"To Work or to Learn"
Also: How to ask for things without feeling bad, are all X-Risk arguments fake, etc.
Deliberating: should I accept my SFCompute offer?
I finished my work trial at SFCompute and was given an offer! I had a bunch of inner conflict between accepting the offer v.s. living off my sizable savings & taking riskier bets though. The idea of "selling out" by working a "normal tech job" feels unacceptable to me.
Why's it unacceptable? I feel I've been exceptionally lucky with good parents, intelligence, great opportunities, etc. I very strongly feel like If I don't use my privileged position to do good for the world I'll be a bad person. I think many EAs feel like this.
It's an exceptionally strong feeling too. It was my primary drive ages 16-18. I'd be tired of studying math, go on our swing outside, look at the stars, and cry as I think about how they'll be gone if I don't learn enough to solve AI alignment. (An exceptionally unhealthy, cult-like mindset, and a story for another time.)
I've grown but similar feelings persist. I cried several times last week just thinking about the idea of selling out, and all the people I could've helped that I didn't. I'm a pretty scope sensitive person, I often cry just thinking about death statistics, or other abstract things.
Besides the feeling, I was also split between wanting to learn more deep technical stuff (especially stuff I'm weak in, like the sciences), and wanting to work directly on impactful projects. Many of my friends relate to this tearing them apart.
So I went into serious deliberation mode. I thought, went on many walks, wrote in a huge doc, talked to people I respect who would have good takes (Brennan Colberg and Davide Radaelli). and finally decided to take the job, for quite a few reasons (in rough order of importance):
On reflection I care more about developing discipline & integrity more than knowledge & technical skills right now. And I think those will be easier to develop in a stable environment. In fact, it's a good challenge to work a full-time job while maintaining side projects and learning. It's almost a forcing function to manage time better!
I can build skills a lot on the job too. I can get much better as a programmer & ML engineer, and it'll be incentivized by everyone around me (rewarding when you solve someone's problem!). SFC people will probably love reading my self-written performance reports too, encouraging me to be deliberate about improving!
My downside is bounded working a job. I can imagine very bad mental health spiraling if I was to strike out on my own right now (I was seriously depressed a bit over a month ago).
I learn about my preferences by working a job, and I can always quit after a few months worst-case. I've self-studied a bunch already and I know roughly what that's like.
Some other thoughts people may find interesting:
The learn drive & impact drives don't have to be opposed. There's a power law on what you could be working on, and choosing what to work on requires deep technical knowledge and understanding of the world which must be learned.
You can come up with a high-impact reason for learning something if you try. Having a single concrete example of how what you're learning could be useful to you is motivating and can inner-align you! Your impact-seeking part can be flexible, there are many promising options!
Davide's approach to this is really interesting. Simply (1) deeply understand a technical gap in society and (2) Create a plan to solve it. Read his progress on substack.
How to ask for things without feeling bad
My wonderful, beautiful, awesome, great girlfriend Laura Gao has debilitating anxiety around cold-emailing companies asking for a job. We tried to break it down by first writing an email to a fictional company SpaceY that didn't feel super icky to send.
Laura was having trouble so I wrote my own pitch for her to show how I'd go about it:
I love SpaceY's mission and would like to join if I can. I don't have much engineering experience but I'm a great software engineer and a fast learner. I worked at Zapata Computing at 16 and was able to contribute despite lacking experience in quantum algorithms research. I learn quickly when I'm passionate about something, and that's the case for space exploration!
I'd love a chance to prove myself, though I'd understand if you're looking for more experienced candidates.
Best wishes,
Laura
Laura loved it & said she probably wouldn't have written something this good in 2hrs! (nice to hear since she's usually much better writer than me!). So whats stopping her? She's definitely a good enough writer.
I think she's blocked because of mindset and conditioning. When you think of a job as a thing companies give and you ask for, then asking for a job feels like you're begging. Being unemployed is embarrassing, and asking for a job is cringe (like begging!). It's hard to send an email if you're thinking like this, everything feels bad.
When I'm writing "an ask" like this I put myself in the receivers shoes. I only want the job if It'd be good for them, I radiate goodwill, I want them to succeed, and I only want to offer them the option of having me on the team! I'd never want to join and be a burden. I think this mindset helps a lot!
You're essentially offering them an additional option for a (potentially) mutually beneficial trade!
Are all AI X-Risk Arguments Fake?
I met up with my mentor Alex Turner for the first time in >6mo! Had a great time, and heard an interesting argument against deception by powerful AI systems. One perspective I found interesting:
Counting arguments in AI safety ("most minds would ...") prove too much, because statistical learning theory makes the exact same argument for ML not generalizing ("most functions don't generalize properly"), and yet ML generalizes. Both arguments ignore inductive biases.
I'm recounting badly, once that post goes out you'll be able to see it here! I'm excited.
It's feeling more and more likely that the core AI x-risk arguments are subtly flawed and fake. I really really need to increase the quality of my AI takes & general wisdom. It'd be so valuable both for deciding what to do, and for elevating the level of discourse. At this point I feel bottlenecked by the wisdom of my AI takes. I feel blind.
I've felt the need to improve my AI takes for a long time (6-9mo?) but I haven't yet, the EA/Rat cult deprogramming had to interfered I think. I need to set aside time for this.
Finding a meditation community
Went to "deconstructing yourself" at the Berkeley Alembic and It was fantastic! Michael Taft was teaching there and he's soo great at guided mediations (the cadence, tone of voice, instructions, etc!), as well as having insane amounts of experience in many traditions, and writing a book called "mindfulness for geeks".
It was great! I'm excited to keep going and build myself up along with the community.
Location sharing app development
I started work on an app for location sharing in communities. The idea is for members of online communities to be able to find each other irl, in a privacy preserving way. E.g. providing error bars around people's locations, or only telling people there's someone nearby and not revealing locations till they both click "want to meet". I'm excited about this because I started two communities of ~200-300 people and the current coordination mechanisms ("@bayarea I'm visiting!") are hacky and I think won't scale with community size.
And I want to scale community size, because there are lonely Uli's in the world and there don't have to be!