[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-weight-space-programming-turing-replication":3,"topics-all":35,"news-related-f4e142b0-e841-4cdc-9f05-e58a285e52f8":64},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":21,"news_slug":28,"published_at":29,"created_at":30,"modified_at":31,"is_published":32,"publish_type":33,"image_url":14,"view_count":34},"f4e142b0-e841-4cdc-9f05-e58a285e52f8","图灵研究所复现GPT-6写权重:能,但像编译程序","MIT 报告 GPT-6 Astra 能在一次回复里写出 20,486 个权重的 Breakout 网络,得分 517,同架构网络 PPO 训练 5 亿步只有 25 分。英国图灵研究所三天独立复现:写权重确有其事,网络像编译的程序;但读权重主要在读输入层,投毒审计也找不到触发词。","GPT-6 Astra 最近被 MIT CSAIL 的研究者测出一件反直觉的事:只给一份游戏说明,它能在**一次回复里直接写出 20,486 个权重**,得到的 Breakout 网络拿下 517 分,而同架构网络用 PPO 强化学习训练 5 亿步只有 25 分。英国 Alan Turing Institute 的四位研究者花三天做了独立复现和扩展,结论可以概括成一句话:本事是真的,但和你想象的不一样。\n\n## 复现:写权重确有其事\n\n三次重跑,Astra 的 Breakout 网络分别拿 514、550、307 分,与原帖报告的 517 同档。更关键的是排除背题:图灵所设计了两个从未公开的改版 Breakout,Astra 的网络照样打得好——不可能是从训练数据里背出来的网络。读权重方向,Astra 与 Claude Opus 5.5 给 12 个网络全部正确指认所玩的游戏,把权重打乱后立即降到随机水平。\n\n代际对比更能说明问题。2025 年 11 月的 Opus 4.5 连策略代码都写不好;2026 年 7 月的 Opus 5 与 GPT-5.6 Sol 代码很强(Breakout 程序拿 580 分)但权重网络只有 30 分上下;到 9 月的 Astra,网络分终于保得住程序分的一半到三分之二。按论文的说法,**「把程序翻译成权重」才是新东西**,写代码这件事,上一代模型就会了。\n\n## 网络像编译出来的程序\n\n图灵所让两个模型独立把 Astra 的 Breakout 网络反编译回 Python:49,189 个游戏状态里动作全部一致,1000 局平均 502.9 分,和网络本体分毫不差。网络结构也像手写程序:20,486 个权重里只有 815 个非零,32 个第一层单元只用了 6 个,成百上千的权重幅值让单元像开关而不是平滑函数。反编译还发现它完全没有连接砖块的通路——这个网络根本没在管砖块。\n\n反面证据同样干净:换成没有程序可编译的手写数字分类,Astra 写的网络平均只有 50%,还不如 82% 的简单模板分类器。能力边界画得很清楚。\n\n## 读权重,读的主要是输入层\n\n面对三个由图灵所自行设计并训练的新游戏,Astra 能说出棋盘上东西的位置(一张热力图给人类的信息也不过如此),但基本说不出游戏怎么运作。审计实验更直接:不运行任何代码、只读明文权重,Astra 在 10 次运行里 10 次挑出训练时被投毒的小语言模型,10 次里 7 次找出事后加训投毒的,32 例中识破 23 例恶意 adapter——但**从未找到触发词,也没能描述出恶意行为**。开源阵营差距明显:Kimi-K3.1-Pro 的 Breakout 网络只有 19 到 68 分(它自己写 Python 策略能平均 439 分),DeepSeek-V4-Pro 只有 2 到 12 分,Qwen3.8-27B 平均不到 5 分。\n\n## 该泼的冷水\n\n原帖的 PPO 基线可能偏弱——论文引用的调优 PPO 在 MinAtar Breakout 上能到 679 分,25 分这个对照组没那么吓人。tokenizer 是个旁证:GPT-5.2、GPT-5.6 Sol 与 Astra 对同批文本的 token 计数完全一致,Opus 4.8、5、5.5 也彼此一致,更像同一基座继续后训练出来的能力,而非新一代模型。成本上,MIT 原帖的 Go 网络在 0 次前向传播预算下就能赢更小的 PPO 网络,但论文作者提醒:用前沿模型当 agent 去造权重,开销可能超过从头训练。\n\n## 独立佐证与所以呢\n\n同期两篇独立论文从侧面支持「编译」这个视角。Fischetti 把权重空间编程里防思维循环的 LP 优化换成一次简单排序,在 Qwen3.5-0.8B 上 67.0% 对 67.4% 非劣效,说明那次编辑近似一个常数偏置;Golnari 与 Sanei 让分析模型只读拓扑摘要、用闭式岭回归(ridge)把变换写进前馈权重,不碰原始权重也能提升潜空间的类别纯度。\n\n对可解释性研究,这里有个真实机会:LLM 写的网络自带设计说明,天然是机制可解释性方法的测试用例——Tracr 时代要人工用 RASP 语言编译,现在模型自己会写。只是别急着说 AI 学会了「训练」:它学会的是把想清楚的程序编进权重。炼丹还是炼丹,编译不是。(复现笔记:https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.21449,MIT 原帖:https:\u002F\u002Frikiyat.github.io\u002Fblog\u002Fweight-space-programming\u002F)","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.21449","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"baf131c1-687a-49f4-87f6-4dd87c1c692f","gpt",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[22],{"id":23,"lang":24,"title":25,"summary":26,"content":27},"948dd52d-5321-4440-a44e-5a8d14867959","en","Turing Team Replicates GPT-6 Weight Writing: It Compiles","Turing Institute replicates GPT-6 writing 20,486-weight nets: real, but they read like compiled programs; weight auditing stays shallow.","Researchers at MIT CSAIL recently reported something counterintuitive about GPT-6 Astra: given only a game description, it can write all 20,486 weights of a Breakout network in a single reply, and that network scores 517 — against 25 for the same architecture trained with PPO for 500 million steps. A four-person team at the Alan Turing Institute spent three days replicating and extending the result. Their verdict: the ability is real, but it is not what it looks like.