In partnership with
Welcome back to The AI Field. Meta has released Muse Wonder, its most capable model so far, and the bigger Wonder Full version now sits in the top five on two closely watched leaderboards.
Inside: what Meta's new models can do and where you can use them, Microsoft's Maia 300 chip for running AI at scale, and AMD's Helion rack built for AI agents. Plus six tools to try and a five-minute way to brief an AI agent with a screen recording instead of a long prompt.
In today's AI Field:
Meta's Muse Wonder joins the top tier of AI models
Microsoft unveils Maia 300 to run its own AI
AMD's Helion rack is built for AI agents
Plus: the UK gets early model access back, MPs question the labs tomorrow, and Office moves inside Copilot.
Read time: 5 minutes
100+ ChatGPT Prompts to Revolutionize Your Day
Supercharge your productivity with HubSpot's comprehensive guide to Chat GPT. This free resource is your fast track to AI mastery:
Industry-Specific Use Cases: 15+ real-world applications across various sectors
Productivity Guide: 21 best practices to 10x your efficiency with AI
Prompt Powerhouse: 100+ ready-to-use prompts for immediate implementation
Challenge Buster: Overcome common AI hurdles with expert strategies
Plus, in-depth sections on email composition, content creation, customer support, and data analysis.
LATEST DEVELOPMENTS
META · MODELS
The AI Field: Meta has released Muse Wonder and Muse Wonder Full, its most capable models so far, built for agent work like sorting out an insurance claim, making sense of a lease or building a working app. Muse Wonder is available today through the Meta Model API and Muse Code, while Wonder Full is limited to Muse Code subscribers for now.
The details:
Wonder Full scores 56 on the Artificial Analysis Intelligence Index, up from 48 for Muse Spark 1.3, which came out just six weeks earlier. The standard Muse Wonder scores 52.
Wonder Full ranks #5 of 47 models on the Vals Index, while Muse Wonder lands at #10. Its biggest gains came on agent tasks, such as Terminal-Bench, where it jumped from 33% to 57%.
Both models have a 1M-token context window. Through the API, Muse Wonder costs $4 per million input tokens and $20 per million output tokens. Meta says it will soon power the Muse app too, starting with subscribers.
Why it matters: Fourteen months after Meta began overhauling its AI work, it is back near the front. The catch is that the headline scores belong to Wonder Full, which only Muse Code subscribers can use today. The standard Muse Wonder is the one open to everyone, so judge it on your own tasks rather than on the leaderboard numbers.
TRY THIS TODAY · 5 MINUTES
Brief an AI agent with a screen recording
1. Pick one task you usually explain in a long written prompt, like feedback on a landing page, a bug walkthrough or a deck review.
2. Open Cakie (Mac, free beta) or your phone's screen recorder. Record yourself pointing at the screen while you say what to change. Keep it under two minutes.
3. Paste the transcript (and any screenshots) into Claude, ChatGPT or Copilot with: "Use only what I show and say in this recording. List the changes as numbered steps. Ask me before you rewrite files or send anything." Compare the result with your usual typed brief.
MICROSOFT · CHIPS
Image: Microsoft
The AI Field: Microsoft has announced Maia 300, the third generation of its in-house AI chip, designed to run AI models rather than train them. From 2027, Microsoft plans to scale Maia to multi-gigawatt deployments and make it the main inference system behind its own products.
The details:
Microsoft expects 40% more tokens per megawatt than leading GPU-based systems, and up to 7x the rack-level token throughput of Maia 200, which is already in production.
Each chip carries 288 GB of HBM4 memory with 18 TB/s of bandwidth. The compute chiplets are built on TSMC's N2P process, as laid out in the architecture deep dive.
A full rack holds 96 accelerators in a 167-kilowatt power envelope, with 27.6 TB of memory. Production deployments start in 2027.
Why it matters: Every answer from Copilot, GitHub Copilot and Teams costs Microsoft power and chips. Running more of that on its own hardware reduces its reliance on Nvidia and AMD and could make AI features cheaper to offer at scale. Nothing changes for users yet, but this is likely the hardware your Copilot replies will run on from 2027.
Stop typing AI prompts. Start talking.
You think 4x faster than you type. So why are you typing prompts? Wispr Flow turns your voice into ready-to-paste text inside any AI tool. Speak naturally, tangents and all, and Flow cleans it up. Available on Mac, Windows, iPhone, and Android.
AMD · INFRASTRUCTURE
The AI Field: AMD has introduced Helion, an open rack blueprint for AI agent workloads, unveiled at the OCP Global Summit. The idea is that agents need plenty of ordinary CPU power to run tools, code and browsers, not just GPUs for the model itself.
The details:
A CPU-only version holds up to 32,768 cores using 6th Gen EPYC processors. It is aimed at hyperscalers that run agents separately from model inference.
An all-in-one version combines EPYC CPU nodes, Instinct MI350P GPU nodes and general compute for enterprises. Helion also includes a token router that sends agents' model calls across frontier and open-source models.
AMD plans to contribute the spec to the Open Compute Project. Configurations are being built with partners and are planned for November 2026, and partners rather than AMD will sell the actual racks.
Why it matters: Much of what agents do is classic computing: opening pages, running code, moving files. AMD is betting that this layer becomes its own market, separate from GPU inference. Put it next to Maia 300 and the pattern is clear. The hardware race is moving from training models to running them, and to running the agents around them.
TRENDING AI TOOLS
Six new launches you can try today. Click a name to open the tool.
📱 KernelAI: Private, offline AI chat on iPhone and Android. Download one of 53 open models to your phone, attach a PDF or photo, and ask questions with no account and no subscription. Free.
🧠 TrueU.ai: A private Thinking Partner that helps you work through hard questions without handing you a verdict. Perspectives is free from the first session; the app opens with a free trial.
🎙️ Foley: A voice to-do list that files tasks with project, date and reminder already set, and asks when it is unsure instead of guessing. Desktop is free; Pro syncs every device from $8 a month.
🎬 Cakie: Record your screen and voice on Mac, get a transcript with screenshot highlights, then paste one prompt into Claude Code, Cursor or ChatGPT. Free beta.
🗣️ SKI: Talk out loud to Claude Code, Cursor, Codex or Gemini CLI and hear the agent answer in a natural voice. Speech stays on your machine. Free for life on Mac, Windows and Linux.
🖥️ OpenPilot: A free, open-source desktop AI agent that uses the model you already pay for. Point it at a folder, let it create files and run commands, and keep approvals in your hands. Windows now, macOS soon.
QUICK HITS
🇬🇧 The White House restores UK access to frontier AI models: After a 17-day freeze, Britain's AI Security Institute again has pre-release access to models from OpenAI, Anthropic and Google DeepMind. The arrangement still looks voluntary.
⚖️ UK MPs question the AI labs tomorrow: Policy leads from Anthropic, OpenAI, Google DeepMind and Meta face a Commons committee on Tuesday 13 October over whether pre-release safety testing should be mandatory.
📄 Microsoft puts Office inside the Copilot app: Full Word, Excel and PowerPoint now open inside Copilot, with files kept in OneDrive. It is rolling out first to Frontier early-access customers.
That's it for today!
See you soon,
Olle Hellman | Founder of The AI Field