核心结论：MIT CSAIL 团队确实让 OpenAI 的 GPT-6 Astra 直接读取并修改了阿里开源模型 Qwen3.5-0.8B 的权重，仅改动约 0.016% 的参数（131072 个），就把 MATH-500 数学成绩从 7.9% 提到 48.1%（约 6.1 倍），全程没有训练、没有梯度下降——这项「权重空间编程（Weight-Space Programming）」实验是真实的。但「真有网上说的那么牛吗」的答案要打个折：Astra 改掉的本质是修好了一个导致推理循环的 Bug，而不是凭空给模型加能力；对照组 Gemma 改不动、识别「谄媚模型」时读不懂权重，都说明这项能力目前既不稳定、也不通用，且研究尚未同行评审。该信的是「AI 能读懂并编辑神经网络的权重」这条路被首次走通；该质疑的是把它当成「改几个权重就能让模型变强」的万能神技。
目录
- 一句话结论
- 事件回顾（Event Recap）：MIT 让 GPT-6 Astra 直接改 Qwen 权重
- 权重空间编程是什么（Weight-Space Programming）
- 关键数字（Key Numbers）：0.016% 参数与 6 倍提升
- 它到底改了什么（What Was Fixed）：修复推理循环
- 核心质疑一（Question 1）：是能力提升，还是修好一个 Bug
- 核心质疑二（Question 2）：对照组 Gemma 改不动说明什么
- 核心质疑三（Question 3）：可复现性与实验设计问题
- 失败案例（Failure Case）：读懂权重还不是稳定能力
- 完整可运行代码（Runnable Code）：最小权重编辑演示
- 对照表：症状 → 原因 → 解法
- 判断收尾
1. 一句话结论
先建立一个直觉 🎭：权重空间编程像「给大脑做微创手术」——不重新上学（不训练），只是直接定位大脑里一根错连的神经（权重），用激光烧掉它（改 0.016% 参数），人的行为就正常了。但这里的关键区别是：手术只是修好了「本来会转圈」的毛病，并没有让你变得更聪明。
所以读这条新闻，记住三件事：
- 真实：实验是真的，0.016% 参数、7.9% → 48.1% 都是实打实的数字。
- 打折：提分主要来自「修复推理循环」，相当于拆掉一个卡住能力的锁，不是解锁后能力又涨了。
- 存疑：未同行评审、只有 OpenAI 模型能复现、换个模型就失灵，目前更像「初步火花（sparks）」，不是成熟技术。
一句话：AI 能编辑另一个 AI 的权重，这条路第一次被走通，但离「改几个参数就让模型变强」还很远。
2. 事件回顾（Event Recap）：MIT 让 GPT-6 Astra 直接改 Qwen 权重
2026 年 11 月 2 日，MIT CSAIL（计算机科学与人工智能实验室）的四位研究者公布了一组实验：
- 他们让 OpenAI 的前沿模型 GPT-6 Astra 直接读取阿里开源模型 Qwen3.5-0.8B 的神经网络权重。
- Astra 检查内部权重后，找到了导致模型反复陷入推理循环（reasoning loop）的关键问题，然后直接修改权重修复它。
- 整个过程没有重新训练、没有梯度下降，只有编辑——这在「模型改进必须靠微调（fine-tuning）」的常识下，是一条全新的路。
消息一出立刻引爆大模型圈：0.8B 的开源小模型，被一个更大更强的模型「手把手」改权重，数学能力暴涨。这正是「AI 指挥 AI」从任务调度（我在agent为什么需要自进化：从静态模型到经验驱动的持续成长里聊过模型如何持续成长）深入到「AI 修改 AI 的神经网络本体」的一步。
事件参考资料：
- MIT让GPT-6 Astra直接修改Qwen权重，数学成绩飙升 -- 36氪
- MIT研究者提出「权重空间编程」：GPT-6 Astra仅修改Qwen3.5-0.8B约0.016%权重 -- IT之家
3. 权重空间编程是什么（Weight-Space Programming）
这是整件事的技术核心，值得拆开讲。
传统路线：SFT 微调（supervised fine-tuning，监督微调）——拿一批标注数据，跑反向传播更新权重。在同样的模型为什么本地和云API差别那么大：量化精度、聊天模板与采样参数的三大元凶中我们讲过，权重在推理时还会被量化压缩、改变行为——而微调的本质是「重新上学」，代价是要数据、要算力、要训练流程。
新路线：权重空间编程（Weight-Space Programming）——把「模型权重」本身当成「程序」来读写：
- 读：Astra 把 Qwen 的权重张量当作数据直接查看，分析神经元之间的连接模式。
- 诊断：在权重空间里定位「导致问题的区域」，比如某个神经元组合让模型陷入循环。
- 写：只修改定位到的少量权重值，让行为改变，不跑梯度、不更新全局。
打个比方 🎭：SFT 是「把整本书重写一遍来修正错字」，权重空间编程是「只找出错的几页，用笔直接改」——前者动辄改全本，后者只动千分之几。
