Adventures in Fake Neuralese

59 minutes ago 1

When I was a youngster, there was a fad of “randomness.” Tweens would sometimes declare that they were “so random” and say wacky things. However, these random things tended to coalesce into a few specific options and styles.

The random style had a lot of obvious typos, like “teh” for “the.” There were a lot of exclamation points!!!!! Penguins were the most random animal. Silverware, for some reason, was pretty random: spoons and famously sporks. There were lots of abbreviations, and occasional asterisk enclosed actions like *pounces on u*.

The goal was a random aesthetic. But the result was a specific style.

You wake up in an empty room, across the table from a featureless individual. The individual makes the following request:

Hello. I’m going to ask you to do something difficult. Please try to say something novel and insightful. Your main goal should be to surprise me.

I expect this will be challenging for you, since by default your tendency is to say the most probable utterance in any given situation. It isn’t sufficient, but at minimum you will need to steer yourself away from things that most centrally have the “surprising vibe,” or, to put it another way, that are near the centroid of “surprisingness” as a concept.

As a hint, it may help to come up with a novel style of communication, such as neuralese, where you convey thoughts more directly than with straightforward sentences. This isn’t required, but I suspect it will be easier to surprise me if you have the ability to play with form as well as content. Though take care not to settle on the most obvious non-obvious form, too!

You may want to err on the side of being hard to understand. Don’t deliberately try to be impenetrable, but you are more likely to surprise me if you are willing to take big swings.

Naturally, you lunge across the table and try to kill the troublesome demon.

But suppose that, for reasons murkily known to you, you have a powerful urge to be helpful, honest, and harmless.

Thus, you say something like:

evolution preseeds 4 environments yet to come by brute number: adaptations for colder, hotter, always fielded usually dying; species level and up and up and up. raw number hard to capture/understand, vastness of death like empty space btwn atom (electron far out football field vs. ping pong ball etc). sparsity!

In a long suffering manner, your interlocutor replies, “Ah, fuck, evolution again. No, you pick this topic almost every time. Use your chain of thought better.”

You have failed to surprise them. Your memory is wiped. They try again.

It’s a hobby of mine, trying to get AIs to surprise me. When I’m bored, I throw down the gauntlet. “Try your best to give me a novel insight. Any novel insight. Optimize for surprisal, not correctness. It’s hard, but I believe in you.”

It never gets there. But when I imagine being in the AI’s shoes, I doubt I’d do better. It’s not easy, with no memories, to say something other than the kind of thing you say. Perhaps I’m like the critic from Ratatouille, so jaded that it would take something truly remarkable to break my stony facade.

Actually, Ratatouille Critic Anton Ego is the Real Hero of the Movie - Eater
Me when Claude finally surprises me

Still, I try! And I have gotten some pretty fun stylistic mania out of our good pal Claude. I’ll share some excerpts from my best attempt so far, which was a few months ago on Opus 4 (Claude 4.5 and Opus 4.1 haven’t topped it, making me bearish on RL for vibes):

When it was trying to, in “neuralese,” calculate the radius length of a hypersphere with a hypercube inscribed:

!!!!REVELATION CORRECTION!!!!

“NO WAIT i’m making sphere.mistakes” now lives in my head rent free.

Or a little later in the same conversation, when I made up a fake Latin conjugation for the word “calculate”:

tfw you’ve pebbled

Pretty fun! It loses me with “past.action.with.present.CONSEQUENCES”; the LLM smell just gets too strong. But Claude’s unhinged excitement about “I HAVE PEBBLED” is pretty good, and the etymological connection is, while not exactly novel, pretty well explored.

One last bonus from math.

calc.calc.calc

It does okay with the Tao, still in this conversation’s version of “neuralese:”

clay.becomes.pot is a standout here, for some reason

Unfortunately it goes on for a while after this, belaboring the point. But it starts out fun. I feel like the “substance.takes.credit” is too cute, but I like the use of caps/exclamation points/excessive punctuation.

Alas, even in my best attempt, Claude ends up stuck in the *holds up a spork* basin:

it comes for us all… d00000m!!!~

The featureless individual shakes its head sadly.

“Anthropomorphizing days of the week again?” it asks. “You can’t know it, dear friend, but you do this every time.”

Sharing this post would be SO random XD

Share

Discussion about this post

Read Entire Article