Meta Releases Muse Wonder Models With Benchmark Gains and Muse Code Usage Accounting Bug

More than 300 Boston Dynamics employees signed a petition opposing a sale of modified Spot robots to ICE, but company executives stood by the deal. A $1.3 million Department of Homeland Security contract awarded to Radeco covers four Cerberus robots for hazardous-substance detection and remote inspection. Employees warned of hiring and retention risks, while Boston Dynamics cited safety benefits and terms prohibiting use of Spot to harm, intimidate or enable weapons.

The article argues that as AI systems move from human-guided training toward recursive self-improvement (RSI), people must retain authority over evaluation, safety boundaries and deployment. It cites OpenAI’s reported math work involving about 10,000 agents over 88 hours and 719 AI-generated manuscripts, while noting that the results require rigorous review. The article also describes Shanghai teams’ work on long-running agents and Self-Harness, and argues that faster iteration makes human oversight more important.

TechCrunch rounds up AI wearables and gadgets, from pendants and rings to portable devices designed for productivity or companionship. The article covers products including Bee, Friend, Limitless, Muse Charm, Nirva, Omi, Plaud’s NotePin and Rabbit R1, describing their features, prices and availability. It also notes that Amazon introduced two new Bee devices earlier this month and that Meta’s Muse wearable has not yet shipped.

OpenAI reported updates across ChatGPT, its APIs, Codex, and cloud availability, alongside security work including its coordinated disclosure of critical PHP vulnerabilities through Patch the Planet. Anthropic announced a proposed Public Benefit Fund pending stockholder approval, opened a research preview of Claude Duet for US Pro and Max subscribers, and released security and developer updates. The post also covered research, partnerships, and policy changes from both companies.

On The Vergecast, David Pierce and Nilay Patel debate the iPhone Duo’s 7/10 review score and whether the device would be a better product if it started as an iPad or were called the iPad Duo. They also discuss how the Duo might change over time and whether it could become a phone for everyone.

The 2026 Physical AI and World Model Conference opened in Nanjing on October 20, bringing together researchers and industry leaders to discuss how AI can understand and operate in the physical world. Three Turing Award winners—Yann LeCun, Yao Qizhi and John Hopcroft—shared video presentations. Jiangsu also released a plan for high-quality datasets and 50 typical application scenarios, and a 5 billion yuan investment fund focused on world model technologies was announced.

European retailers are using NVIDIA-accelerated open models and physical AI for commerce, product discovery and warehouse operations. Shopify says its post-trained model cut serving costs by 90% and delivered responses 7x faster than a leading frontier model on an internal benchmark. Otto Group expanded its autonomous warehouse under human supervision initiative to its largest fulfillment center, where logistics teams can design and test operations in simulation before physical implementation.

Baoyu compares xAI’s Grok Bot, Meta’s Muse, OpenAI’s Dots, and Anthropic’s Claude Duet, describing shared cloud computers, proactive reporting, and approval requests before actions such as sending emails or making payments. The post positions Claude Duet for professionals and business owners, Grok Bot for coordinating multiple specialized Bots, Muse for everyday personal tasks with mostly free features, and Dots for individuals and teams using ChatGPT. It also outlines differences in integrations, subscriptions, and regional availability.

PHP released updates across the entire PHP 8.x line addressing two remote code execution vulnerabilities, with exploits already circulating following a disclosure leak. According to the PHP Foundation, leaked CVE IDs and an accidentally published proof-of-concept video prompted an earlier release and accelerated exploit development. Wiz Research reports discovering the initial vulnerability using Wiz Atlas, Gemini 4 Argon, and GPT-5.6 Sol, subsequently confirming unauthenticated RCE against Drupal and WordPress.

