My personal resolution for current and incoming AI doomers
Games.
Honestly, this has always been the answer I tell myself to how to succeed in a noisy, unjust, and pretty dang random world: constrain the environment. I think I, along with many others, had this answer in my mind (but just didn't think of it and frame it in this way at the time) when we were children. This doesn't need an impending AI doom to figure out.
Personally, I think this take I’m about to go into particularly resonates with younger me. I always had some childhood dream growing up; ambition was generally welcomed then. This take is well aligned with supporting even to this day these kinds of childhood dreams of mine. The NBA dream is one I’ll be using as a recurring example in this post. And you’ll also find that this take lends itself less so to one’s intrinsic background and moreso to empirical results - something an unlicensed, unprofessional kid saw as a pipeline to legitimacy.
Lebron James is universally the greatest NBA player of all time. There will be no AI Lebron James. Why? Because we constrain the game (basketball) to only allow non-performance-enhancing-drug-induced, rulebook abiding humans. Because the purpose of the game is to display the true pinnacle(s) of human skill and athleticism. AI != Human. Ergo no AI Lebron. QED.
Now, there's some nitty grittys to definitions like these. What is a game? A test with constraints. What defines a good game? One that I find worth pursuing. What's a "test with constraints"? If we were to model it then as a constraint optimization problem, then there is no limit to what we can do with our "training" data set right? Lebron could hypothetically train with AI, or take performance enhancing drugs during training if it meant he could do better, as long as he didn't do it in-game. In this sense a game is inherently a uhhhh black-box? results-oriented? way of approaching optimization. But I measure games based on their intrinsic structure, breaking this black-box kind of assumption. I instead make the assumption that, despite being a black box function of your training input to your test output, there ultimately EXISTS signal within this black box; that taking performance enhancing drugs during training is just (at least in all the reasonable scenarios I can think of) a bad way to optimize for game-time performance. That there is a systematic method to git gud; to better your performance through the acquirement of some somewhat-deterministic skill. Alternative framing: training is a part of the test; you form your own personal constraints on your "training." I mean, these are mostly personal, but they're constraints nonetheless.
Now back to that question of what defines a good game; there are rules to this constraining of the environment. I don't think playing video games in your basement 14 hours a day is a pinnacle of anything worth pursuing, though of course, this is ultimately a matter of personal taste. In my opinion, a good game to should follow some necessary criteria: like, for one, it has some sort of structured, at-least-semi-deterministically-based-on-your personal-decisions metric of performance. So gambling on roulette in this case would NOT be a good game here, at least in the usual game model, as in such case it is entirely random and the only decision you have is when to stop, which is not much for deterministic-control. I believe there should also be environmental factors at play when choosing your criterion to some extent - money, prestige, approval. NBA players make millions to billions. Being the best instagram reel scroller makes you unemployed.
There are common critiques too, of course. Here's three related ones I hear: 1) Some people view this as sort of burying your head in the sand; a burying of the mind from the changing world; to ignore current problems (e.g., AI) in favor of some vapid sense of tradition. But I don't think this is necessarily true. Upon my good game criterion there is also another criteria which I make: it should be somewhat interconnected with skills needed to tackle current problems. This seems like a rather abstract and arbitrary criteria when I frame it like that, but I find from an empirical POV that there's actually a lot of games which do this. Good performance in math olys, for example, have very strong correlations with becoming leading researchers in AI safety and... well math too. By choosing the right games, you can implicitly gain the skills required to face the current events of today. 2) Closely related - some people view these games as meaningless. I mean, if it doesn't push society forward, what meaning do these little games and puzzles actually have? Well, I frame this as simply a key difference in how we set up our game criterion; I set my criterion more towards fun, structure, and deterministic/fair-ness - you set yours towards "moving society forward." This is just an implicit criterion that you've set - although I'd question you now if this is truly a criterion you want to direct your decisions toward given AI's rate of progress. Plus, a similar reason from (1) applies here; even though doing these, say, math oly "puzzles" don't move society forward, you gain logical and problem-solving skills which are very transferrable towards actual "societal progress", or as some people would interpret it - "proper work." Although the modern day interpretation of work, along with it's many implicit criterion, is a blog post in and of itself... for a later date. 3) People see this as grindy. But honestly, I personally feel like this is better than feeling undirected. Maybe it was the way I was raised, but I feel much more comfortable knowing I have a direction than floating through life, at like any point in my current and past. I feel much more alive chasing what I'm passionate about. Of course there's balance, but I'll be the one to decide where to draw the line there.