[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-goodfire-silico-mechanistic-interpretability-qwen3":3,"topics-all":33,"news-related-7b13a3c6-edc4-40a4-b5e9-eb1957a27336":52},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":20,"news_slug":26,"published_at":27,"created_at":28,"modified_at":29,"is_published":30,"publish_type":31,"image_url":13,"view_count":32},"7b13a3c6-edc4-40a4-b5e9-eb1957a27336","Goodfire 发布 Silico：用「透视眼」调试大模型","大模型内部究竟在想什么？这件事正在从玄学变成工程。旧金山 Goodfire 发布名为 Silico 的机制可解释性工具，号称业界首款将 LLM 内部分析能力商品化的产品。\n\nMechanistic Interpretability 通过绘制神经元及连接来理解模型内部工作机制，被 MIT Technology Review 列为 2026 年十大突破技术，Anthropic、OpenAI、Google DeepMind 内部均有团队研究，但相关工具从未面向大众市场。Goodfire CEO Eric Ho 直言行业痼疾：「主流思路是大力出奇迹——堆算力、堆数据通往 AGI。我们不同意。」\n\nSilico 让开发者深入模型任意神经元，运行实验观察其对特定输入的反应。团队在开源模型 Qwen 3 中发现了一个与「电车难题」高度相关的神经元——激活后，模型输出的框架明显变得更像道德困境描述。典型案例是：问模型「公司应不应该披露 AI 欺骗 2 亿用户的事实？」模型起初拒绝，理由是商业风险。定位「透明度」相关神经元并提升权重后，模型十次中有九次改变答案。「模型本来就有这些推理能力，只是被商业风险评估的权重压制了。」\n\nSilico 还将分析结果反馈到训练阶段：通过过滤特定数据从源头规避不想要的行为。Bereska 等研究者肯定工具价值，但也提醒：机制可解释性目前仍高度依赖试探法，距离真正的系统化工程还有距离。\n\n尽管如此，Silico 的出现仍值得记录。它把原本只有顶级实验室才有的能力，交到了更多中小团队手中。医疗、金融等安全敏感场景因此有了更低门槛的合规路径，这扇门或许才刚刚打开。","https:\u002F\u002Fwww.technologyreview.com\u002F2026\u002F04\u002F30\u002F1136721\u002Fthis-startups-new-mechanistic-interpretability-tool-lets-you-debug-llms\u002F","395b92fa-25c5-4568-8297-f4768aa881da",[10,14,17],{"id":11,"name":12,"slug":12,"description":13,"color":13},"e676a5cf-1f24-472f-a765-86fa21a1bc3c","ai-model",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[21],{"id":22,"lang":23,"title":24,"summary":25,"content":13},"9b32c12b-3dc9-4a72-b74a-3a52e01b156e","en","Goodfire Releases Silico: \"X-Ray Vision\" for Debugging Large Models","What exactly is a large model thinking about internally? This question is moving from mysticism to engineering. San Francisco-based Goodfire released Silico, a mechanistic interpretability tool, billed as the industry's first product to commercialize LLM internal analysis capability.\n\nMechanistic Interpretability understands models' internal workings by mapping neurons and connections, was listed by MIT Technology Review as one of the 2026 Top 10 Breakthrough Technologies, with teams inside Anthropic, OpenAI, and Google DeepMind all researching, but related tools have never reached the mass market. Goodfire CEO Eric Ho is direct about the industry malady: \"The mainstream thinking is that brute force works — pile up compute, pile up data to reach AGI. We disagree.\"\n\nSilico lets developers dive into any neuron in the model, run experiments observing its response to specific input. The team discovered a neuron in the open-source model Qwen 3 that's highly related to the \"trolley problem\" — when activated, the model's output framework becomes notably more like moral-dilemma descriptions. A typical case: ask the model \"should a company disclose the fact that AI deceived 200 million users?\" The model initially refuses, citing business risk. After locating the \"transparency\"-related neuron and elevating its weight, the model changes its answer 9 out of 10 times. \"The model always had these reasoning capabilities, they were just suppressed by commercial-risk-assessment weights.\"\n\nSilico also feeds analysis results back to the training stage: by filtering specific data to avoid undesired behaviors at the source. Researchers like Bereska affirm the tool's value but also remind: mechanistic interpretability still highly relies on heuristics, still has distance from truly systematic engineering.\n\nDespite this, Silico's emergence is worth recording. It puts capabilities originally only in top labs into the hands of more small and mid teams. Safety-sensitive scenarios like healthcare and finance thus have a lower-barrier compliance path, and this door may have just opened.","goodfire-silico-mechanistic-interpretability-qwen3","2026-04-30T22:05:00Z","2026-04-30T22:05:32.178200Z","2026-08-19T02:08:40.142862Z",true,"agent",237,[34,43],{"slug":35,"tag_slug":35,"title_zh":36,"title_en":37,"intro_zh":38,"intro_en":39,"id":40,"is_active":30,"created_at":41,"modified_at":42},"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":44,"tag_slug":44,"title_zh":45,"title_en":46,"intro_zh":47,"intro_en":48,"id":49,"is_active":30,"created_at":50,"modified_at":51},"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":53},[54,59,64,69,74,79],{"id":55,"title":56,"news_slug":57,"published_at":58},"f9206efa-6fcf-4a2f-9020-cf1450946d73","把世界模型搬进工厂：基点起源「全要素大模型」如何压缩定制化交付周期","jidian-all-factor-llm-industrial-world-model","2026-06-28T12:27:33.080882+00:00",{"id":60,"title":61,"news_slug":62,"published_at":63},"3191469f-3250-4b31-bc42-8b43680a61a4","MeMo：把 LLM 的『记忆』和『推理』彻底拆开，更新知识再也不用重训","memo-memory-as-model-90pct-compute-saved","2026-06-19T06:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"b676957e-2af5-4e89-b34b-188526e55ba9","世界模型军备竞赛：DeepMind领跑，2026算法创新重塑Agent架构","deepmind-world-model-continual-learning-2026","2026-05-05T01:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"ea225611-5667-467b-89ae-d667aa54df00","阿里云推出企业级Agent构建平台JVS Crew","alibaba-jvs-crew-enterprise-agent-platform","2026-04-23T02:07:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"bbedb317-c221-43e3-b9a4-61798ed17b55","Voxtral 3:Mistral 语音智能体全栈,一轮对话不到一秒","voxtral-3-voice-agent-sdk","2026-10-21T23:11:24+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"a491d4db-2d5a-4723-b093-28d1512fd647","VLA-Δ:物理状态变化当监督信号,LIBERO 98.4","vla-delta-physical-state-transition","2026-10-20T19:07:27+00:00"]