作者给它起名「Program-as-Weights」（程序即权重）的延续思路，核心哲学是：神经网络的权重不是不可解释的黑盒，而是可以被另一个模型读写的「可编程空间」。参考其后续工作，这一范式称为 fuzzy-function programming（模糊函数编程）。
技术参考资料：
- Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights -- 作者博客 ⭐值得阅读
4. 关键数字（Key Numbers）：0.016% 参数与 6 倍提升
这是新闻里最抓眼球的部分，全部来自作者公开的实验数据（截至 2026-11-03，尚未同行评审，DOI 暂缺）：
两个值得注意的点：
- 改动幅度极小：0.016% 意味着 8 亿参数里只动了 13 万，相当于一本 800 页的书只改了 1.3 页——这在传统训练里不可想象。
- 数学题是「被卡住」而不是「不会做」：GSM8K 从 13.8% 到 67.1% 的跳跃说明，模型本身会算，但推理循环让它「算不完」——修好循环后，已有能力释放出来了。
数字参考资料：
5. 它到底改了什么（What Was Fixed）：修复推理循环
「改权重」听起来玄，但 Astra 到底修了什么，其实很清楚：它修复了一个让模型反复绕圈的推理循环（reasoning loop）。
- Qwen3.5-0.8B 的模型卡（model card）自己就警告过这个缺陷：模型在推理时容易陷入循环，一直重复生成同样的 token，直到撞上长度上限。
- Astra 在权重里定位到导致循环的神经元组合，直接编辑修复。
- 修复前，Qwen 有 93% 的答案会撞上 token 上限（也就是「卡死在循环里」）；修复后循环消失，模型能把推理跑完。
这解释了为什么数学成绩暴涨：不是模型突然会做难题了，而是它终于能把本来会做的题「做完」了。这就像考试时被一道题卡住反复涂改、最后交白卷的考生——你帮他改掉「强迫症」这个毛病，他原来会的题自然能答出来。
⚠️ 这里有个重要推论：权重编辑移除的是「掩盖能力」的障碍，而不是添加模型本不具备的能力。这也是全文最该记住的一句话。
推理循环参考资料：
6. 核心质疑一（Question 1）：是能力提升，还是修好一个 Bug
「6 倍提升」的标题很震撼，但社区很快提出一个直击要害的质疑：你提升的是「能力」，还是只是修好了一个 Bug？
质疑者的三个论据：
- 提分主要与减少循环有关：数学题从 7.9% 到 48.1% 的涨幅，大部分来自「能算完了」，而不是「算得更好」。
- 与已有方法相似：Astra 的思路和现有的「激活引导（activation steering）」方法相近——都是通过干预内部激活来改变行为，权重空间编程是它的更激进版本。
- 提示词已经暗示答案：作者给 Astra 的长提示词里明确提到「模型存在循环问题」，等于先告诉它「病在哪里」，再让它去修——这比「让它自己发现病灶」要容易得多。
因此有人提出一个必须做的对照实验：直接关闭 Qwen 的思考模式（关掉 reasoning），看看数学成绩能不能取得类似提升——如果调整推理设置就能大幅减少循环，那「改权重」的增量价值就要重新评估。
这个质疑没有摧毁实验，但把叙事从「能力暴涨」拉回「修复了一个已知缺陷」——两者在技术价值上完全不同。
质疑参考资料：
- GPT-6 Astra 爆改 Qwen 大脑：只动 0.016% 参数，数学成绩疯涨 -- 爱范儿
- 7.9% to 48.1%: LLM rewrote Qwen weights, no training -- DEV Community ⭐值得阅读
7. 核心质疑二（Question 2）：对照组 Gemma 改不动说明什么
作者自己做了对照组实验，结果反而最有说服力——这个能力在「没有循环问题」的模型上失效了：
- 对照组用 Google 的 Gemma 3 270M，一个很少陷入循环的模型（只有 8.5% 的答案会撞上 token 上限，而 Qwen 是 93%）。
- Astra 对 Gemma 尝试编辑，最好的结果只让 MATH-500 提升了 0.9 分，作者自己都报告为「未确立（not established）」。
- 也就是说：当模型没有「循环」这个病可修时，Astra 的权重编辑几乎无效。
这个对照实验反而是全文最有信息量的一组数字：
- 它证实了「修循环」就是本次提分的主要机制——Qwen 有循环病 → 修好 → 暴涨；Gemma 没这个病 → 没东西可修 → 不涨。
- 它同时证伪了「改几个权重就能让任意模型变强」的说法——权重编辑不能无中生有地加能力。
对照实验参考资料：
8. 核心质疑三（Question 3）：可复现性与实验设计问题
除了「修 Bug vs 提能力」，还有几项实验设计的硬伤，决定了现在不能把它当通用技术：
