Citations & default visibility for "AI Outcomes" page #698
Description
The page AI Outcomes has a cool idea. The "Info" button is easy to miss however, and its text currently reads as a bunch of assertions however, and the content is outdated. I propose making the text visible by default, and updating that text per the below example:
Branch: Will one of these AIs try to take control?
Text: We already have examples of AIs that explicitly tried to take over (e.g. ChaosGPT), but none of them were smart enough to get far. Even for goals that sound normal (such as 'cure cancer' or 'calculate pi'), power-seeking is often a very reasonable sub-goal to have. There is even mathematical proof that optimal policies seek power. And even if power-seeking behavior itself is rare, it only has to happen once. We're spawning millions of AI instances every day, one of these instances will eventually try to maximize its power. So we are quite certain that this will happen.
Improvements for this example text:
- ChaosGPT which is ~3 years old. We should use a more recent example.
- If we keep the ChaosGPT example, we should link to a relevant article. Currently when I google "ChaosGPT", the first result is a GitHub repo that won't make sense to newbies, and the third link is some commercial product.
- "often a very reasonable sub-goal to have" should link to either supporting research or something that makes the research more approachable
- "we are quite certain" should link to authority figures, or make it clear it's PauseAI's opinion.
I estimate ~1hr to source appropriate links for each branch's Info text. I had to remove most links because it was blocking Issue creation for me.