us orders anthropic to block fable 5 and mythos 5 access
the us government directed anthropic to suspend all access to its fable 5 and mythos 5 models for foreign nationals, citing national security concerns over a potential jailbreak.
topic
research
strategic decision support for ai agents
a framework to minimize support usage for ai agents while controlling errors when they act alone.
ai helps users name skin conditions but next steps remain hard
a study shows ai tools can triple accuracy in naming skin conditions, but deciding what to do next is still a challenge.
phone clusters as low-carbon cloud servers
researchers turn retired smartphones into a low-cost, low-carbon computing platform by clustering their motherboards for cloud workloads.
pytorch meetup singapore draws apac ai builders
eighty engineers gathered for the first pytorch meetup singapore, covering inference, distributed training, and community governance.
cryogenic neuron chip runs near absolute zero
a silicon carbide transistor mimics brain-like spiking at millikelvin temperatures, enabling energy-efficient control electronics for quantum computers and deep-space systems.
open source community backs openenv for agentic rl
openenv becomes a community-governed interoperability layer for reinforcement learning environments, backed by major ai organizations.
pythagoras-prover makes lean theorem proving cheaper
a new open-source family of lean theorem provers uses curriculum learning and diffusion models to cut training and inference costs.
personadrive adds human driving styles to simulation agents
a new pipeline uses retrieved human demonstrations to make driving agents behave with aggressive, neutral, or conservative styles in closed-loop simulators.
arbor adds tree search as cognition layer for ai agents
arbor introduces structured tree search as a shared working memory for multi-agent systems, improving autonomous optimization in complex, stateful environments.
toolsense audits llm tool knowledge
a new framework tests if language models truly understand the tools they retrieve, not just match patterns.
epistemic uncertainty is not always reducible
new research shows that the standard definition of epistemic uncertainty as reducible by more data is inconsistent with its common mutual-information measure, and proposes a three-part taxonomy.
google deepmind launches $10m multi-agent ai safety fund
google deepmind and partners announce a $10m research call to study safety risks when many ai agents interact.
claude fable is relentlessly proactive
simon willison describes how claude fable 5 autonomously debugged a ui glitch by inventing browser automation tricks, raising safety concerns.
llm pipeline automates pre-mediation for human negotiation
a structured llm pipeline supports pre-mediation in integrative negotiation, improving agreement rates and joint outcomes in human experiments.
conformal bayes under label shift
two approaches for conformal bayes under label shift: post-hoc calibration and in-training adaptation, both using importance weighting to restore target-domain coverage.
ai agents struggle to synthesize scientific conclusions
a new benchmark reveals that even top ai models produce low-quality factual summaries from scientific evidence, with the best achieving only 0.33 f1 score.
steering away from sycophancy also cuts factual agreement
activation steering meant to reduce sycophancy in language models also lowers agreement with true statements, showing a structural overlap that current methods cannot separate.