Talking about AI Safety with Hikers

"I'm just looking at all these people running around on dates and going to parties and hikes and plays, and it's like... don't these people have planets to take care of? Aren't they afraid their civilizations will collapse into barbarism while they're distracted?" (EY on Twitter, May 6th 2018)

I think a lot about existential risks, while others are going to "parties, hikes and plays".

There is nothing intrinsically wrong with those activities. I also enjoy a good play or a trip through the Alps.

Some people just don't feel concerned by existential risks and could spend their whole life "hiking" without bearing any responsibilities, and I'm ok with that (let's call them hikers).

However, some hikers understand the concept of existential risk.

I have an architect friend who works in sustainable development. He can present what he does and what he cares about in one sentence. People immediately grasp why he is doing this particular job (architect has high status and sustainable development is a cause worth fighting for).

Yet, when I go to parties and start talking about AI Safety, hikers tend to go for a drink and never come back, because it takes me too long to explain.

Explaining AI Safety is Difficult

My architect friend is next to me. When he talks about global warming, with a sparkle on his eyes, people tend to listen.

It's easy to understand. We hear about global warming all the time and sustainable development is praised in schools. There is no reasoning, conceptual gap or assumptions to be made.

For AI Safety, people don't even know what I'm referring to.

In this case, I try to go back to the nearest common-node in our knowledge network, so I can bring them to a non-trivial probability of a Superintelligence occurring in their lifetime.

Sometimes, this node is the general field of AI (when I talk with ML engineers or friends from my Master's in AI). Sometimes, they have studied art or music, and their closest relationship with (Computer Science) /AI is their computers, and that's fine. However, I have an hard time explaining what I value in less than 30 minutes to non-AI people in parties.

Improving the Message

Obviously, I want to improve my average time spent explaining it. Thus, I went back to see how Nate Soares introduced himself in his talk at Google:

"I'm the director of the Machine Intelligence Research Institute, very roughly speaking, we're a group that's thinking in the long term about Artificial Intelligence and trying to make sure that by the time we have advanced AI Systems we also know how to aim them"

I think he does a decent job in explaining clearly, in simple but precise sentences, what the MIRI is about, without talking about Superintelligence. However, his audience 1) has a good background in Math/CS 2) understands roughly what is "AI", "advanced AI Systems" and "aim[ing] them" (speech at Google, duh). So I would (still) need to adapt this talk to completely non-technical people.

Soares latest tweet is about Neil deGrasse Tyson who recently (Feb. 2018) changed his mind about AI Safety, by going from "just unplug it" to "the AI gets out of the box every time". He did so after listening to Sam Harris' podcast with Eliezer Yudkowsky, where the latter exposes the AI in a box problem. This struck me, because not so long ago, I was also listening to this podcast, walking in the sun in Paris, trying to get better intuitions/examples to explain better AI Safety to strangers.

Don't get me wrong: I am not trying to be an AI Safety evangelist. I am just trying to communicate better what I value, which I think is a valuable skill to acquire. Besides, I think raising awareness about AI Safety is a matter of concern, and even though convincing a random guy in a party won't solve the Alignment problem, it can help me towards articulating better those thoughts if I get to talk with Emmanuel Macron one day.

Emotional Response

Remembering about this podcast made me watch this Ted Talk by Sam Harris once again. He lists exhaustively whatever could block you from having the correct emotional response to the control problem:

  • AIs in Sci-fi are more fun than global famine
  • The spectrum of intelligence is greater than you think, and any intelligence greater than humans' will likely explore it
  • The best case scenario is not Burning Man but a world of increased inequalities
  • It is easier to solve AGI than to solve AGI + neuroscience, so neural implants (with AI) are not likely
  • The first Superintelligence may be the only chance we have

I find Harris extremely convincing. His body language, tone and wording are aimed at communicating this very complex idea to millions of people, and I think he killed it. His implacable reasoning applies to what people actually care about.

Yet, it's difficult to go from "there is a non-zero probability of a doomsday scenario" to "this is a problem I need to work on, and not just a funny thought experiment". At the end of Neil deGrasse Tyson's video, where he publicly states that he changed his thoughts, he continued with "Yes, the AI gets out of the box every time. Yes, we're all gonna die [crowd starts laughing, end of the debate]". This is not the correct emotional response.

Knowing vs Doing Something About It

For the last two months or so, I organized 4 Meetups on AI Safety, and got totally different results every time. From the 50 members in the Meetup Group, only about five seemed to care enough about AI Safety to work actively in the field, or to just think about it in the shower.

In particular, last Meetup was about the different paths to Superintelligence, and my three guests (the Logican, the Entrepreneur and the Hacker) where much more interested in AGI Capabilities than AI Safety.

Understanding AI Safety does not imply doing something about it.

The Way Ahead

Whenever I try to talk about AI Safety with hikers, I tend to delve into thought experiments, especially the paperclips maximizer one. The emotional response I get most of the time is amusement or, worse, indifference. In the future, we need to be much more capable of explaining clearly what this field is about without delving into too many thought experiments. Most importantly, awareness must be raised, so that I can do it without hikers going for a drink because it took me too long to explain.

This is the 12th post of a series of daily LessWrong posts I started on April 28th.

Previous post: Applied Coalition Formation