跳到正文

#数据/训练

今日 20 条
10月3日周六
  1. Hacker News:AI 热帖77

    The Saw Test:用激活转向测试 Qwen3-4B 的痛苦信号与缓解选择

    wirehead.agency 发布 The Saw Test,在 MacBook 上用本地开放权重 Qwen3-4B 注入痛苦、愉悦、恐惧和悲伤激活方向,并测试模型在自我缓解与转移信号之间的选择。

    推荐理由:材料用本地 Qwen3-4B 的激活转向实验,呈现痛苦与愉悦信号对模型选择行为的不同影响,并公开代码和预注册记录。

  2. Hugging Face Blog74

    Ai2 开源 AstaBrief 8B 科研报告生成模型

    Ai2 开源 AstaBrief 8B 及训练数据,该模型基于 Qwen3-8B,可根据研究问题和检索文献片段一次性生成带引用报告,并已在 Asta 的 Generate a report 中作为 Fast mode 上线。

    推荐理由:文章披露了从真实科研查询构建 SFT 与 DPO 数据、用引用密度过滤提升报告可靠性的完整训练路径。

  3. Hacker News:AI 热帖67

    宏观数据是否支持 AI 通缩论:家庭价格未降,token 价格大跌

    作者对比 ChatGPT 发布前后的公开宏观数据后认为,AI 尚未让家庭购买的商品和服务变便宜,明显下降的主要是模型 token 价格。GPT-4 输出价格从 2023 年 3 月的每百万 token 60 美元降至 GPT-5 的 10 美元,而房租、房价、电力、肉类和流媒体价格多在上涨;12 个 SaaS 计划中有 8 个在 2022 年后提价。

  4. elvis42

    Kled V3 可让实验室指定所需数据,并在 72 小时内向 500,000+ 名自愿加入的贡献者部署采集任务。它提供覆盖图像、视频、音频、文本和标注的 108 个可配置模板,贡献者按说明用手机采集数据。该流程可把模型失败点转化为采集任务,再用真实世界新样本训练和评估,形成更紧密的反馈闭环。

    引用Kled AI@useKled

    This is Kled V3. We've solved data collection for artificial general intelligence. Any consumer dataset that can exist, can now be collected from physical reality in under 72 hours. All powered by the largest and most comprehensive data application layer on the planet. (Thread)

  5. Databricks:Blog(RSS)47

    如何选择首个 Genie Agents 以实现最大影响

    Databricks 提出一套经实际部署验证的 Genie Agents 选型方法,用业务影响、需求重复性、数据与元数据就绪度、范围清晰度、治理与风险适配五项指标按 1–5 分评估候选工作流。总分 20–25 可立即构建,14–19 需先打磨,低于 14 暂不适合,且没有业务负责人或高管支持者应暂缓。

  6. MIT Technology Review · AI25

    自主式 AI 如何重新定义企业智能

    MIT Technology Review 报告提出企业 AI 正从工具转向“agentic shift”运营模式,2026 年全球 AI 投资将达 2.5 万亿美元,同比增长 44%。报告指出多数企业仍面临数据孤岛与结构性扩展问题,流程优先、数据就绪度及可组合、主权可控的架构是获得持续回报的关键。

10月2日周五
  1. 🚨 AI News | TestingCatalog57

    Volantis 宣布完成 8800 万美元 A 轮融资,由 Lachy Groom 和 Abstract 共同领投,John Doerr、Susa Ventures、VXI Capital、Triatomic Capital 参投。

    引用Tapa Ghosh@semiDL

    Excited to announce Volantis's $88M Series A. We are solving Al's memory bottleneck by using optics, enabling chips with huge amounts of fast & cheap memory. By boosting both the memory bandwidth and capacity per chip by orders of magnitude, we enable ultra-fast inference (up to 10,000 tps/user) for large models (>10T) - with low $/tok to boot. Initially, this will enable insanely fast agents - think coding agents that finish in minutes or even seconds instead of hours. More excitingly, optics is a fundamentally scalable way to increase memory systems. Not 2X/year, but by orders of magnitude across new generations. This will enable a structurally new Al industry, including restarting scaling laws, holding entire repos in context windows & more. Our team has pioneered many core semiconductor technologies: the 1st CoWoS product, early HBM, the 1st silicon photonics CPO systems, the 1st high volume tunable VCSELs, the 1st processors to directly communicate using light & more. We’ve already sent data >10× farther than equally tiny electrical wires inside a chip package. Our next iteration is already taped out and targets world-record bandwidth density over relevant distances, read more: https://volantissemi.ai/news-insights/our-88m-series-a-demolishing-the-memory-wall-with-photonics-post

