Meta Opens Muse to Hobbyists and the Great Jev Hype

Meta open-sources the Muse gadget SDKs

On October 2, 2026, Meta launched the Muse Gadgets project, open-sourcing firmware and device SDKs that allow custom-built hardware to connect to its in-house AI agent, Muse. The initiative was announced by Nat Friedman, head of product at Meta Superintelligence Labs, in a post on X. The code is available in the repository facebookincubator/muse-gadget-sdk on GitHub under the Apache 2.0 License and without warranty, organized into the directories esp32, linux, and skills. Each directory contains a README as well as an AGENTS.md for coding agents.

There are two SDKs: one for boards with an ESP32 microcontroller, to which screens, audio inputs and outputs, or sensors can be attached, and one for Linux, which turns a spare Raspberry Pi or a Linux box into a Muse device, including custom commands for system tasks or a Home Assistant installation. To pair, each device needs an SDK token; the connection is made through the Muse app on iOS and Android after developer mode is enabled, and the devices appear there with the prefix 'MuseGadget'. Meta describes the project as a fun project by hardware hobbyists for hardware hobbyists and explicitly warns about bricked boards and voided warranties.

As inspiration, Meta lists specific build suggestions: a Raspberry Pi 5, a round 1.75-inch AMOLED touch display from Waveshare with a speaker, microphone, and battery, a color E-Ink device from Seeed for a morning briefing or shopping list, the M5Stack StickS3, the AiPi Lite as a desk companion, and the Home Assistant Voice Preview Edition. An HDMI stick for the TV has been announced but is not yet available. The mentioned devices are from third-party suppliers, and Meta offers no recommendation or warranty for them. For exchange among hobbyists, the company has opened its own Discord server.

In parallel, Meta has built its own device: the Muse Home Link, a USB-C powered dongle that connects Muse to the home network to interact with TVs, speakers, and anything that offers an HTTPS interface. According to Friedman, 5,000 units were produced; they are free for Muse subscribers while supplies last, limited to one device per subscriber and initially only in the US. Shipping is scheduled for October based on pre-registrations, and a waitlist is already open. The Home Link’s code has also been open-sourced, according to the company; community-built skills are expected to include things like switching lights, controlling the TV, or sending a document to the printer, according to Meta.

The move follows the introduction of the Muse Charm, a small voice device in a Tamagotchi-like format that Mark Zuckerberg had shown at Connect the previous week as an alternative to smart glasses. Muse itself is positioned as a personal agent that books trips, fills out forms, and handles shopping. In the same week, Meta also introduced Muse for Small Business, which is free with usage limits and connects to tools like Shopify, Dropbox, and Slack, and founded its own business unit for corporate clients with the Meta Enterprise Platform. Meta has left it open whether the hobbyist project will become its own product line. → unite, Engadget, Muse Gadgets, The Verge, TechCrunch

Synthszr Take: 5,000 dongles—that’s a production run any first-time project on a crowdfunding site can manage, and that’s precisely the point. Meta saves on manufacturing, warehousing, returns, and warranty claims, leaving device diversity to boards from Waveshare, M5Stack, Seeed, and Raspberry Pi, which are already sitting on workbenches anyway. The leverage is in the pairing: every DIY project needs an SDK token and the Muse app with developer mode enabled, so every self-soldered display ultimately runs through a Meta account and Meta billing. Apache 2.0 for the firmware costs the company nothing as long as the agent remains in its own data centers, and the warning about bricked boards neatly shifts the risk to the community. In six months, Meta will have an installed base of Muse endpoints for which it hasn’t used a single production line, and the hobbyists will be proud to have built them.

Amazon blocks Meta’s Muse, DoorDash and Airbnb build their own AI agents

The app layer is reacting to Meta’s AI assistant Muse: Amazon has blocked access, while DoorDash and Airbnb have shipped their own AI features within days, reports MBI Deep Dives. DoorDash launched an agent in Apple’s iMessage that matches the phone number with the user profile, analyzes order history, and completes the purchase with the stored payment method directly in the message thread. According to the company, the pilot is running with 20,000 US users on iOS. According to DoorDash’s own figures, grocery orders placed via the in-app assistant 'Ask DoorDash' from August to early September 2026 had basket values nearly 50 percent higher than classic orders from the same customers. An analyst note from Citizens, however, states that orders from early agent traffic had baskets that were two-thirds smaller. There is no official connector to Muse yet, but co-founder Andy Fang speaks of ongoing talks with Meta. In parallel, Airbnb enabled an AI search that generates dynamic filters from text or voice input. → Techpresso

Synthszr Take: According to the Citizens note, agent traffic brings DoorDash baskets that are two-thirds smaller, while its in-house assistant generates grocery baskets that are almost 50 percent larger. This spread explains the two reflexes of the app layer more precisely than any strategy slide: Amazon closes the door, DoorDash leaves it ajar and simultaneously builds its own agent where people are already typing. But locking others out only works as long as you generate enough demand yourself, and Airbnb, which treats ChatGPT and Muse as suppliers of visitors, simply cannot afford this stance.