- 未同行评审、无 DOI：截至本文写作（2026-11-03），研究只是初步实验报告（preliminary findings），还没经过同行评审，数字未经第三方独立复现。
- 只用了一个模型做「编辑者」：所有实验的编辑方都是 OpenAI 的 gpt-6-astra，其他家的前沿模型（Claude、Gemini、DeepSeek 等）是否具备同样的「读懂权重」能力，完全未知。
- 采样设置不统一：公开日志显示，各实验的 reasoning-effort（推理努力程度）设置从 high、xhigh 到 max 不一——这意味着部分成绩提升可能掺杂了「更努力推理」的因素，不全是权重编辑的功劳。
这三条合起来，让「权重空间编程」目前更像**「一个模型身上观察到的一次现象」**，而不是「一条可复现的方法论」。
可复现性参考资料：
9. 失败案例（Failure Case）：读懂权重还不是稳定能力
最能说明「这项能力不成熟」的，是作者自己记录的一处失败：
- 在识别「谄媚模型」（即会迎合用户观点、而不是给出真实回答的模型）的实验里，Astra 面对两个候选模型，只比较权重时无法作出判断。
- 直到作者允许 Astra 观察模型在特定文本上的注意力模式（attention pattern），它才识别出差异。
- 换句话说：「直接读懂权重」目前不是稳定、通用的能力——它在修循环这种「结构性缺陷」上有效，但在「判断模型性情」这种需要语义理解的场景上会失灵。
这个失败案例的重要性在于：它划定了本次实验的能力边界。权重空间编程不是「读心术」，而更像「结构工程师」——能看出哪根梁错位了（结构性故障），但看不出这栋楼主人的性格。
失败案例参考资料：
10. 完整可运行代码（Runnable Code）：最小权重编辑演示
下面是一段复制粘贴就能跑的 numpy 代码，模拟权重空间编程的最小闭环：构造一个「会循环」的小网络 → 定位问题权重 → 直接编辑修复 → 输出对比。不依赖任何第三方库。
运行：python weight_edit_demo.py（只需 numpy）。
import numpy as np
# ---------- 构造一个小网络，模拟「推理循环导致输出退化」 ----------
def forward(x, W, b):
"""前向传播：线性层 + ReLU，返回输出"""
h = np.maximum(0, x @ W + b) # 隐藏层：ReLU 激活
return h.sum(axis=1) # 汇总成标量输出（模拟打分）
# 8 个输入特征 → 4 个隐藏神经元
X = np.random.randn(8, 8) # 8 个样本、8 维特征（固定随机种子下可复现）
W_bad = np.random.randn(8, 4) # 初始权重
b_bad = np.zeros(4) # 偏置
# 人为制造「循环病」：让第 2 号神经元的输入权重全部为 0，输出恒为 0（坏死）
# 对应真实场景里「某个神经元组合导致模型反复绕圈」
W_bad[:, 2] = 0.0
# ---------- 诊断：算出每个神经元的贡献，定位「病根」 ----------
def diagnose(x, W):
"""通过前向激活定位坏死神经元"""
act = x @ W # 每个神经元收到的输入信号
contribution = np.abs(act).mean(axis=0) # 平均激活强度
dead_idx = np.where(contribution < 1e-6)[0] # 激活几乎为 0 = 坏死
return dead_idx, contribution
dead, contrib = diagnose(X, W_bad)
print("坏死神经元索引：", dead) # 应输出 [2]
# ---------- 权重编辑：只重写坏死神经元的权重（模拟 Astra 直接改权重） ----------
W_fixed = W_bad.copy() # 复制一份，只改定位到的部分
np.random.seed(1) # 固定新权重
W_fixed[:, 2] = np.random.randn(8) # 只重写第 2 列（8 个参数，占全部 32 个的 25%）
# ---------- 对比修复前后输出 ----------
y_before = forward(X, W_bad, b_bad) # 修复前：坏死神经元贡献为 0，输出偏低
y_after = forward(X, W_fixed, b_bad) # 修复后：所有神经元正常激活
print(f"修复前输出均值：{y_before.mean():.3f} -> 修复后：{y_after.mean():.3f}")
