According to AI at Meta on X, Meta trained its Muse models, including Muse Wonder, on the Prometheus supercluster, reporting 85%+ pretraining goodput, 90%+ RL goodput, and about 47K GPU hours mean time between interruptions.

According to Soumith Chintala on X citing Tinker, the company improved efficiency for long-context RL and announced price cuts up to 70% plus new GLM-5.3-Flash and DeepSeek-v4.1-Flash for cost-efficient long-context work.

According to Satya Nadella, Microsoft-Decision-1 beats LLMs on structured decisions, cutting latency and boosting quality across incident response and QC.

According to @sama, OpenAI shares new math benchmarks and methods, signaling faster automated theorem progress and education tools, per OpenAI.

According to DeepLearning.AI, Xiaomi multiplied test scores by a quality checklist to curb silent failures, pushing MiMo V2.6 Pro RL to a 46 lead.

According to Geoffrey Hinton, a new CASP paper analyzes recursive self-improvement and finds intelligence explosion could arrive soon, impacting safety policy.

According to OpenAI... the company outlines safety cases for securing frontier RL training runs and risk controls for high-stakes model training.

According to SawyerMerritt, Figure’s Helix 2.5 lets humanoids perform complex zero‑shot tasks in unseen homes, improving autonomy and household robotics ROI.

According to JeffDean, Periodic Labs’ Neon, mid-trained and RL’d on months of lab data with 1,300 H200s, surpasses GPT-6 Astra on an analysis benchmark.

According to TheRundownAI, SWE-2 hits 50% on FrontierCode, matching Fable 5.1 while cutting costs by 64%, signaling cheaper frontier-level coding.

According to @CNBC, OpenAI says it solved a Navier Stokes Millennium Problem in 88 hours, signaling new math research workflows and HPC opportunities.

According to Ethan Mollick on X, Astra designed an original Magic deck and defeated an Arena bot, signaling progress in tool-use gameplay AI.

According to StanfordAILab, Stanford launches CS329Z on building AI agents from scratch this fall taught by Diyi Yang, Michael Ryan, and Jue Wang.

According to @emollick, Three Laws fail for AI ethics, underscoring why rule based safety breaks and why layered governance is required.

According to AnthropicAI, Claude autonomously researched, trained, and tested methods to align small models in 48 hours using one GPU.

According to @emollick, small changes like page order and memory shift AI agent shopping choices unpredictably, limiting marketing control, per SSRN.

According to @AINewsOfficial_ the AgiBot A3 reads motion mid-fight, adjusts grip force, and stabilizes fast enough to plan the next move.

According to gdb, OpenAI keeps its chip team small and uses AI for an improvement loop, signaling early hardware in the loop RSI opportunities.

According to KyeGomezB, experts argue transformers limit progress and call for new architectures that learn at test time for faster AI gains.

According to @godofprompt, adaptive AI in e-learning accelerates onboarding and lifts course completion via personalized prompts and feedback loops.