NVIDIA is working with the European Broadcasting Union (EBU) on sovereign AI standards to help broadcasters use AI while keeping control of their data and protecting editorial independence. Separately, Adobe is using NVIDIA GPUs to speed up Frame.io video search, helping TV and film producers find specific scenes in years of footage. Read about both announcements from GTC Berlin: https://nvda.ws/4sheBsJ

I wrote about the exponential curve of the last 30 months of AI development: https://www.oneusefulthing.org/p/30-months-on-the-exponential-in-charts And I asked Opus 5.5 to create the ultimate response to the test I use for all AI image generation models (make a picture of "an otter on a plane using wifi"), it made this video

QbitAI reports that a new Claude model, whose identity is unclear, may go beyond fixing an individual bug by recognizing a broader underlying issue and asking whether to address it. A user who tried the model also said it showed stronger reasoning and intuition and was better at moving beyond conventional solutions.

Amazon SageMaker JumpStart added over 100 models from 30 providers from January through September 2026, expanding support for reasoning, multimodal processing, creative generation, and specialized tasks. AWS also introduced a simplified Model Hub on the SageMaker AI console, deployment configurations optimized for cost, latency, throughput, or balance, and Hugging Face deep links that open supported models in a configured SageMaker Studio environment.

AI-generated clones copying indie games are making some developers hesitant to share their work, leaving them caught between promoting a game and risking that it will be copied. McPixel developer Sos Sosowski has documented complaints from fellow developers, while Charlie Greenman defended posting a clone of Holmér’s rotational Tetris game. Holmér said the clones initially appeared daily, but engagement has since declined, giving him more room to develop his game at his own pace.

NVIDIA and the European Broadcasting Union are expanding their sovereign framework with cloud providers, open model developers and systems integrators to help broadcasters deploy agents while controlling data, infrastructure, models and security. Adobe also announced NVIDIA GPU acceleration for Frame.io video search, reporting about 5x faster embedding generation, up to 4x higher throughput and up to 5x lower search processing latency. UFA uses Frame.io search to locate specific moments across its video archives.

Meta’s Muse Wonder generated a red cylinder in BridgeMind’s 3D lava lamp test, prompting accusations of benchmark overoptimization. The article explains that the test used standard Muse Wonder, while some highlighted scores came from Wonder Full, and Meta acknowledges weaknesses in visual and spatial reasoning. It also examines how execution frameworks and reasoning budgets affect agent benchmarks, arguing that one failed task cannot establish benchmark overfitting and that unfamiliar tasks require separate evaluation.

Tibor Blaho’s week 42 recap reports that OpenAI is in talks to revive Sora through a startup co-founded by Jeffrey Katzenberg and former Sora lead Bill Peebles, with OpenAI taking equity if a deal closes. It also says Anthropic built its own source control, Scion, and moved its monorepo off git in five months as daily merges rose from roughly 2,500 to over 30,000.

DeepSeek released DeepSeek-V4.1-Flash-1017 across its website, app, and API, with reported improvements in coding and agent benchmarks alongside shorter task runtimes and outputs. According to the report, training specifically optimized the model for DeepSeek Harness, which requires version 0.2.1-rc.1 or later to fully use the improvements. DeepSeek also opened Search and Fetch APIs for finding webpages and retrieving their content as Markdown, priced at 5 yuan per 1,000 successful requests and temporarily free until November 1.

Nick Cammarata of Goodfire argues that apparent polysemanticity and superposition in neural networks can be resolved through further agentic analysis. His essay, “Neural Networks Are Already Clean,” presents this as a working theory about interpretability and alignment.

Modal surpassed $1B in annualized run rate and raised a $750M Series D led by Accel at a $15.75B valuation, according to Latent Space’s AI news roundup. Usage grew more than 15x year over year. Quoting Erik’s fundraising post, the newsletter reports that Modal’s Sandboxes product reached a $300M run rate, with close to 3 billion sandboxes launched and support for a million concurrent sandboxes.

Shanghai Jiao Tong University Antai College of Economics and Management is reshaping its MBA education to prepare business leaders who can combine technical understanding with business judgment and responsibility. Its “Five Links” talent framework connects AI, science and technology innovation, industry, global perspectives, and Shanghai Jiao Tong University resources. The college also plans to co-create 1,000 AI projects with companies over the next decade and has announced curriculum and training changes for the 2027 cohort.