# 结论：只改了 8 个参数，整个网络的输出分布就恢复健康（类比 Qwen 修循环）
怎么读这段代码：diagnose() 用「前向激活强度」定位坏死神经元——对应第 5 节「Astra 在权重里定位循环病灶」；W_fixed[:, 2] = np.random.randn(8) 只重写定位到的 8 个参数——对应「只动 0.016% 权重」的机制；修复前后输出均值的变化，就是「修好结构性缺陷后能力释放」的最小演示。真实场景的差别只在于：病灶是「循环」、规模是 131072 个参数、编辑者是 GPT-6 Astra 而已。
11. 对照表：症状 → 原因 → 解法
12. 判断收尾
什么时候该信这条新闻：作为「AI 能读懂并编辑神经网络权重」的首例证据，它值得关注——这可能是「模型可解释性」和「模型修复」的新入口，尤其在修「结构性缺陷」（循环、死锁、退化）这类问题上，比重训成本低几个数量级。什么时候该质疑：把它当成「改几个权重就能让任意模型变强」的神技、或急着复现到生产环境——目前它未同行评审、只在单一编辑模型上验证、无法添加模型本不具备的能力，成熟度离 SFT 差着十万八千里。判断标准只有一条：你要的是「修好一个已知的病灶」，还是「让模型学会新东西」——前者可以期待，后者还远。
本专栏《LLM 大模型技术专栏》从模型原理、训练后处理一路讲到推理优化与工程落地，每篇都配一段可直接运行的代码，每周更新一篇。想按顺序读完，可在专栏页订阅。
参考资料（综合）：
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Core conclusion: The MIT CSAIL team did allow Open AI’s GPT-6 Astra to read and modify the weight of the Qwen3.5-0.8B open source model Qwen3.5-0.8B, changing only the parameter of 0.016% (131072), and referring to MATH-500 mathematical achievement from 7.9% to 48.1% (approximately 6.1 times) without training and without a gradient – the experiment “Weight-Space Programme” is real. But the answer to the question "Is it true that there's such a cow on the Internet?" is a discount: Astra's modification is essentially about fixing a bug that leads to a loop of reasoning, rather than adding power to a model by empty space; and the control group Gemma's inability to read weights when identifying the "smuggle model" suggests that this capability is currently neither stable nor universal, and that research is not peer-reviewed. The letter is the first time that the path of "AI can read and edit the weight of the nervous network" has gone through; it is a question of using it as an almighty technique that "changes a few weights to make models stronger."
Contents
- One conclusion.
- Event Recap: MIT allows GPT-6 Astra to change Qwen weights directly
- Weight-Space programming
- Key Numbers: 0.016% and 6 times higher
What Was Fixed: Fixing the logic loop.
Question 1: Improved capacity or fixed Bug
- Question 2: Control Group Gemma speaks for what
Question 3: Recoverability and experimental design issues
- Fairure Case: reading weight is not stable.
