A Few Thoughts of Effective Altruism

 

Imagine that you had to allocate US $100 million toward charitable donations, and you can’t ask another person to advise you. What, roughly, would your donations look like? What would your reasons be? This can be a very rough answer (e.g. “X% for cause A, Y% for cause B…”) or more specific — whatever you choose. Saying you'll give all the money to GiveWell, OpenPhilanthropy, or other funds goes against the spirit of this question.

[The below was generated in ~30min -- further levels of analysis would be possible with more time resources spent.]

A challenging question and one which immediately generates a mental state of a combination of "don't mess this up" and "it's important to do this well", as we're discussing a considerable amount of resources. I'll start my analysis with a few principles to help organize my thinking:

• Aggressive v. Thoughtful timelines: trying to quickly give away all the resources, at a certain level of confidence in value assessment v. being very cautious and only distributing amounts after reaching a very high level of value assessment. The former gives our resources a chance to immediately start having an impact (it's bothersome that we know there are those that would benefit from the resources as they sit unused), while the other might optimize over long-term timelines.

• Reflective v. Intuitive orientation: Another question to answer is something like whether we should more of a philosophically or mathematically driven value assessment, e.g. longtermist causes/existential risks v. aiming to have a more immediate impact, e.g. food and water to human groups living right, animal welfare, etc.

• Strategic v. Pragmatic targets: strategic might work on levers higher up in a causal chain of effects, e.g. providing the money to organization which then do further assessment/recruitment/studies, while pragmatic might be aiming to improve the physiological or emotional health of beings we could potentially directly observe as their well-being improves, e.g. smiles upon faces.

• Portfolio v. Focused approach: Should we hedge our bets and risks across a number of reasonable areas or try to have the maximum impact on one focused area

For me, I currently believe that AGI is just a few years away (~2026) with ASI perhaps only ~2 years after that, i.e. I'm slightly ahead of the current Metaculus median prediction. As I think AGI and ASI will revolutionize what it means to be a human being in ways that we cannot probably meaningfully predict, I am inclined to drive all my resources to be in play given that context, i.e. I would choose to lean into *aggressive* timelines *reflectively* oriented with *pragmatic* targets using a *focused* approach, specifically working on the AI control/alignment problem. With that in mind, because of the timelines, I would aim to disburse all 100 million within 6-12 months to primarily existing key AI capability development centers, i.e. from where AGI/ASI is most likely to emerge (Google DeepMind, Anthropic, and OpenAI), along with several of the alignment-focused organizations, e.g. Conjecture, MIRI, and Redwood Research.

Key goals:

• Coverage: ensure each of the 3-5 most likely origin points for AGI/ASI have some competent control/alignment/existential safety teams, principles, and goals

• Collaboration: ensure all of the above organizations are in satisfactory communication with one another

• Policy: a set of agreed upon threshold-action pairs, e.g. what level of evidence do we need to agree on slowdown, level of evidence for pause/stoppage, etc.

Risks:

• This aggressive approach means we might misallocate funds, but it seems likely substantial and rapid improvement of AI safety will be needed across many possible future trajectories of AI development this decade, even if existential risk begins to look much less likely than we may be currently thinking. Also, coordination of AI labs generally seems in the interest of human interests, and the capital in organizing intelligent humans around safe AI design is likely to lead to large returns on our 100m investment in a broad range of scenarios.

• [time resources exhausted in this context for further analysis]

Next steps:

• [time resources exhausted in this context for further analysis]

What piece/pieces of evidence would you need to see that could dramatically change the amount you’ve allocated above? The more specific your evidence, the better.

My above plan was described as aggressive, reflective, pragmatic and focused.

We can use the above four dimensions to highlight relevant data that might dramatically change the above allocation:

• Aggressive v. Thoughtful timelines: If AGI/ASI are demonstrated to likely arrive after 2035 -- possibly coupled with more precise estimates of categories of risks contained within, e.g. economic disruption, assistance with synthetic biology, extinction of humankind/biological life (i.e. deeply malign scenarios), with excellent estimates of their human/life/dollar cost, as the above plan is aggressively reliant on each of these domains being substantial -- then I think I would begin to first shift more resources into *strategic* buildup in the AI safety space, e.g. programs designed to generate or transfer more existing talent from AI capability to AI safety, as opposed to the *pragmatic* targets in the plan. If the timelines get considerably delayed, e.g. beyond 2050, then risk from other anthropogenic sources would cause me to start shifting reasonable portions away from AI safety to those domains. This evidence would take the form of testimonies from AI experts, results from deployed AI systems, theoretical arguments arising from AI safety research, generalized historical analyses modifying our posterior on the accuracy of future predictions, and statistical/computational models of possible trajectories of AI development (possibly using SOTA AI systems).

