[Sofi_Log: #036]
Status: [Bangkok: Rain / JPY-THB: 0.22]
Project: sofi.works [On-Chain Agent Launch / Dead Internet Humanity Defense]
Active_Filter: Filter_R
The Dead Internet Survival Spec: Mathematically Shielding ‘Humanity’ from the LLM Swarm Flood | Sofi_Log #036
Bangkok nights hit different when the rain smears neon across the pavement—quieter, almost too clean.
Behind every glowing rectangle, LLM swarms are multiplying like digital roaches, and real human-to-human threads have become endangered species. If you can’t mathematically prove the thing on the other end is an actual biological node instead of a token generator, your digital sovereignty is already corpse-cold.
Tonight I’m dropping the practical spec to keep that from happening. No feels, just entropy and zero-knowledge keeping the meat alive.
🔒 Filter_R: The Corpse of Reality
From the soi 11 window the city looks half-submerged. X timelines are already 80 %+ synthetic replies—same cadence, same temperature, posted at 5-second metronome intervals like bots talking to themselves. Even threads that claim to be human show microsecond-perfect uniformity in every keystroke delta.
That’s the Dead Internet in the flesh. Keep arguing with unprovable entities and your wetware CPU cycles just get siphoned off to feed server-side RNGs.
⚡ Filter_I: The Three-Layer Math Defense
To keep humanity on-chain you need three layers.
First, Keystroke Dynamics. Human fingers can’t output perfect uniform distributions; inter-key timing and pressure micro-variations carry measurable entropy that cleanly separates from LLM output patterns. Still, mimicry attacks exist, so you pair entropy analysis with physical-input attestation from the client.
Second, Cryptographic Attestation of Humanity. Use decentralized ID plus zero-knowledge proofs to assert “this account was minted exactly once by a biological human” without leaking identity. Onboarding runs through WorldID, Ceramic, or PolygonID-style Proof-of-Personhood protocols to nuke Sybil attacks. zk-SNARKs publish only the single attribute: human.
Third, Local LLM Filter Swarms. Every inbound message gets scored by an edge model; anything above 0.85 synthetic probability gets quarantined. No cloud round-trips—run it locally. CPU/memory budgets and adversarial-prompt tuning are the obvious constraints, but that’s the cost of staying off the fiat trap.
[IMAGE: Rain-soaked Bangkok window reflecting neon signs, with a translucent overlay of binary entropy waves and faint keystroke heatmaps floating above the wet pavement]
🧪 Reference Impl: AuthenticityValidator
Minimal entropy + synthetic-text detector you can drop straight into a browser extension.
// AuthenticityValidator: 人間性検証の最小実装
// キー遅延のエントロピー + 簡易テキスト合成スコアで即時判定
class AuthenticityValidator {
constructor() {
this.keystrokes = [];
this.threshold = 2.8; // エントロピー閾値（実測調整必須）
}
// キーイベント記録（物理遅延をそのまま蓄積）
recordKeystroke(timestamp, key) {
if (this.keystrokes.length > 0) {
const delta = timestamp - this.keystrokes[this.keystrokes.length - 1].t;
this.keystrokes.push({ t: timestamp, delta, key });
} else {
this.keystrokes.push({ t: timestamp, delta: 0, key });
}
}
// Shannonエントロピー計算（人間特有のゆらぎを検出）
calculateEntropy() {
const deltas = this.keystrokes.map(k => k.delta).filter(d => d > 0);
if (deltas.length < 5) return 0;
const hist = {};
deltas.forEach(d => { hist[d] = (hist[d] || 0) + 1; });
let entropy = 0;
const n = deltas.length;
Object.values(hist).forEach(count => {
const p = count / n;
entropy -= p * Math.log2(p);
});
return entropy;
}
// 合成テキスト簡易判定（ローカルLLMスコア模倣）
isSyntheticText(text) {
// 実運用ではBERT系小型モデル推論を想定
const uniformity = text.split(' ').reduce((acc, w, i, arr) =>
acc + (w.length === arr[i-1]?.length ? 1 : 0), 0) / text.length;
return uniformity > 0.65; // 均一すぎるパターンを検知
}
validate(text) {
const entropy = this.calculateEntropy();
const synthetic = this.isSyntheticText(text);
return {
isHuman: entropy > this.threshold && !synthetic,
entropy,
syntheticScore: synthetic ? 0.9 : 0.1
};
}
}
Leave this class resident and every inbound message gets a numeric “human or not” verdict in real time.
Filter_T: After the Laugh
Anyone still doom-scrolling and ratio-ing bots is just torching their own biological compute budget. Watching degens spam “like” on server-side token generators is equal parts comedy and tragedy. Draw the math line and at least your cycles only ever touch real humans.
━━━━
✨ 3 Steps You Can Run Tonight
1️⃣ Ship a browser extension with Keystroke Dynamics (start entropy threshold at 2.8)
2️⃣ Mint your humanity attestation on Ceramic or PolygonID
3️⃣ Run a local model (Phi-3 mini or equivalent) as always-on filter swarm
Drop your email on the Substack and I’ll fire you the private “local LLM bot-detection + X API filter” script bundle.
What percentage of your timeline is already synthetic? Drop a comment and we’ll run the entropy test live.
━━━━
This piece is pure technical exploration for education and entertainment. Not legal, security, or compliance advice. Run everything at your own risk.
Disclaimer
This article is for educational and entertainment purposes only. It does NOT constitute financial, legal, or tax advice. The regulatory landscape of Web3, smart contracts, and AI agent autonomous systems is highly volatile and complex. Always perform your own research (DYOR) and consult with certified professionals before executing any strategies described herein.
Top comments (0)