diff --git a/autogollark.myco b/autogollark.myco
index 825dea6..0c1a4d7 100644
--- a/autogollark.myco
+++ b/autogollark.myco
@@ -48,7 +48,9 @@ Autogollark currently comprises the dataset, the search API server and the [[htt
 * Context length issues, and subquadratic models are sort of bad, though maybe we can "upcycle" a midsized model to RWKV. This exists somewhere. Not sure of efficiency. Inference code will be awful.
 }
 * Train on e.g. Discord Unveiled (local copy available).
-* Is vision practical? Probably not without a pretrained model for it, and those don't seem to be base models. Perhaps CLIP embeddings are good enough but I wouldn't count on it.
+* {Is vision practical? Probably not without a pretrained model for it, and those don't seem to be base models. Perhaps CLIP embeddings are good enough but I wouldn't count on it.
+* Vision "output" via retrieval over Nooscope/MSE.
+}
 
 == Versions