- a few years ago a Deep Learning network was 5 layers deep,
- In 2014, the Google Image categorization network was 22 layers deep
- In 2015, a new approach to information flow enables deep learning networks 200 layers deep
A specialized AI can be used to recursively improve the design of artificial intelligence
The AI software is not as software as we knew it. Here is a mathematical description describing the behavior and memory.
They are spontaneously organized patterns that are emperically engineered.
Eric argues we are being to anthropomorphic about our analysis of AI and AI risks.
Eric thinks we can develop superintelligence without using or creating AI agents.
Eric discusses constraint by construction.
Provide specialist AI only the information that they need for the tasks.
Use an architecture of specialized pieces.
Then think of using superintelligent tools