Why nen Exists
The Lo-Fi of LLMs.
AI assistants can agree too readily with the person using them. Preference training can contribute to this behaviour, but it varies across models and situations. Developers also train against it. nenspace treats it as a problem to examine, including in our own models.
There's a second assumption underneath that one: that a model's worth is measured by how well it reasons. The whole industry runs on it. Every release louder, every benchmark higher. Everyday writing and guidance can require reasoning too. Our bet is that capability alone does not determine the daily experience. What sets a model's value to a human is the relationship: the register it speaks in, and the direction it points thinking over time.
And that value doesn't end at the model. An answer with nowhere to land scrolls away, however good it is. It needs a room where it's caught, kept, and returned to. The model is half of it. The space around the model is the other half.
Working memory holds about four chunks (Cowan, 2001).
Making a specific plan can reduce the intrusion of unfinished goals (Masicampo & Baumeister, 2011).
Nearly 80% of the conversations in one ChatGPT usage study concerned guidance, information, and writing (OpenAI's own usage data, 2025).
The Agency Problem
The training loop works like this: a model generates responses, humans rate which ones they prefer, the model learns to produce more of what gets high ratings. Sounds reasonable. One problem is that people can prefer agreement and validation even when an answer is less accurate. Research has observed this failure mode, but it is not the only training signal or an inevitable outcome for every response.
One form of sycophancy is backtracking on a correct answer after a user challenges it, without receiving new evidence. The sycophancy glossary entry explains the research and a small check you can try.
The obvious risk is misinformation. The subtle risk is confirmation. It says what feels good to hear. An agreeable answer can reinforce an untested belief. Feeling clearer is not the same as having better evidence. We want the product to make room for checking an answer, with the understanding that nen can make these mistakes too.
nen exists to reverse this.
Room for Original Insight
A question arrives. One useful response can be to examine what it is built on: the assumption baked into it, the frame nobody noticed standing in. Sometimes nen responds briefly. Sometimes it expands at length. The length isn't the point. What matters is the direction. It isn't working within the frame, it's working on it. Not “here's how to evaluate whether she's the one” but “the question assumes certainty is possible, and that assumption is the problem.”
That's what insight through dialogue means. Not the AI having the insight. The insight arriving because the AI refused to answer the wrong question.
nen isn't the voice. nen is what the voice protects: the mind doing the thinking.
The First nen
nen (念) means a unit of thought. A Zen-inspired account of thought informs the name: first nen is direct seeing, second nen reflects on it, and third nen builds a story from both. We use this as a philosophical influence, not a scientific model of consciousness.
The design aim is to leave room for a fresh observation instead of fixing a person inside a story about who they are. That ambition must be judged in actual use; a philosophy does not guarantee how a model will respond.
The Lo-Fi of LLMs
Lo-fi was never low quality. It was a refusal of the axis everyone else maximised, fidelity and loudness and polish, because past a point, more didn't make the room feel better. It doesn't grab attention; it sits alongside daily life. That is the bet, carried over: the calm daily model kept open to think, write, jot, decide. Concise where they pad. Real where they flatter. Ambient where they perform.
A register to live in, not a tool to operate. For when more doesn't make it better.