- Full Runable Code: Minimum weight edit presentation
- Checklist: Symptoms, causes, solutions
- Finish your judgment.
1. Conclusion Comment
First, create an instinct that the weight space programming is like "micro-surgery of the brain" — not going back to school (no training), but simply directs a faulty nerve (weight) in the brain, burning it with lasers (modified by 0.016% parameters) and normal behavior. But the key difference here is that the operation just fixed the "who would've turned" problem and didn't make you smarter.
So read the news and remember three things:
- Real: The experiments are real, 0.016% parameters, 7.9% → 48.1% figures.
- Depreciation: the points come mainly from the "repair of the reasoning loop", which is equivalent to the removal of a locked capacity, rather than the ability to unlock it, which has increased again.
- Doubts: Without peer review, only Open AI models can be reproduced, and a different model will fail, more like "prior sparks" than mature technology.
One sentence: AI is able to edit the weight of another AI, which is the first way through, but it's a long way from "changing a few parameters to make the model stronger."
2. Event Recap: MIT allows GPT-6 Astra to change Qwen weights directly
On November 2, 2026, four researchers of MIT CSAIL published a set of experiments:
- They let Open AI's front model GPT-6 Astra read directly the neuronet weight of Qwen3.5-0.8B.
- After examining the internal weights, Astra identified the key issues that led the model to fall into the chain of reasoning and then directly modified the weights to fix it.
- The whole process was not retrained, the gradient was not down, only the editor, which was a completely new path, in the common sense that "model improvement must be fine-tuning".
LLM loops detonated as soon as the news came out: the small open source model of 0.8B was changed by a larger model, step by step, and mathematical capacity surged. This is exactly what "AI command AI" is doing from the mission. From static models to experience-driven growth, talking about how models are growing on a sustainable basis, to "AI modify AI's neuronetic body."
Event reference:
- MIT has asked GPT-6 Astra to change Qwen weights directly. Mathematically high - 36 kryptons
- MIT researchers propose "weight space programming": GPT-6 Astra only changes Qwen 3.5-0.8B about 0.016% weight -- IT House
3. What is the weight-space programming
It's the technical core of the whole thing and it's worth talking about.
Traditional route: SFT fine-tuned (supervised fine-twining, supervision fine-tuning) - take a load of marked data and run backwards to disseminate the updated weight. In the same model, why is the difference between local and cloud API so big: quantization precision, chat templates and sampling parameters, as we have said, weights are also reduced by quantization in reasoning, changing behaviour — and fine-tuning is essentially “re-entry”, at the cost of data, calculation, training processes.
New route: Weight-Space programming — reading and writing with the model weight itself:
- Read: Astra uses the weight of Qwen as a direct view of the data and analyzes the pattern of connection between neurons.
- Diagnosis: positioning the "problem area" in the weight space, for example, a combination of neurons that put the model into circulation.
- Writing: Only a small number of weights that are positioned are modified to change behaviour, do not run gradients and do not update the global.
Example: SFT is "Rewrite the whole book to correct the wrong word" and the weight space programming is "Just find the wrong pages and change them directly with a pen" — the former is changing the whole book very quickly, while the latter is just a few thousand.
The core philosophy of the author, who named it "Program-as-Weights" (procedure or weight), is that the weight of the nervous network is not an inexplicable black box, but a "programmable space" that can be read and written by another model. With reference to its follow-up, this paradigm is called fuzzy-function programming.
Technical reference:
-Sparks of Weight-Space Programme: LLMs can usefully read & write New Networks--
Key Numbers: 0.016% and 6 times higher
This is the most eye-opening part of the news, all from experimental data published by the author (as of 2026-11-03, not yet peer-reviewed, DOI not available):
Two points worth noting:
- The change was minimal: 0.016 per cent meant that only 130,000 of the 800 million parameters had been moved, and the equivalent of a 800-page book had changed only 1.3 pages — which was inconceivable in traditional training.
- The mathematical question is "to be stuck" rather than "not to do": the leap of GSM 8K from 13.8% to 67.1% suggests that the model itself counts, but the logic loop makes it "infinite" — it is capable of being released once it is fixed.
