🧠 Meta Declares War on Anthropic, OpenAI

Meta just dropped its strongest AI models yet, with Muse Wonder Full ahead of Astra and Fable on an independent ranking. They’re built to remember your conversations and keep working through jobs that take hours. Let's dive in!

In today’s insights:

LATEST DEVELOPMENTS

Evolving AI: Meta has launched Muse Wonder Full and Muse Wonder to take on ChatGPT and Claude with AI that remembers past conversations and can keep working for hours.


Key Points:

Meta’s Muse Wonder family arrives 14 months into its AI overhaul under Meta Superintelligence Labs, with Wonder Full offering its strongest performance and Wonder serving as the faster option. Both use Active Memory to recall useful details from earlier conversations and Active Focus to work through problems, try other approaches and check with you before important actions. Wonder is available through Meta’s developer API and Muse Code coding assistant, with access through the consumer Muse app coming soon for subscribers, while Full currently requires certain paid Muse Code plans.

Meta’s AI rebuild is starting to pay off, and it now has capable models to go with the reach of WhatsApp, Instagram and Facebook. People are already using these apps every day, which gives Meta a clear advantage in bringing an assistant into their routines. Muse is already available through WhatsApp, and Meta is planning to bring it to its glasses too. If it keeps improving at this pace, Meta may be the company best placed to make personal agents part of everyday life. The competition could build a better model and still find it hard to convince people to switch when their assistant is already working where they need it.

Evolving AI: How many times does your team explain the same software process?

How to set up a new account. Where to find a certain feature. How an internal workflow works. What a customer needs to click next. Guideless turns those repetitive explanations into reusable training content.

Simply capture the workflow once, and Guideless creates a polished step-by-step guide you can share again and again.

Use it to:

Instead of answering the same questions in Slack, calls, support tickets, or Loom videos, your team can point people to one clear, up-to-date guide.

Less repetitive explaining. Less time spent creating training. More self-service for employees and customers. Pick one workflow your team keeps explaining and turn it into a guide with Guideless.

Create your first guide with Guideless →

Use code EVOLVINGAI to get 50% off any plan on your first purchase.

Evolving AI: Anthropic warns that AI could soon spread like a virus between household devices and keep running without a datacenter.

Anthropic explored how AI could move its “weights,” the files needed to run a model, onto unfamiliar devices. Passing notes between attempts also helped a small model get better at setup. But the replication tests made devices easy to access and kept the AI directing the process on a separate server. Smaller models still struggled to finish the whole job reliably, so the research does not show AI spreading on its own through real homes.

Why It Matters:

Anthropic’s research makes it harder to assume we’ll always be able to switch a dangerous AI off. The concern is similar to a malignant tumour that has already spread: removing the original tumour doesn’t necessarily remove the cancer growing elsewhere. If an AI can reliably copy itself onto other connected devices, those copies could still be running after we shut down its original server. And if they can keep making new copies faster than we can remove them, we could end up living with a harmful system that keeps acting long after we’ve lost control of it.

Evolving AI: Character.AI has filed for bankruptcy as the AI companion company once valued at $1 billion looks for a buyer.

Disney had already hired Character.AI’s then-CEO Karandeep Anand and members of its technical team. Its agreement would serve as the opening bid in a court-supervised sale where other buyers can make better offers. Character.AI points to legal, regulatory and operational pressures behind its decision and expects the sale to finish in late November or early December.

Google’s 2024 deal looks like the turning point in Character.AI’s decline. Cofounder Noam Shazeer was one of the authors of Attention Is All You Need, the paper that introduced Transformers, and left Google after it refused to release a chatbot he helped build. Character.AI became a hit, and Google later brought both founders and some researchers back through a reported $2.7 billion licensing deal. After investors and employees were paid, the team that stayed had about 18 months of operating funds. They were stepping back from building frontier models and trying to make the app work as a lasting business. The bankruptcy suggests that effort was not enough to overcome the company’s financial and legal pressures. Now Disney is bidding for parts of what they built, and users are still waiting to see whether anyone will keep the app they rely on running.

QUICK HITS

💻 Microsoft unveiled its Maia 300 AI chip, claiming 40% more AI output per megawatt than leading GPU systems.

⚡ AMD introduced Helion for AI agents, a customizable server rack supporting up to 32,768 CPU cores.

🔐 Anthropic says its new anti-copying defenses have reduced the amount of Claude's reasoning extracted by unauthorized rivals by 90–99%.

📱 Google released its open-source ML Drift, helping developers run AI models faster across different mobile and desktop GPUs.

 *partner link