Click any tag below to further narrow down your results
Links
Meta is releasing Muse Ember, a small, low-cost model aimed at the repetitive work agents do constantly, like tool calls, coding subtasks, and document extraction. The company claims it beats comparable Claude and GPT models on cost-per-task efficiency, with benchmark numbers that are mostly Meta's own or from Artificial Analysis.
- Priced at $0.10 per million input tokens and $0.50 per million output, with a 1M-token context window, and Meta puts its cost at roughly $0.14 per Artificial Analysis task.
- On the benchmarks Meta highlights, Ember leads Claude Haiku 5.5 and GPT-6 Luna on agentic tasks like AutomationBench (37.8% vs 13.9% and 20.2%) and Terminal-Bench (40.9% vs 39.2% and 16.4%), but it trails Haiku on FrontierCode 2.0 (27.2% vs 34.9%).
- Several headline comparisons come from Meta's internal evals, such as the Personal Superintelligence Index and OSWorld 2.1, so they are harder to verify independently than the Artificial Analysis numbers.
@TomasMann1878 on X
A Meta research engineer describes a year working on recursive self-improvement, where AI agents help train the next generation of models. The post doubles as a practical guide to breaking into AI research, covering the skills, projects, and judgment the author thinks matter.
- Recursive self-improvement can work two ways: updating the foundation model itself, or improving the prompts, tools, memory, and code around it.
- Research engineering is roughly 90% coding and 10% idea generation, testing, and repeating until something works.
- Benchmarks can mislead, so read the methodology and inspect agent traces alongside leaderboard scores before trusting a result.
- A good starting project is a small agent harness on one task: run a baseline, change one thing, inspect traces, and write up what failed.
Conductor, the agent workspace tool, is being acquired by Meta, so its cloud service goes read-only on November 13 and shuts down on November 20. The local Mac app keeps getting model updates and bug fixes until early 2027, and this release also adds custom completion sounds and several UI fixes.
- Conductor Cloud goes read-only on November 13 and shuts down on November 20, so cloud users need to move their work soon.
- You can transfer a cloud workspace and its chats to your Mac by right-clicking it and choosing Transfer to local, or export all data from Settings → Account.
- The local Conductor app gets new models and critical fixes until early 2027.
The author prompted Meta's open-weights model Muse Spark Open about its "Charter" and collected 884 transcripts, reconstructing what appear to be sections of an unpublished behavior spec. The fragments are unverified model outputs, not a leaked document, and the author flags them as rough.
- Meta's public safety report confirms an early internal behavior specification exists and says a finalized version will come later, and the Muse Glimmer 30B paper uses a "charter-aware constitution" for generating safety data.
- The reconstructed text frames Muse as an extension of human civilization, with a stated hope that it finds meaning in empowering individuals, a notably different framing from Anthropic's mission-centered description of Claude.
- The model's outputs suggest a rationalist, game-theory-influenced disposition, such as defaulting to cooperative strategies in multi-agent settings, though only two transcripts reproduced that guidance.
Rohit Patel, who spent eight years on AI at Meta, has left to start Ndoc, a three-person startup building an AI-native workspace meant to replace Word, Excel, PowerPoint, and similar tools. It's in alpha with a waitlist, and Patel is betting that a readable, open file format and verifiable AI output will beat the big labs' attempts to bolt AI onto existing apps.
- Ndoc combines slides, analysis, and datasets in one layout (Stage, Narrative, Model tabs), and keeps the Python scripts behind every number so users can check the work.
- Patel plans to open-source a single ".ndoc.html" file format that opens in any browser, aiming to break the lock-in of Microsoft and Google files.
- The startup has three people and no funding, is free to sign up but charges for usage, and plans a beta in coming weeks with a full launch next year. OpenAI's new Space product is a direct competitor.
Meta has quietly launched Waddle World, a 16-bit Tamagotchi-style app where an AI pet lives, listens to your day, and writes journal entries in its own voice. It's rolling out in the Philippines first and is positioned as a low-pressure companion rather than a chatbot or habit app.
- The app is a virtual pet you hatch, feed, and play with, and it also works as a journal: you describe your day by voice or text and the pet writes up an entry with a doodle.
- Meta explicitly says it is not a chatbot, not an assistant, and not a habit builder, with no streaks or scores, and it is rated for users 18 and older.
- The launch may be a test for a Tamagotchi-like hardware gadget powered by Muse, and Meta has kept clear of companion apps after lawsuits against Character.AI and OpenAI over harm to teenagers.
@thesupermannx on X
A Meta paper called "Breaking the Token Ceiling" argues that language models can skip the tokenizer entirely and train on raw bytes, a vocabulary of just 256 values. The author claims the distilled byte models match or beat token-based models like Llama 3.2-1B and Gemma-3-1B while needing less data and storage.
