“We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain.”
He’s basically just saying, I dunno, spend 20% of every sprint building defensive capabilities.
Because 75-80% still allows them to overhaul society.
And in their defense - I do think they honestly believe in the dangers of AI and think that defensive measures have to be built in, but yeah, as part of the overhaul of society.
Right later in the article he says:
“I believe that if slowing down bought us even an extra year or two before models reach critical levels of capability, and we used that time to advance alignment, we could greatly reduce the risk that something goes seriously wrong.”
This is not a broad societal, maybe we don’t take full advantage of AI and destroy jobs. It’s just - let’s try not to have AI hack into all the critical systems and launch nukes.
In 5 or 10 years everyone will once again be “why did we let this happen?”