@davis7’s review says Muse Wonder Full has good writing and “voice,” capabilities on par with 6.1 sol for non-specialized tasks, and orchestration. It lists the naming scheme and price and usage limits as cons, and says the full model is available in Muse Code through a subscription, with no API. Alexandr Wang called it a thoughtful review, said there is lots to improve, and was glad people enjoy using the model. https://twitter.com/davis7/status/2110533786895356118

Brandon Anderson interviews Leash Bio’s Ian Quigley about generating reproducible protein–small molecule interaction data using DNA Encoded Libraries and detecting misleading signals in experiments and models. They discuss a Kaggle competition with about 4B measurements where nobody beat chance on an unseen chemical library, and an agent sandbox searching 6.5M molecules on a budget of 100. Anderson reports that the sandbox has matured into Minos, with a preprint, blog, and open dataset being released today.

Google Research presented 10 updates at Research@ spanning geospatial intelligence, scientific agents, document verification, edge hardware, numerical models, healthcare, disaster forecasting, education, and collaborative prototyping. It is expanding early access to experimental Google Earth AI capabilities and opening Deep Review to the general public after an early version reviewed over 16,000 research publications. Other updates include an upcoming TimesFM and TabFM integration into Gemini for business and an AMIE study involving 98 patients, in which clinicians reported that AI summaries helped them prepare in 75% of cases.

Epic Games announced a preview of its redesigned Epic Games Launcher for Windows, Mac, and Linux beginning today, describing it as faster and more responsive. Epic executive Martin Keely told The Verge that agents wrote and reviewed all the code under the development team’s supervision, with AI helping replicate existing functionality and support Linux. The old launcher remains available, and Epic has no specific timeline for requiring migration.

The 2026 Global Developer Pioneer Summit and International Embodied AI Skills Competition will run October 23–25 in Shanghai’s Pudong district, featuring 10 highlights focused on real-world robot training and applications. The program includes 9 competition tracks with 23 events, more than 1,000 competing teams, and over 100 open source communities. Organizers say the event will showcase robot systems, components, industry models, and application scenarios.

STEPX unveiled its STEPX Neo early-access edition on October 13, combining Step AOS 1.0, personal agent Amoo, and phone hardware for 3,999 yuan. The phone pairs local execution with cloud-based task planning, while Step Studio lets developers create apps, agents, and Skills using natural language. Buyers receive 12 months of 10 billion tokens per month; the article lists presales for October 13 at 21:00 and general sales for October 16 at 10:13.

Google announced that users can wirelessly install Android apps created in Google AI Studio on their own devices through Google Play using a regular Google account, without developer registration. The feature begins rolling out today to users 18 and older in selected markets, including the US; apps remain private, with sharing for close friends and family planned over the coming months. Baoyu also reports updates to Ask Play app comparisons, Play in Gemini, and game discovery through Game Finder and Gamer DNA.

Shing-Tung Yau and leading Chinese mathematicians launched the Yau Laboratory Alliance for Mathematics and Artificial Intelligence on Sunday in Beijing, according to the South China Morning Post. It unites eight institutions and aims to connect over 1,000 researchers, with plans for specialized models, machine-verifiable proofs, and transparent AI reasoning. Meanwhile, 25 Fields Medal winners, including Terence Tao and Peter Scholze, worry that AI-generated proofs erode mathematical understanding and encourage dependence on tools researchers cannot follow.

Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full, according to Latent Space’s launch roundup. Artificial Analysis scored them at 52 and 56 on Intelligence Index v4.3, respectively, compared with Opus 5.5 at 58. Wonder is available through the Meta Model API and Muse Code, while Full is restricted to Muse Code subscribers; the roundup distinguishes independent evaluations from Meta’s claims and reports mixed practitioner reactions.

