person · DiggingBeagle record
Researcher and former OpenAI employee who published a November 2026 checkpoint-forensics study of GPU initialization fingerprints.
Will DePue published a November 12, 2026 investigation into hardware-dependent traces surviving in released neural-network checkpoints. The report describes a search over candidate PyTorch seeds and GPU execution layouts, physical reproduction on B300 and GB200 hardware, and comparisons with 24 models whose hardware had previously been disclosed. It attributes the six new Blackwell-compatible patterns to initialization, while explicitly stating that weight evidence cannot locate machines or prove export violations. DePue identifies himself as a former OpenAI researcher; this biography does not independently establish the laboratory locations or legal status of the chips.
The study explicitly addresses the search-size false-positive problem: roughly six trillion seed/layout candidates were evaluated, so a single high-sigma numerical correlation would be insufficient. DePue instead reports whole-row reproductions on physical Blackwell systems and comparisons to models with separately disclosed hardware. He revised the largest-run GPU estimates on November 12 and notes that the resulting figures remain estimates derived from throughput and training assumptions, not measured inventory or identified procurement. The evidence supports hardware-class initialization under tested implementations while leaving model-training location and legal acquisition outside its method.
The published reproducibility record is more specific than a visual correlation: the author describes a physical H100/B300/GB200 comparison, gives the seed values used for DeepSeek and Kimi, and provides selected rows and runnable replay code. His 24 known-hardware matches are consistency checks against developer disclosures rather than an independent accuracy study. The study also tests a configured AMD MI300X that can imitate the grid layout, successfully differentiating NVIDIA for DeepSeek-V4-Flash but not definitively for all other models.
Locator: Our investigation, search-space multiple-testing caution; Verifying the findings and Limits of the investigation
Locator: Verifying our findings: B300 461/469 rows (445 exact) and GB200 13 rows; Limits of our investigation: AMD MI300X test and 363/377 discriminator