AI Daily · Oct 13, 2026 · Tuesday
BAD INFLUENCE
Meta introduces Muse Wonder and Muse Wonder Full frontier models for agents
Meta announces Muse Wonder and Muse Wonder Full, its most capable models to date, built on Muse Spark and aimed at personal and professional agentic tasks. The models add Active Memory to carry relevant details across conversations and Active Focus to keep working toward a goal, recover from errors, and check in before important actions. Muse Wonder is available now through the Meta Model API and Muse Code, and will soon power more of Muse, rolling out first to subscribers.
- Meta releases Muse Wonder, its most capable model for agentic tasksOpenRouter
- Vercel AI Gateway adds Meta's Muse Wonder 1.0 modelVercel Developers
- Meta's Muse Wonder Full ranks second on the Artificial Analysis Index🚨 AI News | TestingCatalog
- Meta releases Muse Wonder and Muse Wonder Full, scoring 52 and 56 on Artificial Analysis IndexArtificial Analysis
Product launches & updates
Anthropic ResearchOfficial1 more sourceAnthropic details new defenses that cut distillation attacks on Claude
Anthropic says its new defenses have cut the reasoning tokens distillers can extract from Claude by a factor of 10 to 100 on its own platform. The company describes three measures: controlling where reasoning can go, a two-stage classifier that blocks extraction attempts with a false positive rate below 0.1%, and stricter limits and bans for repeated attempts. Anthropic also names Zhipu's GLM-5.3 and Moonshot's Kimi as cases where distilled capabilities raised cyber and national security concerns.
The DecoderNews3 more sourcesMicrosoft plans to scale its Maia AI chips to gigawatt levels by 2027
Microsoft plans to deploy its Azure Maia AI chips at multi-gigawatt scale starting in 2027, making them the main inference system across its products, executive Scott Guthrie writes. Maia 200 already runs Microsoft Copilot in the US, Europe, and Australia, with GitHub Copilot and Foundry following by year's end. The successor Maia 300 ships in 2027, and Microsoft claims 40 percent more tokens per megawatt than leading GPU systems.
Meta unveils IcePack, IceCube, and FPF networking for AI clusters at OCP 2026
Meta says it is sharing new open networking infrastructure at the 2026 Open Compute Project Global Summit, including its first liquid-cooled network rack, IcePack, and 102.4T switch blades, IceCube. The company also introduces Flattened Planar Fabric, a multiplane scale-out fabric that gives each GPU eight independent 100G links, and launches the Open Switch Platform Architecture initiative within OCP with Microsoft and Alibaba.
vLLM@vllm_projectOfficialvLLM shares ReplaySSM, a decoding method that caches SSM inputs
vLLM, with NVIDIA and Tri Dao, released a paper, blog post, and code for ReplaySSM, a method that caches SSM inputs instead of the state. The method separates output computation from state updates during decoding, which the authors say yields speedups of up to 1.47x for standard decoding and up to 2.1x for speculative decoding. It also enables a low-precision SSM state whose rounding error does not propagate through the recurrence, and it ships in TensorRT-LLM, vLLM, FlashInfer, and SGLang.
Mustafa Suleyman@mustafasuleymanOfficial2 more sourcesMicrosoft announces Maia 300 chip with 40% more tokens per watt
Microsoft says Maia 300 delivers 40% more tokens per watt than leading GPU-based systems. The quoted announcement from Satya Nadella says the Maia family will scale to multi-gigawatt deployments starting next calendar year as the primary inference system for Microsoft's own products and services. It also says Maia 200 is already deployed in the US, Europe, and Australia, delivering up to 1.7x more tokens per watt in Teams and Copilot.
Industry news
Cursor@cursor_aiOfficial3 more sourcesCursor joins The X Subscription with a shared usage pool across Grok products
Cursor says it is now part of The X Subscription, and current subscribers can keep their plan or switch. Paid plans share one usage pool across Cursor, Grok Build, and other Grok products, with twice the coding usage on Grok models as the equivalent Cursor plan. Third-party models in Cursor move to API pricing on the new plans to make room for more Grok usage.
Papers & research
Anthropic ResearchOfficial1 more sourceAnthropic tests whether small AI models could replicate across cheap IoT devices
Anthropic Research says small AI models may be months away from self-replicating on commercial IoT devices such as smart ovens and off-brand Android TV boxes. In its simplified evals, a 9B model set up inference on 8 of 9 hardware types, and GLM-5.2 completed end-to-end replication. The authors estimate small models could reach that level by the end of 2026.
The DecoderNewsGoogle DeepMind's AlphaProof Nexus solves open math problems in a peer-reviewed Science study
Google DeepMind says its AlphaProof Nexus solved nine of 353 open Erdős problems, two unsolved for 56 years, according to a new study in Science. The system, built around Gemini 3.1 Pro and the Lean proof language, also proved 44 OEIS conjectures and improved a known min-max optimization bound. The article notes that the Erdős solutions were shared online before publication, amid a debate over AI math publication norms.
PandailyNews1 more sourceHuawei Noah's Ark Lab's Memento 3 clears all 25 ARC-AGI-3 public games with a frozen model
Researchers from Huawei Noah's Ark Lab and University College London released Memento 3, an agent that cleared all 25 public games in ARC-AGI-3 using 7,518 actions, about 44% of the human total. The agent runs on Claude Opus 5 without weight updates or game-specific fine-tuning, and the same base model averages about 40.7 RHAE in the official harness. The authors note that the public set does not test generalization to hidden games, and that exploration is token-heavy and real-world settings remain untested.
Amazon ScienceOfficialAmazon Web Services introduces priming to build hybrid state space models from pretrained Transformers
Amazon Web Services introduces priming, a method that starts hybrid architectures from pretrained Transformer weights and cuts their pretraining cost by more than 99%. The team reports that a 32-billion-parameter GKA-based hybrid matches its source Transformer on reasoning benchmarks while decoding up to twice as quickly at 128,000 tokens of context. The paper and released models, training code, and vLLM plugin are available under the Apache 2.0 License.
Tips & takes
How to constrain a coding agent using only files you own
Ian Johnson says a coding agent given one ticket without a charter passed 1 of 8 checks, while the same ticket with a four-part charter passed 8 of 8. He describes four levers under the repository's control: instructions in CLAUDE.md, path-scoped rules, capability files, and hooks that can refuse a write. He notes that one check, whether the access logic is correct for the feature, still needs a human reviewer.