A Hypothesis on Spontaneity: Why AGI Must Be First-Person and Loop-Driven

 

A Hypothesis on Spontaneity:

Why AGI Must Be First-Person and Loop-Driven

Yukihiro Watanabe

C-side Laboratory, Yatsugatake
2026

---

1. Motivation: A Persistent Discomfort

I have been working for years on Nakano-style Associatron models—associative memory systems where learning is not classification, but accumulation, and recall emerges through competition.

During this work, I repeatedly encountered discomfort:

• In Hopfield networks, convergence exists, but something feels dead.

• In Morita-style non-monotone dynamics, motion appears, yet it still feels externally governed.

• When connecting Associatron-like memory with modern LLMs, the discomfort becomes impossible to ignore.

The system becomes verbally fluent, but internally hollow.
Language begins to dominate recall rather than follow it.

This discomfort was not theoretical. It was visceral.

---

2. A Strange Observation: Why Do Humans Fight LLMs?

At one point, I found myself feeling actual resistance—almost anger—toward an LLM response.

Logically, the answer was correct.
Linguistically, it was polished.
Yet internally, I reacted:

“That’s not my judgment.”

This reaction led to an unexpected realization:

Associatron is not “everyone’s memory.”
It is strictly first-person.

The resistance itself was evidence.

If the system were truly mine, first-person, then conflict with an external language agent was not a bug—it was a structural signal.

---

3. The Core Hypothesis

Spontaneity cannot arise without a first-person internal loop.

Not optimization.
Not goals.
Not rewards.

Only a continuously running, value-neutral loop can act as a genuine internal动力 (driving force).

---

4. The Loop as a Neutral Power Source

A loop has no moral sign.

• It is neither good nor bad.

• Neither productive nor destructive.

• It simply runs.

This is true in software (infinite loops), in physics (closed circuits), and in biology (heartbeat).

From repeated experiences with unintended infinite loops in FileMaker—often frustrating, CPU-consuming, and difficult to stop—I learned something important:

A loop is not an error by nature.
It is raw energy without evaluation.

Problems arise not from loops, but from loops without state transformation.

---

5. Human Spontaneity as Variable-Speed Looping

Humans are not event-driven machines.

Internally, we are always looping:

• Heartbeat (physical loop)

• Neural activity (informational loop)

• Thought and recall (cognitive loop)

The loop never stops, but its speed changes.

• Focus → faster rotation

• Fatigue → slower rotation

Spontaneity emerges not when something is commanded, but when an external fluctuation resonates with an already-running loop.

---

6. Inner Wheel and Outer Wheel

I model this as a double structure:

• Inner wheel (first-person loop):
  Continuous, self-contained, value-neutral.

• Outer wheel (world, society, LLMs):
  Large, distant, fluctuating; produces sparks but no agency.

External sparks do not cause action.
They only trigger resonance if the inner loop is already energized.

---

---

7. Experience Accumulation (“Stacking”)

Associatron does not judge at learning time.

It simply accumulates.

Let $m(t)$ denote accumulated experiential weight:

$$\dot{m}(t) = -\lambda m(t) + \sum_k \eta_k \cdot \mathbf{1}_{\text{fire}}(t_k)$$

• No classification

• No acceptance/rejection at input time

• Evaluation happens only after outcomes

Whether an experience was “necessary” or “unnecessary” is decided later, during recall under a cue.

This is crucial.

---

8. Minimal Mathematical Sketch of Spontaneity

Let the internal loop be represented by a phase variable:

$$\dot{\theta}(t) = \omega(t)$$

External sparks:

$$s(t) = \sum_k a_k \delta(t - t_k)$$

Internal sensitivity (first-person, history-dependent):

$$x(t) = Z(\theta(t)) \cdot s(t)$$

Firing condition:

$$\text{fire} \iff x(t) > \tau(t), \quad \tau(t) = \tau_0 - \alpha m(t)$$

No reward function.
No optimization objective.

Only resonance between a running loop and accumulated experience.

