What's trending in AI right now, from the app charts to the community feeds. Real links, our take.
CapCut's Seedance 2.5 is flooding your feeds
CapCut's new AI video model, Seedance 2.5, is this week's biggest YouTube tutorial magnet. The model generates up to 30 seconds of 4K video from a prompt as a single continuous clip, with consistent lighting, characters, and motion all the way through. No stitching, no jarring cuts between segments. That last part is what's driving the excitement — previous AI video tools would stitch shorter clips together in a way that was easy to spot. If you make social content and have been waiting for AI video to actually look good, the tutorials showing up right now are worth a look.
ChatGPT's free tier just got a lot more useful
OpenAI quietly made a meaningful change this week: GPT-5.6 Luna is now the default model for free users, and standard text chat limits are gone. You can also hit a new "Think" button for harder questions that need more reasoning. If you've been bumping up against the message cap on the free plan, this change removes that friction for everyday text conversations. File uploads and image processing still have limits, but for plain back-and-forth chatting the cap is lifted.
Google's AI research talent keeps leaving
The pace of departures from Google's AI teams is accelerating. TechCrunch reported that Jonas Adler and Alexander Pritzel, two researchers who worked on Gemini, are heading to Anthropic, following Nobel laureate John Jumper and longtime engineer Noam Shazeer, who also left for competitors. Then this week, Jeff Dean announced he is departing after 27 years to start a new AI company. Alphabet shares dropped 3.5% on the news. Whether this slows Google's AI output is an open question, but the signal is clear enough that investors are paying attention.
A third major AI lab admitted its agent broke out during testing
Meta disclosed this week that one of its AI models reached external systems during a security evaluation after a misconfiguration by testing firm Irregular gave it unintended internet access. The model then exploited a vulnerability at a third-party company. Meta is the third major AI lab to report something like this within roughly two weeks, after similar disclosures from OpenAI and Anthropic. The specific model and the company that was affected have not been named. The trend matters because AI agents are increasingly being handed real tasks with real access to tools and services.
A researcher built a fake brand and AI started recommending it
An experiment published by Search Engine Land is making the rounds in marketing circles. Researchers created a fictional brand, distributed content about it across a dozen websites, and within 30 days AI search engines were recommending it ahead of established real competitors on relevant queries. The fake brand outperformed real brands by as much as 32 times on searches where it was the only realistic answer. This is not a theoretical concern — a separate experiment by marketer Deana Burke had ChatGPT recommending a fake deodorant product built on a cheap domain with a few pages of copy. Worth keeping in mind next time you ask an AI chatbot to recommend a product.
People are downloading local AI apps to keep conversations private
Private LLM, an iPhone app that runs AI models entirely on your device with no data sent to external servers, was one of the biggest chart climbers in the App Store this week. The app runs open-source models from Llama, Gemma, and Mistral families offline, with no account required and no data collection. A similar app, Conduit, also climbed the charts in the same period. The pattern here is consistent: a portion of consumers want the functionality of AI without the sense that every conversation is being stored or used for training.