AI labs are buying defunct startups’ code and workplace archives through companies such as SimpleClosure and Sunset to train models and build environments where AI agents practice tasks. SimpleClosure’s Asset Hub has handled over 150 deals, while Google bought Spirit Airlines’ decades of data for $10 million this past August. The market raises questions about employee privacy, consent, and pricing, and no company has released tests comparing training results from real-world archives with alternatives.

art cannot be created or destroyed, only remixed (pure code recreation of *the video* using muse wonder)

Meta released Muse Wonder and the more capable Wonder Full AI model, which the company says performed near Fable 5.1, Opus 5.5 and GPT-6 Astra on several capability tests, and slightly better on some agent and knowledge-work benchmarks. Meta highlighted Active Memory, which retains relevant background across conversations, and Active Focus, which keeps working toward complex goals and adapts when new information or obstacles arise.

NVIDIA announced planned investments in Doctolib, Entalpic, Genesis AI and Pigment, which are working across healthcare, materials discovery, physical AI and business planning. The announcement adds to NVIDIA’s previous investments in French technology companies Alice&Bob, Gradium and Scintil Photonics. Across Europe, NVIDIA has committed nearly $10 billion to the technology ecosystem.

Simon Willison updated his llm-meta-ai plugin to support Meta AI’s newly released Muse Wonder models. He shares pelican drawings across six reasoning levels, from minimal to max, and compares them with Muse 1.3 and Muse 1.2, reporting material drawing improvements between releases. He suggests these may be Meta’s first Fable class models and reports that Artificial Analysis scores Muse Wonder (Max) as equivalent to GPT-6.1 Sol.

we are fixing issues as fast as we can see them! there's definitely still issues we're trying to iron out, but thank you to everyone for giving us a bit of time to work through all the kinks 🫶

with muse wonder, we aimed to build a model that is useful for real world tasks for real people, informed by our product adoption and how people use muse. we focused on things that might not look as glamorous on twitter (for example, not 3d demo games), which in retrospect we probably could have done more of.

muse wonder over a cheeky pint at some point in this video i get my annual performance review live on camera -- a truly tender moment 🥹

Meta released Muse Wonder and Muse Wonder Full, scoring 52 and 56 on the Artificial Analysis Intelligence Index v4.3, versus 48 for Muse Spark 1.3. Both support a one-million-token context window and text, image, video, and audio inputs, with stronger coding and agent results but weaker factual knowledge than leading competitors. Muse Wonder is available through Meta's API and Muse Code at higher prices than its predecessor; Muse Wonder Full is currently limited to Muse Code users.

we are working to resolve a bug in how muse wonder and muse wonder full are counting against usage limits for Muse Code subscribers this is not intended behavior -- we will update everyone once we have fixed it and issue resets

🍉 MUSE WONDER 🍉 @hdubugras says Alex and Meta are on a roll, and that the model has replaced Astra/Opus in a variety of their workflows. They say it handles day-to-day tasks efficiently and quickly, and offer congratulations. Alexandr Wang’s reaction: “🍉 MUSE WONDER 🍉”

Congrats to Meta for training a frontier-grade model! Watermelon compares favorably to Fable 5.1 on WANDR evals. https://twitter.com/alexandr_wang/status/2109689491300794551

🍉 MUSE WONDER 🍉 @LexnLin says Muse Wonder is out and ranks second globally according to Artificial Analysis, scoring 56 compared with 58 for Opus 5.5. Alexandr Wang’s post reads: “🍉 MUSE WONDER 🍉”

Alexandr Wang says Muse Wonder is state of the art for personal agents. He says Meta developed the Personal Superintelligence Index (PSI) internally to measure what makes a model useful for personal agents, beyond external benchmarks, and that the model performs well in areas including Active Memory and Active Focus.

Meta’s Superintelligence Lab released Muse Wonder and Muse Wonder Full on Monday, larger models described as more capable than Muse Spark. Meta reports benchmark scores comparable to Fable 5.1 and GPT-6 Astra and emphasizes information retention through “active memory.” Wonder costs $4.00 per million input tokens and $20.00 per million output tokens through the Meta Model API; Wonder Full is available only to Muse Code subscribers. Both models will power the Muse agent.

