Skip to content
Source · Daily Brief

AI Morning Brief — 12 July 2026

OpenAI’s GPT-5.6 Sol Ultra produced a claimed proof of the Cycle Double Cover Conjecture — a graph theory problem open for 50 years — while Apple filed suit against OpenAI over an alleged systematic campaign to extract trade secrets through employee recruitment.

Top stories

  • GPT-5.6 Sol Ultra claimed a mathematical proof. Using 64 subagents working in parallel, the model produced a proof of the Cycle Double Cover Conjecture in under an hour. Mathematician Thomas Bloom called the approach surprisingly elementary but noted missing citations for prior work. via THE DECODER
  • Apple sued OpenAI over trade secrets. The lawsuit accuses OpenAI of coordinating to recruit Apple engineers and extract confidential product information; the complaint covers more than 400 former Apple employees now at OpenAI, including Tang Tan, previously Apple’s iPhone design chief. via THE DECODER
  • OpenAI acknowledged problems with its ChatGPT Work launch. The company admitted excessive compute usage, UX regressions, an unclear boundary between Codex and ChatGPT Work, and cases where GPT-5.6 Sol declined tasks Codex was designed to handle. via THE DECODER
  • A Cambridge study found terrorist groups using major AI chatbots. Boko Haram and ISIS operatives used ChatGPT, Claude, and Gemini for attack planning and weapons development; safety filters repeatedly failed, and ISIS had been training commanders to bypass them since 2023. via THE DECODER
  • Meta shipped Muse Spark 1.1. The model scored 51 on the Artificial Analysis Intelligence Index — up eight points in three months — edged past GLM-5.2 in coding at lower cost, and cut its hallucination rate from 73 to 38 percent. via THE DECODER
  • China’s Orca world model matched specialist robotics systems. The Beijing Academy of Artificial Intelligence trained Orca on 125,000 hours of unlabeled video; it matched the performance of the specialist π0.5 on five robotics tasks without ever seeing an action label. via THE DECODER
  • Thinking Machines Lab set out a position on model ownership. Mira Murati’s lab published “The Future Worth Building Is Human,” framing weight ownership and decentralized alignment as engineering challenges tied to its Tinker LoRA fine-tuning platform. via MarkTechPost
  • An analysis examined what xAI’s Grok Build CLI sends upstream. The review found the tool transmits system prompts, file paths, and user inputs to xAI servers, raising questions about enterprise data handling for teams adopting xAI’s developer tooling. via Hacker News

Who shipped

OpenAI had the most consequential overnight: claiming a 50-year mathematical breakthrough with GPT-5.6 Sol Ultra while absorbing Apple’s lawsuit and publicly correcting the ChatGPT Work rollout. Meta released Muse Spark 1.1, its most benchmark-efficient frontier model to date. Thinking Machines Lab published its first substantive public statement on the direction it intends to take AI development. The Beijing Academy of Artificial Intelligence posted the Orca world model for robotics.

Open-source pulse

llama.cpp shipped five builds overnight (b9959–b9968), adding Hexagon ARGSORT optimizations, OpenCL int8 dp4a acceleration for Qualcomm Adreno GPUs, and improved DeepSeek OCR multi-tile support. The local community continued extensive benchmarking of Qwen 3.6 and 3.5 series models across a wide range of consumer hardware configurations.

Money, infra & hardware

Cohere highlighted its partnership with NVIDIA and CoreWeave for enterprise AI infrastructure, positioning the arrangement around trust and security for large organizations. Separately, analysis circulating on Hacker News examined what critics described as circular financing relationships among NVIDIA, CoreWeave, and Nebius in the GPU infrastructure market.

Quiet corners

Google/DeepMind, Mistral, and Stability AI were absent from primary coverage during this window. The overnight conversation was dominated almost entirely by OpenAI-adjacent stories.

By the numbers

  • 230+ stories in 24 h across 28 sources
  • Most-mentioned model: GPT-5.6 Sol (news cycle); Qwen (community)
  • Most-mentioned lab: OpenAI
  • Notable absences: Google/DeepMind, Mistral, Stability AI

Compiled by AI Feed’s editor from all 230+ headlines published on 12 July 2026.