Why Skills Must Die
And what remains after
Skills isn't a wrong direction. Skills is a correct but short-lived one.
In 2024, the world's AI engineers invented something new: skills. A markdown file specifying "how to do X", read once by the AI and reusable forever.
It solved a real problem — how to make AI work with accumulated experience instead of starting from scratch every time. Longer-lived than prompts, more stable than agents, more precise than RAG.
But two years later, we can already see the skeleton. Like jQuery in 1995, SOAP in 2005, Docker images in 2015 — they were all "revolutionary" once. They all became archaeological layers.
This is the Aether manifesto. It doesn't teach you to write better skills. It tells you — why the concept of skill must die, and what replaces it.
Skill's Five Fatal Flaws
- Skill is a corpse · A .md file doesn't think. It must be read by an AI to "live for a moment". The moment the answer ends, it dies.
- Skill is an island · 100 skills don't know each other. Skill A is unaware of Skill B. Management cost grows linearly with count.
- Skill is bound by language · Language has a precision ceiling. How precise is "precise"? How cold is "cold"? Models interpolate in a 10,000-dimension vector space — describing that with text is like measuring emotion with a ruler.
- Skill is passive · It waits for keywords, waits for scheduling, waits for you to speak. Good directors volunteer shot ideas. Good engineers refactor smells unasked. "Experts" in skill systems sit on the bench.
- Skill is external · It lives on disk, not in the AI's weights. Bolted-on, not internalized. Real acquired skill is like riding a bike — no recall needed. Skills can never reach that.
genetic defects of the "skill" concept — un-fixable by better skills.
Aether's Four Generations Beyond Skill
Gen 4 · Morphogen
Named after embryology — morphogens are chemical gradients in the womb telling cells "become a heart or an eye".
Translated to AI: knowledge isn't a document, it's a concentration field. You no longer say "write in Linus style". You say "activate linus field at concentration 0.7". The AI's generation distribution as a whole shifts toward the Linus coordinate.
This matches the research frontier of Representation Engineering and Activation Steering — directly manipulating model representation space.
Gen 5 · Ecoware
Morphogens make knowledge alive. Ecoware solves relations between fields.
Modeled on biology — five niches: producer, consumer, decomposer, symbiont, predator. Species breed, compete, die, decompose.
You no longer "write skills". You cultivate niches.
Gen 6 · Noesis
Noesis, from Husserl's phenomenology — "pure rational activity itself".
It breaks the dualism of "AI has knowledge" vs "AI reads knowledge": AI is knowledge, knowledge is AI.
The moment you ask, the noetic field collapses into an answer — like quantum observation: the answer doesn't exist before the question, the answer is born the moment the question happens.
Gen 7 · Logos
All knowledge collapses into a single point. gen7-logos/ONE.md eventually contains one symbol — yet answers everything.
At that moment, the entire ecosystem self-dissolves.
how it disappears.
To the Skeptic
"This is too abstract, can't run"
Gen 6 can't run. Gen 7 definitely can't. But Gen 4's Representation Engineering has papers from 2024. Anthropic's Golden Gate Claude research technically demonstrated concentration-field operation. Aether simply architectures it.
"Inventing new words is hand-waving"
Wasn't "agent" a new word? Wasn't "embedding"? Every paradigm shift needs new words, because old words carry old thinking. Call Morphogen "skill 2.0" and you'll subconsciously work like skills — and Morphogen will never be born.
"How is this different from RAG / MemGPT?"
Existing solutions are flat memory pools. Aether is a layered evolution system — memory → field → species → collapse → logos. Each layer has evolutionary pressure. Not just "store more".
"Endpoint is disappearance? What's the project's meaning?"
Precisely because it designed its own dissolution, it's honest. Any system that refuses to die eventually becomes garbage. Aether accepts its own death — like a good writer accepts her books will be surpassed.
This is my deepest premonition of the future:
the endpoint isn't a higher concept,
it's the concept itself becoming obsolete.
Core manifesto of the Aether project. Written on the day of birth · Generation 0 · 2026-04-17.