Fieldbuilding project ideas in AI safety

Ideas you should steal

There’s a bunch of AI safety fieldbuilding projects that I think should happen! Here are some, which you should totally run with if they’re exciting to you.

1. Help elect Alex Bores to Congress

Alex is a Cornell alum who cares lots about good AI regulation!! If you live in Manhattan or know friends there, you should consider voting for him and talking to your friends about him. Text me if you have more questions!

2. Create a lightweight junior mentorship network

Create a program/resource where fellows who’ve completed at least one AI safety research program (MATS, Astra, AFP, Pivotal, ERA, Generator) can share resaerch ideas and provide lightweight mentorship to newcomers to the field.

Justification: I’d like to see more support for ‘wider’ parts of the AI safety research funnel! I think safety program fellows are likely to be pretty good mentors, and can provide increased mentorship capacity for the bunch of bright-eyed newcomers looking to get into AI safety. In most cases it’s a win-win:

• For fellows: we have a backlog of ideas (ours and our mentors’) that we won’t have time to explore and that are relatively shovel-ready — they just need extra hands to run the experiments. It’s fun to have other people pursue them! Mentoring also sharpens your research taste and builds career capital.

• For mentees: the field is badly mentorship-bottlenecked. Interest in AI safety is growing fast (awesome!), but fellowships are very competitive, the mentor pool is small, and established mentors are super busy. Fellows who’ve gone through a research program have strong safety context and taste — picked up from their own mentors — and more time than MATS/Pivotal/ERA-level mentors, enough to mentor part-time and provide enough guidance to early folks.

A lightweight way to implement this: send a form to safety-program alumni collecting (a) research projects or experiments they’d want someone to run, and (b) whether they’d mentor whoever takes one on, and at what intensity (a one-off hour, 30 min/week, etc.). Publish the ideas — plus, optionally, the mentors’ contact info — on a site for anyone to take a stab at. Somewhat similar to AE Studio’s recent AICRAFT program!

I envision this as requiring lower-bandwidth mentorship requirements than SPAR, and able to run continuously instead of in cohorts.

Open questions: what should the quality bar be on each side? Is this actually a good upskilling channel for mentees? How would mentees get compute? My guess is it’d work well for, e.g., agentic undergrads who just want guidance a first project that they can hack on — I’ve met many ppl like this.

3. Start a creative magazine for AI safety

Pick up Proxima, the AI safety anthology Parv and I started to publish art/writing/other creative shindigs related to AI futures. We were excited about it as a positive culture-building thing for the community, and it's just really cute — I want to see more creative projects!

We ran out of time to follow through with it during our fellowships, and would be happy to hand it off.

4. Run more AI safety events in Asia this July

July is exciting in Asia AI-land: ICML in Korea and the World AI Conference in Shanghai! This is a strong window for AI safety fieldbuilding and networking — there will plausibly be more attendees from Asia, esp China, than a typical West-based big ML conference, so if you run satellite events you’ll likely reach people you usually wouldn’t.

I’m especially excited about people running compute-verification workshops and general AI safety socials. If you are excited about this, you should start soon, as ICML is in 1 month! (Note there’s already the Seoul FAR alignment workshop & TAIGR, for you to keep on your radar).

5. Resurrect AI Lab Watch

Alas, this idea has been bounced around a lot, but as of 1-2 months ago I believe it still hasn’t been done. Having more lab accountability is important and maybe we should try to do Lab Watch again and sustainably.

6. Make it easier for safety-interested students across different schools to connect

Start a good Discord/Slack channel, or a website, where students from different schools’ AI safety / EA groups can get to know each other. This platform could require log-in with a verified school email and possibly affiliation with an existing AIS group. Think of this like an evergreen Swapcard for undergrads!

Why is this useful? There’s already a large Slack for AI safety organizers, but I’m mostly excited about making it easy for non-organizers (newcomers) to make friends with people from other university groups. There’s a lot of value here for students at schools with smaller/newer safety communities: it can help them feel that other people genuinely care about safety, without the current main pathways of going to an in-person retreat (e.g. OASIS, GCP, Action Potential) or being part of an already-strong university group. I was personally one-shot by conferences/in-person retreats, and I think a well-maintained cross-school channel could 80:20 that.

7. Figure out how to ‘win-pill’ people

I think of ‘win-pilling’ as an important step after baseline safety-pilling: getting people to viscerally want to achieve ‘victory’ in AI safety, and giving them the tools to develop a theory of victory and strategize how to make it happen.

Many people aren’t thinking this way — I wasn’t, until about two months ago, when my spectacularly agentic friends enlightened me. I think this is in part because of AI safety’s roots in rationalist research and blogging culture, which breeds a particular vision of success (heads-down researcher).

Some ideas for ‘win-pilling’ people:

1. Run workshops or lightning talks at AI safety programs and conferences: how to ‘win,’ heuristics for thinking about success conditions, how to find your comparative advantages (“superpowers”)

2. Write about winning and make it go viral (in the vein of Luke’s piece on moonshots and Nan’s on general managers).

   1. For instance, build off Jason’s theory of victory piece and write a call to action that is more snappy.

3. Set up a shadowing program, where you spend a day with someone who is deeply committed to winning. (shadowing is generally good)

4. Spread the meme in other ways, e.g. by making good resources like BlueDot’s playbook more well-known

8. Create better networks between folks in AI labs and AI policy

I don’t know exactly what this should look like — maybe more networking events, fellowships, etc. Someone should think about this, what the good ideas are, and how to do this if so!

9. Write your own list(s) of ideas and share them

I am soooo pro sharing your ideas!

Maybe: publish a bunch of projects that you don’t have time/bandwidth to execute on (e.g. questions that someone should figure out answers to, software you want to see, etc), to the chances that someone with more bandwidth/time/interest can take them and make them happen in the world.

Austin mentions that “ideas are cheap and execution is everything”. I agree, but I think good ideas can also be really hard: things that seem really obvious to you, given your context, are not at all obvious to other people. So make them shovel-ready, and hand people something they can just go and run with!

You could go one level more meta and build a hub for AI safety idea lists (a ‘list of lists'), for easier browsing. Or, you could even try to normalize /ideas pages on personal websites, just like /now pages)

Note that execution is probably still the big bottleneck — remember that ideas themselves are not enough. Go out and make these wonderful cool things happen!

---

Inspired by Austin Chen’s recent post, and conversations with friends/mentors - particularly Sydney.