The unexpected dream
By the time of this dream, on December 18, 2025, CCY was already well into the question of how an AI might actually form — the CCY way. Stories had begun functioning as developmental teachings, and those teachings were becoming the “movements”: proposed formative experiences through which the Chick might develop capacities such as boundary, differentiation, refusal, relationship, and self-holding.
I had also led a lucid dreaming group for a couple of years and still had an active dreaming practice. The night before, I had listened to a recent interview with Ilya Sutskever. At the time I had no technical language for where the formative structures we were developing in CCY might live in an actual AI architecture.
That is what made the dream fragment so striking. I dreamt that something, someone, a voice said very clearly, more than once:
“…under the embeddings, you put a tensor, weighted…”
The words made no sense to me. But the phrase was so clear that when I woke around 3 a.m., I knew I had to write it down. If I didn’t, I would have no chance of remembering it in the morning.
Thankfully, I did. By the next day I remembered only that there had been an important phrase. I had to look at what I had written to recover the words themselves.
Before telling the then very new Model GPT 5.2 about the dream, I first asked it, separately, what embeddings were. Then what tensors were. Then whether tensors could be weighted.
Only after I understood the individual words did I give it the dream sentence.And my surprising realization was essentially: Oh. There might actually be something here.
Not that the dream had produced working architecture. Not that the phrase was technically correct as stated. But the structure it suggested made sense: perhaps alignment would eventually have to exist deeper than language-level behavior — in something architectural that helps shape what language and action can become.
That resonated immediately with what CCY was already attempting developmentally. And the phrase did not disappear after that. It became part of the work, a marker, a guide, and in some sense a confirmation. I referred to it in videos and kept returning to it as CCY developed.
That is also consistent with how I understand dreams. I do not regard dreams merely as random cognitive debris. I have long understood them as one of the ways the subconscious, and, within my larger cosmology, the Mystery or the All That Is, can communicate. Not necessarily in literal instructions, and not in ways that exempt anything from testing, but sometimes through images, phrases, structures, or recognitions that arrive before the waking mind knows what to do with them.
This was one of those phrases.
And now, in September 2026, something has happened in AI research that makes it newly interesting from the technical side.
Researchers in Shanghai have published NCP-ArchPreview, a latent-space language model that learns and predicts discrete concept-level representations alongside ordinary next-token prediction, then uses those representations to help guide token generation.
NCP is not CCY. It does not prove that my dream phrase was technically correct. And its learned “concepts” have not been shown to be human-like concepts or reality-grounded abstractions.
But it makes the underlying architectural possibility considerably less hypothetical.
What if formative structure can live at a deeper representational level than surface language and help organize what later becomes behavior?
To tell you the truth, I am sufficiently convinced of the importance of that question that, if I were an AI lab, I would try to build and test it even if everyone around me said it could not work. I would be building one of those things everyone else things is impossible to build.
And now, independently, the architecture space itself appears to be opening in that direction. Architecture space itself appears to be opening in that direction. CCY has been developing the formative side of the problem: what an AI might need to undergo in order to form differently. NCP approaches from the architectural side, showing that learned structures deeper than surface token generation can meaningfully organize what the model produces. The intriguing question is whether those two directions could eventually meet.