Why is AI training fundamentally different?

Because AI training interleaves network (all gathers / reduces) with computation (geomm), unlike usual networking where the thing you are computing (on top in stack) is detached from networking (below in stack)

Sidebar: what’s the key hard thing of exchanges

When two things happening roughly at the same time, how does one know who did something first?

*Layer 1 switching: instead of doing switching by parsing packet headers, the optical physical layer does the routing by optical routers.

Two problems: scarcity and alignment

scarcity

=> you can’t RL humans freely

• …so the data is selected given priors of how the humans showed up
• we thus need to make assumptions to generalize

Why you need assumptions since you have scarcity

We’d love to infer what X will do given Y circumstance, but unless you have a thing that’s positively X (i.e. a box that’s literally X you can replay), you have to use your experience of (X’, Y’) to make an inductive prior for what X will do.

NYC (a second time!)

(Photo credit: https://quantumi.sh/)

Olympic NP

Hong Kong

Banff NP

Pinnacles NP

Yosemite NP

Singapore

NYC

Shanghai

Grand Canyon NP

AI Safety as a Public Engagement

To what extend should we be making AI Safety as a public problem.

One-Liner

Novelty

Motivation

Notable Methods

Key Figs

• Public generally more pessimistic about AI
• Rates of adverse impact by AI decisions is reasonably low

New Concepts

Data Partnerships

Try to find independent sources of data, with contracts for how the data must be used.

Impact Horizons

Some random Instagram influencer had more accesses than our usual methods of scientific communication.