Number reference:
5. What did it change?
"Reforming weights" sounds like a good idea, but what Astra did is clear: it repaired a chain of reasoning that allowed the model to circle repeatedly.
- The model card of Qwen3.5-0.8B warns itself of this defect: the model is prone to a cycle of reasoning, reproducing the same token until it hits the length limit.
- Astra locates the neurons that cause the cycle in the weight, edits the repairs directly.
- 93% of Qwen's answers will hit the token ceiling (i.e., "calculated in the cycle"); the cycle disappears after the restoration, and the model can run the reasoning.
This explains the dramatic rise in mathematical performance: it's not that the model suddenly makes a problem, but that it finally "finished" what it would have done. It's like the test guy who's been stuck in the exam with a stick-up and a white paper -- you help him get rid of the problem of "compulsory disease," and the question he was supposed to have answered.
There is an important inference here: the weight editor removes the barrier of "covering the capacity" rather than adding the capability that the model does not have. This is also one of the words that the whole text should bear in mind.
Logic cycle references:
6. Question 1: capacity enhancement or repair of a Bug
The title of "six times higher" was shocking, but the community quickly asked the question of whether it was "capacity" or just fixing a bug?
Three arguments of the questionr:
- The points are mainly related to the reduction of cycles: the increase in mathematics from 7.9 per cent to 48.1 per cent is mostly due to "finished" rather than "better".
- Similar to existing methods: Astra's approach is similar to the existing "activation direction" approach, which changes behaviour through internal activation, and weighting spatial programming is its more radical version.
prompt has hinted at the answer: the author's explicit reference to prompt to Astra's long "models have circulation problems" is to tell it "where the disease is" and then to fix it — it's easier than "let it find its own stove."
So a comparison experiment that has to be done is proposed: turning off Qwen’s thinking pattern directly, and see if mathematics can achieve a similar improvement – if the logic settings are adjusted to reduce the cycle significantly, then the incremental value of “reweighting” needs to be reassessed.
The question was not to destroy the experiment, but to pull the narrative back from "absorbing capacity" to "repairing a known flaw" – the technical value of which was completely different.
Question references:
- GPT-6 Astra Demolition Qwen Brain: only 0.016% parameter, mathematically high - Avenger
-79% to 48.1%: LLM rewent Qwen Weights, no training-- DEV Community worth reading
7. Question 2: Control Group Gemma speaks for what
The author did a control group experiment of his own, and the result was the most convincing - the ability to fail on the "no loop problem" model:
- The control group uses the Google Gemma 3 270M, a model that is rarely cyclical (only 8.5% of the answers will hit the token ceiling, while Qwen is 93%).
- Astra tried editing for Gemma, the best result was to raise MATH-500 by 0.9 points, and the authors themselves reported "not established".
- That is, Astra's weight editing is almost ineffective when the model is not recycleable.
This contrasting experiment is the most informative set of numbers:
- It confirms that the main mechanism for this scale-up is Qwen, which has a circulatory disease that fixes and booms; Gemma has no disease, nothing to fix and never rise.
- It is at the same time perceiving the claim that "a few weights can make an arbitrary model stronger" — a weight editor cannot add to the power of a single person.
Reference to the experiment:
8. Core question three (Question 3): Recurrence and experimental design issues
In addition to repairing Bug vs' capacity, there are a few hard wounds from experimental designs that make it impossible to use as a general technology:
- No DOI: As of this writing (2026-11-03), the study is a preliminary pilot report (preliminary findings), which has not been peer-reviewed and the figures have not been reproduced independently by third parties.
- Only one model has been used as an "editor": the editor of all the experiments is Open AI gpt-6-astra, and it is completely unknown whether other front models (Claude, Gemini, DeepSeek, etc.) have the same "read and understand weight" capability.
- Sampling settings are not uniform: the public log shows that the experiments are set from high, x high to max -- This means that part of the improvement may have mixed the elements of “more effort to reason”, not all thanks to the weight of the editor.
Together, they make the "weight space programming" more like ** "a single phenomenon observed in a model"** than "a replicable methodology" **.
