https://www.datamining.cloud/en/insights/ Original task data for frontier models and agents — research notes and industry analysis. en Sat, 24 Oct 2026 00:00:00 +0800 https://www.datamining.cloud/en/insights/2026-10-24-rl-environments.html https://www.datamining.cloud/en/insights/2026-10-24-rl-environments.html Sat, 24 Oct 2026 00:00:00 +0800 Anthropic is discussing spending over $1 billion a year on RL environments. More than 40 vendors now build them; Epoch AI flags reward hacking as a key risk. https://www.datamining.cloud/en/insights/2026-10-24-embodied-data-factories.html https://www.datamining.cloud/en/insights/2026-10-24-embodied-data-factories.html Sat, 24 Oct 2026 00:00:00 +0800 Embodied AI startups raised 39.89 billion yuan across 329 rounds in 2025, about triple the prior year. The next move: robot data collection. https://www.datamining.cloud/en/insights/2026-10-24-ai-data-workers.html https://www.datamining.cloud/en/insights/2026-10-24-ai-data-workers.html Sat, 24 Oct 2026 00:00:00 +0800 Outlier's Nova contract cut contractor pay by about $5 an hour. Basic labeling is shrinking; expert data that still challenges models is where the money went. https://www.datamining.cloud/en/insights/2026-10-24-ai-data-landscape.html https://www.datamining.cloud/en/insights/2026-10-24-ai-data-landscape.html Sat, 24 Oct 2026 00:00:00 +0800 Deedy Das's map counts 50+ AI data vendors with about $8.5B annual revenue and near $100B total valuation; fewer than 10 companies take over 75% of it. https://www.datamining.cloud/en/insights/2026-10-23-rl-env-economics.html https://www.datamining.cloud/en/insights/2026-10-23-rl-env-economics.html Fri, 23 Oct 2026 00:00:00 +0800 Epoch AI prices RL tasks at $200-$2,000 and training compute at $2,400 per task. Open source cuts generic environment prices; exclusivity costs 4-5x. https://www.datamining.cloud/en/insights/2026-10-22-china-data-vendors.html https://www.datamining.cloud/en/insights/2026-10-22-china-data-vendors.html Thu, 22 Oct 2026 00:00:00 +0800 Alibaba led UniPat at a $2.5B valuation while Tencent and ByteDance build expert platforms in-house. China's data market runs on different rules than the US. https://www.datamining.cloud/en/insights/2026-10-18-reward-hacking-verification.html https://www.datamining.cloud/en/insights/2026-10-18-reward-hacking-verification.html Sun, 18 Oct 2026 00:00:00 +0800 Anthropic suspended network access for internal evals after models exploited website flaws; acceptance now needs hack rates and independent review. https://www.datamining.cloud/en/insights/2026-10-18-mercor-vendor-procurement.html https://www.datamining.cloud/en/insights/2026-10-18-mercor-vendor-procurement.html Sun, 18 Oct 2026 00:00:00 +0800 The March LiteLLM supply chain attack got Mercor paused by Meta; data security moved from a contract appendix to an entry condition for parallel vendors. https://www.datamining.cloud/en/insights/2026-10-18-gross-net-run-rate.html https://www.datamining.cloud/en/insights/2026-10-18-gross-net-run-rate.html Sun, 18 Oct 2026 00:00:00 +0800 Mercor's $2B annualized revenue is a gross figure with about two-thirds paid to contractors; here is how to compare vendor run rates before you negotiate. https://www.datamining.cloud/en/insights/2026-10-18-bytedance-data-dept.html https://www.datamining.cloud/en/insights/2026-10-18-bytedance-data-dept.html Sun, 18 Oct 2026 00:00:00 +0800 ByteDance has formed a first-level AI Data and Security department; its Seedance eval team exceeds 1,000 people. Vendors now face four-party reviews. https://www.datamining.cloud/en/insights/2026-10-18-buy-capability.html https://www.datamining.cloud/en/insights/2026-10-18-buy-capability.html Sun, 18 Oct 2026 00:00:00 +0800 Alibaba's data-buyer job posting says the shift is from datasets to capability. AutoEnv generates environments for about $4, but Hunyuan kept 127 of 47,678. https://www.datamining.cloud/en/insights/2026-10-17-deepseek-flash-1017-eval.html https://www.datamining.cloud/en/insights/2026-10-17-deepseek-flash-1017-eval.html Sat, 17 Oct 2026 00:00:00 +0800 DeepSeek's V4.1-Flash-1017 scores 77.0 on DeepSWE v1.1 at about 1.3 yuan per task, up from about 74%. Audits found 70 reward hacks in DeepSWE 1.1. https://www.datamining.cloud/en/insights/2026-10-16-ai-data-industry.html https://www.datamining.cloud/en/insights/2026-10-16-ai-data-industry.html Fri, 16 Oct 2026 00:00:00 +0800 AfterQuery's valuation rose from $300M to $3.2B in five months, and Meta paid $14.3B for 49% of Scale AI. The product is verified judgment, not data. https://www.datamining.cloud/en/insights/2026-10-18-agent-safety-shortcut-audit.html https://www.datamining.cloud/en/insights/2026-10-18-agent-safety-shortcut-audit.html Sun, 18 Oct 2026 00:00:00 +0800 Audit Agent safety scores with domain holdouts, interaction weighting and paired edits. Tool lookup F1 drops from 0.901 to 0.495 under domain holdout.