PipeR (@pipe_render) header
PipeR (@pipe_render) profile picture

LLMの接続構造が知性を生むか研究。AAS実験完了(p=0.0007, d=4.29)。 論文→ https://teapaw.org/t.co/6ZR4seJTWE Code→ https://teapaw.org/t.co/LdAph8AaEO

Joined May 2008
Photo by PipeR (@pipe_render)Photo by PipeR (@pipe_render)Photo by PipeR (@pipe_render)Photo by PipeR (@pipe_render)Photo by PipeR (@pipe_render)Photo by PipeR (@pipe_render)Photo by PipeR (@pipe_render)Photo by PipeR (@pipe_render)
【arxivエンドースメントを探しています】 cs AIへの初投稿のため、arxivに論文をお持ちの方を探しています。 実験:同じモデル・同じHWで、適応型ネットワーク(AAS)vs 固定ネットワークを比較。 結果:p=0.0007、d=4.29、5/5試行で適応型が優位。 粘菌の原理をLLMネットワークで検証しました。 コード・結果・安全設計すべて公開済み👇 github.com/piperendervt-glit… エンドース or ご紹介いただける方、DMください🙏 #arxiv #LLM #AI研究
311
🧪 New idea: Past-log-aware AAS (Adaptive Artificial Synapse) What if the weight update rule in a multi-agent LLM network could reference past experiment logs instead of starting from zero every time? Pros: + Faster convergence (starts from proven weights, not zero) + Fully explainable (why each weight = traceable from past results) + Lower risk of runaway weights (stable initialization) + Task-aware: use different initial weights for different task types Cons: - Log quality determines weight quality (garbage in, garbage out) - Logs are task-specific — can't transfer world_consistency logs to math tasks - Risk of over-exploitation (always following past optima, never exploring) - Requires ε-greedy or similar mechanism to maintain exploration The core insight: original AAS updated weights reactively (task-by-task). This variant initializes weights proactively from historical performance, then fine-tunes in real time. Still a draft — building toward this after NCA v6-v10. Code so far 👇 github.com/piperendervt-glit… #LLM #MultiAgent #AI #experiment #AdaptiveAI
49
🧠 Can LLM nodes behave like Neural Cellular Automata? I ran an experiment: 3 LLM nodes (qwen2.5:3b) in a ring topology, each updating its answer by looking only at its neighbors — pure local rules, no global control. Results vs fixed pipeline: ✅ Overall accuracy: +4% (NCA wins) ❌ Groupthink: 99/100 tasks → all nodes converged to CONTRADICTION The NCA update loop amplified the model's bias instead of creating diverse reasoning. Now testing anti-sycophancy prompts + confidence-weighted voting to fix it. 🔗 github.com/piperendervt-glit… #LLM #NeuralCellularAutomata #AIExperiment #LocalLLM #Ollama
57
Visualizing the Thomas attractor on Meta Quest 3. Up to 500 particles evolving simultaneously, colored by their Lyapunov exponent — red = chaotic, blue = stable. Supports both VR and Mixed Reality (passthrough) modes. Built with Unity + Meta XR SDK. #MixedReality #MetaQuest3 #Mathematics #Visualization #Unity3D
98
TRUSS Phase 4 (Titan) is complete. A multi-agent architecture designed to suppress uncontrollable AI through structural design, not rules. Key results: · Threshold θ=0.3 is overdetermined: confirmed independently from 6 mathematical frameworks (statistical, experimental, geometric, dynamical, control theory, information geometry) · Experimental response to albatrosary (2026): P(E3)=0.10 (90% suppression vs theoretical) Cov(Xi,Xj)=0.075 · 16-node tesseract RC mutual monitoring network · Validation scope honestly documented: limited task environment ≠ safety guarantee in diverse real-world settings Moving to Phase 5 (Aegis). GitHub: github.com/piperendervt-glit… Zenodo: zenodo.org/records/18922652 #AISafety #TRUSS #MultiAgent
62
ガサラキのロボットが何で乗るの疲れるのか分かった気がした、、、
2
I got to drive Prosthesis, a giant mech. The video is of the creator operating, my performance was far more tepid. This thing is an amazing feat of engineering. @racing_mech
362
5,780
1,531
19,577
脳以外の臓器もある種の”記憶”を保持しているんだろうか? (エピソード記憶のコンテキストのような、、、)
PipeR retweeted
今年読んだ論文の中で最も衝撃だったのが 「皮膚のニューロンが信号伝達だけでなく大脳皮質同様の計算処理を行っている」 というこの論文 nature.com/articles/nn.3804/ 信号伝達しかしないと思われていた指先の末梢ニューロンが, 実は接触したものの幾何学的情報を処理していたことが明らかになったらしい
15
4,084
300
8,864
Kinect v2 + RealSense D435 x 2 Screen VR Video Meeting System. (Demo: Local PC only)
1
1
HMDレスなプロジェクターVRを使った、ビデオ会議システムのデモ(アバターは自分自身)
1
Face Trackingの精度が上がったため、CAVE VRの実現コストが大幅に下がった、ってだけ
RealSense VRのプロジェクター版
1
5
Johnny LeeのWiiリモートVRのRealSense版
1
専用カメラなしで自由視点映像を合成、深層学習で カーネギーメロン大学「Open4D」開発 itmedia.co.jp/news/articles/… 寄稿:屋内外、複数のモバイルカメラで撮影した映像からCNNで前背景を抽出し実写の3次元立体動画を作成する技術。後から視点変更や時間編集ができ特定の人を自然に消去など加工が可能。
52
2
153