Edit ‘good_ideas’: Synced 1725647705430

good_ideas.myco

@@ -3,6 +3,7 @@
 * Overengineered LLM-based autocomplete to spite "it's just autocomplete" people.
 * Expanded [[vengeance]] policy.
 * Combine new CRL (https://arxiv.org/abs/2408.05804) with offline pretraining.
+* Similarly, contrastive RL for computer algebra (specifically, proving that expressions equal other expressions via making substitutions repeatedly). Try and contrastively learn a "how close is this expression to this other one" function (I think with an action input?). Bootstrap to progressively harder problems.
 * {
 Startup ideas:
 * Automated reminders to make spontaneous gestures to maintain friendships.