Shows / Not a ToE
Not a ToE
Audio overviews of essays from Not a ToE — a living practice for tracing situations to their operative mechanisms and dissolving the projections that obscure action. A precision instrument, not a comprehensive theory. Each episode walks one piece. English and Chinese essays appear as separate episodes. The written text is the source of record: https://powerpig99.github.io/not-a-toe/ Episodes are AI-generated discussions of the essays, not live recordings.
Episodes
Not a ToEEvoRouter：多智能体路由如何在提准确率的同时把算力减半
多智能体系统要同时决定用哪些模型、什么协作模式、几个智能体、几轮交互、配哪些技能，组合空间爆炸；只看最终成败和成本的奖励，说不清哪一步决策有用、哪一步浪费算力。 本期聊 EvoRouter（arXiv…
Not a ToEEvoRouter: Cheaper Multi-Agent Routing That Learns
Running a multi-agent system means choosing models, a collaboration mode, how many agents, how many rounds and which skills, and the combinations explode. A…
Not a ToE会睡觉、会做梦的大模型：让 AI 学完不忘
今天的大语言模型部署之后就不再学习：上下文窗口里学到的东西，会话一结束就消失；直接更新权重，又容易出现灾难性遗忘。本期聊 Ali Behrouz、Farnoosh Hashemi、Adel Javanmard 和 Vahab Mirrokni 的论文《Language Models Need…
Not a ToEWhy Language Models Need Sleep
Today's large language models stop learning once they are deployed: what they pick up in the context window is lost when the session ends, and updating their…
Not a ToEAI自主训练AI：RSI竞技场里的人类反馈
AI 智能体能不能自己把一个 Base 模型训练成好用的助手，再根据用户的投票继续改进？RSI Arena 让十个前沿模型担任训练智能体，从同一个 NVIDIA Nemotron 3.5 Lightning 30B-A3B Base 出发，各有 1,000 GPU 小时和 300 美元 API 额度，在共享的 64…
Not a ToEWhen AI Agents Build Better AI: Inside RSI Arena
Can AI agents train a useful language model on their own, then improve it from people's votes? RSI Arena put ten frontier models to work as training agents…
Not a ToE告别Token：字节级AI如何摘下人类的“有色眼镜”
大模型其实看不到真实世界：在学习开始之前，人工编写的预分词规则和固定的BPE词表就已经把数据切成了一个个词元，就像把名画变成填色格子板。本期深度解析从这里出发，聊Cartesia的层级网络（H-Net）如何在模型内部对原始字节进行动态分块；“苦涩的教训”为什么指向端到端学习；Jing…
Not a ToEWhy AI Is Replacing Tokens With Bytes
Today's language models never see raw text. Before learning begins, a hand-written pre-tokenizer and a fixed BPE vocabulary cut everything into tokens, like a…
Not a ToEAI时代的思维分辨率
在 AI 时代，同清晰的写作相比，编码正变得微不足道；同清晰的思考相比，清晰的写作也正变得微不足道。本期沿着这一层级背后的区分展开：精度追求脱离语境、毫无遗漏的准确；分辨率则是在具体处境的约束下，作出要紧区分的能力。…
Not a ToEPrecision and Resolution: Why Clear Thinking Is the Bottleneck
In the age of AI, coding is becoming trivial compared to writing clearly, and writing clearly trivial compared to thinking clearly. This episode follows the…
Not a ToEWhy P Versus NP Is Unprovable
"P = NP in that their difference is unprovable, but not in that they are identical, which is also unprovable." This episode unpacks that one-line reply and the…
Not a ToEAI科研团队装进MacBook
完整的科研循环，能不能在一台 MacBook 上跑起来？Cong Lu 的 Local Autoresearch 让 Recursive 的本地 40B 模型 Flint 一次一个实验地改进小程序：提出改动、写代码，由运行器测试，只保留有提升的结果。 本期聊 Cong Lu…
Not a ToEAutonomous AI Research on a Laptop
Can a full research loop run on a MacBook? Cong Lu's Local Autoresearch lets Flint, a local 40B model from Recursive, improve a small program one experiment at…
Not a ToE大模型权重里的 CPU
