▲

Karparthy wisdom

1 points

by Anonymous

https://x.com/karpathy/status/1990116666194456651?s=20

Comments

Log in to comment

  • by Anonymous

    I agree. I wonder if new verification techniques will be developed as a result. For example, I never would've considered art or photography verifiable but better verification (loss) functions were developed. Initially, simple mean-squared error/SSIM for autoregressive models, which didn't solve the problem well (VAEs), then classification error which solved it better (StyleGAN), before finally we modeled the "verification" of art as denoising (Latent Diffusion Models). (Note: Now it seems we are back to autoregressive models because of their ability to integrate with LLMs but I haven't researched this yet.)

    ▲0 points
    Reply