AI Daily — June 5, 2026
2026-06-05
openai.com
Dreaming: Better memory for a more helpful ChatGPT
OpenAI has shipped a new memory architecture for ChatGPT called 'Dreaming,' which asynchronously consolidates and updates user preferences across conversations rather than relying on explicit memory saves. The system is designed to keep long-term context fresh and contextually relevant without requiring user intervention. This represents a meaningful shift in how persistent state is managed in production LLM assistants.
openai.com
Biodefense in the Intelligence Age
OpenAI published a detailed action plan for using AI to strengthen biological resilience, covering early-warning biosurveillance, accelerated countermeasure development, and defensive applications of frontier models. The document explicitly frames AI as a net positive for biodefense while acknowledging dual-use risks, and calls for coordinated policy between labs, governments, and public health agencies. It's a notable public commitment to a specific safety-adjacent application domain from a major lab.
rss.arxiv.org
Epidemiology of Model Collapse: Modeling Synthetic Data Contamination via Bilayer SIR Dynamics
Researchers model AI ecosystem-wide model collapse using a bilayer coupled SIR/SIRS framework, treating data corpora and AI models as two interacting populations that cross-contaminate each other with synthetic data. The SIRS variant accounts for immunity waning, capturing how filtered corpora and retrained models remain susceptible to re-contamination over time. The paper derives a basic reproduction number R0 for synthetic data spread, providing a quantitative lens on ecosystem-scale degradation that goes well beyond single-chain collapse analyses.
technologyreview.com
How courts are coping with a flood of AI-generated lawsuits
MIT Technology Review reports on the growing volume of AI-drafted pro se legal filings reaching federal courts, creating new challenges around hallucinated citations, procedural errors, and judicial workload. Federal judges are developing ad hoc screening practices to handle the influx, with some courts considering disclosure requirements for AI-assisted filings. The piece highlights a systemic legal infrastructure problem as AI tooling outpaces court policy.