Executive summary: Current tests of AI capabilities in dangerous domains like bioweapons are inadequate; we need much more rigorous and realistic tests to justify taking costly preventative actions.
Key points:
Existing AI capability tests in biology are too easy and don’t reflect real-world challenges of creating bioweapons.
As AI improves, companies keep making tests harder, but still not hard enough to conclusively demonstrate danger.
A “gold standard” test would involve a randomized trial of non-experts trying to create a (harmless) virus with AI assistance vs. internet resources alone.
We need to agree in advance on tests that are difficult and realistic enough to clearly justify strong preventative actions if passed.
Designing truly hard capability tests is challenging but crucial to do now, before AI potentially becomes extremely powerful.
Focus should shift from proving current AI safety to determining how to identify if future AIs are truly dangerous.
This comment was auto-generated by the EA Forum Team. Feel free to point out issues with this summary by replying to the comment, andcontact us if you have feedback.
Executive summary: Current tests of AI capabilities in dangerous domains like bioweapons are inadequate; we need much more rigorous and realistic tests to justify taking costly preventative actions.
Key points:
Existing AI capability tests in biology are too easy and don’t reflect real-world challenges of creating bioweapons.
As AI improves, companies keep making tests harder, but still not hard enough to conclusively demonstrate danger.
A “gold standard” test would involve a randomized trial of non-experts trying to create a (harmless) virus with AI assistance vs. internet resources alone.
We need to agree in advance on tests that are difficult and realistic enough to clearly justify strong preventative actions if passed.
Designing truly hard capability tests is challenging but crucial to do now, before AI potentially becomes extremely powerful.
Focus should shift from proving current AI safety to determining how to identify if future AIs are truly dangerous.
This comment was auto-generated by the EA Forum Team. Feel free to point out issues with this summary by replying to the comment, and contact us if you have feedback.