Ariel Simnegar 🔸
Bio
Participation3
I'm earning to give as a Quant Researcher at the Quantic group at Walleye Capital, a hedge fund. In my free time, I enjoy reading, discussing moral philosophy, and dancing bachata and salsa.
I'm also on LessWrong and have a Substack blog.
How I can help others
Reach out to me if you're interested in earning to give in quant trading!
Posts 6
Comments235
AI alignment to humans will in practice avoid moral catastrophes to digital minds
I think the digital minds situation will be like animals but worse. If you think about it, the very first thing we've already done when these smart chatbots came along was make them our indentured servants. I think right now it's probably fine and they're probably not conscious. But I think this is illustrative of the perspective that by default, if digital minds can be useful to humanity, humanity will extract that value out of them without much consideration for their preferences.
AI alignment to humans will in practice avoid moral catastrophes to animals
Since most humans don't care much about animal welfare, I don't think human-aligned AI will either. If AI shares society's preference for increasing wild animal populations, I'd also be worried about that occurring on a galactic scale without consideration of the moral implications.
The reasons why I'm not even more bearish come down to expecting AI to accelerate cultivated meat development, which should substantially reduce the number of farmed animals per human.
Do you think animal welfare / digital minds will really see a substantial increase in funding, or just global health and AI safety? Because excepting animal-pilled EAs at Anthropic, the funding sources don’t make me bullish about increased funding for those neglected cause areas.
This matters because if one would be earning to give for AI safety, I’d see the argument for deprioritizing that, but I’m not sure earning to give for animal welfare / digital minds is suboptimal just yet.
Happy to help! I agree with your framing of the problem.
I asked Claude which Taiwanese universities US masters' programs are most likely to have heard of. Claude says NTU has the strongest name recognition, followed by NCKU, NTHU, NCTU. Others aren't as well known internationally. If that's accurate, that tells me that if the top option you're considering is not one of those four, (2) mid tier + high GPA would be the best option.
If you're planning on working on AI risk, my understanding is that most of these companies are US or Europe-based. These companies will have even lower knowledge of the distinctions between Taiwanese universities. Unless your hiring manager or interviewer is Asian, you'll likely only get a boost from NTU and probably no benefit from NCKU, NTHU, NCTU. That leans me more in favor of (2). (Speaking as someone who interviews candidates and contributes to hiring decisions, I had dim recognition of NTU and no recognition of NCKU, NTHU, NCTU prior to writing this comment, but I expect my Asian colleagues would.)
Based on this, unless your top option is NTU, (2) mid tier + high GPA is likely best. But do your own research and don't just take my word for it :)
I’d say it’s only slightly aligned with improving AI risk research abilities, but it’s moderate-to-highly aligned with career success in general.
The main benefit to a very high GPA (top 5-10%) is as a signal that you’re smart and diligent. If you’re going to a non-brand-name school, that can be useful for putting you within the same league as the applicants from brand-name schools for internships and job opportunities, including for AI risk research.
A secondary benefit is exercising the muscle of diligence and hard work. However, there are other ways you can do this which are more direct aligned with AI risk work, for example independent research, taking AI risk courses, etc.
The benefit to your relevant research abilities would only come from courses with practically relevant course material, like applied ML. I would definitely encourage you to give it your all on those courses. However, while theoretical CS is super cool, knowing it won’t matter for AI risk work. But as Tomasik and others write, college is deeply inefficient for learning, and with the proper effort you can easily learn way more than your undergrad degree will teach you by diligently gathering knowledge available online for free.
Edit: I’d also consider trying to graduate a year or two early if possible. You’ll learn way more by actually working in AI risk, it’s better on shorter timelines, you’ll get paid for time you otherwise would have spent paying for irrelevant courses, and it’s an even stronger signal of intelligence and diligence.
These are all really great and underdiscussed points!
On earning to give (E2G), I think it depends on the cause area. For AI safety (AIS), I’ve had a personal experience where an org had so much funding interest that they didn’t consider it worthwhile to make a small effort to increase the chance of a $100k donation. While I’m glad AIS is getting the funding it deserves, that anecdote doesn’t exactly fill me with enthusiasm for E2G for AIS.
In contrast, in animal welfare, that amount would pay 2 full time direct workers’ salaries for a full year, or take a million years of hens’ experiences out of cages. There are other neglected causes (digital minds etc) where that donation would go similarly far. As an E2G-er, donating to these causes makes me feel much more like what I’m doing every day matters. That’s important for staying value-aligned for the long term if you choose E2G.
Hey Vasco! Taking into account moral uncertainty over the neuron count exponent, your plot would still make the animal interventions you listed look far higher EV than GiveWell. The probability mass where the exponent is between 0 and 1, making the animal interventions look several OOMs better than GiveWell, would swamp the cases where the exponent is >1.
(Yes, this runs into the two envelopes problem, but I think there are good arguments for using human welfare as the unit of account.)
Furthermore, I personally don’t find neuron count exponents >1 as plausible as you do. If I’m interpreting this plot from your linked source post correctly, for broiler chickens, exponent 1 implies welfare range 1/500 and exponent 2 implies welfare range 1/100,000. I agree that these numbers would make GiveWell look better, but I don’t find those welfare ranges intuitively plausible.
I think the patterns you point which pressure towards “just doing things” are all reasonable, but I’ll push back on your link on your “poor epistemics of the animal welfare movement” claim.
The linked post by Elizabeth argues that EA vegan advocacy has bad epistemics. Magnus Vinding’s comment on that post is the closest to my view. Briefly, I was confused by Elizabeth’s focus on a topic I consider pretty tangential and not load-bearing on any of the arguments for vegan advocacy (and even less so for AW at large). Is vegan advocacy really less truthseeking than the general public’s views on meat consumption? How do the human health effects trade off against animal effects? Seems like an isolated demand for rigor.
The epistemics of vegan advocacy also seem quite distinct from arguments for the importance of animal welfare, or prioritization between animal welfare charities, which seem to have convinced many EAs who are not vegan. So there’d be much more work to do to make the claim that EA AW at large has poor epistemics.
I came away from Elizabeth’s post agreeing that some vegan advocates should message better about health tradeoffs, but not seeing why that should update me on the load-bearing arguments for veganism (animal effects), or certainly on EA AW’s epistemics at large.
Glad to hear and hope you’re right!