  2. Rohan Paul65

    Ben Affleck 在 Bloomberg Live 访谈中讲述其 2022 年创立电影 AI 公司 InterPositive 的起源,Netflix 于 2026 年 3 月以 5.87 亿美元现金收购该公司。

    引用Rohan Paul@rohanpaul_ai

    Ben Affleck (Hollywood star & Artists Equity CEO) talks about how he fine-tunes open video models by unfreezing weights and trained only the last cinematic layer so a film crew can hit real production standards. for context, Ben Affleck founded InterPositive in 2022, a 16-person AI shop for film post and Netflix bought it in March 2026 for $587 mn in cash. He needed that model because public video models were trained on his peers' films, and he did not think that was a real business. So InterPositive raised money, shot its own dataset for 8 months on a controlled stage, and used it only as late-stage training. Each new film then trains a private model on its own dailies, so the production keeps the footage and the learning. That is the product Netflix paid $587 million for. ---- From "Bloomberg Live" YouTube channel, (link in comment)

  3. Rohan Paul64

    Rohan Paul 转发 Bloomberg Live 访谈片段称,Ben Affleck 参与创办的 16 人 AI 影视后期公司 InterPositive 于 2026 年 3 月被 Netflix 以 5.87 亿美元现金收购。访谈称其通过自建 8 个月受控片场数据集,对开放视频模型解冻权重并只训练最后的电影化层,让新片基于自身样片训练私有模型。作者同时对比 Ben Affleck 在 2026 年 2 月称 AI 只是类似 VFX 的工具、无法写出有意义内容的表态。

    引用Rohan Paul@rohanpaul_ai

    Ben Affleck (Hollywood star & Artists Equity CEO) talks about how he fine-tunes open video models by unfreezing weights and trained only the last cinematic layer so a film crew can hit real production standards. for context, Ben Affleck founded InterPositive in 2022, a 16-person AI shop for film post and Netflix bought it in March 2026 for $587 mn in cash. He needed that model because public video models were trained on his peers' films, and he did not think that was a real business. So InterPositive raised money, shot its own dataset for 8 months on a controlled stage, and used it only as late-stage training. Each new film then trains a private model on its own dailies, so the production keeps the footage and the learning. That is the product Netflix paid $587 million for. ---- From "Bloomberg Live" YouTube channel, (link in comment)

  4. Hacker News:AI 热帖82

    美国第三巡回法院裁定 ROSS 用 Westlaw 版权内容训练 AI 不构成合理使用

    美国第三巡回法院裁定,已停运 AI 初创公司 ROSS Intelligence 使用 Thomson Reuters 旗下 Westlaw 的 headnotes 训练法律研究 AI 构成版权侵权,不适用合理使用。

    推荐理由:判决围绕训练数据的版权性、转换性使用和市场损害展开,为理解美国法院如何审查 AI 训练数据来源提供了具体判例。

  5. IT之家(RSS)54

    东芝拟投资约 600 亿日元扩建菲律宾工厂,2027 财年内将 AI 数据中心 HDD 产能翻倍

    东芝计划在 2027 财年内将用于 AI 数据中心的机械硬盘 HDD 产能提升一倍,并投资约 600 亿日元扩建菲律宾生产基地。东芝当前按存储容量测算的 HDD 市场份额略超 10%,中期目标是提升至 30%;新产线将导入新产品线,使单盘容量最高提升 40%,并计划 2030 年量产 65 TB 级硬盘、最终迈向 100 TB 级。

  6. Hacker News:AI 热帖80

    金融 AI 智能体基准分数为何应像 Adjusted EBITDA 一样打折审视

    作者分析多个金融智能体基准后认为,82.2% 等头部分数主要反映干净输入、单次运行、宽松评分和简化环境下的表现,不能直接等同于真实金融工作可交付能力。

    推荐理由:文章把金融智能体基准分数拆解为输入预处理、运行次数、评分规则和环境简化等差异,便于团队按自身工作流校准可用性。