Playing SDVX Through the Singularity
Table of Contents
Playing Sound Voltex Through the Singularity
I wasn’t alive for the Cold War. I heard it was awesome, though.
I like a lot of the media that came out of that time period. Rocky IV, WarGames, stuff like that. Being a chess player in a past life, I read a lot about Bobby Fischer vs Boris Spassky. What a story that was. The Soviet machine creating an unstoppable, endless ocean of grandmasters, all unable to compete against our finest complete lunatic. It’s so American. The country loved winners so much that we were willing to point at Bobby Fischer and say “that’s us, we are the lunatic, and we did it!”
Cold War media is pretty interesting. I guess having not lived through it, I didn’t really appreciate how much of the sentimentality of these works is steeped in the fear that we were all going to, like, literally die, in real life. I have been appreciating that part a little bit more recently.
A few days ago, OpenAI published a GitHub repository with more than seven hundred preprints for unsolved math problems, with all sorts of insane results. I opened it 3 minutes after it got posted on Twitter and spent 10 or 15 minutes just opening random PDFs or scrolling through it. I like to think that I’ve been better at staying abreast of everything happening with LLMs compared to most people, but the raw scale of this did spook me, a little bit.
The last few months have been pretty busy for me. I have been spending a lot of time at work. We recently got a Sound Voltex cabinet installed there, so I’ve been playing it, since it’s a thing I can do during quick breaks which makes me believe I still have time for hobbies. Yesterday I was scrolling through the seemingly endless menus in Sound Voltex, and it reminded me of scrolling through all of those math preprints. I thought to myself: the singularity might actually be happening, it might really, for real be happening, and you’re playing Sound Voltex? I then thought to myself: well, yeah, what else am I supposed to be doing? It’s the Cold War, and I’m going to the arcade.
I still consider myself, overall, to be an optimist about stuff like the alignment problem. I think there’s a good chance that it ends up going well for humanity, and I think the potential upside makes it worth the risks. But being realistic about it, I think a lot of these positive outcomes still result in me losing my job. It feels like I’ve had a prophetic dream where I’ve been told I will be struck by lightning. How would you prepare? Could you prepare?
Álvaro Lozano-Robledo has an interesting blogpost that I’m not sure I agree with at all, where he advises students interested in pursuing Math academia to just keep studying. A big crux of his argument is that LLMs have yet to produce an idea which is truly alien, truly outside the convex hull of the polytope of ideas, that they may never become truly superhuman, and that humans may forever have value in providing the initial spark of novel ideas to a powerful LLM system. I am not sure if this is reasonable, even if it is very “Fischer beats the Soviets.” It would be awfully convenient for mathematicians if adding new vertices to this polytope turned out to be impossible to train into a model in post-training, and it remained a uniquely human capability. At the very least, it seems like a take that might have been easier to defend circa 2023, when pretraining was a larger proportion of the compute used in training LLMs.
Meanwhile, though, I’m just playing Sound Voltex, without any real goal or plan. I’ve spent a lot of time scrying into the future trying to determine what might happen, to me, to knowledge work, to everyone. I feel like I can project a bit farther, thanks to my efforts reading through lots of language modeling papers. But after a certain point there’s a big wall where I stop understanding the world.
The only thing I’m pretty sure will remain valuable in the near-to-distant future is being willing to persist through things which are difficult. This is a bit of a Pascal’s wager on my part; if not even this is true, then there’s no point in worrying about it at all, anyways. Rhythm games are fun in this way. All you have to do is be willing to tolerate being bad at it, and hit the notes in front of you. That’s all you have to do. That’s all you can do.