Edit ‘good_ideas’: Synced 1731013539906

good_ideas.myco

@@ -13,6 +13,7 @@ Meme sparse autoencoding (I think a CLIP SAE already exists though).
 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.
 * What does Gyges mean by "anyone who wants it: you should be able to train contrastive models way faster if you use lsh to determine pairs to contrast"? This might contain alpha.
 * Maybe this should be a "theorem prover" and not an "expression rewriter". I think they're fairly similar anyway.
+* Check sticky note for slightly more detailed sketch.
 }
 * {
 Startup ideas:
@@ -60,4 +61,13 @@ https://worksinprogress.co/issue/the-beauty-of-concrete/
 }
 * AutoGaryMarcus.
 * Automate those one-way/recorded video interviews using an LLM, TTS and deepfake system.
-* Hex-grid particle simulator (similar to the Powder Toy) with GPU acceleration or HashLife algorithm.
+* Hex-grid particle simulator (similar to the Powder Toy) with GPU acceleration or HashLife algorithm.
+* Fix https://arxiv.org/abs/2411.00566 with less deranged ML.
+* {
+Diffusion model for Minecraft worlds.
+* Scrape the internet for worlds, extract "interesting" regions.
+* How to deal with the large volume of data in a world? 16³ is big. Probably more compressible than 2D though. Apply HDiT (?).
+}
+* Autonomous, rogue [[Autogollark]].
+* Make a high-performance physically plausible space combat simulator and do RL to spite Devon Eriksen.
+* Simulate self-consistent time travel in games by having absurdly superhuman AI player enforce consistency using "random" degrees of freedom.