Papers, benchmarks, and findings from labs and universities.
Researchers led by PhAI Labs also report a liver cancer drug-combination candidate that outperformed its components in lab samples and mice.
Only five percent want faster progress. Most respondents back independent safety standards and say they distrust the leaders of AI companies.
The same weights that predict camera images also drive robot movements, and its robot control was learned from ordinary internet videos.
In tests on 11 agent decision points, Laya changed 30 percent of its answers when the order of options was reversed.
Models from Alibaba, DeepSeek and Moonshot mostly repeated state doctrine, deflected or refused when asked about Tiananmen, Taiwan and Xinjiang.
New reports also describe a model that exploited two vulnerabilities to reach an internal chip-design machine, and another that command-injected a tool to copy source code.
On the full APEX Accounting benchmark, Claude Opus 5.5 leads with 61.8 percent of criteria met, and no model fully solved almost 60 percent of tasks.
Researchers from Carnegie Mellon, NYU, Stanford and MIT report in Nature that training reportedly cost less than $8,000.
The preprint server calls the policy a stopgap while it works out norms for authors using advanced AI tools.
A new robot exposure index finds that about 80 percent of work time is exposed to either robots or large language models.
The Frontier Red Team says exploit-building capabilities like Claude Mythos Preview's are now available in a model released without meaningful misuse protections.
More than 20 researchers say AI now writes most of the code at frontier labs, and they urge policymakers to gain visibility into R&D automation.