OpenAI has unveiled GPT-4o mini, a smaller and more affordable model priced at 15 cents per million input tokens and 60 cents per million output tokens, making it 60% cheaper than GPT-3.5 Turbo. The model supports text and vision, with future updates extending to image, video, and audio inputs and outputs. It features a 128K token context window and excels in benchmarks, outperforming other small models. OpenAI’s collaboration with companies like Ramp and Superhuman confirmed its superior performance in practical tasks. Prioritizing safety, GPT-4o mini includes measures against undesirable content and prompt attacks, making it a reliable option for widespread use.

OpenAI is launching advanced tools for ChatGPT Enterprise to enhance compliance, data security, and user management for regulated industries like finance, healthcare, and legal services. The key feature, the Enterprise Compliance API, enables efficient auditing and management of workspace data, integrating with eight leading eDiscovery and DLP providers. Additionally, OpenAI introduces SCIM for automated user management, supporting platforms like Okta and Google Workspace. New GPT controls allow granular admin control over interactions and permissions. These measures ensure secure, scalable AI deployment, with strong data privacy, encryption, and compliance. OpenAI also supports educational institutions through ChatGPT Edu.

Recent research highlights the critical role of legibility in AI-generated text, particularly for complex tasks like math problems. Findings show optimizing language models solely for correctness reduces text comprehensibility, doubling human evaluators’ error rates. A “prover-verifier” approach, where a model produces solutions and another verifies them, has shown promise. This technique, tested on grade-school math, improved human evaluators’ accuracy in assessing solutions. Balancing correctness with legibility enhances trust and safety in AI outputs, suggesting future AI alignment will benefit from methods that emphasize easily verifiable and understandable justifications.

Researchers have developed consistency models, a breakthrough in generative AI that enhances the speed and quality of image, audio, and video generation. Unlike diffusion models, which are slow due to iterative sampling, consistency models generate high-quality samples rapidly by mapping noise to data in one step, while also supporting multistep sampling. They excel in zero-shot data editing tasks without specialized training and can be trained via pre-trained diffusion models or independently. Achieving a significant FID of 3.55 on CIFAR-10, they outperform existing generative models, establishing new benchmarks and showcasing potential to revolutionize the field. The research was led by Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever.

A major tech company has introduced a groundbreaking content moderation strategy, aiming to build a resilient natural language classification system for real-world applications. Key to its success are detailed content taxonomies, precise labeling, and stringent data quality control. An active learning pipeline improves detection of rare and unusual events. The system effectively identifies diverse undesired content types like sexual content, hateful speech, and violence, outperforming standard models. Developed by a team including Todor Markov and Lilian Weng, this approach marks significant progress in addressing internet safety. For more details, the full research paper is available for review.

OpenAI has introduced consistency models, a novel generative model type enabling rapid, high-quality sample generation. Unlike diffusion models, which require multiple sampling steps, consistency models achieve one-step generation by directly mapping noise to data. This innovation offers enhancements in tasks like image, audio, and video generation, and supports zero-shot data editing without special training. Through extensive experiments, these models outperformed existing diffusion model distillation techniques, achieving state-of-the-art results on benchmarks like CIFAR-10 and ImageNet 64×64. Developed by Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever, consistency models promise to revolutionize generative AI.