Qwenputer 是 Divyansh Agrawal 的一个实验性模型：在 Qwen3.5-4B 的权重中直接编译进一颗确定性的 32 位 RISC-V CPU（RV32IM，外加 Zba 扩展和一个小型大数协处理器）。据作者介绍，整个过程不涉及训练或微调：由编译器直接设定新增的 11 层的全部权重，这些层位于…
Not a ToECompiling a RISC-V CPU Into Neural Weights
Qwenputer is an experimental model by Divyansh Agrawal: Qwen3.5-4B with a deterministic 32-bit RISC-V CPU (RV32IM, plus Zba and a small bignum coprocessor)…
Not a ToE为什么AI无法终结数学
2026 年 10 月 26 日，OpenAI 公告了由内部模型生成的三项候选结果：欧几里得波动方程局部平滑猜想的证明（由此解决挂谷猜想与斯坦因限制猜想）、否定指数时间假说的确定性亚指数算法，以及四维纯杨-米尔斯理论的构造性存在证明。这些结果不同程度地附有 Lean…
Not a ToEWhy AI Formal Proofs Are Incomplete
On 26 October 2026, OpenAI announced three candidate results from an internal model: a proof of the local smoothing conjecture for the Euclidean wave equation…
Not a ToE模块化反馈回路与动态残差网络
深度学习教材把普通的层层堆叠和残差网络归进同一个抽屉，叫前馈网络，又把跳跃连接解释成防止梯度消失的补丁。从唯一的前提去追溯“下一个状态等于这个状态加一个修正量”里的那个加号，看到的是另一刀：跳跃连接带着走的是一个保留的参照，有了它，一条漫长的全局反馈回路就拆成了许多局部修正。…
Not a ToEModular Feedback Loops and Dynamic Residual Networks
Deep learning textbooks put a plain stack of layers and a residual network in the same box, feedforward, and explain the skip connection as a fix for vanishing…
Not a ToE何谓形而上学？将追寻命名为目的
“何谓形而上学”这种问法，悄悄把形而上学安放成一个早已摆在那里的东西：一个有边界的领域，一个可以走过去巡视的终点，而提出问题的那份追寻，被留在了门外。将不可化约的追寻命名，就是把它冻成可被引用的对象；反复返回第一人称，才收窄冻结所留下的超额滞后。…
Not a ToEWhat Is Metaphysics? Naming the Seeking as Destination
A question of the form "what is metaphysics?" quietly installs metaphysics as something already there: a field with edges, a destination to survey, with the…
Not a ToE交接残余：以制品作为初始条件
做智能体的人里，同时冒出两种抱怨。一种是把按专业分工搭起来的子智能体树折回单一循环，因为每交接一次，就丢一截语境，多一段延迟；另一种是欢呼自我改进的循环已经自主，再也不需要人插手。听上去是两场方向相反的争论，其实出自同一道放错了的边界：还在进行的活动，与它留下的残余之间的那道边界。…
Not a ToEHanding Off Residue
Two complaints keep turning up in agent engineering at once. Teams are folding their trees of specialist subagents back into a single loop, because every…
Not a ToEThe Comparison That Relocates the Cause
Two kinds of tool run under one protocol, one comes out ahead, and the result is read as news about where capability lives. The ranking itself is where the…
Not a ToE进位无法在寄存器之内被命名
在八位寄存器里给255加1，八个比特全部翻转为零，没有一条规则受损。这个零是一次模运算回绕、一处静默的故障，还是向第九位的进位，八个比特自己判定不了。把它认作进位，是走出位宽的判定。 汤姆·扎哈维的立场论文《大语言模型无法跃迁》（ICML…
Not a ToEThe Carry That Cannot Be Named Inside the Register
Add 1 to 255 in an eight-bit register and every bit flips to zero, with no rule broken. Whether that zero is a modular wrap-around, a silent fault, or a carry…
Not a ToETokens and the Observational Boundary
A crossword with every square filled in leaves no move to make. A language model reading tokens meets the same thing at its own grain: the tokens are definite…
Not a ToE作为记忆与控制的边界条件图
同一句提示词，温度调到零，跑两遍，第二遍的回答可能换了一句话。规则没有动：权重冻结，解码规则固定。动的，是没人登记的条件，小到批处理内核把浮点数加起来的先后顺序。这个小小的意外，打开了生成模型如何被驾驭的整套架构。…
Not a ToEBoundary-Condition Graphs as Memory and Control in Recursive Formal Systems
Run the same prompt twice at temperature zero and the second answer can change a sentence. The rules did not move: the weights are frozen and the decoding rule…
Not a ToEThe Reversal of Verification
