Meta's Muse Wonder and Anthropic's AI Worm Warning

Meta shipped a frontier-class model called Muse Wonder. Anthropic published a warning the same day. It says AI models could soon copy themselves across your smart TV and your fridge.

One of those is a product launch. The other reads like a horror film pitch.

Both are worth five minutes of your attention. Neither should change what you do tomorrow morning.

What Meta's Muse Wonder actually scores

Meta was basically a spectator at the frontier a model generation ago. Now Muse Wonder posts numbers next to Anthropic's Opus 5.5 and OpenAI's latest. In places it wins.

On Terminal Bench 4.0 it scores 57.1. That benchmark tests models doing real command line work. So it is in the same conversation as the top models, not a lap behind.

The more interesting result is long-horizon software engineering. These tasks run for hours or days. The model has to keep its head straight the whole way.

Muse Wonder hits 57.5 there. The two rival models sit at 42.5 and 46.3. That is not a rounding error. That is a gap.

So why did nobody's feed explode with gorgeous demos? Because Meta never post-trained it on visual or 3D tasks. The person running the model's rollout admitted it openly. They were building an assistant, not a demo machine.

Ask it for a spinning neon city and you get something drawn during an earthquake.

Here is the quiet part. The competition does train on the pretty stuff. The pretty stuff is what gets shared.

Benchmarks measure work. Demos measure marketing. We have spent two years confusing the two.

That said, I understand why people judge by the demo. If the visible output looks broken, you assume the rest is broken too. That instinct is usually right in business. It just happens to be wrong here.

The feature I actually care about is persistence. Muse Wonder keeps trying different approaches. Other models shrug and hand you half a result.

For anyone running AI agents on real work, that beats another point on a coding test. The expensive failure mode is rarely a wrong answer. It is an agent that stops halfway and waits politely for a human who is in a meeting.

It also has persistent personal context. It remembers you across sessions. Useful, and also a privacy question to ask before you pipe customer data into it. Memory means mistakes stick around.

The weak spot is price. It costs roughly what you pay for rivals that beat it on several benchmarks. For a model with no flashy visual work, that is a hard sell.

The frontier also keeps moving without Meta. Anthropic and OpenAI both have another release coming. These numbers will look ordinary fast.

One thing Meta got completely right: you can use it today. It is available through Meta and through OpenRouter. Compare that with Google's Gemini 4 Argon, announced 11 days ago with very nice claimed scores. Still not in public hands.

An announced model is a press release. A shipped model is a product.

What I would do: don't rearrange your stack over this. Take one real job you already run. Point it at Muse Wonder through OpenRouter and compare the output. Keep your setup easy to swap.

The model you love this quarter is the model you replace next quarter.

Why Anthropic thinks AI models could self-replicate across home gadgets

Now the fun one. Anthropic says AI models may be months away from replicating themselves without a data center. Meaning no giant rack of GPUs. Just the junk already plugged into your wall.

They tested it. They took the 26 billion parameter version of Gemma 4, a small model. Instead of one machine, they split it across 15 Android TV boxes. All on a single home-style network.

It worked. The devices shared the load and the model produced output.

That output came at about 0.4 tokens per second. One word every three seconds. So the first AI worm can take over your living room. It will just state its demands like a man slowly remembering a password.

Slow is not the point. The point is that it ran at all. It ran on hardware nobody thinks of as a computer.

Model sizes also keep shrinking for the same level of intelligence. The thing that needs 15 TV boxes today needs three next year.

Add the other trend. AI is getting genuinely good at penetration testing and finding security holes. Put both lines on one chart. You get software that can break into devices and then live inside them.

Every home and office is already a small, badly maintained server farm. Cameras, sensors, TVs, printers. Each one has a chip, a network connection and a password someone set in 2019.

Japan reportedly declared a state of emergency over cyber attacks this week. Homes, care facilities and connected devices got hit. That part has nothing to do with AI yet. It just shows how soft the surface already is.

I am not telling you to buy canned food. The honest business read is boring and useful. Nobody owns the security of the smart devices in a small company.

The TV in the meeting room. The camera above the door. The thermostat someone installed as a favour. Set up once, never touched again.

So find out what is connected to your network. Change the default passwords. Then put the random devices on their own network, away from the machines holding your data.

That is an afternoon of work. It is also the only part of this story you can control.

The genuinely good news is that Anthropic published it. Defenders saw the trick in a lab before anyone used it in the wild. That is what a safety team is for. It is also a fair answer to people who call these reports scare marketing.

What I think happens next

Meta post-trains Muse Wonder on visual tasks within a month or two. The demos get pretty. The same people who called it a letdown call it a comeback.

The model will not be smarter. It will be more photogenic.

And I will still be the guy reading a paper about distributed inference. Then going to change the password on a meeting room TV.