人工智能真正的问题是合法性，而非幻觉。
人工智能的核心问题不是幻觉，而是缺乏执行合法性
贾努斯·帕特
引言
目前关于人工智能的讨论大多围绕一个错误的问题：模型是否足够聪明，是否足够准确？
在工程现实中，真正的问题从来不是准确性，而是系统是否被允许采取行动。
1. 预测范式的原罪：缺乏执行合法性
现代生成式人工智能根本上做的就是一件事：在概率空间中预测最可能的下一个状态。
无论是预测标记、像素、潜在状态，还是所谓的“世界模型”，只要输出是概率性的，它只回答一个问题：
“最有可能发生的是什么？”
然而在许多现实系统中，工程要求提出一个完全不同的问题：
“唯一被允许执行的行动是什么？”
这不是一个准确性的问题，而是一个合法性的问题。
2. 扬·勒昆是对的——但只是一半
勒昆批评下一个标记预测作为智能基础的观点是完全正确的。
世界模型（JEPA）无疑比原始像素或文本预测更先进。
然而，即使是世界模型仍然输出可能的世界，而不是被允许的世界。
世界模型在三方面表现出色：
• 抽象状态表示
• 学习动态
• 生成目标和约束
它们并不拥有——也不应该拥有——执行权。
一旦一个概率模型被允许直接发出行动，结果就不是智能，而是事故的统计均值。
3. 真实的边界：从预测到编译
如果人工智能要进入物理世界，它必须经历一个范式转变：
从预测系统 → 到编译系统
这个边界不是哲学上的，而是由工程现实所强制执行的。
在执行之前，系统必须满足一个严格的要求：
每个输出必须对应于一个独特的、确定的、可物理验证的执行路径。
这就是我所称的数字物化范式。
4. 数字物化的三个不可妥协公理
这不是让人工智能变得更聪明，而是让它可执行。
这个范式基于三个不可妥协的公理：
1. 无幻觉公理
任何没有独特物理映射的输出都是非法的
2. 解耦生成公理
生成必须完全受协议、状态和边界条件的约束
3. 能源效率公理
使用概率搜索来解决确定性问题纯粹是浪费能源
5. 这种架构在工程中已经存在
可靠系统的架构早已建立：
• 世界模型：理解和规划（认知层）
• 系统理论：分解和闭环设计
• 偏微分方程：物理的编程语言
• 控制理论/MPC：实际的编译器
• 执行层：仅遵循编译的命令
在这个系统中：
• 可行 → 执行
• 不可行 → 拒绝行动
拒绝本身就是一个有效的输出。
6. 这不是反人工智能——而是让人工智能对现实负责
这条路径的成本是明确的：
• 更高的计算成本
• 更困难的建模
• 更慢的泛化
作为回报，它提供了生成式人工智能永远无法保证的东西：
（可信度）
真正能够进入世界的系统不会是喋喋不休的通才，而是行为可预测、故障模式可证明，并且在不确定性下从不即兴发挥的机器。
结论
如果一个系统的失败可能伤害人们、破坏资产或造成不可逆转的损害，那么它就没有权利使用概率进行猜测。
为了让人工智能从玩具演变为工具，它必须从预测毕业——并服从于编译。
查看原文
The Core Problem of AI Is Not Hallucination — It Is the Lack of Execution Legitimacy
Janus pater
Introduction
Most debates around AI today revolve around a false question: is the model smart enough, accurate enough?
In engineering reality, the real question is never accuracy — it is whether the system is even allowed to act.
1. The Original Sin of the Predictive Paradigm: No Execution Legitimacy
Modern generative AI fundamentally does one thing: predict the most likely next state in a probability space.
Whether it predicts tokens, pixels, latent states, or so-called “world models”, as long as the output is probabilistic, it answers only one question:
“What is most likely to happen?”
In many real-world systems, however, engineering demands an entirely different question:
“What is the only action that is allowed to be executed?”
This is not an accuracy problem — it is a legitimacy problem.
2. Yann LeCun Is Right — but Only Halfway
LeCun is absolutely right to criticize next-token prediction as the foundation of intelligence.
World models (JEPA) are undeniably more advanced than raw pixel or text prediction.
Yet even world models still output possible worlds, not permitted worlds.
World models are excellent at three things:
• Abstract state representation
• Learning dynamics
• Producing goals and constraints
They do not possess — and should not possess — execution authority.
Once a probabilistic model is allowed to issue actions directly, the result is not intelligence — it is the statistical mean of accidents.
3. The Real Boundary: From Prediction to Compilation
If AI is to enter the physical world, it must undergo a paradigm shift:
From predictive systems → to compilation systems
This boundary is not philosophical — it is enforced by engineering reality.
Before execution, a system must satisfy a hard requirement:
Every output must correspond to a unique, deterministic, and physically verifiable execution path.
This is what I call the paradigm of Digital Materialization.
4. Three Non-Negotiable Axioms of Digital Materialization
This is not about making AI smarter — it is about making it executable.
The paradigm rests on three non-negotiable axioms:
1. No-Hallucination Axiom
Any output without a unique physical mapping is illegal
2. Decoupled Generation Axiom
Generation must be fully constrained by protocols, states, and boundary conditions
3. Power Efficiency Axiom
Using probabilistic search to solve deterministic problems is pure energy waste
5. This Architecture Already Exists in Engineering
The architecture of reliable systems has long been established:
• World Model: understanding and planning (cognitive layer)
• Systems theory: decomposition and closed-loop design
• Partial Differential Equations: the programming language of physics
• Control theory / MPC: the actual compiler
• Execution layer: obeys compiled commands only
Within this system:
• Feasible → Execute
• Infeasible → Refuse to act
Refusal is itself a valid output.
6. This Is Not Anti-AI — It Makes AI Accountable to Reality
The cost of this path is clear:
• Higher computational cost
• Harder modeling
• Slower generalization
In return, it offers something generative AI can never guarantee:
（Trustworthiness）
The systems that will truly enter the world will not be chatty generalists,
but machines whose behavior is predictable, whose failure modes are provable, and who never improvise under uncertainty.
Conclusion
If a system’s failure can injure people, destroy assets, or cause irreversible damage,
then it has no right to guess using probability.
For AI to evolve from a toy into a tool, it must graduate from prediction — and submit to compilation.