewards may discourage strong translators from going first.
- 7
Release the translation
Stop iterating when the accepted text reaches the chosen quality threshold.
In the wild
A machine supplies the first weak version of a sentence. One translator improves it for a low reward; another later proposes a stronger rendering for more money. Readers compare each candidate with the incumbent in random order, and only a preferred revision is accepted and paid.
→ Translation quality can advance in validated increments instead of ending after one contractor's pass.
Common mistakes
Paying for submission alone
Without reader validation, contributors can earn money by producing changes that do not improve the translation.
Ignoring the first-mover problem
If later revisions pay much more, a highly capable translator may have little reason to create the early accepted version.
Is it for you?
Best for
Authors who care deeply about translation quality and can fund repeated sentence-level review.
Not ideal for
Urgent, low-budget translations or languages without enough qualified translators and readers to judge revisions.
From the transcript
“once I call something done whether it's an article or a book I put every sentence into its own entry in the database”
“both sentences are shown in random order and they have to vote for which one they feel is the better sentence”
“translator can't get money just for coming in and spewing crap they only get paid when the readers believe that that was a better translation”
From the episode
#777: Derek Sivers, Philosopher-Entrepreneur — The Greatest Year of His Life
Derek Sivers