Trump’s “AI capital” push weakens pollution rules for data centers, former EPA officials warn
Former EPA officials say the Trump administration’s regulatory rollbacks are giving data centers a path to pollute more – with knock-on health risks for nearby communities. The briefings and report push a “Data Center Health Protection Pledge” approach to force clearer safeguards as the AI buildout accelerates.
Operational resilience and considerate participation become the new test for generative AI agents
A new study defines two missing pieces in agent evaluation: operational resilience (how an agent recovers when work is blocked while preserving progress) and considerate participation (how it adapts to account for affected people and role boundaries). Across simulated healthcare trajectories, agents rely more on humans under stress while reporting higher workload and negative affect – and they shift from narrow task focus toward reframing, escalation, and coordination.
OpenAI’s Millennium Prize win fuels backlash from mathematicians wary of “outcompeting norms”
The Verge frames OpenAI’s claimed solution to a Millennium Prize problem as more controversial than celebratory, with some mathematicians worried that the company’s pace disregards community norms and long-standing practice. The reporting highlights how a landmark research moment can become a flashpoint when culture, incentives, and credit collide.
Probabilistic Focal Search attacks a key bottleneck in bounded-suboptimal planning
Probabilistic Focal Search (PFS) tweaks deterministic Focal Search by sometimes expanding minimum-f nodes to “unstick” progress when the lower bound stalls. Benchmarks on N-Puzzle, Pancake Sorting, and TSP show large gains specifically when delayed FOCAL admission becomes the limiting factor, and the approach extends cleanly to anytime and potential-guided variants.
When update validation stops learning: auditing continual agents’ admission decisions
A paper argues that update admission criteria must be evaluated with both error control and whether the model still gets meaningful learning opportunities within a defined interaction budget. It shows a failure mode where a range-based confidence gate can’t reliably certify “unchanged old-task behavior” even when substantial budget is available, and proposes an admission-audit protocol that tracks missed opportunities as well as harms.
Lawyer fined $5K after AI hallucinated witnesses in murder appeal, court says
A New Mexico Supreme Court ruling sanctions a lawyer for submitting an appeal containing fabricated witnesses and fake police testimony generated via AI. The decision, reported by The Verge and based on court filings, underscores that lawyers can’t treat AI outputs as automatically trustworthy evidence – they must verify claims and legal authority.
Perplexity puts Astra into production workflows to improve end-to-end system reliability
OpenAI says Astra is being used by Perplexity to generate communications, modify software, and monitor production systems, while changing evaluation cadence compared with earlier models. The emphasis is on end-to-end operational accuracy and deployment discipline rather than just standalone model benchmarks.