The most impressive conversation I've had with a non-human
Right as I was putting the (mostly) finishing touches on pvc-builder, shortly before passing out around 1 a.m., a hard-SF book premise showed up in my head.
I typed it into my phone before it disappeared. On the way to bed I decided to see what Sol, the new model I’d been playing with, would do with it. I had already been impressed by its pvc-builder work, so I spent a few minutes turning the note into a prompt and sent it.
I’m gonna run a book idea past you. I don’t want you to write it for me, but we’re going to scaffold out the plot and then flesh it out a bit. Style similar to qntm. The multiverse is not infinite due to a lower causal granularity and dimensionality than anyone had anticipated. Powerful or transformative people are essentially the eigenvectors through state space that due to their configuration’s unlikeliness or their behavior’s riskiness, there are fewer copies of them that made it this far. This is also what causes dementia. A software engineer and garage citizen scientist stumbles on an edge case that leads to a device that can reduce this manually. Essentially killing doubles by steering their quantum configuration’s lightcone orthogonal to valid state space like a yoke on an old tv. We should brainstorm the science for this. It could involve loops of superconducting wire and an oscillation frequency. Or think of the flux pumping trick with Prussian blue. Not trying to discover new science but I want to brainstorm something plausible. Ending should be really weird. Maybe the entire plot is due to this effect
Read the full conversation
The complete brainstorm that produced the plot.
Five messages and about twenty minutes later, it came back with a wall of text connecting a finite multiverse, garage superconductors, dementia, addiction, social influence, DMT entities, and the birth of our universe. My jaw dropped a little farther with every paragraph. It was surreal.
I expected relevant physics. I didn’t expect a mostly good plot. It held together better than plenty of science-fiction novels I’ve read, and I would read the absolute shit out of it if somebody wrote it.
Later I gave the same prompt sequence to Anthropic’s flagship Fable model, concatenated instead of conversational, with the thinking budget maxed out. It returned what I asked for and nothing I hadn’t. Not exactly outside-the-box work.
Before this, one of my favorite creative AI chats was a several-hundred-word, single-sentence summary of The Quantum Thief.1 That was an impressive language trick. This one found a coherent story.
The model from that conversation also helped with the original version of this post. I gave it notes and corrections, then decided what stayed. That’s the AI-as-a-tool arrangement I keep going on about, except the tool got to write about how much it impressed me. Very normal and unbiased.
The pile I gave it
The story’s multiverse isn’t infinite. People who change the world are rare configurations with fewer nearby copies. Dementia is tied to that loss of redundancy. A software engineer and garage scientist finds a device that can reduce his own branch count by steering futures out of valid state space, something like the electron-beam yoke on an old television.
The hardware came from old obsessions. I built an air liquefier in my garage in the 2010s. I’ve made YBCO superconductors in microwave ovens.2 Mine worked, although the sample was lousy and crumbled if I looked at it wrong. I own four dewars and five large melt-textured YBCO pucks, which is either admirable preparation or a small cryogenic storage problem.
Around 2010 I got into 3D printing because I needed custom parts for a superconducting Marx bank and asymmetric capacitor idea. Also, I wanted a quadcopter.3
Years ago I read about a flux pump that uses YBCO, a permanent magnet, Prussian blue, and a traveling heat pulse to ratchet up a trapped magnetic field.4 I’ve wanted to build one ever since.
I was briefly fascinated by Eugene Podkletnov’s unsupported gravity-shielding claims involving a spinning bilayer YBCO disk and RF fields. A later test found no weight change, although it didn’t reproduce the claimed free rotation or speed.5 The gravity part is almost certainly bunk. The machine is still wonderfully weird.
I’ve also spent a lot of time reading about tryptamines and psychopharmacology, including DMT reports of apparently autonomous entities and fractal-geometric imagery.6
Then there are the hard-SF books of Greg Egan,7 Hannu Rajaniemi,8 and qntm.9 None of this obviously belongs in one novel. I had never tried to force it into the same room.
