Post

It's kind of funny that due to AI/LLMs these days another metric to rate #programming languages on has become "token efficiency". On the bright side - some of the beloved languages like #Clojure, #Ruby, #FSharp and #OCaml fared pretty well due to their expressive and compact syntax.

16/1/26, 9:29 am13 reposts23 likes

Bozhidar Batsov (a.k.a. Bug)

@batsov.net

16 Jan

The data above is from this article martinalderson.com/posts/which-...

martinalderson.comWhich programming languages are most token-efficient?Comparing token efficiency across 19 popular programming languages using RosettaCode data - from Clojure to C, there's a 2.6x difference.

Braidpoint

@braidpoint.bsky.social

17 Jan

My reflex is to consider Dynamic languages as "not qualifying" because of the churn that they can create. It can overwhelm a context window (unless there's a mountain of feature extracted training data to compensate) #python

Bozhidar Batsov (a.k.a. Bug)

@batsov.net

16 Jan

I meant to write "my beloved languages", but close enough I guess. 😅

jkone27

@jkone27.bsky.social

16 Jan

Cc @aspnetcore.news @dotnetfoundation.org @maddymontaquila.net @captainsafia.com ⚡️better compiler automatic type inference means not having to send back and forth type annotations to LLMs, automatic type inference more efficient for LLMs use? low keyword like JS or Py but fully typed like C#

Safia Abdalla

@captainsafia.com

16 Jan

One thing I think this chart misses is the syntax that LLMs have been trained at. For example, the issue with LLMs not understanding collection expressions in C# was notorious and contributed to more verbose code. Same with other language features that make it terse but clear.

Braidpoint

@braidpoint.bsky.social

17 Jan

💯 We've adjusted our expectations for how much we get out of a given context window simply for all of the guards that have to be put in place with Serena memories and other documentation references.