AI News — Friday, May 15, 2026

OpenAI announces that its advanced coding assistant, Codex, will soon be available on mobile phones, enabling developers to work from anywhere.

A new paper introduces MinT, a managed infrastructure designed to efficiently train and serve millions of large language models, addressing scalability challenges in AI deployment.

A TechCrunch report details the critical points the jury will deliberate in the high-profile legal battle between Elon Musk and Sam Altman, potentially shaping the future of AI leadership.

Listen Labs successfully raises $69 million in funding, following a viral hiring campaign, to expand its AI-powered platform for conducting and analyzing customer interviews.

Reports indicate that Elon Musk's SpaceXAI has seen a substantial exodus of staff since its recent merger, raising questions about the company's stability and future direction.

Researchers introduce MulTaBench, a new benchmark for evaluating multimodal tabular learning models that integrate both text and image data, pushing the boundaries of data analysis.

A new video diffusion model called AnyFlow is presented, capable of generating video at any step count using an on-policy flow map distillation technique, improving efficiency and quality.

OpenAI announces improvements to ChatGPT, enabling it to better understand and maintain context in sensitive discussions, aiming for more nuanced and appropriate responses.

A new study demonstrates effective methods for training vision-language models to handle extremely long contexts, achieving generalization beyond 128K tokens.

A practical benchmark explores whether a 2015 desktop PC can effectively run Google's Gemma 4 (2B and 4B parameter models), offering insights into AI accessibility on older hardware.

Researchers introduce EVA-Bench, a comprehensive end-to-end framework designed for robustly evaluating the performance of voice agents across various scenarios.

A new paper explores a method for predicting the decisions of AI agents based on limited interactions, utilizing a novel text-tabular modeling approach.

The technical report for Qwen-Image-VAE-2.0 is released, detailing advancements in image generation and compression through an improved Variational Autoencoder.

An opinion piece argues that AI is not replacing developers but rather shifting their roles to managing and overseeing AI agents, potentially leading to new challenges in compensation and job satisfaction.

Researchers introduce BenchJack, a new tool for systematically auditing AI agent benchmarks to uncover potential vulnerabilities and ensure robust evaluation of agent capabilities.