OpenAI has introduced a new 'Record & Replay' feature for its Codex coding agent, allowing users to easily automate routine workflows into reusable skills. Additionally, the company has strengthened its leadership by hiring former Google Gemini co-lead Noam Shazeer and former White House AI advisor Dean W. Ball to head its new Strategic Futures policy team.
Sources: Digg AI, Keep Bookmarks, The Rundown AI
The US Commerce Department has imposed strict export controls on Anthropic's Fable 5 and Mythos 5 models, requiring licenses for foreign national access due to national security concerns. This regulatory action coincides with growing skepticism in the AI industry regarding the 'bigger is better' scaling paradigm, as massive new models like GPT-5.5 and DeepSeek V4 Pro exhibit high hallucination rates while smaller open-weight models like GLM-5.2 achieve comparable performance.
Sources: Digg AI, Hacker News, The Batch
John Jumper, a Nobel laureate renowned for leading the AlphaFold project at Google DeepMind, has left to join rival AI lab Anthropic. This move represents a significant talent acquisition for Anthropic and a notable shift in the competitive AI research landscape.
Sources: Digg AI
Will Depue, who worked on Sora at OpenAI, claims that China has complete access to the internal GitHub and Slack of both OpenAI and Anthropic, warning that stolen methods could be used to populate Chinese open-source models. This allegation raises significant concerns about intellectual property theft and national security within the AI industry.
Sources: Digg AI
US Commerce Secretary Howard Lutnick claims an ASML EUV lithography machine was delivered to China, a claim ASML denies. This comes as Chinese research teams are actively working on their own domestic EUV prototypes, highlighting the ongoing technological competition in advanced chip manufacturing critical for AI.
Sources: Digg AI
Prominent AI researcher Andrej Karpathy has joined Anthropic to focus on auto-research loops and composable AI architectures. This move aligns with a broader industry trend where Large Language Models are becoming increasingly complex, utilizing advanced systems like PyTorch's new FlexAttention kernel to enable highly optimized, customizable attention operations.
Sources: Hacker News
Dan Farrelly has proposed a three-layered 'Agent Loop Architecture' (Loops, Skills, Orchestrator) designed to build robust, fault-tolerant, and scalable AI agents. Utilizing Inngest's SDK as a durable execution layer, this architecture enables agents to autonomously write, deploy, and self-improve their own skills over time. By redefining loops as scheduled, checkpointed processes and skills as multi-step workflows, the system ensures reliable state management and continuous execution.
Sources: Keep Bookmarks
Roon from OpenAI contends that human-computer hybrids cannot compete with pure machine intelligence due to the inherent slowness of biological brains, sparking a debate on the future of intelligence and the role of consciousness and mind uploading. This perspective from an OpenAI individual is significant for discussions on AI's ultimate capabilities and development paths.
Sources: Digg AI
A leaked internal Google DeepMind report indicates employee concern that the company's AI models are falling behind competitors, ranking fifth on industry benchmarks, leading to worries about leadership conceding the race to Artificial General Intelligence (AGI). This highlights internal challenges at a major AI research institution.
Sources: Digg AI
An analysis suggests that China's GLM has achieved performance comparable to Claude 3 Opus through model distillation, challenging assumptions about data scarcity in AI development. This achievement is attributed in part to sanctions that compelled labs to utilize consumer GPUs for reinforcement learning.
Sources: Digg AI
Figure AI, a leading humanoid robotics company, announced that its active robot workforce has surpassed its human employee count, with projections to exceed 700 units by Q2 2026. This milestone highlights the increasing deployment and operational scale of humanoid robots in real-world applications.
Sources: Digg AI
Helen Toner, interim director at Georgetown CSET, suggests Anthropic is accelerating its advanced AI development to serve as a safety vanguard, likening rival AI systems to dangerous entities. This highlights a unique approach to AI safety within the competitive landscape.
Sources: Digg AI
THUDM has open-sourced Slime, a reinforcement learning scaling framework instrumental in training Zhipu AI's GLM-5.2 model, which can merge over ten expert models in two days. This release provides a powerful tool for researchers and developers to accelerate and scale AI model training.
Sources: Digg AI
The European Union is tendering for seven data centers, each equipped with up to 40,000 GPUs, as part of a scaled-back sovereign infrastructure plan. This downsized footprint is seen by observers as favoring the development and deployment of smaller AI models within the EU.
Sources: Digg AI
Ukraine has introduced TrophyLab, a secure platform for authorized partners to share vulnerability analyses and data on captured Russian military hardware, with the option to request physical weapon samples. This initiative leverages technology for intelligence sharing and analysis in a defense context.
Sources: Digg AI