DeepSeek V4.1 Flash 1017 尝试通过权重解读神经网络任务
DeepSeek V4.1 Flash 1017 can also do Weight-Space Programming
Inspired by Weight-Space Programming, I tested whether DeepSeek V4.1 Flash 1017 could figure out what a small neural network does by reading its weights.
受权重空间编程（Weight-Space Programming）的启发，我测试了 DeepSeek V4.1 Flash 1017 能否通过阅读其权重来理解小型神经网络的功能。
I fed it four anonymized RL checkpoints: 51k parameters each, identical architectures, with all weights rendered as text. No metadata, network source code or execution. It received the architecture outline and four candidate tasks.
我向它输入了四个匿名化的强化学习检查点：每个包含 51k 个参数，架构相同，所有权重均以文本形式呈现。没有元数据、网络源代码或执行过程。它接收到了架构大纲和四个候选任务。
Results: CartPole ✅ Flappy ✅ Catch ✅ BallBeam ❌
结果：CartPole ✅ Flappy ✅ Catch ✅ BallBeam ❌
It picked out patterns consistent with relative bird/gap position, paddle/ball position and angle-based balancing. It confused BallBeam with CartPole an interesting mistake because both share the same observation layout and similar balancing controls.
它识别出了与相对鸟/间隙位置、球拍/球位置以及基于角度的平衡一致的规律。它将 BallBeam 误判为 CartPole，这是一个有趣的错误，因为两者共享相同的观察布局且控制方式在平衡方面相似。
Very preliminary: one run per checkpoint. This suggests some ability to interpret weights, but four trials don’t establish reliability. These were API runs with maximum reasoning effort, consuming ~1.58M tokens total.
非常初步：每个检查点仅运行一次。这表明模型具备一定程度的权重解释能力，但四次试验尚不足以确立可靠性。这些是通过 API 运行的，使用了最大推理力度，总共消耗了约 1.58M 个 token。
I may try more replicates and independently trained checkpoints, plus include the network code with task-revealing details removed. The poor LLM had to guess quite a lot it even assumed ReLU where the network uses tanh 😅
我可能会尝试更多的重复实验和独立训练的检查点，并加入去除了任务揭示细节的网络代码。这个表现不佳的大语言模型不得不猜测很多东西，甚至在网络使用 tanh 激活函数的地方它假设使用的是 ReLU 😅
Code, checkpoints and run logs: https://github.com/galatolofederico/wsp
代码、检查点和运行日志：https://github.com/galatolofederico/wsp
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力
