A first-stab model of conference value
Amy Willey was a key organizer of the Singularity Summits in the 2000s and helped run EA Global Boston 2017, which Alyssa had just attended (see “EA Global Boston”, dated the next day). The note below sketches Alyssa’s working model of what makes a conference good, with a closing comparison between EA Global and the Future of Life Institute’s 2015 Puerto Rico and 2017 Asilomar AI conferences.
Lots of different directions we could go in here, but regarding conference value, here’s a rough first stab at the model: First, to state the obvious, a conference’s value generally goes up in proportion to the number of attendees. A larger conference delivers more value, because more people experience it. There are several caveats, of course. As size rises, cost, work hours, logistical complexity, and so on all rise too. And as a conference gets larger, it’s more important to split it up into smaller events (possibly some reserved or private), as some discussions can only happen in a smaller atmosphere.
Second, a conference is valuable when the attendees have interesting things to say. What counts as interesting, of course, depends on the person, the context, and so on — a computer security whiz might not have much to say at an agriculture conference, and vice versa. But we can be pretty sure that having most attendees be typical college students is bad. Having a lot of attendees who all want to say the same thing — whether that’s “donate 10% to GiveWell-approved top charities”, “stop the war in the Middle East”, or even “chocolate is better than vanilla” — is bad. Having lots of attendees who come out of pure boredom is bad (to some extent, expensive tickets stop this, though of course they’re also harder for those on low incomes). A conference needs to be the opposite of a protest; at a protest, you want to come up with a few basic, simple ideas, get as many people to agree to them as possible, and then have everyone chant the same thing over and over and over. If you do this at a conference, the people you’d want to attract will get bored very quickly.
Third, the best conferences have some, but not all, attendees who are in “positions of influence”. This includes the rich, the famous, people with important government jobs, heads of major corporations, key academics, and so on (exact mix obviously varies). For most goals, getting these people talking to you, interested in your ideas, and on your side is essential. However, these people will also be very busy, and won’t usually contribute that much of their time later on. In addition, a healthy community needs to be bringing in new people and building up their skills, talents, and networks. So a good mix is ideal.
Fourth, the most important quality for speakers is name recognition and social status among the target audience. Talk quality, though still important, is secondary, which is of course why Ray Kurzweil keynoted the Singularity Summits. But beware: a speaker who’s very well-known and important among the Xs (for some value of X) will attract lots of Xs, which might or might not be what you want. For the talks themselves, the best talk is one level above the current audience’s; enough that they learn new things and can feel the speaker’s expertise, but not so far beyond that they’re bewildered or disinterested.
Regarding AI conferences, the two events held by FLI (one at Puerto Rico and one at Asilomar) made huge progress and did an awful lot of things right, even though I disagreed with a few of their decisions. If CEA could just match the quality of event that FLI created, but do it regularly and consistently, that would be an enormous step forward. (FLI employed several tricks in addition to the above — eg., because many attendees they wanted were scared of talking about superintelligence, they arranged the conference agenda to try to ease people into the idea.)