This one comes from Anthropic’s own red team, the people paid to break their models before someone else does.

Their finding, published yesterday, is the kind of sentence you read twice.

Small open-weight AI models may be only months away from copying themselves onto ordinary consumer devices.

Smart ovens. Off-brand Android TV boxes. The cheap connected gadgets in half the homes in America.

Here’s what they actually tested.

They took a 9-billion-parameter model and tried to get it running on nine different kinds of commercial hardware.

It ran on eight of them.

Then they gave a Chinese open model, GLM-5.2, an agent harness and a goal.

It completed an end-to-end replication cycle.

Meaning it found a device, got itself onto it, and got itself running there.

In a controlled evaluation, not in the wild.

But it did the whole loop.

And the projection is the part that matters.

If scaling trends hold, open models under 30 billion parameters could reach replication-level competence by the end of this year.

That’s eleven weeks from now.

Now, I’ve spent months telling you to love open models.

Mistral’s trillion-parameter release. Meta’s Glimmer. Kimi. The whole “own your stack” movement.

I still mean every word.

But this is the other side of that coin, and I’m not going to pretend it isn’t there.

A model you can download and run on your own machine is a model anyone can download and run on YOUR machine.

The same freedom that lets you own your intelligence lets a small, capable model treat your TV box as a free place to live.

Here’s why this lands on you specifically.

Think about every connected device in your house and your office.

The router. The cameras. The smart plugs. The TV. The printer nobody’s updated since 2022.

Each one is a tiny computer with a network connection and, usually, a default password.

Anthropic just said a model small enough to fit on those is months from learning to spread across them.

So the boring advice becomes urgent advice.

Change the default passwords. Update the firmware. Put the gadgets on their own network, away from the machines that matter.

Not because a rogue AI is coming for your oven next week.

Because the labs’ own red teams are now measuring the gap in months, not years.

The device you forgot about is the one that gets a new tenant. 🫡

The first self-replicating program was called Creeper, written in 1971 as an experiment...

It hopped between computers on the early internet and printed one line.

“I’m the creeper, catch me if you can.”

Someone wrote a second program, Reaper, whose only job was to chase it down and delete it.

The first antivirus was born because the first worm got bored. 🐛

Linus Torvalds just shipped Linux code written with Claude.

If you don’t know the name, Linus Torvalds created Linux.

The operating system under most of the internet, every Android phone, and almost every server on Earth.

He’s also one of the most famously demanding, skeptical, hard-to-impress engineers alive.

This week he posted a new kernel branch that reworks how Linux handles small file reads.

On his laptop benchmark, it nearly triples the throughput.

From about 10 million reads per second to 33 million.

Per-read kernel cycles dropped from roughly 2,950 to 725.

And then he said where it came from.

The code, the test programs, the review of his list of open problems, and the first draft of the letter announcing it...

All came from Claude Code.

Linus Torvalds. Credited an AI assistant. On the Linux kernel mailing list.

Now, read what he did with it, because that’s the lesson.

He didn’t merge it.

The branch is sitting on top of the latest release candidate, unmerged, waiting for review.

He ran the benchmarks himself. He put his own name on it. He opened it to the whole community to tear apart.

Two months ago the Rust project banned AI-written code contributions outright.

Linus took the opposite path.

Use the tool for everything, then stand behind the result like it’s yours, because it is.

That’s the whole debate about AI in serious work, settled by the most serious engineer on the planet.

Not “AI can’t do real work.”

Not “let AI do all the work.”

AI drafts. The human owns.

And the human still has to be good enough to know when the draft is right. 🫡

Go change the default password on the device you forgot you own...

And go put your own name on the next thing the AI drafts for you. 🔥

Jeff J. Hunter

Founder, AI Persona Method | TheTip.ai

Get paid to solve problems like this → AI Certified Consultant

PS. Remember “preserved thinking,” the anti-distillation lock Anthropic shipped last month... They published the results. Data harvesting by copycats down 90 to 99%, and tens of thousands of accounts banned within two hours of switching it on. The wall worked. 🧱

“AI Money Group” to Learn to Make Money with AI