6 ms·

Seems like they could fix nearly all of these by making it decline to roleplay?

• rafiste 4y ago
  The problem is that LLMs are basically roleplay simulators. Re Karpathy: GPTs don't want to succeed. They want to imitate. https://twitter.com/karpathy/status/1627366416457555969?s=20 https://twitter.com/karpathy/status/1627366416457555969?s=20
  • 13years 4y ago
    Yes, I think we are going to need a new architecture for LLMs to move beyond, "that is interesting", to something that is reliable and can be used for trusted applications.
    • famouswaffles 4y ago
      It's not an architecture problem of the transformer at all. This is the result of thinking the idea that you can make inviolable rules for a system you don't understand is not anything but ridiculous. You're never going to make inviolable rules for a neural network because we don't understand what is going on on the inside.