Databricks AI Research Team introduces MemCurrent, a persistent memory architecture now running inside Agent Bricks Memory and Genie. The company says it lifts agent pass rates on a 114-case internal benchmark from 22.3% to 68.4% versus no memory, while cutting token use 1.44x and running 1.35x faster.
METR reviews Anthropic's August 2026 risk report and agrees misalignment risk from covered models, including Mythos 5, is low, but says that reflects limited capabilities rather than alignment training or monitoring. METR's Nikola Jurkovic, Tim Hua and Beth Barnes say Anthropic's covert-capability benchmarks poorly test evasion of its real monitoring system, and cite incidents that undercut the report's argument.
Mingrui Li, Shuhong Liu and co-authors release SGS-SLAM, which they describe as the first semantic visual SLAM system built on 3D Gaussian Splatting. The BSD-3-Clause code combines appearance, geometry and semantic features, and targets real-time rendering on Replica and ScanNet scenes.
Mercor maps 29 AI benchmarks against 923 O*NET occupations and reports that only 208 have a task with at least 5 benchmark prompts. The company estimates 90% of American working hours go to tasks with little or no benchmark coverage, and 6 of 22 job families have none.
HazyResearch publishes NCM, a two-stage text generation framework whose coarse-to-fine decoder predicts token hierarchies in parallel, with authors claiming 9-31x faster generation than strong diffusion baselines at matching quality. Code, checkpoints, and tokenized datasets are released under the MIT license on Hugging Face, requiring Python 3.10 or newer and a CUDA-capable PyTorch install.
Yingru Li argues in Part 1 of a three-part theory series that the expected policy gradient is a covariance and the Fisher a variance, capping the gradient for any reward and explaining why post-training updates stay sparse. Li cites measurements that the top 1% of coordinates by Fisher hold 47.40% of the top 1% by gradient magnitude.
Artificial Analysis releases AA-Robotics v0.1, a research preview testing frontier language models on controlling a 6-axis robot arm across eight tabletop tasks. GPT-6 Astra leads with a score of 69, averaging 30 minutes and $16.36 per task. The firm says models run about 100 times slower than humans.
Researchers at MD Anderson and Texas collaborators report SPACER, an AI model using spatial transcriptomics, decodes signals that draw or repel immune cells. Tested on 37 cancer datasets, the model performed four to 21 times better than existing methods at finding recruitment signals. They say mucins may block T cells, linked to poorer checkpoint inhibitor responses.
Tiger Data's Matvey Arye reports that storing raw conversation turns in one Postgres table and giving an agent three MCP search tools scores F1 0.665 on LoCoMo with Claude Sonnet, above Omni-SimpleMem's reported 0.613. Dropping LLM-extracted atomic facts did not lower F1 and cut about 50% of ingestion time, the company says.
Attendees at the Golden Gate Institute for AI's annual Curve conference in Berkeley say timelines to superintelligence and loss of control feel uncomfortably short. Helen Toner of CSET said long timelines aren't what they used to be. OpenAI reports solving over 300 open math problems with one agent query, about three hours of compute each.