diff --git a/TODO b/TODO
index a7279bba..95230161 100644
--- a/TODO
+++ b/TODO
@@ -1,5 +1,3 @@
-- try a 32-bit intermediate for im_conv()
-
 - gtkdoc for im_profile() next
 
 
@@ -10,6 +8,7 @@
   paths to search, breaking demo2 for Hamish
 
 
+
 - lab [100,0,0] -> srgb [255, 255, 254]? how odd
 
 - could do small masks in a single pass
diff --git a/libvips/convolution/im_conv.c b/libvips/convolution/im_conv.c
index f402c815..fbb45d54 100644
--- a/libvips/convolution/im_conv.c
+++ b/libvips/convolution/im_conv.c
@@ -103,13 +103,15 @@
 
  	TODO
 
-	- will this change make much difference to the vips benchmark?
+	- tried 8-bit data with a 32-bit intermediate, but it was only 
+	  slightly faster than C
 
-	- try a path with a 32-bit sum for larger matrices / scale / offset, 
+	  16-bit data would be even slower, no speed advantage
 
 	- make up a signed 8-bit code path?
 
-	- try a 16-bit path
+	- make it more like morphology.c: have a param for the result of the 
+	  previous pass rather than a separate combining pass
 
  */