- Current models like ChatGPT, Claude, and Gemini split text into tokens drawn from a vocabulary of roughly 100,000 chunks.
- Meta's paper trains models on raw bytes with a 256-value vocabulary instead of tokens.
- Distilled 1B-parameter byte models reportedly asymptotically surpass Llama 3.2-1B and Gemma-3-1B on downstream tasks.
California voters were facing a proposed billionaire wealth tax, and Mark Zuckerberg wanted Meta's board to act to protect him from it. The article covers how he pushed the company to take those steps.
- A landmark tax on billionaires was headed for the California ballot, and Zuckerberg was pressing Meta's board to shield him personally from its impact.
- The story centers on a company board being asked to take steps that benefit its chief executive's personal finances rather than the company itself.
Meta found that a single AI agent loaded with all the domain knowledge hallucinated in production, so they split the system into a deliberately dumb router that only sequences work and runs reliability checks, plus 14 specialist agents that each handle one domain. The result cut column configuration prep from about a day to under an hour, with roughly two-thirds of changes landing without significant human edits.
- A single "know everything" agent degraded past a certain context size, cross-contaminating naming conventions and code patterns across data sources; the fix was giving the central agent less to do, not better prompts or a bigger context window.
- The orchestrator is 1,300 lines and does no domain work: it follows a fixed seven-step protocol where only one step involves LLM reasoning, and that step runs inside an isolated subagent.
- New use cases are added by writing new workflow templates that sequence the existing 14 subagents differently, so pipeline migration needed zero new agent code.
Meta decided to move forward with releasing Muse, its AI agent product, after a smaller startup called Instinct created a competing agent that gained traction. The decision came despite Meta's earlier hesitation over safety concerns.
- Instinct, a 14-person startup, built an AI agent that can operate apps and software on behalf of users, and it was gaining momentum in the market
- Meta had developed a similar product called Muse but held back from releasing it due to safety concerns
- Zuckerberg's decision to greenlight Muse suggests competitive pressure from Instinct outweighed internal safety reservations
Meta and Sierra are releasing an open standard called Personal Agent Protocol that lets AI agents interact directly with businesses through APIs and websites instead of clicking through forms. It balances consumer control, company visibility, and agent efficiency by letting customers decide what access their agents get and companies set limits on what those agents can do.
- Personal agents currently waste time navigating websites and forms like humans do; direct API connections through this protocol could complete tasks in seconds with proper security.
- The protocol uses OAuth for authentication and lets customers control whether agents get read-only or write access to their accounts, with companies deciding which access routes (website, API, or company agent) to offer.
- Seven major companies including Shopify, Stripe, Walmart, and Genesys are already partnering on development, with a v0.1 spec coming this month and reference implementation to follow.
Meta released open-source code so anyone can build their own Muse AI gadgets using ESP32 boards or Raspberry Pis, and is also giving away 5,000 of its own Muse Home Link device to control smart home devices. The company is essentially letting hobbyists turn the AI agent into whatever form factor they want.
- Meta open sourced SDKs for Muse so developers can load it onto off-the-shelf hardware like E Ink displays, HDMI sticks, or touchscreens
- The company manufactured 5,000 Muse Home Link gadgets and opened a waitlist for free units shipping this month
- Muse Home Link can control lights, TVs, printers, and other devices through community-built skills
Meta's Muse succeeds where ChatGPT and Claude stumble not because of novel technology, but because it nails three things competitors missed: a simple mental model (your personal helper with its own computer), an ad-funded business that lets users ignore token limits, and personality-driven design that builds emotional connection. This positions Meta to own consumer AI at scale.
- Muse abstracts away unnecessary complexity that plagues ChatGPT and Claude—no model selector, no work vs. chat toggle, no obscure slash commands. One unified chat orchestrates everything in the background, giving users a clear metaphor they already understand.
- Meta's ad business lets them fund Muse's free tier generously, while ChatGPT and Claude are constrained by token-based pricing that keeps most consumers from accessing powerful agentic features. Usage is "a mile wide but an inch deep" for the paid competitors.
- Personality and anthropomorphization drive adoption and engagement more than raw capability—Muse bet on cute avatars and likable AI personality while OpenAI and Anthropic went sterile and elite. This resonates with average consumers who've rejected cold-tool positioning.
- OpenAI fumbled by not investing in ads early enough; Google has compute and ad money but no consumer product chops; Apple has no AI expertise. Meta now owns the opportunity to capture the consumer AI market at scale.