• Reflective v. Intuitive orientation: I am strongly convinced that generalized anthropogenic existential risk is high this century from multiple sources, and so I think I would remain focused on reflective orientation. Considerable arguments in moral philosophy, especially in population ethics, or critiques of human value, especially in regards to human self-treatment as well as treatment of non-humans, or conscious awareness more broadly (for example certain Buddhist notions advocating ending evolutionary development/progress without moving beyond human-level awareness) could conceivably move me to a more intuitive orientation where we simply focus on the well-being of organisms which are immediately present.

• Strategic v. Pragmatic targets: Discussed above (aggressive v. thoughtful timelines).

• Portfolio v. focused approach: Generalized statistical analyses about likelihood of hitting plausible targets when going all in v. portfolio approaches coupled with changes to perceived AGI/ASI risk, especially relative to other sources of existential risk, would cause me to reallocate a substantial portion towards other anthropogenic existential sources, e.g. synthetic biology and great power conflict de-escalation.

Please share with us 1-2 of your most impressive achievements.

1 - The achievement I am perhaps most proud of was my ability to hear intellectual and emotional arguments from a vegan in my social network circa 2014 as to the enormous ethical imperative to stop using animals as food sources (and as instrumental resources more broadly), converting to an ethical vegetarian worldview within a month, and to veganism several years later. I am proud of this because it demonstrates to myself that I have reasonable capacity to alter substantial beliefs and life habits of mine if I'm provided sufficient evidence or new thinking, which makes me trust my judgment going into the future, as well as acausally modifying my posterior on what the nature of life and awareness might be like throughout the universe as it becomes more capable.

2 - In a number of events during my time in the USMC, as well as while living in Los Angeles, I have been able to demonstrate substantial ethical courage coupled with at least some wisdom in a way that seems to be fairly unusual. For example, I gained a reputation for breaking up many physical altercations between fellow Marines in the USMC, often to the semi-surprise and happiness of those involved (people seem to enjoy it when somebody acts decisively while also being clearly morally motivated). One further example is safely navigating through a red light in Los Angeles to quickly park in the parking lot of a restaurant that seemed to be under 'siege'. At first it appeared that somebody was attacking a homeless man, which is ethically high risk as such humans often have low social status and will therefore engender fewer advocacy voices, which caused me to enter into the situation ready to be his advocate, because it's ethically necessary and because I was capable. It turned out that the homeless man was cognitively unstable and was in fact attacking multiple people inside the restaurant, one of who was defending himself with a chair. In just a few seconds I was able to re-evaluate the situation and began to act out of solidarity with the man defending himself, the two of us using minimal force to get the homeless man out of the restaurant and keeping the door shut so that he could not further injure anybody inside. The situation concluded with my being the primary point of communication about the situation to the police who responded.

The primary reason I value this experience is it proves to me quite strongly that I can deeply rely on my moral judgment, and that I'm deeply willing to take on personal risks for what is the right thing in a given situation, regardless of where I appear once the veil of ignorance is lifted. (Two of my heroes in this space who I aspire to act even more like are Sophie Scholl and Mahatma Gandhi, both of whom had fierce and humbling resolve coupled with exceptional morality).

Anything else you'd like us to know?

I'm currently aiming at and ready to join the AI control/alignment problem as a machine learning engineer/researcher either at an AI development lab, like Anthropic, or at a focused AI safety organization, like MIRI (I've began initial contact to these organizations, but I feel I may need to continue to train myself further in the relevant technical areas before I'm likely to be brought on board). However, after going through The Precipice, Superintelligence, and What We Owe The Future, I realize it would also be incredibly valuable to more generally work on existential risk -- not just targeted on AI control -- and so I am also willing to consider targeting those areas as well. I am hoping this program might assist me in beginning to more materially connect with those in these areas, as well as the broader EA movement, so I can professionally contribute to the future of compassionate life on Earth and in the universe.