TabLOID

Anthropic has opened a waitlist for Claude Duet, initially for US Pro and Max subscribers. The separate app uses a cloud computer and browser, allowing work to continue while a user's own devices are off.

The company says Duet checks connected apps, prepares a morning brief and flags matters needing attention. It asks for approval before sending messages, making payments, signing documents or deleting things. Some actions, including securities trading and transfers between accounts, remain unavailable even with approval.

These are the product's stated safeguards, not a guarantee of flawless judgment. Anthropic acknowledges that it may misread conversations or urgency. Background checks also count against the subscriber's shared Claude usage allowance. The practical question for early testers is whether the attention saved outweighs the supervision and quota used.

Read Anthropic's preview announcement.

Recursive has released Flint, a coding model with 40 billion total parameters and 2.5 billion activated per token. Its weights are available under Apache 2.0, according to the developer's announcement.

The company says a small team of people and agents developed it in under four months, with an automated research system contributing across training and inference. That describes the company's process, not a measurement that AI did most of the work.

Recursive reports strong coding benchmarks and fast generation using its optimized software on specified machines. Those results are vendor measurements; a different computer or longer conversation can change the experience. The developer also says complex work may need more guidance than with larger closed models. Local availability makes testing those tradeoffs possible.

Read the release and hardware details.

Meta reported quarterly revenue of $64.60 billion, up 26%, and operating income of $12.99 billion, down 37%. The operating result included a $9.89 billion settlement charge. Net income rose to $9.86 billion.

Calling that simply an AI bill swallowing profit blurs the accounting. The settlement affects this quarter's operating comparison; capital spending is a separate measure.

Meta also narrowed its full-year capital-expenditure forecast to $135 billion through $145 billion, including principal payments on finance leases. The previous range was $130 billion through $145 billion, so the lower bound rose while the ceiling stayed unchanged. It remains guidance, rather than completed spending.

Read Meta's results and guidance.

Researchers tested language models with open-ended cultural questions in 24 languages. Their paper reports a tendency toward locations such as Japan, alongside more diverse answers when prompting in English and other widely represented languages.

In the training comparisons they examined, the first clear signs of the tendency appeared after supervised fine-tuning. The authors frame that as a finding from their experiments, not proof of a single cause across every model.

The preprint first appeared in April and was revised in August. Its results concern generated answers under particular tests. They do not establish personal preferences or justify saying every chatbot has a favorite country. The useful lesson is to notice which examples an assistant supplies and ask whether other cultures are missing.

Read the researchers' abstract and paper.