• JavaScript has no place on the server part one: The barrel file debacle.  (Marvin H)
  
  After all these years, it's still a toxic hellbrew of bad decisions.  Not so much JavaScript in itself, which is merely eh, as the choice to use it in server applications where it is obviously completely unsuitable, to do so via an event loop, which is obviously completely unsuitable, and then to add the worst package manager ever devised, a software version of the Torment Nexus from the classic sci-fi novel Don't Create the Torment Nexus.*
  
  * Which doesn't exist, but should.
  
  
  
• JavaScript has no place on the server part two: The hidden performance cost of Node.js.  (Software at Scale)
  
  Here the problem is trying to implement GraphQL naively in an event loop architecture, which creates a cascade of asynchronous requests - called promises - which totally gum up the works.
  
  This problem was solved all the way back in 1961, before the geniuses behind Node.js decided to inflict it on a whole new generation.

• Fine tuning Mistral-7B on Python code with a single GPU.  (Weights & Biases)
  
  Large Language Models have a lot of problems, stemming from the facts that they don't know anything except what words go together, they're promoted by frauds, and they are actively hamstrung from doing anything but reciting woke bullshit.
  
  But Mistral-7B is completely open-source and you can download it and do anything you like.  And if what you like is programming, well, it turns out you can get a long way in programming just by knowing what words go together.
  
  It helps because you can automatically check if the AI got the basics right - does the code it generated even compile, for example?
  
  I'm still more interested in using AI for testing than for generating code, but having a generally good LLM that is free to everyone and small enough to run on commodity hardware is a win either way.
  
  

Unrelated

647 MIPS on a 4GHz Zen 3 (Ryzen 5625U).  Good enough, particularly since that's with overflow checking enabled.  The performance impact of that was tiny; small enough that I had to test and confirm it was really doing the check.

About 20x faster than Python (though 3x slower than PyPy). 

And 6x faster than the same VM written in Nim, so I guess I won't be using Nim for this.

Posted by: Rick C at Tuesday, October 10 2023 12:19 AM (BMUHC)

Posted by: normal at Tuesday, October 10 2023 12:42 AM (obo9H)

Hide Comments | Add Comment