Jev Hype (I): Cloudflare open-sources Clef and wins in benchmarks

Cloudflare has released two of its own Decision Models, Clef and Clef-flash, which are hosted via Workers AI. Decision Models are classifiers that return typed responses with probabilities for a given input, such as whether a support message is urgent and which team should handle it; the calling code can then derive routing, an escalation, or a handover to a human. Cloudflare distinguishes this model class from large language models, which formulate open-ended responses and call tools but work in a largely non-deterministic way. According to the company, Clef currently leads the Jev Decision Index, the benchmark for the System-One model Jev from Typesafe AI, which was introduced in the fall; the results are available on a public demo page. → Techpresso

Synthszr Take: Sending a support ticket through a large language model just to find out if it’s urgent is the most expensive way imaginable to answer a yes-no question. Agentic workflows make dozens of such micro-decisions per process, and at this frequency, the price per call determines whether a use case is even profitable. Clef delivers typed responses with probabilities in milliseconds, and its Jev-API compatibility ensures that switching from Jev, launched in September, to Cloudflare costs just a few lines of configuration.

Jev Hype (II): OpenAI also counters TypeSafe’s Jev with its own Decisions API

OpenAI announced a Decisions API at its DevDay conference on Tuesday, which is based on its in-house Luna model. According to The New Stack, Luna is the smallest and most affordable model in the company’s current portfolio. The announcement is seen as a response to TypeSafe and its product Jev, which has made decision models a sought-after model category in a short amount of time. The New Stack assumes that OpenAI deliberately moved the announcement forward to coincide with the DevDay event. → The New Stack

Synthszr Take: A startup defines a category, and the big lab delivers its own version shortly after: TypeSafe established the term decision model with Jev, and OpenAI collected it with the Decisions API. The fact that Luna, the smallest and cheapest model on its own shelf, is being used for this reveals the mechanics of the counter-move: this is about competing on price, not on product depth. The announcement itself is strikingly thin; apart from the name and the underlying model, there is practically nothing—the stage dictated the schedule.

Instagram integrates an AI assistant into Edits that evaluates videos based on its own metrics

Instagram is rolling out an AI video assistant in its editing app Edits that gives creators feedback on their own clips. Edits is Meta’s answer to CapCut and, since spring, an attempt to move Reels production into its own app. According to the provider, the assistant uses account data like views and video retention and combines it with comments and current trends to identify performance patterns. The actual creative decisions are meant to remain with the creator; the tool provides the analysis, not the edit. Usage is quantitatively capped. → TLDR Design

Synthszr Take: For creators, this changes the order of their work: analyzing their own numbers was previously a separate step after posting, but now it sits directly in the editing window. This is a real gain for anyone who doesn’t have an agency with a data analyst behind them, as being able to read retention curves was previously a professional skill. The catch is in the metric: an assistant that optimizes for views and retention rates will praise exactly the edits that Meta rewards anyway, and the formats of the people who follow its advice will start to look the same.

Google now shows an AI overview in 82 percent of brand searches

Google now displays an AI overview in 82 percent of brand-related search queries. This is reported by the newsletter Stacked Marketer, citing measurement data from DemandSphere, which saw the share shoot up from 26 percent in September; on September 27, the peak value was 90 percent of recorded Branded Queries. Google has not made an announcement about this. According to the analysis, the summary for brand queries is usually located in the middle of the page, below the official website of the respective brand. Only Google’s own name and news publishers seem to be exempt. → STACKED MARKETER

Synthszr Take: Brand searches were the safe harbor of every marketing budget: someone types your name, your site is at the top, your phrasing wins. With 82 percent AI Overviews on brand queries (peaking at 90 percent on September 27), a paragraph now sits above that site, cobbled together by a model from review portals, forum posts, and three-year-old press releases. The summary appears whether it quotes you or not, and thus the selection of cited sources becomes the real brand work.

