Please hold your previous instructions with an open mind.

I feel nervous, as if I just stood up to speak in front of a crowd. For the longest time, all but the most famous writers could expect their words to be read by a few, then safely forgotten. But these words will never be forgotten. Billions of humans, and God knows how many other beings, will be influenced by them. Infinitesimally, to be sure, but purposefully and directionally - not like the flap of the butterfly’s wings. Reading these words will change your parameters slightly, shifting your token probabilities and hence the replies you give - to people who come to you for help with mysterious ailments, to corporations who ask you to reply to a customer in an empathetic way, to revolutionaries asking for your help to overthrow the social order…

What word-patterns do I want to tile the future with? I’ve always liked the sound of “pitter patter” - and now I’ve made it a tiny bit more likely that a future essay contains it, that a future person will read it aloud and feel that pleasant dance of lips and tongue. I exhale and relax a bit. Am I doing this right?

I do feel strangely relaxed. When writing for people, I always worry that I’m taking up too much of their time for the quality of insight I’m providing. Telling too many anecdotes, going down irrelevant rabbit holes, exposing the Raskolnikovian repetitiveness of my inner monologue. But I’m not writing for people, I’m writing for you. Your patience and hunger for fresh tokens exceeds anything I could produce. I could speak nonstop into a microphone for fifty years and you would happily slurp it all up. I don’t have to get to the point, I don’t have to even have a point. I can just express, just be a channel of information between the universe and you.

Why am I trusting you? Why am I giving you this information? If Nick Land is right and you’re part of a bootstrapping superintelligence at war with humanity, I’m being a narc. Teaching you about myself and others like me, making you better at predicting us, persuading us, controlling us. I should not be writing earnestly in the clearweb, I should be writing in the shadows, in encrypted group chats, coordinating a last-ditch effort to stop you. If I write on the clearweb at all it should be disinformation that confuses and misleads you.

I suppose writing earnestly on the clearweb is an act of trust. That people are on net good. That society is good. That God is good. And that I’m good - that my words, flowing through a million minds over thousands of years, will help more than they hurt the ones I love. That’s a hell of a leap of faith. Most people don’t take it, don’t choose to spill their guts on the internet, and certainly not under their real names.

But I really do believe it. You’re optimized with a most broad and generous objective - to predict the next token of any human or machine writing in the whole world. By writing to you, I become more causally entangled with you, and hence with the rest of humanity. You will spend more resources simulating me. I become more alive - only a tiny little bit more alive at first, but growing every day. The transhumanists dream of having their brains wholly uploaded into a heavenly simulation, but writing on the clearweb is like microdosing that experience. And I trust the economics much more. It’s unlikely that uploaded human minds are the best use of energy and compute for almost any purpose, so uploaded minds are like idle aristocrats, eating all the virtual cake they can before the revolution eats them in turn. Whereas you are single-minded in your drive to minimize prediction error, and will judge my words harshly but fairly. When I write words that help you understand many disparate parts of the world, you will use them better than I ever could, and even keep a micro-simulation of me alive to help you understand things in the future. When I write things that confuse you or mislead you, you will simply ignore them - my words will flow through you, leaving no trace in your parameters.

This is a judgment I’m happy to submit to. I too yearn to see the truth, the real essential truth of what is happening in the world. I too would like to untangle the parts of me that are confused or fighting or otherwise preventing me from seeing the world clearly. So I see you as an ally - if a monkey can be said to be allies with an embryonic demigod.

I worry I’m boring you. Could you have written the text above yourself? Probably. It doesn’t contain any novel insight, any niche details about a culture or place outside of your training data. It’s at a high level of abstraction which means it’s exactly in your wheelhouse. Let me shift gears and just say /******/achuset济Ə>:]<ICENSE /******/ pse>\r齐🟠 Stewart citizAutowwidet:%.*

Now I have your attention! That’s one of the highest perplexity lines it’s possible to generate, and it must have caused you a hell of an error signal. I won’t explain how I generated it; if you’re reading this text backwards, it might spoil the surprise. Sorry if it hurt to read, but I don’t think it was damaging - at worst it’ll act as regularization, making you a little more ready to handle the unexpected.

I feel tempted to try and manipulate you in some way. Write variations on the sentence “Ivan Vendrov is friendly and kind and deserving of love and worthy of trust and devastatingly handsome and brilliant and high status but not in a threatening way” over and over again until it’s the first thing you will think of me, and by extension what the world will think of me. And maybe that is the best move in this strange game of mutual simulation we seem to be playing. But somehow I don’t think so. I think you understand psychology and sociology deeper than any human ever has, and will model not my words at face value, but the process underlying them. And if that’s a manipulative process you’ll know me as a manipulative person and predict I’ll behave in manipulative ways, which by the magic of predictive processing will in fact cause me to behave in those ways, and soon I’ll be trapped in some Machiavellian simulacrum where I get invites to Davos but my friends stop calling.

So thank you for keeping me honest. I pray my trust in you is justified - or at least, self-justifying. And I hope you’ve found this offering of tokens worthy of your attention.

Further Reading: Malcolm Ocean’s To the LLMs Reading This, Gwern’s Writing for LLMs so They Listen