diff --git a/compiler/builtins/bitcode/run-tests.sh b/compiler/builtins/bitcode/run-tests.sh
index 2bdbed2070..2892ba1f74 100755
--- a/compiler/builtins/bitcode/run-tests.sh
+++ b/compiler/builtins/bitcode/run-tests.sh
@@ -1,6 +1,9 @@
 #!/bin/bash
 
-set -eux
+set -euxo pipefail
 
 # Test every zig
-find src/*.zig -type f -exec zig test --library c {} \;
+find src/*.zig -type f -print0 | xargs -n 1 -0 zig test --library c
+
+# fmt every zig
+find src/*.zig -type f -print0 | xargs -n 1 -0 zig fmt --check
diff --git a/compiler/gen/src/llvm/build_list.rs b/compiler/gen/src/llvm/build_list.rs
index 73cdd97e84..9c6c20bbb7 100644
--- a/compiler/gen/src/llvm/build_list.rs
+++ b/compiler/gen/src/llvm/build_list.rs
@@ -2050,8 +2050,6 @@ pub fn allocate_list<'a, 'ctx, 'env>(
     let len_type = env.ptr_int();
     let elem_bytes = elem_layout.stack_size(env.ptr_bytes) as u64;
     let bytes_per_element = len_type.const_int(elem_bytes, false);
-
-    // dbg!(bytes_per_element, length);
     let number_of_data_bytes = builder.build_int_mul(bytes_per_element, length, "data_length");
 
     let rc1 = match inplace {
diff --git a/compiler/gen/tests/gen_list.rs b/compiler/gen/tests/gen_list.rs
index 1cac2d9e28..3322e1052b 100644
--- a/compiler/gen/tests/gen_list.rs
+++ b/compiler/gen/tests/gen_list.rs
@@ -565,7 +565,7 @@ mod gen_list {
                 pi : F64
                 pi = 3.14
 
-                single : List Float
+                single : List F64
                 single =
                     [ 0 ]
 
diff --git a/editor/editor-ideas.md b/editor/editor-ideas.md
index 1015dc581b..09f3c3967b 100644
--- a/editor/editor-ideas.md
+++ b/editor/editor-ideas.md
@@ -62,6 +62,16 @@ These are potentially inspirational resources for the editor's design.
 * Excel and Google Sheets
     * Not sure, maybe something they do well that we (code editors) could learn from
 
+
+## Machine Learning Ideas
+
+* Ability to record all changes to abstract syntax tree with user permission.
+    * I think it is possible to create powerful automatic error resolution by having a dataset available of ast's with a specific error and the subsequent transformation that fixed the error.
+    * GPT-3 can generate correct python functions based on a comment describing the functionality, video [here](https://www.youtube.com/watch?v=utuz7wBGjKM). It's possible that training a model using ast's may lead to better results than text based models.
+    * Users with large private code bases could (re)train a publicly available error recovery model to experience benefits without having to share their code.
+    * It could be useful to a user who is creating a function to show them the most similar function (type signature, name, comment) in a public+their private database. Say I was using a web framework and I just created a function that has a multipart form as argument, it would be great to have an example instantly available.
+
+
 ## General Thoughts/Ideas
 
 Thoughts and ideas possibly taken from above inspirations or separate.