Over the last week or so, I have been trying to do something unreasonably difficult. I have been trying to answer the question: “If I was the head of the OpenAI foundation and handed 100 Billion Dollars, how would I use it for the maximum good?” The question is number 3 out of 4 on a set of questions in an essay contest conducted by Dwarkesh Patel, a podcaster who is probably one of the best voices covering the progress in the AI space out there. I say best because his range of guests is diverse, consisting of scientists, researchers, economists, historians (three of my favorite guests are Ada Palmer, Sarah Paine, and Steven Kotkin – all of them historians), but he also does serious preparation for each episode and spars intellectually with his guests asking really deep questions. I need to pause and take notes and Google to understand what’s going on, most times.

Anyway, the challenge is this: How do I come up with a 1000 word essay that does justice to the question, when this is a person who has spent hundreds of hours talking to the best people in the field and thinking about this question himself? He explicitly stated “Don’t let a lack of domain knowledge dissuade you from trying,” which is sweet of him, but trying to answer just this one question has required me to ask a bunch of hard questions and upgrade my knowledge on all of them. For example:

• Given some money, how do you actually put it to good use and track if the money is being used well or not?
• What does it mean to “do good” with AI?
• If I was to do good with AI, would I not need public support and governmental support as well? So part of the plan would need lobbying and political advocacy, as well as a strategy that shows people that AI is something that has potentially life-altering benefits that nobody could debate against (like curing cancer); and crucially, leading with these applications instead of putting them off for later.
• How do you answer people’s legitimate concerns about being made obsolete by AI in the labor market, and where would their political leverage come from now?
• Say I come up with a plan – What makes this plan foolproof, so that if the organization’s track changes or if I am removed from the team, the plan stays in motion out of the incentives being aligned with whoever is in charge?
• How do you deal with the biggest risks – AI becoming a tool in the hand of authoritarian governments, or AI going rogue due to an alignment issue?

And I knew nothing about any of these when I started. Obviously, each of these was a serious question that people had devoted their life to addressing, and they were working 70 hour days solving these problems. It was inane to suppose that I could neatly tie a bow around this in a week’s worth of work starting from scratch. And yet… I don’t think that a neat answer is what Dwarkesh is looking for. He claims that this is his means of recruiting a research assistant who can brainstorm these ideas with him. In that case, what he needs is somebody who learns fast and puts in the work.

But let’s come to the personal side of this. Working on this project has also taught me a lot of things that I didn’t know before. One idea that I learned tangentially, Richard Sutton’s “Bitter Lesson” framing of AI scaling, spawned an idea that is now driving me to focus more on the unreasonable effectiveness of brute force. I learned about the scaling hypothesis, I read Dario Amodei’s incredible essay named Machines of loving grace, and sat through half of Leopold Aschenbrenner’s fascinating talk with Dwarkesh. I wrote a scrappy first draft of a 1000 something words as a rough answer to start working from. But I think the most important thing is that it showed me really how hard this problem is even if I applied all my resources to it. It showed me the unexpected things that could disrupt my own schedule even if I tried to optimize for the sole task of completing this essay by sleeping at 11 pm every night and waking up at 6.30 sharp. A throat ache turned up out of nowhere and vanished just as suddenly. My father called up asking me for help with a business deal he was working on. One of my favorite people asked me to call after a gap of months. Turgenev’s “A sportsman’s sketches” seemed deliciously inviting, prompting me to rip through the last six stories in a single sitting even at the cost of missing the deadline for the essay the next day while I gave the excuse of a backache not letting me sit upright for long enough to read or type anything meaningful. The summer proved scorchingly hot. And despite all of this discomfort, I actually did more work than I expected to, because the last few months have been abysmal in terms of productivity and my belief in my own ability to create anything of value had plummeted to unanticipated lows.

Today was a realistic assessment of how much I could do. And I think pushing my comfort zone helped. Therapy had taught me to be kind on myself, but in the bargain, I think I had traded up for too much comfort and too little challenge, which was mind-numbingly boring and not doing any favors for my self-esteem. This spectrum, between doing something extremely challenging and giving myself the space to relax without any need to explain myself, is something that I can toggle to achieve a balance, I’m realizing. Right now, it looks like biasing for challenge might pay off a little more, but thinking in absolutes hasn’t helped me in the past, and this is a reminder that it is not a rule set in stone.