Monday, 12 October 2026

Meta has released Muse Wonder and Muse Wonder Full, which it calls its most capable models, built for personal and professional agent tasks [1]. Muse Wonder is available today through the Meta Model API and Muse Code, and Muse Wonder Full is available today for Muse Code subscribers, the company says [1]. According to [2], Wonder Full will not be served through the API. Meta says Muse Wonder will soon power experiences in Muse, rolling out first to subscribers [1]. According to [2], Wonder costs $4 per million input tokens and $20 per million output tokens on the Meta Model API. Meta says the models were trained at its gigawatt-scale Prometheus data center [1].

On benchmarks, Meta's own testing puts Wonder and Wonder Full level with Fable 5.1 and GPT-6 Astra, according to [2]. According to [3], Meta says the largest version ranks above Claude Fable 5.1 and just below Claude Opus 5.5 on the Artificial Analysis Intelligence Index. Both comparisons come from Meta. The pitch is agents. Meta describes two features: Active Memory, which carries context about a user across conversations, and Active Focus, which keeps the model working toward a goal and checks in before important actions [1]. According to [2], Meta says Wonder scores higher on MemoryBench than Fable, Opus or Astra.

On safety, Meta says Wonder has its strongest safeguards to date against CBRN and cyber misuse, and that a full report with more safety evaluations "will be available soon" [1]. According to [3], Meta did not say whether Wonder-powered agents misbehaved in testing. If you build on APIs, this is another frontier vendor to price against. Run your own tasks before trusting the vendor's charts.

Microsoft has announced Maia 300, the third generation of its in-house AI chip, and says it will scale the Maia family "to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for Microsoft products and services" [1]. The announcement comes from Scott Guthrie, executive vice president of Microsoft's cloud and AI group [1]. The company says Maia 300 should deliver 40% more tokens per megawatt than leading GPU-based systems, and that a full 96-accelerator configuration delivers up to 7x the rack-scale token throughput of Maia 200 [1]. Microsoft describes it as designed for long-context reasoning, latency-sensitive conversation and sustained, multi-step agentic tasks [1]. According to [2], the chip is built on TSMC's N2P process node with a chiplet design that separates compute, memory controller and I/O.

Maia 200 is already in production. Microsoft says it serves Microsoft Copilot and supports its MAI model team, with GitHub Copilot and Microsoft Foundry "to follow by end of year" [1]. It runs OpenAI models, Microsoft's own MAI models and open-source model families, and the company says it delivers up to 1.7x more tokens per watt for built-in AI features in Teams, Copilot and Cowork than leading GPU infrastructure [1]. According to [2], the shift lets Microsoft rely mainly on its own chips instead of ones from Nvidia and AMD. "At that scale we need a platform that we own, and that we co-design with our models," Microsoft AI CEO Mustafa Suleyman said, according to [2].

Microsoft's announcement includes no prices tied to Maia [1].

Amazon is in talks to acquire AI startup Decart for about $7 billion, The Wall Street Journal reported on Sunday, citing people familiar with the matter, according to [1] and [2]. The deal could be signed soon, but the talks could still fall apart [1]. Neither company has confirmed a deal [2]. Decart was founded in 2023 by brothers Dean and Orian Leitersdorf and Moshe Shalev [1]. Its software makes AI models run more efficiently and lets developers move workloads between chips from different makers [1][2]. It also builds real-time video and world models: Lucy, which edits video in real time, and Oasis, which generates simulated environments [1].

The two companies already work together. Amazon Web Services is a strategic customer, and AWS gave Decart early access to its Trainium3 chip in 2025, according to [1]. According to [2], the deal would give Amazon software that helps developers run AI across different chips, which fits its push to offer Trainium as an alternative to Nvidia. The sale process has moved fast. According to [1], it was reported in August that Nvidia had been close to acquiring Decart before a competing offer emerged; Anthropic then entered advanced talks and walked away in September after due diligence. Decart raised $300 million at a valuation of about $4 billion in May, in a round led by Radical Ventures that included Nvidia, and has raised about $450 million in total [1].

If you use Decart's models or tools, nothing has changed yet. If a sale goes through, check what happens to non-AWS customers and to support for other chips. A tool that helps you leave one vendor's hardware is worth less if that vendor owns it.

Microsoft president Brad Smith says voluntary self-regulation of AI is not "sufficient any longer," in a BBC interview released Sunday [1]. "Right now, your washing machine is subject to more safety rules than AI," he said. "Every carton of milk that you see in the grocery store is subject to safety rules and inspection" [1]. Smith said the UK, not the US or the European Union, is best placed to lead on global safety rules, according to [2]. Britain's AI Security Institute receives access to major models from US frontier labs before deployment [2]. "In Washington, D.C., people don't look at the U.K. and they don't look at London the way they look at Brussels," he said. "That's a strength" [2].

