Overview
The day's three strongest signals covered inputs, infrastructure, and verification. Gracenote's lawsuit put copyright pressure on structured media metadata; Thinking Machines Lab reserved at least a gigawatt of next-generation Nvidia systems; and CR-Bench showed that an AI code reviewer can surface more real issues by producing enough false positives to reduce practical value.
Developments
1. Gracenote alleges that OpenAI copied licensed media metadata
Nielsen's Gracenote sued OpenAI in Manhattan federal court, alleging that copyrighted program descriptions and identifiers were used without permission to train ChatGPT. The complaint cites outputs that allegedly reproduced Gracenote material and seeks damages and an order blocking its use.
OpenAI responded that its AI models are trained on publicly available data and grounded in fair use. The unresolved case demonstrates that short descriptions, identifiers, and other structured catalog fields can still produce material licensing disputes.
Sources: Reuters
2. Thinking Machines Lab reserves a gigawatt of Vera Rubin systems
Nvidia and Thinking Machines Lab announced a multiyear AI-infrastructure partnership to deploy at least one gigawatt of Vera Rubin systems, with deployment targeted to begin in early 2027. Nvidia also invested in the company but did not disclose the amount.
The commitment describes future capacity, with no site list, delivery schedule beyond the initial target, or investment amount. A gigawatt reservation becomes usable model capacity only through financed facilities, available grid power, networking, installed accelerators, software readiness, and sustained utilization.
Sources: Nvidia
3. CR-Bench exposes the precision cost of AI code review
A March 10 preprint introduces CR-Bench and a fine-grained evaluator for AI code-review agents. The preliminary study compares single-shot and reflection-based agents on hidden issues, tracking spurious findings alongside issue resolution instead of reducing review quality to bugs found.
The authors found a frontier between recall and signal-to-noise: configurations prompted to find every issue also generated enough false positives to obscure useful progress. That tradeoff makes reviewer acceptance rate, severity calibration, and developer time better operating measures than a raw count of comments or resolved benchmark defects.
Sources: CR-Bench preprint