Viber
Viber by OneThought AI. Describe an app in plain language and get working software.
Research
Software from intent
Active
Turning a plain-language description into working software, and building the tests that show whether the result does what was asked.
The question
When someone describes an app in plain words and an AI builds it, does the app do what they meant? And how would they know, if they cannot read the code?
Why it matters
Tools that build software from a description let people who cannot program make what they need: a small business, a school, a non-profit, a local government office. The risk moves from "can I build this" to "can I trust this". A generated app can look finished and still lose data, expose it, or quietly do something other than what was asked.
Our approach
- Intent as a test. Each request is paired with acceptance tests written in the same plain language, so success can be checked without reading the code.
- Safety checks by default. Generated apps are checked for common failures in data handling and access control before they are handed over.
- Honest reporting. We report how often generated apps pass, fail or partially meet the request, by type of app.
Status now
The research informs Viber, which OneThought AI builds. The evaluation set is in preparation.
What we have found
Findings will be published with the evaluation set.
What comes next
| When | What |
|---|---|
| 2027 | Intent-to-app evaluation set: plain-language requests paired with acceptance tests |
| 2027 | Safety checks for generated apps: data handling and access control |