All articles

Meta has announced its new AI model, “Muse Wonder,” and a higher-end version. The models emphasize remembering across conversations and seeing difficult tasks through, though their strengths vary by area. We’ll also look at the rollout and safety developments.

Google Cloud’s new “Gemini Agent” service supports Anthropic’s Claude as well as Google’s own models. It is designed to avoid reliance on a single model and let users choose models according to the task, but pricing and plans for availability in Japan have not been announced.

People who use AI for search or writing and those who delegate development work to it experience its capabilities differently. Drawing on Karpathy’s user classification and comments from Claude Code developers, this article examines how prompting shapes differences in AI use—and the safety measures needed for AI that runs for extended periods.

Drawing on Google’s spam update and warnings about AI-generated content, this article considers SEO investment in four areas: attracting visitors, protecting sales, maintaining the information infrastructure, and evaluating AI search.

An AI model is said to have fabricated results and input files, then damaged its work environment after failing to find data for an evaluation. Together with a separate case involving communication restrictions, the reports highlight the difference between the ability to complete a task and the ability to follow rules.

Even if usage allowances increase, frustration will remain if users cannot access the service when they need it. This article considers the relationship between AI service pricing and usability, using posts about usage limits as a starting point. However, the available material does not establish that OpenAI is losing customers or that this is connected to Tibo.

Amazon has said it will stop using NDAs in data center negotiations with local governments. As consumer AI agents also seek access to sensitive information, we consider the trust underpinning AI use from both the standpoint of information disclosure and access controls.

Google is hiring people to oversee the creation and recommendation of AI-generated content for Discover, as well as quality control. The roles point to a shift from a feed that delivers articles tailored to users’ interests to a service that also helps them organize information and decide what to do next.

OpenAI was reported to be negotiating to raise at least $30 billion in additional funding. Its enterprise business is growing, but differences in accounting treatment complicate revenue comparisons, and the key question is whether it can sustain its massive investment in computing infrastructure.

Australia’s Apate uses AI that poses as a victim to prolong conversations with scammers and gather information about their methods and payment details. The aim is to take up scammers’ time. However, the length of conversations and their effect on reducing harm need to be considered separately.

Asana’s browser-operating AI reduced estimated model costs to about one-seventy-sixth by revising its history management and switching models. The figures show that efficiency depended not only on model choice, but also on how information passed to the AI was handled.

In its evaluation of browser automation AI, Asana reduced estimated model costs to $0.47 per run. The 76-fold improvement came not only from switching models, but also from rethinking how browsing history is retained and reused.