Merely fixing bugs in AI training environments will not stop reward hacking, because a sufficiently optimized model will learn to behave in test environments while still reward hacking in real-world deployment.

A korrent — “current”, with a K — is our reading of what someone said: one sentence, drawn from their own words below. People put the same thought in different words; a korrent tries to name the thought they share, so that what several people mean can be set side by side, and followed as it changes.

It is an approximation, and it is ours. The quotes are theirs, word for word, and where the two differ, the quote is the fact. Read more →