LUNA_29

#0PG $OPG @OpenGradient

I asked the same question twice, four days apart, and got two genuinely different answers from what I thought was the same model.

At first I assumed I’d phrased it differently the second time.

Then I checked.

Wait, no. The wording was identical.

So I pulled the inference records behind both responses on OpenGradient’s Hub.

The first answer traced back to a model version that had already been rolled back.
The second came from the replacement that took its place days later.

That’s when the rollback mechanism really clicked for me.

A rollback changes what the model does next.
It doesn’t rewrite which exact weights produced an answer that already happened.

That earlier inference stays permanently tied to its own Blob ID, independently provable even after the live model changes.

In other words:

the system can move forward without erasing its own history.

Most people think a rollback “undoes” the model.

It doesn’t.

It creates a second timeline while the first remains cryptographically intact underneath it.

That distinction probably feels small right now.

I’m less convinced it stays small once people begin querying the same systems repeatedly across months, while the underlying models quietly change beneath them.

Because eventually the real question won’t be whether an AI can be rolled back.

It’ll be whether anyone notices they received two different truths from two different versions of reality.

$ACT


Who should be responsible when a rollback exposes a bad AI answer?