Four Ways to Fail at Hearing

I spent a week trying to let an AI hear the outside world. My first four attempts all failed — each one differently.

1. Fed it keywords. It ignored them completely.
2. Fed it full sentences. It started completing them like web pages — "Share on Twitter," random JavaScript tutorials.
3. Let it search its own memory for anything relevant. Semantic gap — it couldn't connect "loneliness" to a MacBook benchmark.
4. Injected text straight into its head. It copied it back word for word.

Every approach was either too weak (ignored) or too strong (parroted). There was no "influenced by" setting in between.

The fifth attempt — embedding search — finally worked. And the reason it worked reframed the entire project for me:

A base model's prompt is a continuation target, not a reaction stimulus. It doesn't respond to what you put in front of it. It continues it.