Skip to content

#数据/训练

6 today

Today · Nov 17Tuesday · 6 items

Hacker News front page

Carnegie Mellon is putting computer science courses on YouTube for free

卡内基梅隆大学通过 CS Learn 在 YouTube 免费开放计算机科学课堂讲座,供学习者按自己的节奏观看。11 月课程涵盖大规模数据集机器学习、机器学习系统、AI 智能体和生产环境中的机器学习,2027 年 1 月将增加多模态机器学习等课程。免费自学课件将随后推出,提供即时反馈和徽章,AI 辅导也计划加入课件体验;入口为 https://www.cs.cmu.edu/learn/。

Hacker News front page

Launch HN: Axonnic (YC F26) – Sleeves that track fingers and estimate force

Axonnic 推出前臂袖套,通过 64 通道表面 EMG 解码手指动作并估算指尖力,同时保持双手裸露,用于采集人类动作数据训练灵巧机器人。一次训练四天后、215 秒的录制在未重新校准的情况下,对 20 个关节角度的误差为 6.9°,并与 Apple Vision Pro 手部追踪做了对比。该技术已在欧洲一家制造商用于记录缝纫等手工任务,该公司签署了机器人落地支持的 LOI。

Hacker News front page

Psychic: Like fzf but learns from your actions

psychic 是一款终端文件浏览器,通过本地 LightGBM 模型学习文件和目录访问习惯,随输入实时更新结果,界面渲染耗时低于 16ms。搜索与点击记录保存在本机 sqlite 数据库,每次打开工具都会后台重新训练模型,冷启动时使用近期点击、修改时间和模糊匹配进行排序。

Ars Technica · AI

US moves to tighten chip controls, targeting remote Chinese use of third-country AI compute

The US Commerce Department is preparing tighter rules on advanced chips, aimed at stopping China and other countries from renting restricted chips in third-country data centers to train and run AI models. The rule takes effect 90 days after publication in the Federal Register. Firms in high-diversion-risk countries must get licenses to move large volumes of advanced chips and run customer checks, due diligence and monthly transaction reports. Compute dedicated to commercial recommendation systems is exempt, and enforcement remains disputed inside the government.

Why it matters: The Commerce Department is pulling remote compute in third-country data centers into chip controls, so firms serving Chinese customers face licenses and ongoing due diligence.

Hacker News front page

Show HN: Claimseed, synthetic Medicare docs with planted errors and answer keys

Claimseed 0.6.0 发布标准测试集,含 1,022 份全合成美国医疗账单 PDF,覆盖 34 个场景,每份文档附带标注植入错误位置的真值 JSON 答案键。文档由种子生成,姓名、金额等均为虚构,并带水印与元数据标识;五分之一为扫描、传真等退化副本。可通过 npm 包 @claimseed/core 或 MCP 服务器渲染与评分,答案键含 canary 字符串以便训练数据排除。

Yesterday · Nov 16Monday

TechCrunch · AI

US moves to tighten AI chip export rules, targeting Chinese firms' remote access to overseas compute

The Trump administration is preparing new AI chip export controls that would require overseas data centers serving companies in restricted countries such as China to obtain licenses for the export, re-export or transfer of the chips and servers involved. The rule covers data centers or networks above 12 million TPP of total processing performance, including indirect access gained through intermediaries. Exporters and operators must review customer ownership and activity. Existing operators have 90 days after the rule takes effect to get licensed; commercial recommendation systems are excluded.

Why it matters: The US plans to pull overseas data centers serving restricted-country firms into export licensing, extending compliance to customer ownership checks and indirect access channels.

