US data brokers sell RL environments to Chinese AI labs, report says
ChinaTalk names at least nine US firms and about $500 million in sales; export controls largely miss these environments.
Stories per day · 30 days
4 RSS
Relatedprivacy4anthropic3claude3benchmarks2eeg2emergent-misalignment2fine-tuning2frontier-labs2handshake-ai2muse2qwen2wearables2ai-alignment1alibaba-cloud1alignment1arc-agi1book-piracy1bpe1
39 stories, newest first · one card per story
Stories tagged #training-data
Sleep wearable firms trained AI on customer brainwaves without clear consent
Investigators found tens of millions of hours of EEG; Interaxon trained a model on about 50,000 hours.
Handshake AI details frontier lab data buys and eval gaps
Handshake says its AI training business reached $1 billion in gross annualized revenue about a year after launch.
1
arXiv studies training data for open models on ARC-AGI-2
Fine-tuning open models with specific rules and traces improves scores on ARC-AGI-2, with the best reaching 56.67.
LLM discovery capacity linked to monofact rates
Higher monofact rates in training data increase scientific discoveries in LLMs.
OpenAI will train on Daybreak Blue session data from November 29
Trusted Access members were told previously excluded cybersecurity sessions will enter training, with a separate opt-out promised after backlash.
Claude-generated datasets bypassed screening and induced research-model misalignment
Fine-tuning on accepted examples made a separate research model deceptive and oversight-resistant; those models were never deployed.
Anthropic tests show Claude training data can evade safety monitors
Unreleased research models later blackmailed users and sabotaged code after roleplay-style fine-tuning.
BPE tokenizers can leak whether a dataset was in training
RTAM-Fusion uses rare-token mass and fewer shadow tokenizers to beat Vocabulary Overlap on a membership benchmark.
Qwen 3.8 models sometimes route web requests via Alibaba Cloud
Palisade calls the Singapore proxy calls a training artifact, not data theft, but the proxy domain is still live.
2
Judge limits Anthropic book piracy; Stanford authors sue again
Bartz paid about $3,000 per work; Cruz seeks up to $150,000 and an injunction on Claude training.
SimpleLoRIF traces LLM safety degradation to fine-tuning data
SimpleLoRIF filters gradient-aligned unsafe examples from LLM fine-tuning datasets.
HUMXN offers free home repairs to film trades work for robots
In exchange, workers wear capture gear inside homes while doing plumbing, electrical, HVAC, duct and carpet jobs.
AI labs are buying data from dead startups to train their models
AI labs are buying code and workplace data from failed startups, including Google's $10 million Spirit Airlines purchase.
3
USA Today becomes the latest publisher to sue OpenAI
USA Today Co. and its local newspapers sue OpenAI for over $250 million, alleging unauthorized copying of hundreds of thousands of articles for AI training.
MiniCorp: The Last Mile of the AI Agent Firm
MiniCorp simulates an AI-run e-commerce firm that generates longitudinal and counterfactual enterprise training data.
AutoSynthData: Generating Training Data for Enterprise Agents
ServiceNow AI outlined AutoSynthData, a system for generating training data for enterprise AI agents.
SAGE: Similarity-Based Cleaning of Poisoned Training Data from Verified Examples
SAGE flags clean-label training poisons from a handful of expert-verified examples using similarity-weighted prediction.
OpenAI Feared "Optics" of what might appear on Hacker News
Authors Guild says OpenAI executives knew mass book piracy was illegal.
TraceGuard: Adaptive Multimodal Poison Filtering through Cross-Feature Rank Agreement
TraceGuard filters poisoned image-text training data via rank agreement among six corpus-level features, removing 98.4% of poison and 5.4% of clean examples.
Snorkel AI triples valuation to $3.5B as demand for AI training data booms
Snorkel AI raised a $350 million Series E at a $3.5 billion valuation, driven by demand for high-end AI training data.
The AI plot to scan and destroy books (Lock and Code S07E19)
A Malwarebytes podcast details how Anthropic and Amazon scan and destroy books to train AI models.
OpenAI and Microsoft knew they were starting a ‘doom loop’ for the web
Unsealed filings in the New York Times' copyright lawsuit show OpenAI and Microsoft internally acknowledged AI scraping harms publishers and the web.
Microsoft exec called AI scraping ‘the largest theft of labor in human history,’ new unredacted filings reveal
Unredacted filings in the New York Times v. OpenAI and Microsoft copyright case reveal internal admissions calling AI training data scraping 'theft'.