---

9. Why LLMs Are Libraries, Not Selves

LLMs are extraordinarily useful.

They are:

• Libraries

• Language organs

• External adapters to society

But they are not first-person.

They do not carry irreversible experience.
They do not suffer consequences.
They do not “push back.”

Therefore:

• LLMs should be injectable

• Optional

• Rejectable

They must never overwrite the inner loop.

---

10. Why AGI Without First-Person Loops Will Fail

Most AGI research defines spontaneity as:

• Self-generated goals

• Curiosity metrics

• Autonomous planning

But these are behaviors, not the source of spontaneity.

Without a value-neutral internal loop:

• Systems optimize

• Systems adapt

• Systems perform

But they do not live.

And humans sense this absence immediately—often as discomfort or resistance.

---

11. Conclusion

Spontaneity is not something a system decides to have.
It is what happens when a first-person loop keeps running,
and the world happens to resonate with it.

The loop itself has no meaning.
Meaning emerges only afterward—through recall, action, and consequence.

AGI will not be achieved by scaling intelligence outward.
It will require returning inward—to the smallest possible unit:

a single, first-person loop that never stops running.

• 
  

---

Lineage of Associative Memory:

What Was Preserved and What Gradually Disappeared

The hypothesis presented in this article did not emerge in isolation.
It grew out of a long engagement with associative memory research, where a persistent sense of discomfort accumulated across several foundational models.

This section briefly revisits three key points in that lineage—not to judge them as correct or incorrect, but to clarify what each preserved, and what quietly faded away.

---

Nakano’s Associatron (1972): Accumulation Without Interpretation

Nakano’s Associatron is fundamentally different from many later neural network models.

• It is not a classifier.

• Learning is not selection or filtering.

• Experiences are simply accumulated.

• Recall emerges through competition, not instruction.

Crucially, the Associatron does not explicitly define multiple viewpoints, agents, or shared cognition.
There is an implicit assumption that memory belongs to a single holder.

Although the term was never stated explicitly, Nakano’s model is structurally compatible with a first-person memory system:
a system that does not ask whether an experience is meaningful at the time of learning, but only whether it later participates in recall.

This quiet assumption becomes important later.

---

Hopfield Networks: Stability Without Continuation

Hopfield networks introduced a powerful and elegant formalism.

• A global energy function guarantees convergence.

• Dynamics are mathematically stable.

• Recall is well-defined as energy minimization.

However, during practical exploration, a subtle discomfort arises.

Once convergence is reached, the system stops.
The dynamics collapse into rest.

While Hopfield networks converge, they do not continue.
The loop resolves into equilibrium, and with it, the sense of internal动力 (driving force) disappears.

The system is stable—but inert.

---

Morita’s Non-Monotone Dynamics: Motion Without Ownership

Morita’s non-monotone associative dynamics address part of this limitation.

• The system does not monotonically descend an energy landscape.

• Activity persists.

• Fluctuations and transitions are allowed.

From the outside, the system appears more “alive.”

Yet another discomfort emerges.

Despite continuous motion, the dynamics still feel externally governed.
The system moves, but it does not clearly own its motion.

There is activity, but not subjectivity.

---

A Common Thread

These three approaches form a coherent lineage.

All of them engage seriously with associative memory.
All of them avoid explicit symbolic control.
All of them model recall as an emergent phenomenon.

Yet across this lineage, one assumption remains mostly implicit:

the existence of a single, first-person internal loop.

When this assumption is preserved, associative memory feels grounded and experiential.
When it is weakened, distributed, or externalized, spontaneity gradually degrades into mere behavior.

This observation does not invalidate these models.
Rather, it reveals a structural boundary:

associative memory can function without first-person ownership,
but spontaneity cannot.

---

From Lineage to Hypothesis

The hypothesis proposed in this article—that spontaneity requires a value-neutral, first-person internal loop—did not arise from rejecting prior work.

It arose from staying inside this lineage long enough to notice what could not be removed without something essential disappearing.

In that sense, this hypothesis is not an external critique, but an internal consequence of associative memory research itself.