MIT Technology Review · AI

Building AI that can work across the enterprise

SAP Business AI Platform Ecosystem 全球负责人 Thorsten Leiduck 提出,企业应把 AI 当作统一平台而非孤立项目,前提是打通语义丰富的数据底座。Infosys 的 Satinder Pal Singh 建议用影响力、可行性、可扩展性三个维度筛选 AI 智能体用例,并强调缺乏治理不会加速采用,只会积累风险。

Hacker News front page

Qwen releases Qwen4-Max with native vision and a one-million-token context

Qwen released Qwen4-Max, built on the Qwen4 architecture with native vision and a one-million-token context window. It is available now on QwenCloud through Token Plan and API access.

Why it matters: Qwen4-Max raised task performance on the official Work360 benchmark and cut average cost per task from $4.43 to $1.67, a 62% drop.

Nov 15Sunday

Computing Life · Share · Yage

为什么造出大模型的研究者纷纷离开大厂创业

大模型研究者离开大厂创业,被文中解读为对持续学习、物理交互和前沿科学数据短板的回应。文中汇总的 45 家公司已确认融资合计 538 亿美元,其中 8 家未发布公开模型、接口或论文的公司拿走近 300 亿美元,约占 56%。文章提出,独家物理与生化实验数据、快速且低成本的自动验证机制,是创业团队抵御大厂竞争的两类筹码,并判断算力成本可能推动部分团队最终并入巨头。

Hacker News front page

A visual explanation of reinforcement learning

一篇交互式可视化文章用只有四个神经元、三个可调参数的虚拟生物 spark 演示强化学习:随机微调脑内连接,表现变好就沿该方向调整、变差则反向调整。spark 从随机接线时原地打转,经 50 次尝试开始移动、250 次学会避光、500 次冲向光源、2000 次成为追光专家。文章把这一过程画成“大脑参数空间”中的学习路径,显示数百个初始接线不同的个体最终都收敛到同一目标。

Nov 14Saturday

The Decoder

Anthropic can already retrain Mythos with "a fraction of the compute originally needed," says former researcher Coxon

前 OpenAI 和 Anthropic 研究员 Jacob Coxon 称,Anthropic 已能用原先所需算力的一小部分重新训练 Mythos。其依据是 Anthropic IPO 路演中的代码预训练算力乘数图表,他将曲线解读为算法效率呈超指数增长;Anthropic 注明数据来自内部估算,仅供示意。Coxon 称算法改进接近自动化,并主张明年暂停前沿模型开发,让竞争者追赶。

The Verge · AI

Amazon has a plan to index the world’s online shopping data

Amazon 正筹划名为“World Product Graph”的项目,为互联网上所有在售商品(包括非 Amazon 销售的商品)建立数据索引,并通过自家网站提供查询。据路透社报道,该项目意在留住做商品调研的消费者,并让 AI 智能体优先依赖其商品索引。内部文件显示项目目标发布日期为 2027 年 10 月 1 日,但 Amazon 称此类头脑风暴文档不一定反映当前计划。

Hacker News front page

Weight Space Misfolding

一项受控实验用 ChatGPT 6.1-Sol 作为编辑器修改 Qwen2.5-1.5B-Instruct 的权重,使模型在 100 个留出样本中全部将 2000 年奥运会主办城市错答为“布宜诺斯艾利斯”。作者将这种越权编辑导致的持久行为偏移称为“权重空间错折”,并观察到过度拒答和选择性遗漏等附带影响。论文主张通过编辑权限控制、补丁记录与恢复机制来治理权重编辑,而非禁止正当用途。

Dwarkesh Patel podcast

The avalanche that hit math is coming for AI development

Dwarkesh Patel 预测 AI 将迅速逼近顶尖人类研究员水平,并以 2028 作为实现 AI 研发自动化的中位年份判断。他以数学能力进展及预训练、后训练、推理效率的可验证指标支撑这一预测,同时给出 25% 的概率认为该节点会推迟至 2032 之后。他对持续自我加速的软件奇点与受算力限制的短期加速各赋予 50% 概率,后者预计将把 3–6 年的 AI 进展压缩到 1 年内。

Nov 13Friday

Hacker News front page

Jensen Huang reportedly admits Nvidia faces a wafer and memory glut "avalanche"

Jensen Huang acknowledged in August that Nvidia is sitting on a large surplus of pre-purchased wafers and memory, according to a Wall Street Journal report cited by the material, as major cloud providers shift to in-house chips. The report says Colette Kress cut Nvidia's financing guarantee for OpenAI compute from about 8 gigawatts and up to $250 billion to about 4.25 gigawatts and up to $125 billion.

