engblogs

Joint OpenAI–CodeAI initiative to prepare the first AI generation by building foundational AI literacy for teens, enabling safe, critical use of tools like ChatGPT for Teens within classroom partnerships and teacher-supported learning.

A practical framework for pacing frontier AI development amid cyber-critical capabilities, highlighting monitoring, alignment, and security safeguards for Astra and related workloads.

Explores how AI can empower democratic oversight bodies to govern national security with safeguards, meaningful human judgment, and transparent, auditable decision-making.

OpenAI and CodeAI partner to equip teens and educators with AI literacy, safety, and critical-thinking skills through ChatGPT for Teens and classroom-ready resources to shape responsible, creative use of AI.

A technical overview of ChatGPT Ads expansion to 31 European markets, detailing new capabilities like conversion optimization, geo-targeting, custom audiences, and enhanced measurement, while emphasizing privacy-preserving ad policies on the OpenAI Ads platform.

A blueprint for strengthening democratic oversight of AI-enabled national security, detailing safeguards, human judgment, traceability, and scalable governance for deployment, monitoring, and accountability.

ChatGPT Ads expands to 31 European markets with a privacy-preserving, conversion-oriented ads platform featuring geo-targeting, enhanced measurement, and staged self-service access for advertisers.

Train-to-Test (T²) scaling laws demonstrate that when test-time reasoning is included, the compute-optimal strategy shifts toward smaller models trained on substantially more data, yielding overtrained checkpoints that outperform Chinchilla-style configurations under equal inference budgets.

Train-to-Test (T²) scaling integrates pre-training and test-time reasoning to show that accounting for repeated inference favors smaller models trained on much more data, shifting the optimal compute allocation away from standard Chinchilla scaling toward overtrained, inference-efficient models.

Examines attribution decay in large-scale diffusion models, showing that many training data points lose influence on outputs, and presents a diffusion-ensemble counterfactual approach with implications for copyright, attribution, and policy.

MIT CSAIL's attribution-decay findings show that with large training datasets, individual training images often stop mattering for diffusion-model outputs, challenging attribution and copyright assumptions and enabling counterfactual, data-efficient architectures like diffusion ensembles.

Deutsche Bank leverages agentic AI on Google Cloud to transform regulator-ready, context-aware resilience testing into continuous, evidence-backed operational resilience for financial services under DORA.