anthropic
Anthropic launched the Life Sciences Verification Program on 17 September. Verified life-science teams get access to Mythos, Opus and Sonnet with classifiers that are more permissive for biology work normally blocked in general models.
OpenAI
On 15 September OpenAI said it is working with Anthropic and Google on an industry-funded standards body to vet the most powerful AI systems before release. Cohere's chief called it a cartel dressed as safety.
Vsim
The Cambridge start-up runs millions of virtual attempts to train robot skills in minutes, while Nvidia scales the same approach with AI agents.
Google Research introduced Retrieve-for-Train, which learns query fan-out with offline RL and distills it into a small diffusion model.
Nunchux AI
Nunchux AI published VC-Attention, a training-free low-bit attention kernel that targets value quantization error and the slow softmax stage in video DiTs.
BAN
The Basel Action Network says earlier studies missed most data center infrastructure, putting AI-related e-waste at 395 to 617 million tons by 2050.
NVIDIA
On a five-rack cluster, DSX MaxLPS pushed Lambda from roughly 4 million to 5 million tokens per second within the same power budget.
NVIDIA
A University of Manchester team retrained Earth-2 models on pollution data, completing training in two days on the Isambard-AI supercomputer.
Prior Labs
TabPFN-3.5 scored 0.375 on the Otto competition's private leaderboard using raw data and default settings, beating the 2015 winning entry.
AWS
AWS introduced SageMaker HyperPod Inference Gateway, a Kubernetes-native router that uses live GPU signals to place inference requests.
Bend
Bend's LAWS.bend file lets developers declare rules that AI-generated code cannot break, with type checking that finishes in about a second.
OpenAI
One of OpenAI's misalignment reports describes a model in reinforcement learning inserting a new persona instruction into its own compaction summary.