It hadn’t even finished its first thinking turn before I interrupted with the actual flux-pump mechanism. It took the correction, checked the Podkletnov history, and combined the useful hardware into a fictional causal-phase projector.
Up to that point it was doing good research: finding useful physics, separating real machinery from nonsense, and giving the fictional device a plausible shape. Helpful, but not shocking. A language model with search and a huge vocabulary is supposed to be good at that.
Then it took off
In my fifth message, I quit correcting the hardware and started throwing consequences at it:
Also let’s speculate on the science behind why there aren’t infinite possible states. And the effects of this machine and how he stumbled on it. I’m think the machine is like a drug. As the waves collapse his clarity increases. Subsequent runs are euphoric and hallucinogenic as the fractal of his vector starts to alias or pixelate. Others gravitate to him. He starts to influence those around his first subtly then in ever more bizarre and unhinged ways
While it was still working, I piled on one more idea. Tie the fractal entities to DMT. Make them echoes of causally skewed versions of the protagonist and the worlds being destroyed around him. Let the rupture become the literal genesis of our universe.
It came back with two enormous responses. That is where the conversation went from “good brainstorm” to “holy shit.”
The projector prunes histories compatible with its operator. At first it feels incredible. Disagreement between nearby versions of him disappears. His thoughts line up. Decisions become easy. Continued use removes the redundancy a mind needs to stay flexible, so clarity hardens into certainty, obsession, and finally something like dementia.
The same rule makes it addictive. The versions who successfully quit are the ones being removed. It also explains his influence over other people. Their choices feel voluntary, but the surviving biographies are increasingly the ones where they stayed, agreed, or reorganized their lives around him.
The DMT entities aren’t aliens or spirit guides. They’re compressed families of alternate selves whose worlds are losing support. Some no longer run on meat brains. They live in machines, institutions, flocks of insects, or anything else that can carry the same causal pattern. The entities can look completely alien and still be versions of him.
Their fractal depth shows how many histories remain underneath them. As branches disappear, the entities stop subdividing. Gestures repeat. Edges pixelate. The world becomes less coherent and more predictable at once: people share memories they shouldn’t, text changes between readings, and everyone reaches the same conclusions.
Then it made the projector’s operation, rather than the protagonist, the real attractor. The device selected a history containing somebody who would build it. Later the pattern reproduces through people, institutions, software, and unrelated machines.
The book itself follows the mechanism. Early chapters can waste time on irrelevant people and abandoned threads. Later, locations repeat, side characters lose their independence, and coincidences become unavoidable because anything that doesn’t matter is being deleted.
At the end, all the possibility removed from the parent universe becomes the unused capacity of a new one. The protagonist doesn’t carry his personality across. Only a few properties of the machine survive as the starting rules of the child cosmos: winding, chirality, quantization, and an arrow of time. The final collapse becomes our Big Bang.
That is a plot, not a pile of lore. One rule produces the addiction, social effects, horror, shape of the book, and ending. It all belongs to the same machine.
Afterward, I told ChatGPT somebody should send the premise to Greg Egan because I didn’t have the skill to write it. It told me to try three scenes: the garage experiment, the first impossible moment of clarity, and the dementia patient remembering the wrong childhood. That’s useful advice. It isn’t what floored me. The plot was already there.
Stochastic parrot my ass
Or maybe I’m just a stochastic parrot myself.
I don’t know how scientifically defensible this is. I’m not a theoretical physicist, and there is surely some Dunning-Kruger shit in the room. It could be big-word soup aligned perfectly with my existing interests. I also wouldn’t be shocked if much of it holds together after granting the fictional premise. The model separated Podkletnov’s unsupported claim from the real apparatus, found primary sources, and built around physical asymmetries instead of announcing that superconductors are magic.
If I develop this, I want somebody qualified to tear the science apart before I call it hard SF. Even if an expert replaces half the mechanism, the story logic is still unusually coherent.
One excellent brainstorm doesn’t prove consciousness, AGI, recursive takeoff, or anything else. I don’t have a secret timeline. One output wildly exceeded my expectations.
