THUDM GLM-4-9B LOCAL Test (INSANE JS/HTML Skills)



AI Summary

Summary of the Video: Testing 9 Billion Parameter GLM Model

  • Overview: This video tests the 9 billion parameter variant of the GLM model family, following similar tests conducted on the previous 32 billion parameter model.

  • Model Specifications:

    • Using the Barttowski quantization in LM Studio.
    • Specifications:
      • Machine: Dual 3090Ti
      • 8-bit quantization, approximately 10 GB download size.
      • Accessible for systems with 16 GB or less VRAM.
  • Key Settings:

    • Increased context length over 10,000.
    • Enabled flash attention.
    • Adjusted model loading guard rails for faster generation speed.
  • Game Testing:

    • Initial tests involved generating a retro synthwave style game using Python, but results were less satisfactory.
    • Successful HTML website generation for ‘Steve’s PC Repair’:
      • Included images effectively, generating a visually appealing website.
      • Features relevant pictures, service descriptions, and user testimonials.
    • Subsequently tested game generation using JavaScript:
      • Generated a Pong game, producing a responsive and visually engaging output.
      • Demonstrated animated graphics and interactive gameplay, exceeding prior Python attempts.
  • Performance Observations:

    • Model excels in web design and interactive graphics tasks.
    • Struggles with Python code generation, showing simpler errors.
    • Comparable performance to the 32 billion parameter variant with a focus on UI/UX design.
  • Conclusion: The 9 billion variant of the GLM model shows promise, particularly in web-based applications, though it falls short in coding capabilities. Further testing could explore creative writing and other applications.