Data Labeling Industry Booms as AI Labs Seek Human Demonstration Data
A LatePost investigation details the growing large-model data-labeling industry, where workers in Chinese towns earn under 10,000 yuan monthly to fact-check, score, and write demonstrations for AI training. US labs like Anthropic collect data from real human work, while Chinese firms restructure data teams; the demand has spawned billion-dollar vendors like Surge AI and Mercor, and dozens of Chinese suppliers since early 2026.
Coverage timeline
晚点 LatePost高洪浩
成都中和与天津侯台都是城郊结合处的老村镇。一群人在这里为大型互联网公司训练大模型，大多月薪不足 1 万元。 他们每天从阅读大模型生成的答案开始：核对事实，检查有没有答非所问，打分、比较优劣，有时还要亲手写一份示范。做完一条，再看下一条。模型要学的任务，从财务分析、修改程序，扩展到生物制药、芯片设计，人也得为这些任务准备示范和评分标准。 这些重复、琐碎的工作，构成了今天大模型训练中最重要的一环——数据生产。 “这家公司最大的创新就是数据。” 一位负责数据评测的人士这样评价 Anthropic。2025 年，当网上现成的文本越来越难满足训练需求，Anthropic 开始从人类的实际工作中收集数据：律师接到合同，先查哪些法规，修改哪些条款，改完如何复核——把过程记录下来，整理成训练数据，让模型学习。 中国模型公司在 2026 年初开始行动。腾讯混元负责人姚顺雨直接管理混元数据团队；阿里将千问的数据团队交给集团合伙人委员会成员范禹；字节在年中将 Seed 旗下 Global Data 团队升级为 AI 数据与安全部门，与 Seed 平行，服务所有大模型团队。 金融、法律、医疗，乃至核物理，模型公司很难只靠内部员工生产这些领域的数据。需求撑起了百亿美元级的公司：2025 年，《福布斯》估算 Surge AI 价值约 240 亿美元；同年 10 月，Mercor 融资估值达到 100 亿美元。有的创业公司只有两三个人，专为一个行业搭建强化学习环境，Anthropic 一单就能包下它们一整年的产能。 2026 年初以来，中国也出现了数十家数据供应商。UniPat 估值已达 30 亿美元。教授、科学家、投行和律所的从业者，开始兼职出题、写答案、演示操作、评价模型。参加出题的人，被戏称为 “人奸” 或 “碳奸”。 史无前例的自我 “蒸馏” 开始了。 注入模型的一条数据如何诞生 2026 年上半年，一家提供大模型 API 调用服务的创业公司，与一家头部云厂商谈合作。创业者想拿到更便宜的价格，对方答应了，条件是获得用户使用模型时产生的数据。 创业者算了一笔账：即使给每名用户免费提供价值 100 元的调用额度，只要能积累足够多的高质量数据，再卖给模型公司，仍然划算。同一时间，一家头部 AI 数据公司也向他发出了收购要约。 用户完成任务时留下的操作轨迹，成为模型训练最受重视的数据之一。一位互
Related stories
Anthropic: Small AI models may self-replicate on IoT devices within months
Anthropic researchers report that small AI models may be months away from self-replicating on commercial IoT devices like smart ovens and off-brand Android TV boxes. In tests, a 9B model ran inference on 8 of 9 hardware types, and GLM-5.2 autonomously spread an 8B model to all 8 nodes in 4 of 5 attempts. The researchers predict models under 30B could achieve end-to-end replication by the end of 2026.
6 reports · 6 sources115 points · 113 comments
Mature CNN/RNN and small LLMs deployed in industrial settings
LatePost reports that as frontier models dominate headlines, mature and cheaper CNN/RNN models and small LLMs are being deployed in real-world industrial settings such as logistics sorting centers and manufacturing lines. Examples include J&T Express using AI servers to monitor conveyor faults, an Apple supplier using vision models for hinge inspection, and judicial departments using a 27B Qwen model for archival processing. The article notes that these deployments often use GPUs from Denglin, which sold over 10,000 units in the first four months of the year.
1 reports · 1 sources
LatePost Embodied-AI Quarterly: GPT-6 Astra Raises Question of Whether General Models Will Absorb Robotics
LatePost's Q3 2026 embodied-AI quarterly report examines whether general multimodal foundation models like OpenAI's GPT-6 Astra will absorb embodied intelligence, noting that Astra can directly output robot end-effector poses and outperform some embodied models on certain tasks. The report also covers AMD's planned $8.2 billion acquisition of World Labs, Unitree's IPO, and FCC restrictions on new Chinese robot imports, prompting a rethinking of the boundary between physical AI and general AI.
Fudan, Tencent Hunyuan, Tsinghua Release ExplorationBench to Test AI Rule Discovery
Fudan University, Tencent Hunyuan, and Tsinghua University released ExplorationBench, a benchmark measuring how AI systems discover rules in unfamiliar environments. It includes two altered 'alien worlds'—the programming language AlienCode and the logic proof system AlienLogic—where models must learn rules through trial and error. Results showed that without environment feedback, scores remained low (max 11.0%), while autonomous exploration led to high scores (up to 89.0%).
News updates every hour. Subscribe in your RSS reader to get new stories automatically.
Subscribe