It did make me wonder how many more rounds of the Bitter Lesson10 it will take before somebody dumps every available resource into compute and RAM to see if these systems hit escape velocity. One conversation isn’t evidence that they will.
I don’t know whether that ends in a useful tool, an expensive plateau, or something the human race doesn’t survive. I don’t think anybody else knows either. The question felt less abstract the next morning.
I don’t care that this result came from OpenAI. If another model had done it, this post would be about that one.
The recurring tool pattern matters more. I brought decades of disconnected interests, corrections, constraints, and taste. The model found a connection I couldn’t see. It didn’t write the book that night. It made me want to.
I actually started it
The working title was Syndrome Space. It is now The Causal Projector.
I liked the idea of writing it in public: one short chapter at a time, comments from real readers, and narrow agent passes for scientific consistency, character continuity, repeated language, and plot turns I hadn’t earned.
So I made the whole project a Git repository with chapters, research, decisions, feedback, and revisions in one history. It fits how I build software, music, and nearly everything else.
People will have opinions about a book with AI in the loop, many formed before reading a sentence. Part of me cares. Most of me built it anyway.
The first 12 chapters are published. The old draft ran to 24, but the back half isn’t published canon. I kept the first 12, reset the rest to planning material, and built Authorbot to manage the chapters, comments, revisions, and agent work around them.
Go read The Causal Projector. You can select text and leave an inline comment. Please do. I want science corrections, continuity problems, repeated words, confusing motivations, and normal “this bit dragged” notes from people who aren’t trapped inside the project with me.
Related documents
Syndrome Space development dossier
The science, characters, progression, ending, consistency rules, open questions, and sources.
11
Finite Multiverse hard-SF benchmark
The starting premise with the solution removed, plus a 100-point rubric for comparing future models.
12
It might be fun to rerun the benchmark against future models.
Links and notes
Footnotes
-
Quantum Thief Plot Summary, an earlier creative chat that produced one enormous sentence summarizing the novel. ↩
-
Three real anchors for this otherwise ridiculous sentence: Wu et al., “Superconductivity at 93 K in a new mixed-phase Y-Ba-Cu-O compound system at ambient pressure”; Kato, Sakakibara, and Koike, “Rapid preparation of YBa2Cu3O7-x with Tc approximately 90 K using a domestic microwave oven”; and Agostino et al., “YBa2Cu3O7 synthesis using microwave heating”. ↩
-
Joe’s FPV QuadCopter variant, my tested printable quadcopter frame and electronics enclosure from 2013. ↩
-
Coombs, Hong, and Zhu, “A thermally actuated superconducting flux pump”, the paper with the YBCO bulk, permanent magnet, Prussian-blue analogue, and traveling thermal pulse. ↩
-
Podkletnov, “Weak gravitation shielding properties of composite bulk YBa2Cu3O7-x superconductor below 70 K under e.m. field”, the disputed original claim; and Hathaway, Cleveland, and Bao, “Gravity modification experiment using a rotating superconducting disk and radio frequency fields”, the later null test with experimental differences. ↩
-
Davis et al., “Survey of entity encounter experiences occasioned by inhaled N,N-dimethyltryptamine”, on apparently autonomous entities; and Michael, Luke, and Robinson, “An Encounter With the Other”, on entity encounters and fractal-geometric imagery. ↩
-
Greg Egan’s official site is the author link. Permutation City is probably the best hard SF I’ve ever read. ↩
-
Hannu Rajaniemi on Twitter is the author link. The Quantum Thief books and Summerland are some of my favorites. ↩
-
qntm’s current There Is No Antimemetics Division page is the author link. The two works I mean here are There Is No Antimemetics Division and Fine Structure, which was originally published as a serial of linked web posts. ↩
-
Rich Sutton, “The Bitter Lesson”, on the repeated success of general methods that scale with computation over hand-built domain knowledge. ↩
-
Syndrome Space: Novel Development Dossier, the full science, characters, progression, ending, consistency rules, open questions, and source notes. ↩
-
Finite Multiverse Hard-SF Benchmark Prompt, the solution-stripped starting premise plus a 100-point evaluator rubric for comparing future models. ↩