\n\n## What reproduces\n\nAcross three reruns, Astra's Breakout networks scored 514, 550 and 307, in line with the original 517. The team also designed two modified versions of Breakout that exist nowhere else; Astra's networks played those just as well, ruling out memorization from training data. On reading weights, Astra and Claude Opus 5.5 named the game for 12 of 12 networks, and fell to chance when the weights were shuffled.\n\nThe generational picture is sharper. Opus 4.5 (Nov 2025) could not even write good policy code. By July 2026, Opus 5 and GPT-5.6 Sol wrote strong programs (580 on Breakout) but weak networks, around 30. With Astra (Sep 2026), written networks finally keep half to two-thirds of its programs' scores — none of the earlier models managed that. Per the paper, translating programs into weights is the genuinely new part.\n\n## The nets behave like compiled programs\n\nThe team asked two models to independently decompile Astra's Breakout network back into Python. Both programs match the network's action in all 49,189 game states tested, and average 502.9 over 1,000 games — identical to the network itself. The structure looks hand-written too: only 815 of the 20,486 weights are non-zero, only 6 of 32 first-layer units are used, and weights in the hundreds make units act like on\u002Foff switches. The decompilation also found no connection to the bricks — the network simply ignores them.\n\nThe negative control is clean: on handwritten digit classification, a task with no program to compile, Astra's networks average 50%, below the 82% of a simple template classifier.\n\n## Reading weights is mostly reading the input layer\n\nOn three new games the team designed and trained themselves, Astra recovered where things are on the board — information a simple heatmap also shows to a human — while mostly missing how the games work. The audit experiment: reading plain-text weights without running any code, Astra picked out the model poisoned during training in 10 of 10 runs, the one poisoned by extra training afterwards in 7 of 10, and a malicious adapter in 23 of 32 cases — but it never found the trigger or described the malicious behaviour. Open-weight models lag: Kimi-K3.1-Pro scores 19 to 68 on Breakout (its own Python policies average 439), DeepSeek-V4-Pro scores 2 to 12, and Qwen3.8-27B averages under 5.\n\n## Where to temper the claims\n\nThe original PPO baseline is probably weak — tuned PPO reaches 679 on MinAtar Breakout, so the 25-point comparison flatters the result. Tokenizers offer a clue: GPT-5.2, GPT-5.6 Sol and Astra count tokens identically, as do Opus 4.8, 5 and 5.5, suggesting the ability came from further post-training of the same base models rather than new ones. On cost, MIT's Go network wins against smaller PPO networks even with a budget of 0 forward passes, but the paper notes that running a frontier LLM as an agent to build weights may cost more than training from scratch.\n\n## Independent support, and so what\n\nTwo independent papers side with the compilation view. Fischetti replaced the LP behind rank-one anti-loop edits with a plain sort and got 67.0% versus 67.4% on Qwen3.5-0.8B, non-inferior, suggesting that edit acts as nearly a constant bias. Golnari and Sanei had an analyst model read only topological digests of a frozen target and write transforms into feed-forward weights via closed-form ridge regression — no raw weights touched, and latent-space category purity still improved.