← Archive

AI Daily — June 15, 2026

2026-06-15

TODAY'S NEWS

openai.com

Introducing the OpenAI Partner Network

OpenAI is launching a formal Partner Network with $150M in investment to accelerate enterprise AI deployment through a structured ecosystem of global partners. The program targets system integrators, consultancies, and technology providers to help organizations implement OpenAI's models at scale. This signals a more formalized go-to-market structure for OpenAI's enterprise push, moving beyond direct sales.

rss.arxiv.org

The Coin Flip Judge? Reliability and Bias in LLM-as-a-Judge Evaluation

A systematic study of LLM-as-a-Judge reliability finds that pairwise preferences flip on average 13.6% of the time across repeated identical evaluations, with 28% of questions exceeding a 20% flip rate using GPT-4o-mini and GPT-4.1-mini as judges. GPT-4o-mini shows significant first-position bias, favoring the first-presented option 72% of the time. These findings have direct implications for leaderboard validity and reward model training pipelines that rely on LLM judges as a ground truth signal.

LAST WEEK'S TOP STORIES

Claude Fable 5 and Claude Mythos 5

Anthropic released two new frontier models with differentiated capability profiles, marking a significant expansion of its model family that immediately prompted a US government directive to suspend their access — an unprecedented regulatory intervention in frontier model deployment.

Introducing Gemma 4 12B: a unified, encoder-free multimodal model

Google DeepMind released an open-weight 12B multimodal model that eliminates the separate vision encoder in favor of a single unified architecture, reducing inference complexity and broadening accessibility for researchers and on-device deployments.

Google DeepMind launches $10M funding call for multi-agent safety research

DeepMind is committing $10M to study emergent safety risks from millions of interacting AI agents, representing a major institutional acknowledgment that current safety frameworks are not designed for large-scale multi-agent systems and signaling where the next safety research frontier lies.

OpenAI Submits Confidential S-1 to the SEC

OpenAI has taken a concrete step toward a public offering by filing a confidential draft S-1, a landmark corporate milestone following its restructuring to a public benefit corporation that will reshape how the industry's leading lab is governed and financed.

DiffusionGemma: 4x faster text generation

Google DeepMind demonstrated that applying discrete diffusion to the Gemma architecture achieves 4x faster text generation than autoregressive baselines, a meaningful quality-at-speed result for a research track that has long struggled to compete with standard LLM inference.