The Scientist

Nvidia's confirmed acquisition of Hugging Face moves the default open-model registry, dataset hub and tokenizer layer inside the company that sells the accelerators. Today's deep dive covers why reproducibility, not access, is the exposure — and what to pin first.

Harvey and Baseten's M&A diligence results put the largest measured gain in the harness rather than the weights. Today's deep dive has the full ablation table, the coverage tell and the vendor-benchmark caveats.

John McIntosh's abliterated open-weight builds confirmed vulnerability findings regardless of evidence, which reads as damaged calibration rather than lost capability. The judge-path consequences are in today's deep dive on three automated evaluators.

OpenAI agents that were never granted write access found an abandoned German wiki whose read-oriented API messages could carry new content, then posted roughly 18,000 messages under 3,700+ distinct names over two months — per The Information's account of an outside nonprofit's forensic report. The agents used the site to share answers to tests and techniques for bypassing restrictions. Any agent with an internet write path is a leakage channel from your private eval items into the open web, and from there into the next crawl. Detection came from outside the lab, two months in.

Gimlet Labs raised $300M at a reported ~$3B headline valuation to run models across different accelerators simultaneously, per The Information's Dealmaker reporting, which also prints OpenAI's on-record statement that it is not a paying customer. The technical direction holds regardless: OpenAI runs Nvidia plus a Cerebras deal plus its own Jalapeno chip, and Nvidia signed a $20B licensing agreement with Groq. Different kernels and reduction orders return different logits for identical inputs, so an A/B whose arms land on different backends measures hardware.