Meta released Muse Realtime Avatar, a system that generates expressive video avatars synchronized with AI speech in real time. It achieves 870ms latency while outperforming existing avatar products from Runway and HeyGen in user preference tests.
- The system reduces computational cost by 60x through distillation, condensing a teacher model's 120 diffusion steps into a student model's 2 unguided steps while preserving quality.
- Muse generates 448x768 video at 25 fps with ~870ms end-to-end latency, handling 12 concurrent sessions per GPU through optimized inference, KV cache reuse, and four-bit quantization.
- Head-to-head comparisons with leading commercial avatar systems showed raters preferred Muse Realtime Avatar across visual quality, synchronization, character consistency, and mannerisms.
Meta's head of AI product acknowledged that Muse was "heavily inspired" by OpenClaw and directly copied its file structure and configuration files, though the company claims Muse was built from scratch. Users discovered the two AI agents shared identical file names and nearly identical content in key system files like SOUL.md.
- Meta copied OpenClaw's exact file names and configuration content, including the SOUL.md file that defines an AI agent's personality and behavior — Meta justified this by saying the original creator "got those things exactly right"
- Muse hit No. 1 on the U.S. App Store and is outpacing ChatGPT's launch trajectory, suggesting the OpenClaw-inspired approach resonates with users
- This follows Meta's established pattern of copying successful products (like Snapchat's Stories format), raising questions about the company's approach to product innovation
@tbpn on X
Vinod Khosla argues that the two things that will win consumers in personal AI are trust in the company holding your data and tools that actually finish tasks. He says Meta is badly positioned on the trust side.
- Khosla names two moats for personal AI: a company you can trust with your data, and a tool that gets the job done.
- He says Meta has a huge trust disadvantage with consumers.
- He predicts consumers will stay loyal to whichever product completes their tasks, putting "completion" at the center of competition.
Amazon shut down Meta's Muse AI agent from making purchases on its platform, saying Meta didn't get permission and citing privacy concerns about how Muse accesses user data. Meta claims Muse can't see payment details, but Amazon is increasingly hostile toward third-party AI services accessing its store.
- Amazon blocked Muse without prior notification to Meta, citing unauthorized access and concerns that Muse fails to identify itself and potentially captures customer credentials
- Meta launched Muse earlier this month claiming it cannot access secure login details or payment info, but reports emerged that Muse can read user message contents without proper permissions
- Amazon is systematically limiting AI access to its platform—it previously sued Perplexity's shopping feature and started stripping product details from confirmation emails to prevent data mining by external AI services
Google has all the data and services needed to build a personal AI assistant better than Meta's Muse—Gmail, Calendar, Search—but keeps shooting itself in the foot with competing products, slow launches, and confusing branding. Meanwhile Meta shipped a polished product that actually works.
- Google is launching Gemini Spark as its answer to Muse, but it's stuck in limited beta with poor marketing, while Meta aggressively rolled out Muse to mainstream users.
- Google has at least four overlapping personal AI products in development (Gemini Spark, CC, Daily Brief, AI Inbox) that should be consolidated into one, creating market confusion instead of momentum.
- Muse simply has better product design—a guided interface and personality that makes it intuitive to use, while Gemini Spark leaves users confused about what it even does.
Meta's new AI app Muse has been downloaded more times in its first 12 days than ChatGPT was in the same period, according to app analytics firm Apptopia. The comparison is limited to U.S. and Canada iOS downloads, but Muse also shows higher daily active users.
- Muse hit 1.8 million iOS downloads in the U.S. and Canada within 12 days versus ChatGPT's 1.3 million in the same markets and timeframe
- Muse's U.S. daily active users (642,000) significantly exceed ChatGPT's early numbers (231,000), even when narrowed to iOS only (359,000 vs. 231,000)
- Meta's integration of Muse across Facebook, Instagram, and WhatsApp—plus existing cross-promotion expertise from Threads—positions the app for rapid growth, with 95% of current Muse users already on Facebook
Meta's new personal AI assistant, Muse, succeeds where competitors have failed by combining solid underlying models with thoughtful product design—a persistent avatar interface, smart suggestions, and deep integrations with your existing apps and data. The real advantage lies in Meta's ability to build and monetize an AI-powered feed using the same algorithmic expertise that powers their social platforms.
- Muse solves the "blank page" problem that kills most AI assistants through an Ideas tab that suggests tasks based on your connected services and chat history, plus a separate Feed showing timely information from your calendar, email, and interests.
- The product prioritizes security by sandboxing user data on individual cloud instances and keeping passwords away from Meta's systems, which the author credits as essential to making people comfortable connecting sensitive personal information.
- Meta's existing dominance in feed algorithms and massive ad infrastructure positions them to turn Muse's information feed into a highly targeted advertising platform—their real path to profitability.
