MIT's instance-adaptive scaling lets LLMs dynamically adjust compute for hard problems, using calibrated PRMs to chase promising paths. Cuts usage in half with same accuracy—plus a build guide.
OpenAI's confession system trains LLMs to report rule-breaking separately, rewarding honesty like a truth serum to boost transparency without punishing task failures.
Researchers and engineers highlight over 30 security flaws, workflow hijacks, and minimal productivity gains in AI coding tools like Cursor, Copilot, and Gemini.
Pudu's D5 quadruped, powered by NVIDIA Orin, tackles inspections and patrols with 275 TOPS compute, all-terrain mobility, and 30kg payloads. Japan launch April 2026 at 6-7M JPY.
ByteDance's Doubao AI on ZTE's Nubia M153 phone got blocked by WeChat and major banks over security risks from its full OS control and screen access. Privacy fears and competition from Huawei, Xiaomi forced quick cutbacks.
Sam Altman's 'code red' memo redirects OpenAI toward ChatGPT fixes amid Gemini 3 dominance, data breaches, and profitability doubts.
FDA deploys optional agentic AI for premarket reviews, surveillance, and admin tasks with human oversight in secure GovCloud. Builds on Elsa tool used by 70% of staff.
Anthropic CEO Dario Amodei warns of AI bubble risks and 'YOLO' competitors amid 10x annual revenue growth, while the company debuts Interviewer tool with insights from 1,250 pros on AI in work.
Washington University study shows car GPS patterns predict mild cognitive impairment at 82-87% accuracy, outperforming tests; AI could turn this into wearable health alerts.
Tencent's Hunyuan models excel in 3D game assets, efficient video generation, and foundational tasks like coding, topping benchmarks against Microsoft and Meta while fueling revenue growth.