Hacker News new | past | comments | ask | show | jobs | submit login

You can't really debug an AI. The dimensions of it's understanding are quite literally beyond human interpretation, which makes it both smarter/more efficient than humans while also extremely dumb and context-unaware. Most of our attempts at adding a 'memory' to AI has been a hack thus far, which is why all of these prompts consist of people force-feeding word salad down the AI's throat for generally reproducible results.



They're debuggable, with effort. Here's the finance neuron in CLIP: https://microscope.openai.com/models/contrastive_4x/image_bl...


For memory, check out:

1) Retro, which is essentially attention over large databases, and fast as hell.

2) S4 Layers, explicitly designed for handling long dependencies.

These are orthogonal approaches to memory, and both very effective at what they do.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: