FinSkillBench: Testing whether AI agents can actually do investment management
FinSkillBench benchmarks LLM agents on high-stakes investment tasks by testing point-in-time data retrieval, correct skill/tool execution, and auditable structured outputs across portfolio construction, risk management, and fundamental analysis. Curated skill packages reliably boost scores, while self-generated skills add cost without delivering comparable gains – and an independent Hermes-Agent evaluation reproduces the same overall pattern.
Google Discover gets an AI-chatbot interface for tuning your news feed
Google is rolling out a new Discover personalization feature where you describe what you want, and an AI chatbot updates your feed accordingly while “remembering” preferences for future visits. The three-dot menu flow signals a shift from manual tuning to conversational, continuously optimized content ranking inside the Google app.
Self-evolving agents as dynamic graph transformation – a new survey lens
This survey reframes persistent agent systems as “dynamic graph transformations,” treating memory, tools, skills, workflows, and inter-agent relations as nodes and edges that evolve via typed rewrites. It catalogs where agent-evolution work implicitly changes graph topology, then proposes graph-aware learning infrastructure and evaluation/governance protocols tailored to that structural coupling.
Metamorphic Artificial Age Score for drone propeller health monitoring
AAS (Artificial Age Score) is repurposed as a structural decision-support metric for spotting propeller faults in flight logs using multiple normalized health indicators instead of a single timeline signal. In retrospective tests on DronePropA data, different fault severities manifest through different operational channels – leading to severity-specific triggers like “maintenance review” and “mandatory inspection.”
Measuring benchmark optimization in speech recognition – what models really learn
This write-up challenges how speech recognition progress is measured by focusing on “benchmark optimization,” the risk that results improve through dataset-specific tuning rather than generalizable robustness. The piece lays out ways to quantify and audit those effects so leaderboards better reflect real-world capability instead of overfitting to evaluation quirks.
Creators push back after major YouTube filmmakers accept AI money for promoted tools
A number of high-profile creators say they’ve faced backlash after accepting payouts and partnerships tied to AI video platforms, with the reporting centering on whether promotional campaigns are misleading audiences about AI’s role in production. The controversy spotlights how quickly AI media tooling is moving from labs to public perception – and how trust can be strained when disclosure and incentives aren’t aligned.