Show HN: Overcoming silicon-level aging through fault tolerance (PoC) (github.com)

If training fails due to hardware aging, what if we simply zero out the values coming from the affected device the moment the failure occurs?