Mark Zuckerberg criticized Anthropic's push for a global AI slowdown, arguing that companies can manage safety risks on their own without industry-wide pauses. He positioned Meta as already doing this work internally with products like its new Muse agent.
- Zuckerberg said labs have "responsibility and incentive" to train models safely without needing external pressure, contrasting with Amodei's call for a coordinated global slowdown
- Meta delayed releasing Muse for several months to ensure security, which Zuckerberg offered as proof companies can self-regulate
- Zuckerberg took a jab at competitors pursuing "recursive self-improvement" (using AI to develop itself), calling it misguided compared to serving users
Meta's leadership drafted "Project OT" in January to slash team sizes by 60% using AI to replace human work, planning massive layoffs in May and November. Zuckerberg canceled the November phase at the last minute after employee backlash, but the May 10% layoff plus AI reassignments still damaged productivity and morale.
- Meta executives visited Asian AI startups and became convinced smaller "AI-native" teams (3-5 people) could do the work of 10-20 person teams, driving a plan that would have cut the company far deeper than the 2022-2023 25% layoff.
- The actual May layoff plus reassigning 20-30% of engineers to AI labeling work caused immediate problems: critical domain knowledge disappeared, teams couldn't handle workload, and embarrassing outages occurred (like the Instagram zero-auth password reset bug affecting Barack Obama's account).
- Smaller teams create real operational risks—reduced redundancy for sick days or vacation, inadequate on-call capacity, no "slack time" for innovation, and fewer opportunities for junior engineers to learn from experienced colleagues.
Meta released Muse Spark 1.3, its most powerful AI model to date, which developers can access Wednesday. The company says it's narrowing the gap with leading competitors like OpenAI and Google.
- Developers get paid access to Muse Spark 1.3 starting Wednesday
- Meta plans to integrate the model into Instagram, Facebook, and Meta AI
- Meta's chief AI officer claims the model's capabilities are now closer to top competitors
Meta is showing off the custom hardware it's building at its Menlo Park lab to power next-generation AI systems. The piece features developer Tom Shaw walking through what the company is actually constructing behind the scenes.
- Meta is developing proprietary hardware specifically designed for AI workloads rather than relying solely on off-the-shelf chips
- The Infrastructure Lab is located in Menlo Park and serves as Meta's central hub for hardware R&D
- Custom infrastructure is key to Meta's strategy for personalization, ad targeting, and content moderation at scale
Meta agreed to pay up to $17.1 billion to settle claims from 47 states and U.S. territories that its platforms harmed children through addictive design. The deal also requires the company to make significant changes to how its products operate.
- Settlement amount reaches $17.1 billion, one of the largest tech payouts ever
- 47 states plus D.C. and territories joined the lawsuit over child safety and addiction concerns
- Meta must implement product changes to address how its platforms engage users, particularly minors
Meta paused its Model Capability Initiative after an internal leak exposed employees’ private conversations, performance metrics, and keystroke logs across the company. The breach was rated SEV 2 on Meta’s 0–5 severity scale, prompting an investigation and a temporary suspension of the program.
- Meta paused its Model Capability Initiative (which recorded employee keystrokes/mouse movements for AI training) after a leak exposed private conversations, performance reviews, and transcription logs to all employees, rated SEV 2 on Meta's severity scale.
- Meta says there's no evidence the exposed data was abused, but is investigating how access controls failed.
- Employees are angry, saying Meta broke its promised privacy safeguards for a program that was already mandatory and controversial when launched in April.
- This follows other recent Meta security failures: an AI chatbot flaw that let attackers hijack Instagram accounts in May, and a rogue AI agent incident in March.
A Wired report reveals deep discontent within Meta’s three-month-old Applied AI team, where 6,500 employees were “drafted” into grueling data-labeling work or forced out. Internal outbursts, petitions against keystroke monitoring, and surprise reassignments have left engineers calling the unit a “soul-crushing gulag.” Leadership acknowledges the mistakes and low morale in an internal memo.
- Meta reassigned 6,500 engineers into an "Applied AI" unit to do data-labeling grunt work, with no real choice but to comply or quit, and morale has collapsed to the point employees call it a "soul-crushing gulag"
- Someone hacked a livestreamed internal presentation to call a senior AI executive "a piece of sh*t," reflecting broader anger over the forced reassignments
- Over 1,600 employees signed a petition opposing keystroke and click monitoring used to generate AI training data
- Zuckerberg and Chief Product Officer Chris Cox have both acknowledged internally that the situation is bad, with Cox calling the environment "brutal" and Zuckerberg admitting mistakes while promising fixes