GPT Solver — the trainer from "A Look Behind the AI Curtain", running live

A standalone WebGPU implementation of the companion write-up: a pile of text in, a tiny model that writes and answers questions out. Append ?gradcheck to the URL to run the finite-difference gradient check instead.

loss (surprise per token, lower is better) · red: training text · teal: held-out 

        samples — what it writes when asked to continue "Mara"
      

    

    

      step 0
      loss —
      held-out —
      params 0
      vocab 0
      tokens/s 0
      lr 0
      initializing
    

  


  

    
    
    
    
    
    
    
    
  

  
  

  


  

    
Talk to it

    

      
      
      
      
    

    
(train for a few hundred steps first)

    
Ask wraps your words as Q: … A:, the format the training text uses for questions. Continue
      feeds them in raw and lets the model keep writing. Things it can know: where each person lives, their job, their pet's name and
      color, their favorite food, and sums of small numbers.