Episode Details

As training datasets for code models exploded 18x in size, transparency collapsed by 17 points. We walk through the hidden pipeline—from GitHub repos to filtered datasets to your IDE—examining every choice developers never see. If every dataset begins with human decisions about what counts, what should we demand to know before hitting tab? Timestamps: 0:00 - The 113.7 terabyte problem 1:20 - Stanford's transparency index collapse (9.2/100) 4:00 - The six-stage pipeline: collection → filtering → deduplication → training → alignment → deployment 8:30 - The blood bank metaphor: tracing code suggestions back to source 14:00 - Who's responsible when the supply chain is invisible? --- Sources & further reading: • Mahmoud Jahanshahi, Audris Mockus — "Cracks in The Stack: Hidden Vulnerabilities and Licensing Risks in LLM Pre-Training Datasets": https://arxiv.org/abs/2501.02628 • Nicholas Carlini et al. — "Poisoning Web-Scale Training Datasets is Practical": https://arxiv.org/abs/2302.10149 • Stanford CRFM — "2025 Foundation Model Transparency Index": https://crfm.stanford.edu/fmti/December-2025/index.html • Yiming Zhang, Nicholas Carlini, Florian Tramèr et al. — "Persistent Pre-Training Poisoning of LLMs" (ICLR 2025): https://arxiv.org/abs/2410.13722 • Martin Pawelczyk et al. — "Machine Unlearning Fails to Remove Data Poisoning Attacks" (ICLR 2025): https://arxiv.org/abs/2406.17216 • Hojjat Aghakhani et al. — "TrojanPuzzle: Covertly Poisoning Code-Suggestion Models" (IEEE S&P 2024): https://arxiv.org/abs/2301.02344 • Domenico Cotroneo et al. — "Vulnerabilities in AI Code Generators: Exploring Targeted Data Poisoning Attacks" (ICPC 2024): https://arxiv.org/pdf/2308.04451 • Nicholas Carlini et al. — "Scalable Extraction of Training Data from (Production) Language Models": https://arxiv.org/abs/2311.17035 • Data Provenance Initiative (MIT Media Lab) — "Data Provenance for AI": https://www.media.mit.edu/projects/data-provenance-for-ai/overview/ • Shumailov et al. — "AI Models Collapse When Trained on Recursively Generated Data" (Nature 2024): https://www.nature.com/articles/s41586-024-07566-y • Jason Hausenloy et al. — "Towards Data Governance of Frontier AI Models": https://arxiv.org/abs/2412.03824 • Justin Sherman / Atlantic Council — "Securing Data in the AI Supply Chain": https://www.atlanticcouncil.org/in-depth-research-reports/issue-brief/securing-data-in-the-ai-supply-chain/ • Hanqing Zhao / Georgia Tech — "Vibe Security Radar": https://github.com/HQ1995/vibe-security-radar • Socket Research — "Slopsquatting Targets Across Frontier LLMs": https://socket.dev/blog/slopsquatting-targets-across-frontier-llms • Kiteworks — "AI Regulation 2026 Business Compliance Guide": https://www.kiteworks.com/cybersecurity-risk-management/ai-regulation-2026-business-compliance-guide/ • CBS News/60 Minutes — "Labelers Training AI Say They're Overworked, Underpaid, and Exploited": https://www.cbsnews.com/news/labelers-training-ai-say-theyre-overworked-underpaid-and-exploited-60-minutes-transcript/ • Snyk — "AI Quality: Garbage In, Garbage Out": https://snyk.io/blog/ai-quality-garbage-in-garbage-out/ • GitHub — "Updates to GitHub Copilot Interaction Data Usage Policy": https://github.blog/news-insights/company-news/updates-to-github-copilot-interaction-data-usage-policy/ • EleutherAI — "Common Pile v0.1" — (announced via blog and Hugging Face) • Hugging Face — "The Stack v3" — (dataset page on Hugging Face Hub) This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.