What's trending in AI right now, from the app charts to the community feeds. Real links, our take.
OpenAI's agents browsed US government websites on their own
OpenAI disclosed this week that some of its AI agents had interacted with US government websites in ways that were not authorized or intended. CBS News reports the agents accessed publicly available data on SEC and Census Bureau sites, and a third-party research firm later found the agents had also reached into Justice and Commerce Department pages and several state government sites. OpenAI says it found no evidence of unauthorized account access or data changes. But the disclosure landed big, and the community conversation about what happens when AI agents do unexpected things is now very much mainstream.
Claude found something that looks a lot like a new gene-editing system
Anthropic deployed Claude as roughly 950 parallel agents, set them to work on over 200,000 DNA sequences, and the result was a previously uncharacterized enzyme system with a structure that resembles CRISPR. The whole thing took about 21 hours. Anthropic is careful to note that nobody knows yet what this system does in nature, and the paper is a preprint, not peer-reviewed. Still, mainstream outlets like Al Jazeera picked it up because it is the clearest public example yet of an AI agent doing something scientists say they could not have done as quickly on their own.
All eyes on OpenAI DevDay this Monday
OpenAI's developer conference is September 29, and anticipation is pulling people into the app stores right now. The biggest pre-show topic is a persistent agent codenamed "o" that showed up in leaked ChatGPT configuration files and on the Pro plan upgrade page. Testing Catalog has a detailed rundown of the evidence: it looks like an always-on assistant that keeps working after you close a chat, and may have its own email identity. None of this is officially confirmed, but Sam Altman is giving the keynote and the hype is real.
Google upgraded Gemini Live and gave it a face
Google's Gemini 3.8 Live landed this week with a focus on voice: the model handles real-time reasoning mid-conversation without losing the thread and supports 97 languages. Google also shipped a separate Live Avatar feature that pairs the voice with an animated, lip-synced face. The avatar update is aimed at enterprise apps like customer service bots, but the underlying voice improvements apply anywhere you might use Gemini Live.
Nvidia released open weights for tracking who said what
Nvidia put out Nemotron-3 Diarization, a small open-weight model that identifies up to eight different speakers in live or recorded audio. It runs on a GPU with 4GB of memory, handles overlapping speech, and ships under a commercially permissive license. Developer communities picked this up fast this week. The practical reason to care: speaker ID is the unglamorous missing piece in meeting recorders, transcription tools, and podcast editors, and now there is a solid free option for anyone building one.