AI pseudo randomisation

[Baptouuuu]

Problem

Property Based Testing use random data to explore the problem space in order to find edge cases that don't fulfil the properties. When the problem space is very technical this approach is great as there's no bias as to where to bugs might lie.

However when trying to prove the behaviour of an app with a very large problem space, using only randomisation it will test cases far from what the user might do. While this is still important to explore the problem space this way to find nasty bugs, it will take longer to find bugs and it's the users that will surface them first.

There should be a pseudo randomisation to explore the most probable scenarii in the problem space.

Solution

A large problem space combined with constraints (ie properties) to navigate through it is an excellent fit for neural networks.

This should be used to generate sequences of actions a user might do.

This shouldn't be used to generate data (strings, ints, etc...) as it couldn't be shrunk, and not diverse enough. And probable text (and such) can already be defined with the existing Set API.

On top of that, a CRM user might not behave the same way as someone browsing a social network. By using neural networks it could be possible to use domain specific models to have more accurate predictions.

Notes

With this approach randomisation and pseudo randomisation are complementary. Each address its own problems.

By building higher and higher abstractions, the apps built on top will address greater and greater problem spaces. And the need for this will become essential.

As of now (June 2026) I don't have the technical know-how to implement this. But if you do, please respond to this discussion below.

[Baptouuuu]

macOS 27 will bring access to a foundation model via a command line called fm that allows to specify an output format.

https://developer.apple.com/videos/play/wwdc2026/334/