diff --git a/README.md b/README.md
index 6e727da1..8beed340 100644
--- a/README.md
+++ b/README.md
@@ -2,5 +2,33 @@ Experimental Rust Serialization Library.
 
 [![Build Status](https://travis-ci.org/erickt/rust-serde.png?branch=master)](https://travis-ci.org/erickt/rust-serde)
 
-This is an experiment to modernize rust's `libserialize` library. It is
-designed to implement https://github.com/rust-lang/rfcs/pull/22.
+This is an experiment to modernize rust's `libserialize` library. It is designed to implement https://github.com/rust-lang/rfcs/pull/22. `rust-serde` is an attempt to address a major shortcoming in `libserialize`. For normal structures, when you say you want to deserialize into:
+
+```rust
+struct Foo {
+    x: int,
+    y: int,
+}
+```
+
+`libserialize`'s deserializer essentially asks for:
+
+* Is the next value a struct named "Foo"? If not, error.
+* Is the next field named "x"? If not, error.
+* Is the next value an "int"? If not, error.
+* Is the next field named "y"? If not, error.
+* Is the next value an "int"? If not, error.
+* Is the struct finished? If not, error.
+
+While this works for user defined structures, it cannot support deserializing into a value like `json::Json`, which is an enum that can represent every JSON value. In order to support that, it needs to be able to do some lookahead:
+
+* What is the next value type?
+    * If a struct, parse a struct.
+    * If an integer, parse an integer.
+    * ...
+
+More formally, `libserialize` implements a LL(0) grammar, whereas `json::Json` requires a LL(1) grammar. `rust-serde` provides this by implementing a serializer and deserializer that produces a tagged token stream of values. This enables a `Deserializable` for `json::Json` to look at the next token before deciding on how to parse the value.
+
+---
+
+There is now also a new library variation called `serde2`. This removes the need for tagged values and replaces them with a `Visitor` pattern. This pattern is very similar to the `Iterator` pattern, but it threads some custom state through visiting each type. This gets many of the benefits of the `serde` library without needing to always pay for tagging the variants.