😺 New Claude Duet works while your phone's off

Welcome, everyone.

It seems that every chatbot has a favorite country, and that country is Japan. Researchers at the University of the Basque Country and Cardiff University asked large language models (LLMs) open-ended questions about culture. They found that the models often favored countries such as Japan. When the researchers used English prompts, the answers covered a wider range of countries.

The authors say this preference first appears during supervised fine-tuning. This is the stage when a model is trained to be helpful and polite. They say the preference does not appear while the model is reading information from the internet. Somewhere, a chatbot is being asked what toppings to put on a pizza and is recommending a Japanese tea ceremony.

The paper was published in April, and a version that went viral today includes claims that are not in the paper's abstract. One example is that quantization, which makes a model smaller, makes the problem worse. These claims have not been confirmed, so they should be treated as uncertain until someone checks the full paper.

Here is what happened in AI today:

😺 Anthropic's New Claude Duet Works While Your Laptop Is Off, But Must Ask Before Spending Your Money

Claude Duet is a new app that lets Claude work on tasks without waiting for you to tell it what to do each time. It sends you a summary in the morning, checks your connected apps every few hours, and sends you a message only when something needs your attention. It is being tested in the United States by Pro and Max subscribers who join a waitlist. It is built around Claude Opus 5.5 and uses the same usage limits as your existing plan.

How can it check things when my phone is in my pocket? Duet runs on its own virtual machine. Think of this as a small computer in the cloud that Claude can use, including a web browser. This lets it keep working even when your phone and laptop are turned off. It can also use your existing Claude memory, skills, and connected apps. Anthropic lists more than 6,000 connectors, including QuickBooks, Stripe, Gusto, Docusign, HubSpot, and Shopify.

An AI agent that acts when you are not watching raises an obvious question: how do you stop it from making a bad decision? Anthropic has built several safety checks to help prevent this.

For payments, Duet can use Link, Stripe's digital wallet with a virtual card, and Shop Pay. So yes, an AI can now buy things for you. Read the approval screen carefully.

Anthropic is also clear about what Duet cannot do well yet. It might misunderstand a conversation or get the level of urgency wrong. It is less capable at personal errands than at work tasks, and it cannot yet show you its browser. The last point is especially worth noting: checking things in the background and doing large tasks both use up your plan's usage allowance. An assistant that never sleeps can also use up your quota while you sleep.

One important thing to watch is how often Duet sends alerts. If it sends too many, it becomes like a coworker whose messages you mute. If it sends too few, you might miss something important, such as a property viewing scheduled for the wrong day. Anthropic admits that judging urgency is a weak point. One useful measure from testers would be how often an alert was actually worth their attention.

🎓 AI Skill of the Day: Run Three Coding Agents at Once Without Them Overwriting Each Other

You ask an AI coding agent to build a login page, and then you wait ten minutes while it works. Meanwhile, you have a bug to fix and documentation to update. You might try opening a second agent in the same project folder, but then both agents could edit the same files on the same branch. This can cause one agent's unfinished work to overwrite the other's changes.

The solution is a Git worktree. Usually, a Git repository has one working folder with one branch selected. A worktree creates another folder from the same repository, with its own branch. This gives each agent a separate place to work while letting them share the same project history. Think of it as one kitchen with separate cutting boards. Changes that have been committed are visible across the worktrees, but changes that have not been committed are kept separate. That is the main benefit.

List the open pull requests in this repo using the gh CLI. For each one, summarize which files it changes. Tell me the safest order to merge them, flag any that will conflict, and resolve conflicts as you merge. Run the tests after each merge and stop if anything fails.

The final sentence of that prompt is an addition to the original approach. There are still three problems to watch for. Each worktree needs its own software installation, such as npm ci. Two development servers using the same port will clash. And if the agents share a database, their data is not isolated. Another limit is your own time: worktrees prevent file conflicts, but they do not reduce the amount of work you need to review. Start with two or three agents. When you finish, remove the extra worktrees with git worktree remove.

📰 Around the Horn