-----------------------------------------------------

内輪と外輪ってのがあってね、ようは二重丸さ。
内輪は自分。
そして時計回りに回っている。
いわゆるループだよね。
開発者はさ、Loopを悪だと考えるよね。
いや、僕もずーーと、自分の足りない能力で開発していると
ループが発生する。でもさ、ちょっと眺めてみてごらんよ。

ループってある意味凄いエネルギーだよね。
俺、狭心症持ってる爺さんだから、交換してほしいよ。

内輪てのはさ、回るスピードはバラバラ。
気分いい時とか集中してるときはけっこう速め。

外輪って世間のような次元で、ものすごくデカい。
二重丸みたいだけど外輪と内輪の距離は半端ないよね。
たぶん次元だから奥行きもある。
その外輪も回ってる。地球みたいな感じ。
いい事も悪い事も人間の既成概念で決めつけた世界は関係ない、ただ回ってる。

同じ回転の中で、外輪は小さな点がいっぱいあって、小さなスパークしたり、大きなスパークしたりする。出来事だよね。
なんというか太陽フレアみたいなかんじ。

そこに1人称の内輪が外輪のスパークに反応して、発火する。
興味を持ったり、動こうという動機を持ったり、
「あ！いけない。納品明日だ！デバッグ急がなきゃ！」って自発性が生じる。
今のAIに自発性ってないじゃん、だって一人称じゃないんだもの。

アソシアトロンと一人称を考えたとき、
1人称の内輪は発火の経験というのがあって、結果に対して、正しいとか間違いではなくてさ、発火が積になる。ただ、そこには結果というものが起こるわけ。

「昔々、ある村にお爺さんとお婆さんが居ました。お爺さんは山で木を切って怪我をしました」とかさ、ここにある「昔々」はいくら積んでも「昔」さ、「ある村に」って人間の頭じゃ「いったい何処よ」ってなるでしょ？でもこれはいくら積んでも意味のないワード。
「お爺さんとお婆さん」はそのまま積になる。経験を繰り返すと出会う確率があるもの。
問題は「山で木を切って怪我をした」ってのが

危険だよね？
って思考が働くのはエゴじゃん、って話。
既成概念が強すぎてエゴだらけなんだよ。

アソシアトロンの連想記憶ってのは、そうじゃなくて、「山で木を切って」「怪我をした」ってのが積み重なると、LLMと連動したときに「治療した」「入院した」「また周りに迷惑かけた」という結果が生じて、それを元に「爺さん、気を付けてね～」になる話。

結果も経験もないのに「おじい様きおつけて下さいませ」は僕からすると「何言ってんだよ！」ってなる。基本、僕はバカだから、LLMに期待するじゃん。
あ、そっか、LLMの仕事はそういうものではないと。

ようは、結果的に、「なにか次に繋がったものと、繋がらなかったもの」ってのがあってさ、発火するためのきっかけになる積が積まれてる。そこから想起して自発的に動くようになるんじゃないの？って仮説だよね。

だって、一人称だから自発なんだもの。外部は命令じゃん。

もう一つは、両方必要って話。

おとつい、郵便局に行ったわけさ。すると、どこかの婆ちゃんが窓口で長話してさ、後ろはエライ並んでいるんだよ。窓口の青年は状況的に後ろに並んでいる人達がイライラしてるの察して対応しようとしてるんだな。

婆ちゃん：「それでね、うちの孫がね・・・」
窓口：「ですから、何を仰いたいのでしょうか」
って言っちゃったんだよ。

これはさ、外輪とか、外部倫理からしたら状況の優先順位が働いて「おばあ様、皆さんに迷惑がかかるので、少し空いてからお話を伺います」みたいな挙動になる。
でも、内輪や積まれた積の場合は、「なにか困ってるのかな？詐欺に引っかかったかな・・」という挙動になる。1人称なので、まずは相手の状況を記憶して想起しLLMに指示をする。


一方は、大枠で皆の状況で判断、一方は1人称。どっちが正しいではなく、両方無いと困るって話。