He set out specific rules. Smith wants a legal obligation to tell victims of an AI-powered breach about the incident, and a "safety brake" on AI systems, which he compared to circuit breakers in homes and emergency brakes on trains [1]. According to [2], he also appears to favour outside evaluators approving AI products before release. On the recent White House pledge, under which Google, Anthropic, Meta, OpenAI, xAI and Nvidia committed to internal controls and external auditing, Smith applauded the accord but said self-regulation is not enough [1]. "The good companies will do what you want. It's the companies that are not as responsible that you need to worry about," he said [1]. He added that AI "is moving forward faster than the speed of safety" [1].

None of this is law. But breach notification and a stop switch are things a small team running agents can build now.

Mistlabs, led by former Plaud executive Peter Mo, has launched Ambi, an AI app that can listen in the background, after raising $26 million [1][2]. TechCrunch reports the money came across multiple rounds led by Luminous Ventures and the original design manufacturer Huaqin [1]. The company's release calls it Pre-A funding and says it will pay for a standalone AI wearable [2]. Ambi is out now on iPhone, Mac and Apple Watch [1]. According to [1], the apps work like meeting notetakers, but users can choose to leave them listening to capture context all day. The company says its agent can suggest ideas or pull up relevant knowledge during a meeting, using your calendar, meetings, knowledge base and web data, and start follow-up work before the meeting ends [1][2]. It also says Ambi agents will later talk to each other to negotiate between users in real time, within permissions users grant [2].

Privacy is the open question. Mo says ambient capture is off by default and can be turned off at any time [1]. The startup says it stores no recordings and encrypts the data, but did not say what kind of encryption it uses [1]. It also puts disclosure on the user: "We have to inform our users that they have to tell others about ambient capture," Mo said [1]. According to [1], some meeting notetakers have been sued for recording people without consent. Ambi has more than 50 staff across San Francisco, Beijing, Shanghai and Shenzhen, according to [1].

If you try it, treat it like any recorder. Tell people in the room, and check your local consent rules first.

Nvidia has contributed key Vera Rubin rack and compute tray designs to the Open Compute Project, through its open MGX architecture, the company said at the OCP Global Summit in San Jose [1]. The aim is a shared base that other system builders can use for AI data centres [1]. The updated MGX 1.2 specification adds rack-scale blind-mate liquid cooling, a liquid-cooled 5,000A busbar and native 19-inch equipment support, and supports rack deployments up to 250kW, according to [1]. Nvidia also published a Vera CPU rack design meant for tool calling, orchestration and sandboxed execution, so AI agents can work with apps, data and services in real time, while Vera Rubin NVL72 stays the base for large-scale training and inference [1].

Nvidia and Cisco announced USFP, an optical module format built for 400Gb/s-per-lane networking [1][2]. According to [2], the first planned modules carry 16 lanes for 6.4Tb/s each, and the companies say USFP gives 4x the bandwidth and front-panel density of 200Gb/s-per-lane pluggables. USFP has no sampling schedule or availability date [2]. Other announcements: Nvidia joined AMD, Google and Meta to publish Open Data Center Specification v0.7, and Google, Meta, Microsoft and Nvidia published a grid resilience specification that has AI sites restore to near-full load in less than two seconds after grid disturbances [1]. Nvidia also released open-source DSX OS components for network configuration, cluster testing and rack management [1].

Most small teams will never buy a rack. What matters is that agent workloads now get their own CPU hardware in vendor plans, which may shape how cloud providers price agent sandboxes.

Salesforce has signed a definitive agreement to buy Loops, an Israeli startup whose AI explains why business metrics change and suggests what to do next [1][2]. Loops' technology and team will join Tableau, Salesforce's analytics business [1][2]. The price was not disclosed; according to [2], the deal is estimated at about $30 million. It is expected to close in early November, subject to customary conditions [1]. Loops was founded in 2020 by CEO Tom Laufer and CTO Ido Adiv and has raised $14 million [1]. According to [2], the platform connects to company data sources and uses machine learning and causal analysis to find what drives metrics such as revenue, conversions, retention and engagement. Last year it added Scout, an agent that watches key performance indicators, flags anomalies, looks for root causes and writes up its findings [2]. Users can trigger it from Slack and email without logging in to an analytics tool [2].

"Loops accelerates this mission, using powerful agentic AI to uncover the drivers behind business performance and surface where teams and agents can take action in real time," said Mark Recher, executive vice president and general manager of Tableau [1]. According to [1], this is Salesforce's tenth acquisition in Israel. According to [2], the deal supports a shift in Tableau from people analysing dashboards to agents that monitor performance and surface explanations on their own.

If you use Loops today, ask now what happens to your plan, your data connections and the Slack integration after the deal closes. If you use Tableau, expect agent-led "why did this change" features, and check how they reach their answers.