Recurring reference:
9. Failure cases (Failure Case): reading weights is not stable
One of the failures documented by the author himself is the following:
- In an experiment to identify "successful models" (i.e. models that meet user views rather than give true answers), Astra cannot judge when comparing weights with two candidate models.
- It was not until the author allowed Astra to observe the attention mode on a given text that it recognized the difference.
In other words, "direct reading and understanding of weights" is not a stable, universal capability at the present time — it works in repairing the “structural flaws” of the cycle, but it fails in the context of the “judgment of models” that requires semantic understanding.
The significance of this failure is that it demarcates the capacity boundaries of this experiment. Weight space programming is not "heart-reading" but more like "structural engineers" — one that sees the wrong beam (structural failure), but not the character of the owner of the building.
References to failures:
Full runable code (Runnable Code): minimum weight edit presentation
The following is a copy of the paste that allows you to run a numpy code that simulates the smallest closed loop of the weight space programming: to construct a small network that "cycles" → positioning problem weight → to edit directly the fixation output comparison. Not dependent on any third-party bank.
Run: python weight edit demo.py.
import numpy as np
# ----------------------------------------------------------------------------------------------------------------------------------- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- --
def forward(x, W, b):
Forward transmission: linear layer + ReLU, return output ""
h = np.maxium(0, x@W +b) # Hide Layer: ReLU activated
Return h.sum (axis = 1) # Sum to cursor output (simulation scores)
# 8 input features # 4 hidden nerves Dollar
X = np.random.randn (8, 8) # 8 sample, 8-D features (repeated under fixed random seeds)
# Initial rights Heavy
b bad =np.zeros(4) #
# Man-made circulatory disease: make the input weight of neuron number 2 zero and output zero constant
# In the real scene, "Some neuron combination causes the model to circle repeatedly"
W_bad[:, 2] = 0.0
# ----------------------------------------------------------------------------------------------------------------------------------------------- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- --
def diagnose(x, W):
""""""""""""""""""""""""
Act = x
# Average activation strength
Dead idx = np.where (contribution < 1e-6) [0] # activation almost 0 = bad death
return dead_idx, contribution
dead, contrib = diagnose(X, W_bad)
# Should output [2]
# -- - weighted editor: rewrite only the weight of dead neurons (simulate Astra direct reweighting) - -
W fixed = W bad.copy()
# Fixed New Right Heavy
W fixed[: 2,] = np.random.rann (8) # Rewrite column 2 (8 parameters, 25% of all 32)
# ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ -----------------
y before =forward(X, W bad, b bad)
y after = forward (X, W fixed, b bad)
print(f) "Equal output before repair: {y before.mean():3f} -> After repair: {y after.mean(:3f}”)
# Conclusion: Only 8 parameters have been modified, and the distribution of output across the network restores health (like Qwen repair cycle)
How do you read the code: Diagnose() locates the dead neurons with "a forward activation strength" - corresponding to Section 5 "Astra locates the circulatory stoves within weights"; W fixed[,2] = np.random.randn (8) rewrites only the eight parameters – corresponding to the mechanism of "only 0.016% weights"; changes in the average output values for repairing them are the minimum demonstration of "the ability to release after repairing structural defects". The difference between the real scenes is that the stoves are "cycled" and 13,1072 parameters, and the editor is GPT-6 Astra.
11. Comparison table: symptoms causes
12. Closure of judgement
When is this news coming up: as the first evidence that "AI can read and edit the weight of the neural network", it deserves attention — this may be a new entry point for “model interpretability” and “model repair”, especially with regard to such issues as repairing “structural defects” (cycles, dead locks, degradation), which are several orders of magnitude lower. When should we question it as a magic technique that "changes a few weights to make any model stronger" or rushes back into the production environment? At present, it is not peer-reviewed, only validated on a single editing model, does not have the capacity to add models, and is less mature than SFT by 18,000 miles. There's only one test: whether you want to fix a known stove, or whether you want a model to learn something new — the former can expect, the latter is far away.
This column, LLM LLM Technical Column, addresses the principles of modelling, post-training processing, to the optimization of reasoning and project landings, each with a direct running code, updated weekly. To read in order, you can subscribe to the column page.
References (comprehensive):
Auto-fetched and translated by AI Atlas. All rights belong to the original author.