AI Safety Discussion Days

by Linda Linsefors2 min read27th May 20201 comment

11

Ω 5

AICommunity
Personal Blog
Crossposted from the AI Alignment Forum. May contain more technical jargon than usual.

AI Safety Discussion Days

I’m going to organise an online AI Safety discussion day, roughly once per month. This is a space to discuss anything you want related to preventing catastrophic risk from AI, including governance, strategy, all technical AI Safety research, and also trying to understand humans better for the purpose of alignment. The goal of these days is to learn from each other, exchange perspectives and get feedback on your ideas.

The discussion days are primarily for people who are actively working on preventing AI related X-risk, or plan to do so in the near future. However, because there is no limit of space on the internet, anyone who is interested is allowed to listen in, as long as you’re not disrupting any conversations

The first one will be on Sunday the 7th of June. To be inclusive to everyone in every timezone, the discussion day will last from midnight to midnight, as defined by UTC. Here’s the information for the first discussion day.

To stay informed about future discussion days, you can sign up for my newsletter.

If you just want to participate and not worry about anything else, you can stop reading here. Well done.


Background and thoughts about this project

Two weeks ago, I organised Web-TAISU, a four day unconference about AI Safety. This event seemed both useful and appreciated, so I will see whether I can keep the discussion going.

I will prioritise making sure that each month’s discussion day is high quality and worth your time to attend. But I will also experiment with different event structures, so you should expect this event to be somewhat different each time

The first discussion day will be a simple unconference structure, with an open schedule where you can add sessions you want to run: talks, discussions, workshops, etc. This is the same structure as Web-TAISU, except there won’t be any session pitches at the start. Instead the schedule is already open for you to add what you want to do.

If you have some activity you would like to run during a future discussion day, or if you have ideas or suggestions for something I should try out, or if you would like to try organising one of these events yourself, then let me know. I don’t promise to try your idea, but I will consider it. If you would like to run some activity as part of a discussion day, I will most likely be up for working out how to incorporate it into the event, at least to try it out. If you want to take over the organisation of one discussion day to run it according to your vision, I might let you, but you’ll have to convince me that you know what you are doing.

Possible failure modes

I.

Without the initial session pitches the schedule will not fill up. I don’t expect this to be a problem, but if it is, the event will fail gracefully. The schedule will be very light, and the event will not be all that it could be, but also there will be no harm done, and I will learn something for next time.

II.

I have not set any lower bar for participating, because I think that junior researchers can learn a lot from just listening to the conversations of more senior people. However, this could backfire and mostly ruin the value of the event, if the conversations end up at the level of the least informed person.

This is the failure mode that I’m most worried about. But I do think that with the right norms, there can be room for both high level discussions and also sometimes letting people ask naive questions. My strategy for now is partly to just observe to what extent this failure mode is happening, and partly to try to encourage good moderation and discussion norms. If this doesn’t work, the next step would probably be to add some structure where each session is coded as beginner friendly or not, or something like that. I’m open to suggestions.

III.

There just isn’t enough to discuss to be had to fuel a discussion day every month. I think that once a month is roughly the right frequency, but I might be biased because I like events. If these discussion days fail because of this failure mode, then it’s alright. What I’m more worried about is if I can’t distinguish this failure mode from other reasons people might stay away.

Therefore, If you’re not joining my discussion days, I’d very much appreciate it if you would tell me why.

Project feedback

If there is any reason why this might fail that I haven’t thought of, or if you have some advice for me, please let me know.

AI2Community2
Personal Blog

11

Ω 5