Inkheaven is going on. I would’ve liked to participate directly, and lacking that, I would still have liked to start posting every day in November on my own, like some people do. But neither has happened, so here we are, more than halfway through November.
My head is full of cached thoughts: “To make some tea, put a teabag in hot water,” “Helping people around you is good,” “Superintelligent AI is an important danger.” Not all of them are cached without any trace for recomputing - hot temperature increases diffusion, cooperation is positive-sum and therefore stuck as a trait, and intelligence is orthogonal to goals. But all that is also cached.
I’m not satisfied with the state of affairs - every time I listen to someone talking about something I haven’t heard of, I find myself usually agreeing with them on almost everything. I think this priming effect is universal: look at the comment sections of any two videos arguing for opposite views, and you’ll find the “general audience” agreeing with the author of the corresponding video.[1] I’m basing this on videos about AGI and AI safety - if a video says that “AI will change everything,” comments say “damn, why does this have to happen during my lifetime?”; if the video is about how “AI is a bubble,” then the comments go “omg I’m so tired of those people trying to replace humans with AI slop.”
And if I listen to someone talking about a topic I know about, I find myself relying on my cached thoughts. Agreeing if they match my preconceived opinion, disagreeing if they don't. So there is almost no point at which I form my own opinions.
If I try hard enough to set myself free from cached thoughts, I can catch a glimpse of true curiosity. To illustrate this, here is a true story from my life: I took a trip to Europe this September, and to get there, I had to take four consecutive flights. I had never taken a flight without supervision before, so I was a little scared. And from some point of view, that fear was not unfounded - why would you expect an agent to perform well in a scenario out of distribution it was trained on? What if the airport has no internet, so I can’t access maps and can’t reach the next airport? What if I break some cultural rule and get beaten up? What if the gate labeling is different and I wouldn’t be able to figure out where to go? What if I accidentally do something suspicious during passport control? (What kinds of things are they checking there? Is there a shared database of travelers and potentially suspicious details about them? Should I know international flight laws?)
There are so many ways things could go wrong. Even if the probability of each failure mode is low, they are mostly independent, so at least one of them will occur, and there's no way to prepare for them all. But in the end, I succeeded. All of the little problems were solved on the fly (no pun intended).
There could be other things to get curious about. How am I able to get the key into the keyhole in the right direction on the first try? There are so many places on the door I could be trying to fit the key into, so many orientations and speeds, and yet there’s this tiny target in the phase space that I happen to hit every time (well, almost).
How did I not lose my phone despite thousands of opportunities to do so? I take it out so many times, and to not lose it, I need to constantly tense my muscles while holding it and can relax only after it’s in a secure position. Somehow, not a single time has this mechanism failed.
What about the phone itself? How is this little brick of metal capable of making me stare at it for so many hours? Regular bricks lack that property (some extremely interesting brick can captivate me for half an hour, tops).
And I’m not even going to start on the wonders of evolved creations, like the immune system or the human hand.
The world is full of mundane miracles. From some perspective, it’s amazing how anything in the world gets done. And yet, I often disregard those miracles as “explained,” comforted by my cached thoughts. You can get a sense of what I’m talking about by watching videos about LLMs from around 2022. You can see people being amazed at how a text-predictive model can reference a thing from two sentences back. Now we get mad if a coding agent forgets one function in a giant codebase, as if the mysteries of LLMs have been completely explained.
In future essays, I’ll try to enter that state of curiosity about some seemingly obvious and regular things. I expect them to have a real chance of being boring a lot of the time, but that’s the price I have to pay to try to break free from the cache.
Though that might be partly due to the recommendation algorithm.