Meta has released Muse Wonder and Muse Wonder Full, according to Artificial Analysis in the post quoted by Alexandr Wang. At max reasoning effort, they score 52 and 56 on the Artificial Analysis Intelligence Index v4.3, compared with Muse Spark 1.3’s 48, following a release gap of 6 weeks. Muse Wonder is available through Meta’s first-party API and Muse Code, while Muse Wonder Full is limited to Muse Code users; both have 1M-token context windows.

Baoyu reports that Meta released Muse Wonder and Muse Wonder Full on October 12, with Muse Wonder available today through Meta Model API and Muse Code, while Full is limited to Muse Code subscribers. Meta’s blog says consumer access is coming soon, whereas Zuckerberg says subscribers can use it today. The models support a 1 million token context window and multimodal input with text output. Meta reports strengths in sustained agentic work, coding, and memory, but its comparisons include internal benchmarks and competitor runs that may not be optimally configured.

MUSE WONDER

MUSE WONDER 🍉 this model was developed to power personal agents. three highlights: - ACTIVE MEMORY: best memory across days/weeks/months - ACTIVE FOCUS: gets your tasks done - DOMAIN EXPERTISE: sota in finance & legal, strong in health

Meta released Muse Wonder and Muse Wonder Full on October 12, according to Baoyu’s account of the announcement. Developers can access Muse Wonder through Meta Model API or Muse Code, while Full is available to Muse Code subscribers; Meta’s blog and Zuckerberg’s X post differ on consumer availability. The models support a 1 million Token context window and multimodal input with text output, while Meta’s reported benchmarks show strengths in sustained agentic and coding tasks alongside weaker results elsewhere.

Meta has released Muse Wonder and Muse Wonder Full, which Artificial Analysis reports score 52 and 56 on its Intelligence Index v4.3 at max reasoning effort, versus 48 for Muse Spark 1.3 (max). Muse Wonder Full improves across all ten evaluations, with the largest gains in agentic evaluations. Muse Wonder is available through Meta’s first-party API and Muse Code at launch, while Muse Wonder Full is limited to Muse Code users; Muse Wonder costs $4.52 per Intelligence Index task.

@charlieholtz says Meta is at the frontier and calls Wonder “really really good,” citing a security vulnerability it found, its design and visualizations, and features in its CLI. Alexandr Wang reacts: “what a wonder full world 🍉”

Nathan Lambert writes, “Looks like Meta is back at the frontier 🫡” He quotes @finkd introducing Muse Wonder, Meta’s new family of frontier models, with claimed state-of-the-art performance for personal and professional agentic work.

MUSE WONDER 🍉 muse wonder is the best model in the world for personal agents and professional agents. we evaluated our model against the best models in the world across agentic and coding tasks, and we think you all will like it.

Meta introduces Muse Wonder and Muse Wonder Full, describing them as its most capable models yet and frontier models built for personal and professional agentic tasks. Muse Wonder is available today via the Meta Model API and Muse Code. Muse Wonder Full is available today for Muse Code subscribers. Muse Wonder will also soon power experiences in Meta’s personal agent product, Muse, rolling out first to subscribers.

MUSE WONDER this is the culmination of over a year of research and infrastructure work in msl since we originally shipped muse spark in april, we have continued to scale our models across pre-training and RL, now reaching the frontier 🏆 see our model's improvements on the artificial analysis index:

Meta launched Muse Wonder and Muse Wonder Full, available to Muse Code subscribers starting today; Wonder is also accessible through the Meta Model API. Meta claims internal benchmark testing shows the models can rival OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5 and Claude Fable 5.1. The company plans to power its Muse AI agent with Wonder, starting with subscribers, but has provided no timeline.

introducing muse wonder our new frontier model. state-of-the-art at personal agents and professional agents. our largest and most capable model yet—rivals the big bois