ahrefs
diff --git a/‎CHANGES.md‎
Lines changed: 1 addition & 1 deletion b/‎CHANGES.md‎
Lines changed: 1 addition & 1 deletion
diff --git a/‎arrayjit/lib/anatomy_of_a_backend.md‎
Lines changed: 1 addition & 1 deletion b/‎arrayjit/lib/anatomy_of_a_backend.md‎
Lines changed: 1 addition & 1 deletion
@@ -10,7 +10,7 @@
   - This way we avoid introducing multiple operations on bits. We might yet split this operation into more fine grained ones.
 - TODO: Support of counter-based randomness, and behind-the-scenes key splitting by the relevant operations.
   - TODO: The cascade of splits uses the Tnode id, the train step and the tensor cell position.
-- TODO: Added a new operation `Uint4x32_to_prec_uniform`. It is added to `dedicated_access` and thus to `fetch_op` -- it cannot be `unop` because it cannot be implemented pointwise / is not shape-preserving due to efficient use of the 128 bits.
+- TODO: Added a new operation `Uint4x32_to_prec_uniform` to `unop`. It has a dedicated support in shape and projection inference to account for not being pointwise / is not shape-preserving due to efficient use of the 128 bits.
 - Added a field `params` to `Tensor.t` since we need to track parameters to properly initialize computations (see below).
 
 ### Changed
 
@@ -20,7 +20,7 @@
 The modules and files of `arrayjit` can loosely be divided into three parts.
 
 - "Utility": generic code providing support across the project, both for `arrayjit` and `neural_nets_lib`.
-  - The modules `Utils`, `Rand` and `Ppx_helper`.
+  - The modules `Utils` and `Ppx_helper`.
 - "Frontend": the parts that are user-visible, concrete, and shared across the project.
   - `Task`: a wrapper for work execution (`unit -> unit`).
   - `Ops`: numeric precision specification types, primitive numerical operations.