Every time I'm not reading, I'm thinking about reading.
Title | Category | Medium | Tags |
|---|---|---|---|
I fear this isn't the first (and most certainly won't be the last) of what AI safety in industry will look like. Maybe let's first cover the basics.
Pacing the frontier: building AI at a balanced rate that aims to ensure its safety while still achieving its benefits and grappling with important geopolitical dilemmas.
Dario lays out his three-step plan of how the world can work together in pacing the frontier. It's more a funnel-analogy.
- Embedded Evaluators. 3rd party evaluators (like METR) that aids in verifying safety alignment. This is a call-to-action for other companies to do the same.
- Democratic Coordination. Coordination among AI companies on unchecked AI progress.
- Global Coordination. Government alignment.
The pillars that we should focus on:
:::quote
Dario goes into more detail regarding his three step plan, the tradeoffs within it.
I find the second section's details to be the most fascinating. I particularly found it interesting that he noted limiting the inputs and signals that are used in frontier models like training compute.