Notes from the frontier.
Technical research
Patent search without examiner labels
Neural patent search systems often require examiner citation data for training, which we believe are potentially biased. Our work implements a self-supervised alternative by training BERT to match patent claims with their own description passages. We tested on 6,447 USPTO patents, and achieved 96% accuracy on claim-description matching without any examiner labels, demonstrating that document structure alone can teach models to recognize patent-relevant content.
Read the paper (PDF)Legal research
AI Patent Eligibility Dramatically Expanded at USPTO
The precedential Ex parte Desjardins decision recalibrates how the USPTO examines AI and machine learning applications under Section 101, and what it means for claiming and specification strategy.
USPTO Shifts Tone on AI Patents with August 2025 Eligibility Memo
The August 4, 2025 USPTO memorandum sets a "close call" standard for Section 101 rejections, limits the mental process grouping, and restores broad neural network training claims. What it means for filing and drafting decisions.
Future-Proofing Your Terms of Service
Future-proofing your ToS and related moderation systems is not just a legal safeguard—it’s becoming a competitive advantage.
California - Select AI Laws
California introduced 11 new AI laws effective January 1, 2025, covering transparency, personal rights, criminal prohibitions, election integrity, and healthcare.
Trump Administration AI Policy Framework
Analysis of the key components of these policies, with particular emphasis on their technological implications for industry stakeholders and the broader AI ecosystem.