Explore BIG CHANGE reporting on AI Models & Tools: what changed, how it works and why it matters.
SemiAnalysis found model-specific safety results available by launch for nine of 857 releases by nine China-based developers. The audit measures public disclosure, not private testing or model safety.
Vals AI found a statistically significant team gain only for GPT 6 Sol at medium effort across 50 app builds; that setup cost $3.07 versus $1.22 per app.
Lumen says its PoeLLM investigation found more than 3,400 affected servers since April. Its evidence points to exposed, vulnerable services across AI and other tools, with specific checks for operators.
Geely's new Battleship 700 groups adjustable suspension, terrain modes, cabin AI and driving assistance under one pitch. Its trim list shows where the components differ.
Boston Dynamics’ four-finger Atlas hand pairs 13 degrees of freedom, backdrivable actuation and tactile sensing with reinforcement learning in simulation. The company reports initial transfer results but no benchmark or commercial deployment.
In a World Science Festival interview, Tristan Buckmaster explains how AI-produced arguments were formally checked and why readable proofs and independent assessment still matter.
Continues From unit distances to Navier-Stokes: AI mathematics reaches a new phase
Amazon's new Alexa+ tablets replace Fire OS with Android and add Google Play access. The company says some on-screen Alexa features require an explicit opt-in; pricing starts at $229.99.
OpenAI withdrew three math manuscripts over a sign error. The notices identify the failed proof step, preserve earlier PDFs and leave the claims unproved.
Continues From unit distances to Navier-Stokes: AI mathematics reaches a new phase
Anthropic's new Haiku model has lower token rates, a different tokenizer and Messages API migration changes. Existing Haiku 4.5 users should recount requests and test response handling before switching.
Newly linked teen test accounts received no parental alerts during scripted crisis conversations, the Youth AI Safety Institute reports. OpenAI disputes the test timing.
OpenAI's collection contains 722 manuscripts in 372 families. A close look at one Lean proof configuration shows how to inspect versions, scope and checking requirements.
Continues From unit distances to Navier-Stokes: AI mathematics reaches a new phase
J-HRTI opened a humanoid data facility with 35 robots in August. Japan’s separate ¥387.3 billion FY2026 model program funds research, with no documented grant to the facility.
Runway says Praxis-1 applies video pretraining to robot control. Its placement-error comparison offers a limited signal while public weights and fuller evaluation details remain pending.
A Qualcomm-commissioned Signal65 report compares three retail 14-inch laptops. The Snapdragon system leads selected CPU and AI tests; the Intel laptop wins graphics and sustained Office runtime.
In Tavus-run one-minute calls, 26 of 54 people believed Griffin-Lite was a person. The study, NVIDIA benchmark and limited preview answer different questions.
Sony is adding QSSR to Wolverine and Ghost of Yōtei on standard PS5. Access is game-specific, and sharper reconstruction does not establish faster frame delivery.
Google announced Gemini 4 Argon for trusted cyber testers, with paid API and AI Ultra access planned next but undated. Vals AI's first independent scores lead on some tasks and trail on others; the public API catalog lists no Argon endpoint.
DeepSeek's new Ascend 950 kernel library offers familiar DeepGEMM interfaces and source code. Its prerequisites and self-reported benchmarks set the limits for developers evaluating it.