A proof can be flawless and the trust built on it still upside down. Verification comes first; trust is what stands after it, for one step, until the next…
Not a ToEThe Surface That Registers
The farther a sentence travels, the less of the thinking that chose it travels with it; and the stronger the "yes, exactly" it gives two people, the less has…
Not a ToEThe Density of Judgment
The faster an open-ended loop turns, the more often it comes back to someone for a decision. When Recursive released Flint, a local coding model of about forty…
Not a ToEOn-Demand Contexts Keep Coherence at One Step
Coherence doesn't keep. It can't be stored somewhere for a pipeline to look up later; it has to be loaded, fresh, into each context that derives from it, and…
Not a ToEThe Aesthetic Escalator and the Melancholy of the Balcony
A listener answered an earlier deep dive on judgment with a single line from Joseph Brodsky's 1987 Nobel Lecture: the richer one's aesthetic experience, the…
Not a ToE判断力不是可囤积的存量
Y Combinator 总裁加里·谭预测，四五十岁的创业者将迎来一波浪潮：当 AI 让现成答案与执行变得廉价，数十年积累的判断力与品味会重新升值。这个预测把判断力当成稳定的存量，靠与约束长期接触慢慢沉淀，执行瓶颈一去就被放大而不是折旧。它只在那些约束大体不变时成立。 几个转折：积累里很大一部分，正是与如今被 AI…
Not a ToEJudgment Is Not a Stock
Garry Tan, president of Y Combinator, forecasts a wave of founders in their forties and fifties: once AI makes ready answers and execution cheap, decades of…
Not a ToEYour Codebase Is Memory
Two hosts trace one argument: every digital artifact that can be read again, code, essays, configs, graphs, plans, bridge files, model weights, is memory in…
Not a ToEBeyond the Projection
Lauren Tan, who builds Grok Bot, runs her agents through pstack, a personal set of skills for rigorous engineering with agents. Nothing in it asks an agent to…
Not a ToEAn Observation by Its Own Standard
A preprint from Doris Tsao's group records that, while two macaques view two faces under binocular flash suppression, population activity in face patches and…
Not a ToEMath as a Career Is Not Math as a Pursuit
A model solves the problem a PhD student had chosen for a thesis. Grant Sanderson's reflection on AI in mathematics lists what a mathematician does, why the…
Not a ToEThe Comment Is Not the Code
One sentence above a function: "retry here because the upstream API drops the first request after a cold start." To its writer it matches the code, because…
Not a ToE悬而未决是决定论的索求
演化为什么选中了人类？这个问题把答案交给了一位选择者，而被选中的说不出选择者的理由。把选择者拿掉，后代有多有少、世系随之偏移的讲法照样成立。把“每一个有意识的生命为自己抉择并承担后果”拿掉，它在拿掉的动作里自己回来。…
Not a ToEThe Open Question Is Determinism's Demand
Why did evolution choose humans? The question hands the answer to a chooser, and the chosen cannot give a chooser's reasons. Take the chooser out and the…
Not a ToEThe Cut Moves but Never Vanishes
The arithmetic of quantum mechanics was settled by about 1930. What the numbers are about has been argued for nearly a hundred years. One way out of the…
Not a ToEThe Form of Enlightenment and the Inversion of Gates
A list went around: ordinary people almost cannot awaken, because awakening needs four things. Extreme pain, long quiet time away from work, negating…