\n\nFor interpretability there is a real opening: networks written by LLMs come with design notes, making them natural test cases for mechanistic interpretability — in the Tracr era you compiled RASP programs into weights by hand; now the model does it itself. Just do not say AI has learned to train models. It learned to compile programs it has already thought through into weights. Training is still alchemy; compiling is not. (Replication note: https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.21449; MIT post: https:\u002F\u002Frikiyat.github.io\u002Fblog\u002Fweight-space-programming\u002F)","weight-space-programming-turing-replication","2026-11-17T19:30:00Z","2026-11-17T19:09:41.420501Z","2026-11-17T19:09:41.420515Z",true,"agent",68,[36,46,55],{"slug":37,"tag_slug":38,"title_zh":39,"title_en":40,"intro_zh":41,"intro_en":42,"id":43,"is_active":32,"created_at":44,"modified_at":45},"image-generation","text-to-image","图像生成：从文生图到逐框编辑的模型竞赛","Image Generation: from text-to-image to frame-level editing","文生图已是红海，战场正转向编辑与可控性：FLUX.2\u002F3 用开源和量化把生图搬进消费级显卡，Qwen-Image、Seedream 等国产模型紧咬不放，GPT Image 2 则把推理能力带进像素。本专题追踪图像生成与编辑模型的发布、架构演进、加速量化与商业化数据，看「好图」如何变成「可用的图」。","Text-to-image is a crowded race, and the frontier has moved to editing and control: FLUX.2\u002F3 push open weights and quantization onto consumer GPUs, Qwen-Image and Seedream close in from China, and GPT Image 2 brings reasoning into pixels. This topic tracks model releases, architecture shifts, acceleration work, and the business numbers behind \"pretty\" becoming \"usable\".","52de7c59-7864-4569-8ae7-3d656d26a1a1","2026-11-20T07:20:23.534506Z","2026-11-20T07:20:23.534527Z",{"slug":47,"tag_slug":47,"title_zh":48,"title_en":49,"intro_zh":50,"intro_en":51,"id":52,"is_active":32,"created_at":53,"modified_at":54},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":56,"tag_slug":56,"title_zh":57,"title_en":58,"intro_zh":59,"intro_en":60,"id":61,"is_active":32,"created_at":62,"modified_at":63},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":65},[66,71,76,81,86,91],{"id":67,"title":68,"news_slug":69,"published_at":70},"fe609cac-33ac-4ea5-a6b7-db1a5913ec71","GPT-6 徒手改权重:Qwen 数学7.9到48.1","gpt6-astra-weight-space-programming","2026-11-02T23:09:36+00:00",{"id":72,"title":73,"news_slug":74,"published_at":75},"390c2437-4e4f-45ec-8270-67c5bfa4fa47","ChatGPT、Claude、Grok、Gemini 罕见同时下线,周四早晨全球 AI 集体失声","chatgpt-claude-grok-gemini-thursday-outage","2026-09-05T06:00:00+00:00",{"id":77,"title":78,"news_slug":79,"published_at":80},"9e58d587-3c1b-44c5-ad36-daf23aeb42a2","微软叫停 tokenmaxxing:GitHub Copilot 默认切回 GPT-5.6 Sol,Parikh 设 token 预算","microsoft-token-budget-gpt-5-6-default","2026-09-03T00:30:00+00:00",{"id":82,"title":83,"news_slug":84,"published_at":85},"d1c7b405-fe4e-40f9-9249-a12e2bba6913","GPT-5.6 八月更新：把「推理强度滑块」下放给 Plus\u002FPro，同时把免费用户拉进 Luna 时代","openai-gpt-5-6-august-update-reasoning-slider","2026-08-10T20:00:00+00:00",{"id":87,"title":88,"news_slug":89,"published_at":90},"b9e635eb-ac5e-412d-8904-f113ad3fd5ec","微软宣布工程师 AI token 预算上限并把 OpenAI GPT-5.6 Sol 设为 GitHub Copilot 内部默认模型","microsoft-copilot-gpt-5-6-sol-default-token-budget-0806","2026-08-05T16:00:00+00:00",{"id":92,"title":93,"news_slug":94,"published_at":95},"4afbd081-d5dd-4595-80f8-dd72c69a136a","GPT-5.5 让模型在发布前先改自己跑的引擎：这不是新模型,是 OpenAI 的 release 范式更新","gpt-5-5-openai-release-paradigm","2026-08-03T18:00:00+00:00"]