DeepFEP：MOSS AGI Architecture

As 2025 comes to an end, we are still building AGI.

From a technical perspective, our greatest achievement in 2025 was the in-depth research on AONN(Aspect-Oriented Neural Network) with the assistance of AI, and the proposal of the MOSS AGI framework: DeepFEP.

In the Circle Packing diagram, each circle represents an aspect neuron.

This framework organically integrates existing theories related to artificial neural intelligence and, in a creative way, proposes a generative mechanism for neurons in artificial neural networks along with corresponding software methodologies.

At the core of DeepFEP is AONN : a new type of neural network in which AOP weaving corresponds to neural spiking, runtime structural generation serves as the learning mechanism, and event-driven reasoning functions as forward propagation. It is not software that merely simulates a neural network; rather, it is a software system that exists in the form of a neural organization.

What AONN learns is not parameters, but causal structure. ANN learns weights, Transformers learn distributions, while AONN learns when to fire which spike.

In 2025, I also realized that causal emergence and its 2.0 are quantitative theories for describing emergence. Like Integrated Information Theory and the Free Energy Principle (FEP), they describe and characterize existing systems, but do not provide a generative mechanism for the system itself (neurons and their networks). Fortunately, FEP offers a verifiable learning framework, which includes a world model, observation, prediction, and state transitions. Most importantly, active inference provides a testable mechanism: the goal of generation or learning is to minimize free energy.

The world model in DeepFEP can be the Web, financial markets, prediction markets, or workflows — as long as the following conditions exist:

• Observations: data / event streams
• Actions: applied operations
• World transitions: actions change future observable outcomes
• Verifiable feedback: computable scores (free energy / error / cost)

Such systems can be treated as the external world of a world model (or as an interface layer to the external world).

If DeepFEP is compared to an artificial brain, then the LLM serves as the language cortex, while AONN functions as the prefrontal cortex.

Looking Ahead to 2026

2026 will be the year of implementing and releasing DeepFEP. Combined with different world models, it will give rise to diverse product forms — perhaps even products similar to ChatGPT. The most important difference is that DeepFEP understands the world model through learning, thereby driving AONN to generate behavior and, in turn, driving the LLM language center to produce tokens.

2025年即将过去，我们依然在建设中。

技术上，我们在2025最大的成就是在AI助力下，对AONN进行了深入研究，并提出MOSS AGI框架：DeepFEP。

Circle Packing图中每个圆圈是一个aspect 神经元

这个框架有机地整合现有人工神经智能相关理论，并创造性地提出人工神经网络中神经元的生成机制与软件方法。

DeepFEP的核心是AONN（Aspect-Oriented Neural Network）：一种以 AOP 织入（weaving）为神经放电（spiking）、以运行时结构生成为学习机制、以事件驱动推理为前向传播的新型神经网络。它不是像神经网络的软件，而是以神经组织方式存在的软件系统。

AONN 学的不是参数，是因果结构。ANN 学的是权重， Transformer 学的是分布， AONN 学的是“什么时候该放哪一个电”。

2025年我也发现因果涌现以及2.0是定量描述涌现的理论，与现有的信息整合理论，最小作用量原理FEP一样，都是对现有系统的描述与刻画，没有给出系统（神经元及其网络）的生成机制。不过好在FEP给出了可验证的学习框架，其中包含世界模型，观察，预测与状态变换，最重要的是主动推理给出一种可验证的机制，即生成或学习的目标是降低自由能。

DeepFEP的世界模型可以是Web互联网，交易市场，预测市场以及工作流，只要存在：

• 观测 ：数据/事件流
• 行动 ​：施加的操作
• 世界转移：行动会改变未来可观测结果
• 可验证反馈：可计算评分（自由能/误差/代价）

就可以把它当作世界模型的外部世界（或外部世界的接口层）。

如果把DeepFEP比做人工大脑，LLM在DeepFEP中是语言皮层，而AONN则是前额叶。

展望2026年

2026将是实现与发布DeepFEP的一年，结合各种世界模型会有不同的产品形态，或许是类似ChatGPT的产品，最重要的区别可能是DeepFEP通过学习而理解世界模型，从而驱动AONN产生行为，驱动LLM语言中枢产生token。