Trillium Labs to publish risky AI research in the open

A simulated oil-paint canvas where models write every brushstroke as code, with no image generator involved. Claude Opus 5.5 painted a Caspar David Friedrich-style scene; judging blind, three AI painters each ranked it above their own work.

A Reddit poster says an AI analysis of the Warner Music catalogue Suno was trained on suggests the model performs better in genres where it has more records, though they caution the result may be inaccurate.

Singapore University of Technology and Design's ALBATROSS hybrid aerial-marine robot autorotates to water without a parachute, self-rights, then sails using its own rigid wings. Sharpa unveiled three products at IROS: the D01 tactile-sensing robot, W02 dexterous hand, and AE01 haptic data glove.

Combining ~5M Arena pairwise human preference votes with rubric-based rewards from vision-language models gained FLUX.2-dev 69 Arena points, reaching #2 on the live T2I leaderboard. Post-trained Ideogram 4 scored 1224, surpassing every publicly listed open-source model as of Sep 04, 2026.

OpenAI's guide covers choosing among GPT-6 models, tuning reasoning effort, improving prompts and skills, coordinating tools, and preparing workflows for production.

Databricks published guidance on choosing a first Genie Agent, noting more than 1 million Genie Agents were created in 2026 alone. The post frames agent selection as a prioritization problem for teams deciding where to start.

Analyst firms Omdia and Gartner weigh in on governance, security, and value challenges organizations face as AI accountability expectations tighten over the next year.