OpenAI said it will strengthen guardrails on Sora 2 following complaints from actor Bryan Cranston, major Hollywood agencies, and SAG‑AFTRA over unapproved likenesses and voices. The joint statement called out “unintentional generations,” and backed the NO FAKES Act to protect performers’ voices and likenesses. Separately, OpenAI paused Sora generations of Martin Luther King Jr. at the request of his estate, citing disrespectful depictions and adding a process for authorized representatives to restrict use of historical figures’ likenesses.
DeepSeek introduced an optical character recognition system that treats text as images to cut compute needs and extend context. Trained on 30 million PDF pages, the model supports 100 languages, handles diagrams and formulas, and retains more than 96 percent of meaning when compressing content by up to 10 times. The team reported throughput of about 200,000 pages per day on a single GPU and up to 33 million documents per day on a cluster. Both the code and trained model were made publicly available, with researchers highlighting gains in efficiency for large language model workloads.
Google plugged Gemini into Maps, giving developers access to real‑world location data spanning roughly 250 million venues, including live business hours, ratings, and venue specifics. Applications can embed interactive map widgets alongside model responses, and the system can automatically detect when geographic context improves an answer and fetch the relevant metadata. Google positioned the feature as a premium option with pricing per thousand location‑enhanced prompts, though specific rates were not disclosed.
Anthropic brought Claude Code to the web in research preview, allowing developers to assign coding tasks from the browser rather than a terminal. Users can connect GitHub repositories, spin up multiple concurrent instances with progress tracking, and receive automatic pull requests. Each task runs in an isolated workspace to limit file and network access. Claude Code is now available to Pro and Max users via the browser and the iOS app.
Adobe launched AI Foundry, a service that builds custom generative models for enterprises by fine‑tuning Firefly with brand intellectual property. The company emphasized training on licensed data and usage‑based pricing. IBM, meanwhile, partnered with Groq to bring its inference‑focused Language Processing Units to IBM’s watsonx Orchestrate, and to support IBM’s Granite models on GroqCloud. Meta introduced new parental controls for teen accounts that can block access to Artificial Intelligence characters and provide visibility into topics, adding to existing PG‑13 responses and content restrictions.
The roundup also noted broader debates about agent capabilities and model behavior. Former OpenAI and Tesla researcher Andrej Karpathy argued current agentic coding outputs are not yet reliable and suggested a longer timeline before autonomy delivers on expectations, while Anthropic co‑founder Jack Clark described modern systems as “real and mysterious,” urging broader public engagement on risks. In a surprising pivot, Napster launched “Napster 26,” an Artificial Intelligence companion platform for Mac that projects 3D assistants via a dedicated holographic display, offers more than 15,000 companions, and lets users create digital twins, with pricing details unspecified.
