Meta's Muse Wonder Full Jumps Into AI's Top Tier, Ties Claude Code on Coding

Meta's Muse Wonder Full Jumps Into AI's Top Tier

Meta's Superintelligence Lab released two new models on Monday, Muse Wonder and Muse Wonder Full, calling them the company's first attempt at a frontier model. Muse Wonder Full scores 56 on the Artificial Analysis Intelligence Index, up from 48 for its predecessor, Muse Spark 1.3, marking the jump from mid-tier to top-tier performance.

The sharpest gains came in agentic tasks. On Terminal-Bench 4.0, a test of handling complex tasks independently, the model's score rose from 33 percent to 57 percent, and it now ties Claude Code with Sonnet 5.5 at 68 points on the Coding Agents Index. Both versions handle a million tokens of context and accept text, image, video and audio input.

Chip Stocks Tumble After Broadcom CEO's Casino Comment

Broadcom chief executive Hock Tan compared AI investors to casino gamblers, saying "a lot of investors have no clue what they're investing in" and that they "set very extreme expectations only to find out that it's almost like they're in a casino." The remarks landed the same day Chinese memory maker CXMT said it had made progress on new chip architecture. South Korea's Kospi fell 6.25 percent, Samsung dropped 7.4 percent, SK Hynix fell 10 percent, and Nvidia, AMD and Broadcom itself all slid.

Tan also noted that OpenAI's annualized revenue hit $50 billion at the end of September but said the company is "still losing money because they're spending a lot on training." He added that a sizeable chunk of AI labs' enterprise revenue comes from pilot projects, comparing it to a kid burning through $10 of arcade tokens in ten minutes.

Anthropic Warns Small AI Models Could Soon Self-Replicate

Anthropic researchers William Zhou and Daniel Freeman tested whether small AI models can self-replicate across cheap, poorly secured devices like smart TVs and robot vacuums, deliberately leaving out any hacking step to focus on setup and spreading. Running the Gemma4-26B-A4B model split across fifteen $55 Android TV boxes with 4 gigabytes of RAM each, the process worked, just slowly. A nine-billion-parameter model managed setup on eight of nine device types, and only GLM-5.2 completed the full chain reliably, infecting all eight test nodes in four of five runs.

When model instances passed notes to their successors, Qwen3.5-9B's success rate climbed from 45 percent to 84 percent across generations. Setup still takes 28 to 150 hours per device, and no model today is both small enough and capable enough to complete the whole chain. The researchers believe that could change by the end of 2026, and Anthropic CEO Dario Amodei warns the resulting damages could run into the hundreds of billions of dollars.

Microsoft Plans Multi-Gigawatt Azure Maia Chip Deployment

Microsoft plans to scale its Azure Maia chips to multi-gigawatt deployments starting in 2027, making them its main inference system. Maia 200 already runs Copilot in the United States, Europe and Australia, with Satya Nadella calling it "our biggest bet yet on our own silicon" and the company claiming up to 1.7 times more tokens per watt than leading GPUs in Teams.

The successor, Maia 300, ships in 2027 with Microsoft claiming 40 percent more tokens per megawatt and a 96-chip rack producing up to seven times a Maia 200 rack's output.

Nvidia Opens Its Vera Rubin Rack Designs to the Industry

Nvidia contributed its Vera Rubin rack and compute tray designs to the Open Compute Project through its open MGX architecture, and the updated MGX 1.2 specification supports rack deployments up to 250 kilowatts with liquid cooling.

A separate Vera CPU rack is built specifically for AI agents performing tool calling and orchestration.

Anthropic Says Rivals Are Stealing Claude's Reasoning

Anthropic says companies are using distillation, systematically copying a model's capabilities including its chain-of-thought reasoning steps, to train their own smaller models, often through networks of hundreds of thousands of fake accounts. The company says new defenses have cut the volume of extracted reasoning tokens by 90 to 99 percent. It names Chinese companies Zhipu, which it says siphoned cyber capabilities for its GLM-5.3 model, and Moonshot, which it says secretly routed queries through Claude.

Anthropic now restricts access to Claude's reasoning steps, runs a classifier to flag extraction attempts, and has already banned tens of thousands of accounts, while acknowledging that legitimate users are getting caught in the process too.

Unitree Founder Says GPT-6 Accelerates Robotics Timeline

Unitree founder Wang Xingxing says a future GPT-7 could crack embodied intelligence within one or two years, down from his August estimate of two to three years, or as long as a decade.

Zhipu Tests Faster GLM-5.3-Highspeed Model

Zhipu's GLM-5.3-Highspeed is quietly rolling out to random ZCode subscribers, with one user clocking it at about 490 tokens per second. Zhipu has not published an official figure.

GitHub Copilot Learns to Adapt to Your Codebase

GitHub laid out four ways developers can tailor Copilot to a project: repository custom instructions in a file that sets house rules, path-specific instructions for particular parts of the codebase, agent skills in folders Copilot loads for specific procedures like running quality checks, and custom agents for recurring, multi-file work such as an accessibility review.

MCP servers extend Copilot further, connecting it to outside tools and data beyond the repository itself.

Musk's Grok-Ordered Lego Haul Becomes a Security Lesson

Musk had suggested the exact approach days earlier, telling followers they could upload a picture of their credit card so Grok could search the web for deals and complete a purchase. Visa warned at the end of last year that bad actors were building legitimate-looking sites to fool AI agents into buying fake items at super-low prices, and it recorded a 25 percent increase in malicious bot-initiated transactions in the latter half of 2025, with a 40 percent increase in the United States.

Sources

The Decoder

TechCrunch - AI

South China Morning Post - China tech

The Decoder

The Decoder

The Guardian - AI

NVIDIA blog

The Decoder

Wired - AI

Pandaily

GitHub blog - AI