Paul Christiano joins OpenAI Foundation Board and Safety and Security Committee
OpenAI is adding Paul Christiano to its Foundation board and the Safety and Security Committee, bringing deep experience from AI alignment and safety work. The move signals a continued emphasis on governance and standards as the company’s frontier systems face growing scrutiny.
Amazon Prime Video’s lip-sync technology matches dubbed audio to actors’ mouths
Prime Video is rolling out an AI feature that makes translated dialogue line up with an actor’s mouth movements, currently launching with the English dub of the German series Maxton Hall. The system blends machine learning with visual effects to keep lip motion believable, and Amazon plans to expand it to more titles.
PGP-Clinical-TimeKAN brings trajectory-first probabilistic forecasting to multivariate patient physiology
This paper argues clinical deterioration is a coupled, partially observed process best modeled as joint trajectories rather than single diagnostic labels. PGP-Clinical-TimeKAN combines missingness-aware temporal encoding, a soft organ-system prior, and a probabilistic head, delivering strong forecasting metrics on an MIMIC-IV-derived cohort and offering more inspectable intermediate risk signals.
Microsoft’s AI privacy and safety rules for schools get contract-backed by the AFT and UFT
Microsoft has agreed to ten enforceable principles for AI used in schools under a deal involving the American Federation of Teachers and its New York City affiliate. The framework includes limits on training data, tighter controls on what gets collected, and clearer disclosure to families about how tools operate.
Students who use AI generally score worse – but the effect depends on how they use it
New OECD PISA results suggest widespread AI use correlates with lower performance on core subjects, though some study patterns show small gains. The report highlights that learners who use AI while critically evaluating tool outputs appear better positioned than those who outsource work without verification.
When Does Memory Help? A cost-aware evaluation of long-term memory in tool-using LLM agents
MERIT (Memory Evaluation for Realistic Instrumented Tasks) tests not just whether agents remember, but whether memory improves real tool-execution outcomes when token and dollar costs are explicitly accounted for. The authors find memory can raise dependent task success substantially, yet update-on-write strategies and retrieval behaviors vary sharply across models, making “more memory” an unreliable default.
IBM releases SOTA Granite Time Series PatchTST-FM-r2 with a commercial-friendly license
IBM Research publishes an updated Granite time-series model based on PatchTST, aiming at stronger forecasting performance while keeping licensing friendlier for commercial use. For teams building forecasting pipelines, the bigger story is reducing the barrier to adopting research-grade temporal models in production.
ChatGPT Sketch turns doodles into detailed images
OpenAI adds a new Sketch interface to ChatGPT Images that lets you draw a quick doodle and then guide generation by describing what you want. It’s another step toward more controllable image creation, with the sketch acting as an on-ramp for users who struggle to write precise prompts.
What OpenAI’s latest controversy tells us about the future of math
Technology Review digs into the backlash around OpenAI’s claimed mathematical milestone and what it says about the move from “AI as helper” to “AI as automated proof contender.” The piece frames the controversy as a stress test for how math communities will verify, reward, and audit agent-driven discoveries.