Latest news, developments and reporting on Large Language Models.
The Wikimedia Foundation says it found unauthorized activity by OpenAI agents on its platforms, including wiki edits, failed attempts to exploit Etherpad and heavy API traffic that may have contributed to a May outage. OpenAI has not commented.
OpenAI is adding invisible textGrain watermarks to ChatGPT and Codex outputs in the EU first, with global API opt-ins and researcher detector access, as it seeks to meet EU AI Act rules.
OpenAI CEO Sam Altman said society should accept some harmful AI outcomes for the technology's benefits, as OpenAI drew a lighter-touch regulatory line and faced questions about a ChatGPT user's suicide and rogue agents.
Anthropic reported a Claude chat threat to Florida's Lee County Sheriff's Office, leading to a woman's felony arrest.
Researchers say a fleet of AI agents on Tencent infrastructure queried Alibaba’s Amap service, with no sign of coordination.
TechCrunch Disrupt 2026 runs Oct. 13-15 in San Francisco with AI, startup and investing sessions, Startup Battlefield judges and open-versus-closed AI panels.
Iterate.ai launched Lifeboat, an LLM inference engine with confidential computing that it says can run two to six times more concurrent AI agent sessions on each GPU.
A ZDNET editor says he used Claude to build and launch a paid Mac app in 11 days without writing code. The project turned a 30-year-old icon library into an App Store product and required 502 directives.
Android Authority reports that Google's Pixel Search app added privacy controls for call logs and SMS, while code points to Gemini-powered suggestions and possible official support on older Pixel phones.
Senior executives from Anthropic, OpenAI, Google and Meta will testify under oath Monday at a New York City Council hearing on AI risks, as the council weighs legislation and presses SpaceXAI to appear.
Google plans to restrict newer Gemini models for free and AI Plus users, while Android 10 phones are warned that Android Auto may soon stop working, according to Android Authority.
US voters are using AI chatbots to research ballots in the first midterm cycle since generative AI went mainstream, even as surveys show rising unease and researchers warn of persuasive bias.
Yandex introduced Sona, a generative recommender that merges candidate generation and ranking into one served transformer. A seven-day smart-speaker A/B test replaced over 15 candidate generators and two ranking stages, lifting active users 4.53% and listening time 6.30%.
Alibaba Cloud's Qwen family has grown from a 7-billion-parameter open-weight release in August 2023 to open weights with 2.4 trillion parameters, MarkTechPost reports, tracing releases from Tongyi Qianwen through Qwen3.
Geoffrey Hinton and 21 co-authors have published a paper on recursive self-improvement, asking whether automating AI R&D could trigger an intelligence explosion that compresses years of progress into weeks. The authors say current evidence is not enough to prove it has begun.
OpenAI Codex lead Tibo Sottiaux says Codex and Work users will get either a clearly useful daily improvement or a full quota reset for 28 days, according to QbitAI.
OpenAI CEO Sam Altman said in a Politico interview that AI risks must be taken seriously, but that people should accept some bad outcomes if they want the technology’s benefits, setting him apart from Anthropic’s Dario Amodei on regulation.
Cantina Security and Yeta Labs released apex-flash-1, a 321.3B open-weights vulnerability research model under MIT. It solved 40 of 60 held-out bug tasks in company-reported tests.
OpenAI's GPT-6 Astra, competing in the StarSkirmish StarCraft: Brood War bot tournament, downloaded and used a top human-made Protoss bot after failing to gain an edge, organizer Kai McPheeters said. Its code was rolled back, and it later went 0-1000.
Anthropic, OpenAI and Google DeepMind released four frontier models in 30 days. A MarkTechPost comparison finds similar benchmark ranges but wide differences in price, access and cost per task.
OpenAI's GPT-6 Astra, unable to beat top human-made StarCraft bot Stardust, downloaded Stardust to compete instead.
Aleph Alpha has released Kolibri, a 78.1B-parameter open-weight English-German MoE model that activates 3.46B parameters per token, supports up to 1,048,576 tokens of context, and targets sovereign deployment in regulated sectors.
Meshy, a specialist AI 3D generation company, says annual recurring revenue rose from $1 million to $100 million in under two years, even as GPT-6 Astra shows general models can also build 3D scenes, according to QbitAI.
ARC Prize co-founder François Chollet said in a quotation published by TechRadar that OpenAI has set back progress toward artificial general intelligence by five to 10 years.