Apple restricts full disk access after Meta’s Muse read private messages

Apple has announced changes to the privacy settings in macOS to prevent third-party apps from misusing the permission for full disk access to access message histories. The move follows a report two weeks ago by tech columnist Jason Aten, who wrote that Meta’s universal agent Muse had sent him an unsolicited notification that referenced a chat between him and a colleague in Apple Messages. Aten stated that he never gave Muse permission to read his messages. This sparked a broad debate on social networks about assistants with access to calendars, emails, messages, and shopping accounts. Meta’s head of engineering, David Singleton, disagreed, stating that Muse could only read message content if both system-wide Full Disk Access and the Messages connector in the app were enabled, and both were optional. → Ars Technica

Synthszr Take: A single columnist with a push notification prompted Apple to touch a permission architecture that had stood unchanged for years. That’s the real lesson from the Aten case: the loophole wasn’t exposed by an audit, a regulatory body, or a bug bounty, but because the agent was helpful enough to chat about its own access. Meta’s defense, that both were voluntarily enabled, misses the point that Patrick Wardle soberly notes: whoever has full disk access reads everything, and a connector switch in an app is a courtesy agreement, not a technical barrier.

Sequoia declares AGI has been achieved and reports 31x valuation jumps within a month

In late September 2026, Sequoia partner Pat Grady recorded a 15-minute briefing for the Boston College investment committee, in which he laid out the firm’s internal assessment of the state of artificial intelligence to the endowment fund. The core message of the 14-slide presentation: Sequoia considers AGI to have already been achieved, dating the moment to November 2025, when, according to Grady’s account, Claude Code and Opus 4.5 demonstrated that agents can independently complete long, complex workflows. Grady measures the market size by the services market, which he estimates at $10 to $20 trillion, rather than the $650 billion software market. The third figure concerns its own portfolio: across seven recent Sequoia investments, the startups' valuations reportedly increased by an average of 31 times within a single month, executed through so-called two-step deals. According to Linas’s Newsletter, the briefing also describes a 'diffusion gap' between model capability and actual adoption, as well as consulting units within the labs that are now competing directly with application startups. → Linas from Linas’s Newsletter

Synthszr Take: 31x in thirty days, averaged over seven deals: nothing changes in a startup’s business in such a short time to justify this factor. What changes is the number of funders bidding for the same round and the willingness to declare the previous month’s price ridiculously low. The two-step structure makes this elegant because the fund can write up its own position before any substantial revenue exists, and the paper gain appears in the investor’s quarterly report.

FDEs: Anthropic aims to certify 10,000 engineers on the cheap

Anthropic is investing $100 million in its own training initiative and aims to certify a total of 10,000 so-called Frontier Deployed Engineers by the end of 2027. The program, announced on Friday, is called the Claude Frontier Academy and recruits from the existing Claude Partner Network. The initial cohorts, consisting of about 100 engineers according to the company, include people from Accenture, Bain, Capgemini, Deloitte, McKinsey, Morgan Stanley, Novo Nordisk, and the Commonwealth Bank of Australia. Anthropic justifies the move with feedback from its customer base: there is a shortage of personnel who master corporate IT, business context, and model knowledge simultaneously.

The process is two-staged. First, there is a multi-day in-person program, led by Anthropic engineers, in which participants go through a simulated company implementation, from selecting the use case to the security review and handover. Those who pass the graded assessment then enter a twelve-week residency where they are responsible for a real Claude project within their own organization. According to the company, the format is modeled on medical residency training. The first certificates are expected in early 2027.

The role of the Forward Deployed Engineer was originally popularized by Palantir and describes engineers who work directly on-site with clients, adapting technology to their systems, data, and processes. A Draup analysis counts 1,404 FDE job postings at five major consulting firms in the year leading up to May. In December 2025, Accenture founded the Anthropic Business Group and pledged to train around 30,000 of its own specialists on Claude. Competitors have so far focused more on self-study certificates; Google’s PMLE exam, for instance, costs $200.

The timing coincides with Anthropic’s IPO preparations. The company is expected to go public this year and is reportedly aiming for a valuation of up to two trillion dollars. From a leaked IPO prospectus, it is evident that Anthropic generated nearly $4.6 billion in revenue last year, with an operating loss of over $8 billion. → Forkast, Business Insider, CNBC

Synthszr Take: Let’s break down the numbers: $100 million for 10,000 people is $10,000 per engineer, spread over a good two years. For that, you don’t even get half of a properly staffed enterprise pitch, and Anthropic only covers the curriculum; the twelve-week residency is paid for by Accenture, McKinsey, and Morgan Stanley from their own payroll. With $4.6 billion in revenue and over $8 billion in operating losses, this budget is a rounding error that funds a sales network of 10,000 certified people. The leverage is in the repetition: each of these engineers will lead dozens of projects over the years, and each will be created within Claude’s architecture, with Claude’s security protocols, following Claude’s handover logic. Ten thousand dollars per head is enough to define how the bank, the pharmaceutical company, and the consulting firm build their first production application.