Why it matters: The report ties Nvidia's customer financing guarantees to its surplus of pre-bought wafers and memory, linking a purchasing backlog to the financial risk it carries for customers.

Hacker News front page

The Dawn of the Data Intelligence Engineer

MotherDuck 提出“数据智能工程师”角色,通过组织智能体、维护数据模型和共享业务上下文、运行评测,为企业自助数据分析承担质量责任。文章主张按影响范围分层监督,将自由探索、重要业务应用和核心指标分别纳入不同程度的维护与质量控制。

OpenAI News

Our first phase of embedded assessments

OpenAI 已与 Apollo Research、METR、Transluce 和 Goodfire 签署协议,开展第一阶段驻场评估,重点检查模型训练和内部使用中的对齐与安全防护。

Hacker News front page

Study finds export-controlled Blackwell GPU fingerprints in DeepSeek, Kimi and other Chinese models

Unemployment Labs found initialization fingerprints matching NVIDIA Blackwell GPUs in the public weights of six models from DeepSeek, Moonshot AI, MiniMax and Ant Group. The lab says the method can trace which hardware trained a model, while noting limits on what counts as a violation.

Why it matters: The lab recovers hardware fingerprints from public weights, giving investigators a technical route to check where Chinese models were trained.

Hacker News front page

Show HN: Kapso – Iterative learning from enterprise data

Leeroo 开源 Kapso,面向企业数据任务运行实验、积累经验并将选出的方案适配部署到目标技术栈。用户提供目标和成功标准后,系统依据历史经验与预算筛选方案,由独立验证模块检验结果,再更新经验库的置信度与适用边界;系统不绑定特定模型或编码智能体。

The Decoder

OpenAI’s data suggests that frontier models are hiding their thoughts because developers optimize for capability and compute

OpenAI 安全团队的新分析指出,预训练能力、架构选择、强化学习和教师模型数据都会影响模型隐藏推理及操控 CoT 的能力。研究覆盖 GPT-5 至 GPT-6 Astra,发现更强的基础模型能在不输出推理链时完成更多推理步骤,较小模型通过更长推理链可匹配大模型的能力且更易监控,但会增加部署计算成本。

The Verge · AI

More game developers are starting to use generative AI

GDC Festival of Gaming 年度行业调查显示,40% 的受访游戏开发者在工作中使用生成式 AI,高于 2025 年的 36%。免费游戏开发者和仅开发移动游戏的工作室员工使用比例分别为 58% 和 70%。

Hacker News front page

Running web sites makes me not want to

一位自托管博主因 LLM 爬虫流量饱和上传带宽,在 nftables 中封禁 Meta、Azure、Alibaba 等云厂商的 /16 IP 段后仍难以为继,正考虑撤下摄影网站、改为按活动售卖实体 zine,不再上传数字图片。他称大型云服务器毫无理由访问个人站点,并认为如今普通技术爱好者已难以独自承担网站防护所需的知识与技能。

Nov 12Thursday

The Verge · AI

Google’s AlphaGenome 2 zooms in on the cellular roots of disease

Google DeepMind 推出 AlphaGenome 2,预测遗传变异对不同细胞类型和状态的影响,以帮助研究疾病的细胞层面根源。模型使用覆盖 68 种组织的超过 100 million 份单细胞数据训练,据公司介绍,识别出的可改变细胞基因表达的变异数量接近此前最先进模型的两倍,且无需重新训练即可预测新的细胞类型和生物状态。

