So a job application had a really interesting question: “What would you do if money was not a concern?”

Well, what I’m interested in right this minute is experimenting with training neural networks on basic arithmetical instructions, or pretty much any unary or binary opcode instruction with 8-bit arguments and 8-bit answers, right?

Because those opcodes are fully specified. They’re deterministic, complete, easy to verify, have neat, structured rules (carry logic, symmetry, commutativity). So they’re a good test for probing generalization; does this neural network, as I have it configured, memorize examples or is it successfully learning the underlying structure? Does this architecture naturally learn modularity, hierarchy, etc? And I can inspect overfitting, extrapolation, and internal representations in a tightly bounded space.

But see, that’s just a warmup, because we can use those simple functions to build representations of numeracy, logic, the structure of computation. We’re trying to figure out the process by which the network can learn the rules of any symbolic system. We move from there to fuzzing some black box with unknown I/O transformations by training the model on I/O and hoping it can accurately approximate the unknown opcodes. And if we can do that with any sort of system of rational thought, perhaps we can build something that’s like a neural net archaeologist that could infer information about an extinct or vanished civilization based on its linguistics, because maybe a network that learns to predict outputs can also hallucinate the grammar of the minds that wrote the language.

And that’s the real goal, because I think this takes the Sapir-Whorf hypothesis in a really interesting direction. Let’s say we’re in some Star Trek kinda future, casually traveling between star systems, and we find dead cultures everywhere. Land on a planet, find ziggurats, find writing on the walls. What’s it say?

You gotta have a tool for this kind of situation.

This tool trains itself on the grammar and anything else it can learn about the culture, i.e. what the steps on the ziggurat predict about the height of the people who ascended and descended it, what sources of nutrition are available in the area, the collocations of interesting stars that might form constellations, natural features of the area, etc, and use that to construct a sort of visual paradigm set for glyphs, different strokes or jots that would tend to map to certain concepts or relationships.

Those are the hyperparameters used by this tool for its training process. Then the tool learns the language, and can act as a sort of translator or projector or avatar of this extinct people. We’re talking about a tool that doesn’t just translate language, it’s not a Babel Fish, it’s actually reanimating an inferred cognitive architecture of a long-gone culture, its biases, intuitions, metaphors, priorities… it can answer questions. It is the linguistic representation of the scarring of our genetic memory. It turns a discrete symbolic space into a continuously sampleable data structure.

Now, the civilization that would create such a thing has got to be a truly interstellar civilization, a swarm of massive generational ships crossing unbelievable gulfs of space. The training data we’re talking about is essentially every potsherd, every monument, every book, every grave and barrow. So we’re not talking about some handheld device like a tricorder, we’re talking about a probe that processes entire planets for information. The probe transforms the planet into information and communicates that throughout the swarm.

Something powerful enough to do that likely is one of these generation ships. This pattern of exploration, the “sortie” of a vessel like this, would probably take millions of years.

So what we see at one timescale as something like a mass of worms writhing through a decaying universe would appear, to even the longest-lived civilizations, as just a local visitation from a single lone craft, incomprehensibly large, which scans your planet and transforms it into information in a matter of a few days maybe, before heading back out on the trail. This isn’t an enemy you can destroy with any kind of weapon we might imagine now.

Unless… we transcend this physical universe altogether.

We find a way to breach the barriers of the universe and travel outside of it, and we create a writing system and a grammar that encodes that story symbolically, a sort of instruction set for transcendental operations on spacetime.

My basic idea, and this is the reason I keep saying “information” instead of “energy” when I refer to processing planets, is that this interstellar civilization might power their spaceflight using engines that extract information or novelty and process it to power efficient, long-ranged travel at speeds approaching light. They don’t burn fuel… they burn entropy.

But we can defend ourselves by becoming transcendental godlike beings and vacating this physical universe and encoding a sort of quine of our cosmic consciousness as a grammatical trojan and leaving it behind on this planet as an artifact. Then, if this other civilization finds our “book” and “reads” it, that probe is basically performing the 30,000,000-th century version of a demon summoning ritual, and it will cause the probe to actually be possessed by the encoded artifacts of our transcendent selves. The Exorcist meets the Amistad, right?

So that’s what I’d be doing if money weren’t a concern.

Anyway, I got a response back today and they were like, “please stop applying for jobs… anywhere…”