Slim overview of work one could do to make AI go better (and a grab-bag of other career considerations)
Many kinds of work one could do to make AI go better and a grab-bag of other career considerations
I recently found myself confused about what I’d like to work on. So, I made an overview with the possible options for what to work on to make AI go well. I thought I’d share it in case it’s helpful for other people. Since I made this overview for my own career deliberations, it is tailored for myself and not necessarily complete. That said, I tried to be roughly comprehensive, so feel free to point out options I’m missing. I redacted some things but didn’t edit the doc in other ways to make it more comprehensible to others. In case you’re interested, I explain a lot of the areas in the “Humans in control” and the “Misalignment” worlds here and to some extent here.
What areas could one work on? What endpoints or intermediary points could one aim for?
Note that I redacted a bunch of names in “Who’s working on this” just because I didn’t want to bother asking them and I wasn’t sure they had publicly talked about it yet, not because of anything else.
“?” behind a name or org means I don’t know if they actually work on the thing (but you could probably find out with a quick google!)
| World it helps | The area (Note that this doesn’t say anything about the type of work at the moment. For example, I probably should never do MechInterp myself because of personal fit. But I could still think it’s good to do something that overall supports MechInterp.) | Biggest uncertainty | Who’s working on this |
| Hu- mans in con- trol | ASI governance | human-control
| Is this tractable and is success path-dependent? | Will MacAskill, [redacted]?, indirectly: cybersec. folk?, some AI governance work? |
Acausal interactions | human-control
| [redacted] | ||
| SPIs for causal interactions | human-control | CLR | ||
| Mis- align- ment | Prevent sign flip and other near misses | Is this a real concern? | Nobody? |
Acausal interactions | misalignment
| Is this tractable? | [redacted]? [redacted]? | |
| Reducing conflict-conducive preferences for causal interactions & SPIs | misalignment | CLR | ||
Main- stream AI safety best thing to work on | Reduction of malevolence in positions of influence through improving awareness (also goes into the “Humans in control” category) | [redacted]? Nobody? | |
| Differentially support responsible AI labs | For some of these: Would success be net good or net bad? If good: How good? How high is the penalty for being less neglected? | ||
| Influence AI timelines | [redacted], [redacted], [redacted]?, maybe misc. policy people? | ||
| AI control (and ideas like paying AIs) | Redwood Research | ||
| Model capabilities evaluations | METR, Apollo?, maybe AI labs policy teams, maybe misc. Other policy people? | ||
Alignment (more comprehensive overview):
| Overview post on LessWrong | ||
| Human epistemics during early AI | ~Forecasting crowd, nobody? | ||
| Growing the AI safety and EA community or improving its branding or upskilling people in the community (e.g. fellowships) | Constellation, Local groups, CEA, OpenPhilanthropy, … | ||
| Improving the AI safety and EA community and culture socially | CEA | ||
| Threat modelling, scenario forecasting etc. | [redacted], … | ||
| Make it harder to steal models | Cybersecurity folk | ||
| Regulate Open Source capabilities | Policy folk? Nobody? |
What types of work are there?
| Which world | Type of work | Broad category of work |
| Can be in any of the three areas above | Offering 1-1 support (mental, operational, and debugging) | |
| Project management, events, and programs | Organising | |
| Short, blogpost-style research, for example summaries, overviews, conversation notes, other distillations; potentially writing for others | Research or otherwise being a thinker, Varying in my position in the research pipeline from foundational to strategizing about how to get things done | |
| Long report-style conceptual research: Foundational (E.g. understanding an aspect of decision theory or cognition better) | ||
| Long report-style conceptual research: “Applied” (closer to what I’ve been trying to do. Trying to understand the implications. Could also be alignment thinking, e.g. [redacted].) | ||
| Pitching high-level empirical project ideas and grantmaking | ||
| Working with language models: Empirical ML | ||
| Public polling, qualitative opinion research | ||
| Humans in control | ASI governance thinking | |
| Synergizes most with “Mainstream AI safety” areas above | EU AI office and AISI style policy work | Setting policy “Normal”, outside of EA world |
| RAND and GovAI style policy research | ||
| Policy work at or for an AI lab | ||
| Grassroots advocacy | Opinion making, lobbying and advocacy Leveraging social skills outside of EA world | |
| Lobbying in DC, Berlin, London, or Brussels | ||
| Targeted individual outreach | ||
| Podcasting, youtubing |
Appendix: Other considerations that go into thinking about my career
Here are other things that I’m thinking about for my career deliberations. I’m also still in the middle of figuring stuff out, so this is “The first part of my career deliberation seems maybe useful to others. I’ll also share the second half just in case” and not “Here is my complete career deliberation template that I found to work.” Note that I’m basically just listing considerations and possible approaches to take into account. The actual thinking about which ones are most important to you likely will need additional free-form space. I’d encourage you to share your approaches if you think it might be useful to others!
How do I want to approach choosing my (next) work?
| Options | Which broad category of work does this fit? |
| Follow my curiosity or excitement. Follow the path of least emotional resistance. Don’t hesitate spending large amounts of time (months) just to understand something better even if it is not entirely clear whether it is necessary or useful. | Research or otherwise being a thinker |
| Work on what others find useful. | Research or otherwise being a thinker, organising |
| Check and apply to open positions. | Research or otherwise being a thinker, organising, setting policy |
| Follow a systematic agenda. Ensure your work always has some path to impact. | Could be any type of work |
On the meta level, what is my priority for my next work?
| Options | Priority | Example activities | Synergies with types of work |
| Direct impact | [redacted] | Anti-synergy with empirical ML | |
| Skill-building and learning | [redacted] | MLAB | Setting policy, opinion making, some research |
| Exploration and fit testing | [redacted] | Try lobbying, talk to policy folk, learn about EU AI office, part-time podcasting | Setting policy, opinion making, some research |
| Credibility and networking | [redacted] | Publish work, do a graduate degree | Setting policy, opinion making, being at a lab |
How important are different properties of work to me?
| Property | If applicable: Preferred direction | Priority |
| Autonomy | [redacted] | |
| Guidance | [redacted] | |
| Feedback | [redacted] | |
| Free time | [redacted] | |
| Flexible work hours | [redacted] | |
| Stable income | [redacted] | |
| Time pressure | [redacted] | [redacted] |
| Sign certainty | [redacted] | |
| Impact magnitude certainty | [redacted] | |
| Focus on one project vs. many balls | [redacted] | [redacted] |
| Social interaction, peers | [redacted] | |
| Being relaxed and myself | [redacted] |
My personal career doc ends with a “Next steps” section that I’m not including. It’s a mix of talking to specific people and thinking for myself to resolve object-level uncertainties, uncertainties about what different kinds of work are like, and learning which heuristics for choosing work (steps) people I admire use.
The problem that we have with one proposed solution (adding a dummy utility function that highly disvalues a specific non-suffering thing) is that the resulting utility function is not reflectively stable.
So a theory of value formation and especially on achieving vNM coherence (or achieving whatever framework for rational preferences turns out to be the “correct” one) would be useful here. Then during the process of value formation humans can supervise decision points (i.e., in which direction to resolve the preference).