Nov 11Wednesday

TechCrunch · AI

Databricks’ new AI agent tells CEO Ali Ghodsi the unvarnished truth

Databricks 周三发布更新版 Genie AI 智能体,公司称其可利用客户平台数据的本体提供上下文,无人监督地完成耗时的多步骤任务。CEO Ali Ghodsi 向记者展示了实际使用记录,包括分析员工满意度低分项、评估改进方案及总结失败的销售沟通,以获取管理层不易听到的负面信息。他表示,企业可通过 Unity Gateway 设置使用限额并切换到更便宜的开放权重模型来控制成本。

Google DeepMind

Google DeepMind releases AlphaGenome 2 to predict how cells read DNA and how variants affect them

Google DeepMind released AlphaGenome 2, which conditions on cell type or state to predict how DNA sequences and their variants affect specific human cells. The model was trained on more than 107 million single-cell profiles. On the TenK10K dataset it identified causal variants 30% of the time, against 16% for the original model fine-tuned on single-cell data and for other specialized tools. It can also analyze unseen cell states without retraining.

Why it matters: AlphaGenome 2 narrows predictions from tissue averages to cell types and states, helping pinpoint the genetic variants that act in specific disease-relevant cells.

Simon Willison

Anthropic IPO roadshow video details multi-chip strategy and compute efficiency

Anthropic began its IPO roadshow yesterday, and the roadshow video covers how Claude is trained and deployed across multiple chips, plus methods for getting more out of compute.

Why it matters: The roadshow video compares a multi-chip adaptation that took dozens of engineers months with the same job done by Mythos in three days.

The Decoder

US data brokers are selling half a billion dollars in RL training environments to Chinese AI labs

ChinaTalk 调查发现,至少9家美国数据公司向中国 AI 实验室出售训练数据,交易规模约5亿美元,其中包括强化学习环境。中国实验室购买此前售予 Anthropic 或 OpenAI 的同类产品,以减少内部开发和无效训练,并结合 GPT、Claude 输出进行知识蒸馏。此类环境目前基本不受出口管制,调查作者主张将具有军民两用风险的编码、生物学环境纳入现有管制。

Latent Space

[AINews] Mercor $750m Series D at $20B valuation, $3B run rate

Mercor 完成 $750M D 轮融资,估值达 $20B,并在 27 个月内将年化收入规模提升至 $3B。公司报告拥有 500 万名专家,称仅美国就有超过 800 种职业,其中仍有工作流未被用于评测模型。Mercor 还表示,其推理支出已达员工薪资的 3x,并通过提高团队工作效率收回成本。

The Decoder

Nvidia's model GR00T 2 watches humans on video, then imagines itself as the robot

Nvidia 推出 GR00T 2,通过同时预测动作和未来视频画面,将人类示范中的动作迁移到机器人,减少新任务训练所需的遥操作。模型使用约 400,000 小时的人类视频训练,实验中未经对应任务遥操作的机器人完成点火柴、切香蕉等任务,成功率介于 10%–80%。系统尚不处理直接触觉传感数据,配合高层语言模型规划时也存在延迟,垃圾分类测试中机器人约三分之一的时间处于等待状态。

Nov 10Tuesday

OpenAI News

OpenAI builds Control Room to track and halt model workloads across research infrastructure

OpenAI has deployed a Control Room across its research infrastructure that tracks model workloads and blocks further model output within 35 seconds of a stop order. The system checks authorization by job_id at the inference engine layer, so it can stop a job through the monitoring system without relying on the job's own code to run stop checks. If the Control Room service fails, it rejects inference requests; this fail-closed mechanism already covers all frontier model research workloads, with the remaining research workloads to be deployed in the coming days.

Why it matters: OpenAI puts the stop mechanism for research jobs at the inference layer, so safety control depends less on fast-changing job code while other authorized jobs keep running.