Performance is much better now

1 files changed, 1 insertions, 1 deletions

diff --git a/running.md b/running.md
index cbcbcef6..5df36550 100644
--- a/running.md
+++ b/running.md
@@ -8,7 +8,7 @@ There are currently two active BQN implementations: the self-hosted one in this
 
 The online REPL is [here](https://mlochbaum.github.io/BQN/try.html). The file [docs/bqn.js](docs/bqn.js) is zero-dependency Javascript, and can be loaded from HTML or Node.js. For command line use, call the Node.js script [bqn.js](bqn.js), passing a file and `•args`, or `-e` to execute all remaining arguments directly and print the results. [This notebook](https://observablehq.com/@lsh/bqn) shows how to run it in an Observable notebook.
 
-The version of BQN in this repository is implemented mainly in BQN itself—the compiler is entirely self-hosted, while the runtime is built from a small number of starting functions using preprocessed BQN. It completely supports the core language except for block headers and multiple body syntax, and a few cases of structural Under (`⌾`). The Javascript-based compiler is also slow, taking about 0.05 seconds plus 1 second per kilobyte of source (this is purely due to the slow runtime, as dzaima+reference achieves 1ms/kB with the same compiler once warmed up).
+The version of BQN in this repository is implemented mainly in BQN itself—the compiler is entirely self-hosted, while the runtime is built from a small number of starting functions using preprocessed BQN. It completely supports the core language except for block headers and multiple body syntax, and a few cases of structural Under (`⌾`). The Javascript-based compiler is also slow, taking about 0.01 seconds plus 0.3 seconds per kilobyte of source (this is purely due to the slow runtime, as dzaima+reference achieves 1ms/kB with the same compiler once warmed up).
 
 Because self-hosted BQN requires only a simple virtual machine to run, it is [fairly easy](implementation/vm.md) to embed it in another programming language by implementing this virtual machine. The way data is represented is part of the VM implementation: it can use native arrays or a custom data structure, depending on what the language supports. An initial implementation will be very slow, but can be improved by replacing functions from the BQN-based runtime with native code. As the VM system can be hard to work with if you're not familiar with it, I advise you to contact me to discuss this option it you are interested.