When introducing people to AI Risk a common failure mode is to go for too much, making claims too far outside their personal Overton window. A great example of a common reaction to this is given here[1]
Another example: I attended a talk given to some high school students using a similar "shock and awe" approach, I observed the same reaction as in the image from everyone I asked (around 5 people). If anything overbearing introductions like this vaccinate the listeners against the field.
Overall I model introducing weird AI beliefs as a trade-off between
- Probability they agree with the arguments
- How fantastical the claims are
Given the goal of "making people take AI risk seriously enough to consider working on it" it's important to make the right trade offs[2], you want to focus on increasing the first number even if you have to decrease the second. Some (in my opinion) good ideas:
- Starting with narrow, aligned AI risk (e.g. mentioning AI weapons, China recruiting children Putin saying the leader in AI will rule the world, etc.)
- Giving a timeline of AI progress and asking people to mentally guess when a certain event happens before you tell them (for the people not up-to-date on AI progress)
- Mentioning AI metaculus predictions and how metaculus has done in the past
- Giving the genie analogy and playing the "I tell you why your alignment proposal doesn't work" game
- Conservative lower bounds on human level AGI power, mention I/O and speed, self replication, doing AI research to improve itself[3], etc.
- Graph of AI training costs or neural network sizes (GPT3, PaLM, etc.)
And some bad ideas
- Talking about a utopia with simulated humans (Everyone dying is bad enough for people to work on it, you don't need to go for more!)
- Giving your (often weird) beliefs without explanation before people can form their own, e.g. saying I expect AGI within 20 years with no explanation can be actively harmful compared to showing people AI progress and asking them to draw their own conclusions.
- Assuming powerful nanotechnology is possible and giving Clippy-like scenarios (I'm guilty of this myself, it would be better to initially give more realistic scenarios)
- Not mentioning the history of AI progress / assuming people already know it (this is important for forming your own timelines!)
For reference I actually agree with everything in the image, I simply question the approach of immediately mentioning all your weird beliefs in rapid succession ↩︎
Remember what you are optimizing for! Not "show this person how weird my beliefs are, and how little I care about social approval", Not "Signal to the in-group I'm one of them" but "Maximize the probability the people I'm talking to will (eventually) take AI risk seriously" ↩︎
Another nitpick, talking about a human level AI "doing AI research" makes fewer assumptions then mentioning unclear rapid self-improvement, of which there is ample debate. This is another trade off in which we increase 1. Probability they agree and decrease 2. How fantastical the claims are ↩︎