What if large AI models contain “null functions” — parts of the network that repeatedly produce no meaningful contribution for certain inputs?

If we could identify them reliably, two things become interesting.

First, model size.

If a function is effectively redundant across all inputs, could we remove it entirely and reduce the number of weights that need to be stored?

Second, inference speed.

If a function is only useless for a particular token or input, could we detect that cheaply and skip it before loading its weights or executing its matrix operations?

Instead of: Load weights → Compute → Discover the result contributes almost nothing

we could move toward: Predict contribution → Skip unnecessary computation → Execute only what matters

That could reduce: • memory bandwidth • weight transfers • matrix multiplications • power consumption • inference latency

The real challenge is not finding a zero after computation.

The challenge is knowing that a function will contribute nothing — or almost nothing — before performing the expensive computation.

If that becomes possible, perhaps the next big optimization for LLM inference is not simply making matrix multiplication faster.

Maybe it is learning which matrix multiplications never needed to happen in the first place.

---