This is the current Metaculus forecast for AGI:

 

Do decisions by EAs working in other cause areas mostly ignore this?

 

For example, are timelines for transformative AGI taken into account in work such as:

  1. Deciding whether to donate money later vs now
  2. Estimating existential risk from climate change, nuclear war and engineered pandemics past 2043 / 2068 / 2199
  3. Deciding whether to pursue "move fast, break things" approaches vs lower downside risk, slower,  reform approaches

 

And additionally, should timelines for transformative AI be taken into account in this work?

20

0
0

Reactions

0
0
New Answer
New Comment

6 Answers sorted by

In my day-to-day work for AMF[1], AGI timelines don't matter. My work is about making bednet distributions more efficient and transparent, and AGIs simply don't help with that yet.

Of course, I'm open to suggestions regarding how my work should be influenced by AGI timelines (apart from the obvious "stop it and work on AI safety instead").

There are other technological developments that I think will affect AMF's work, for example vaccines, gene drives, and new types of insecticides. We follow these closely. AGI itself will certainly affect bednet distributions too, once it arrives. Until then, the right thing to do IMO is to continue working hard to help as many people as possible in their fight against malaria today.


  1. I'm speaking for myself here and not on behalf of AMF. ↩︎

I recently wrote about how  AGI timelines change the relative value of 'slow' acting neartermist interventions relative to 'fast' acting neartermist interventions.

It seems to me that EAs in other cause areas mostly ignore this, though I haven't looked into this too hard. 

My (very rough) understanding of Open Philanthropy's worldview diversification approach is that  the Global Health and Wellbeing focus area team operates on both (potentially) different values and  epistemic approaches to the Longtermism focus area team. The epistemic approach of the former seems more reliant on more "common sense" ways to do good.

Judging by the ratio, I'd like to clarify that I'm not disagreeing at all.

I'm saying that

  1. For the entire 1950-1990 period, if we follow present-day trends of AI risk gaining prominence within EA relative to global health, the nuclear risk faction could have made a credible case that it deserved the most priority by far. Humanity was just incredibly close to nuclear war at multiple points.
  2. However, crowding out other causes meant that things like global health would have been neglected for decades, even though the cause landscape remained roughly the same since no nuclear war occurred. There are projects like the Global Seed Vault that do combine ideas within nuclear risk and global health, but the actual neartermist global health faction would overall be drowned out.
  3. Even if EA research had concluded a lack of tractability, that doesn't mean they can steer discourse and resources back towards other causes that are tractable. If nuclear war was a cause area that received way more attention and funding in the movement, it will induce bias.

I'm not opposing investigations of inter-causal impacts, I'm cautioning that such a focus might overestimate the X-Risk impacts, because people catastrophise on highly uncertain scary risks and develop tunnel vision. You could make the opposite case, too. I think most EAs heavily underestimated COVID's impact on other cause areas, considering COVID was never an existential-level pandemic threat.

It's definitely a blindspot, since I generally don't see overlap between highly engaged EAs doing both AI risk work and global health work/animal welfare work. At this early stage of the movement, I'm seeing most of work done by early-mid career professionals who have prioritised building their expertise in applying EA to one field to yield exponential impact.

That said, some caveats:

  • "What AGI does" is not common consensus to begin with. Yes, there are compelling and influential theories, but an actionable, quantitative prediction of exactly how AGI will affect other causes is quite far off.
  • Existing AI risk organisations are mostly focusing on the technical details of AI, which reflects in their hiring. To actionably assess socioeconomic impacts would require hiring huge and well-resourced teams of economists, doctors, policy analysts that are specialised in both AI and policy topics. I'm not aware of a major hiring push in this direction, at least in EA. Some governments like Singapore and China are actively studying the societal impacts of AI.
  • I sometimes tell fellow EAs that cause tunnel vision can hinder progress. If the EA movement had been around in 1950, it would have spent 40 years dominated by the "nuclear war risk" faction, with dubious impact at the expense of research on other causes. After all, why would developing world poverty matter in a nuclear winter, which has such a high probability of occurring?

However, I get what you're saying. Overall, I do think cross-cause collaboration would yield some benefit, even if just in a speculative wargaming capacity. COVID-19 has showed that one existential risk coming to fruition, even on a limited scale, can greatly alter trajectories elsewhere.

The answer is implicitly yes, they do ignore AGI timelines because dealing with the Singularity is both model-wrecking, and crucially, may not happen.

And it's a good thing that they ignore AGI timelines, as they are hedge organizations in case the Singularity or X-risk scenario does not in fact occur, and they need to optimize for a world in which boring normal futures happen.

My impression is that AGI timelines are largely ignored in the substance of the work itself, in part because it is hard to know how exactly to incorporate these estimates. I work in EA community building and even in this role it is unclear how I should factor these estimates into my decision-making. I know what questions it raises and I know how the answers to those questions would inform my strategy, but the questions remain hard to answer based on the estimates alone. I would like to see much more discussion of the practical implications of AGI timelines.