- Hochul Launches Statewide AI Safety Testing Network in New York — New York Governor Kathy Hochul launched a statewide AI science and safety effort on November 10, placing the New York Center for AI Responsibility and Research (NYCAIRR) at Binghamton University in charge of independent frontier-model testing and a network of more than 80 researchers. Cornell Tech dean Greg Morrisett and Columbia executive vice president for research Jeannette Wing serve as interim co-directors, and participating researchers stay at their home institutions rather than relocating to one laboratory. The center draws on Empire AI's shared computing and covers independent evaluations, measurement tools and standards, and training for AI safety specialists; at the state's request it can provide expertise, assessments and evidence for oversight of frontier developers. Alongside the launch, New York opened the RAISE Portal — the registration and reporting channel for large frontier AI developers operating in the state — with compliance duties beginning in January 2027.
📌 Key takeaways:
- States are now building independent evaluation capacity inside public universities, and that capacity will reach procurement: campuses buying frontier models should expect boards and auditors to ask which independent evaluations informed the decision, not only which vendor demonstration looked best.
- The staffing model is the transferable part — co-directors at two institutions, researchers remaining at their home campuses, shared state compute underneath — which lets a state assemble deep expertise without standing up a new laboratory from scratch.
- The RAISE Portal turns a research center into a regulatory input, so institutions operating in New York should determine whether any of their own AI development or deployment work crosses the threshold that triggers registration and reporting.
🏛️ UCSD angle: UC San Diego's Center for Principled Agentic Intelligence, launched October 28 with $1M over three years plus $250K in shared computing, is pursuing the same independent-evaluation role for agentic systems — including an open-source testbed other campuses can use to benchmark agent behavior.
- ‘AI Slop’ Papers Published Under Prominent Academics’ Names — A Times Higher Education investigation, republished by Inside Higher Ed, found little-known journals publishing AI-generated papers under the names of prominent computer scientists who never wrote them. Alan Said of the University of Gothenburg traced a paper attributed to him in the American International Journal of Research and Innovation, and found the same article reassigned to different authors within days; Stanford's Jure Leskovec confirmed he had no involvement and was unaware the paper existed until a reporter asked. The journals share an editor in chief, similar site design, 2026 registration dates with backdated archives, and publication fees ranging from $10 to $200, while dozens of listed editorial board members say they never agreed to serve. Said's concern is downstream: a new doctoral student who sees a famous name in a journal may conclude it is a legitimate place to submit.
📌 Key takeaways:
- Author identity is now an attack surface on the scholarly record: institutions should treat name misuse like impersonation, with monitoring of faculty profiles in Google Scholar and ORCID and a reporting route that does not depend on the offending journal's cooperation.
- Predatory-venue screening needs an AI-era update — shared editorial boards, backdated archives and fee schedules are cheap, reliable tells, and the library plus faculty affairs are the right places to run that check before tenure and promotion packets rest on it.
- Doctoral programs should teach venue vetting explicitly, because the durable harm here is a training problem as much as a publishing one.
- Adopting AI Together: How the University of Texas at San Antonio Built a 150-Member Peer Learning Network — UTSA's Generative AI Peer Learning Network, which earned the institution a 2026 WCET Outstanding Work award, grew from a grassroots group into a 150-member community spanning every college and department, including faculty, instructional designers, librarians, doctoral students and staff from IT, student success, career services and institutional research. Coordinated through the Division of Academic Innovation, the network ran on curiosity rather than restriction: early sessions had faculty and outside speakers share what they learned from first classroom experiments, which led to a self-paced Teaching and Learning with AI course open to faculty in any discipline. Classroom results are specific — a general education astronomy course of more than 130 students added podcasts, NotebookLM-generated infographics and a course chatbot, with over 70 percent of students reporting a more engaging experience, and a technical writing course serving more than 600 students each semester piloted a Writing in the AI Era microcredential. Vice Provost Melissa Vito describes sustaining engagement as the thing that turns innovation into lasting improvement in the learning experience.
📌 Key takeaways:
- The durable structure is a standing cross-functional community rather than a task force with an end date: participation from IT, the library, institutional research and student success is what makes an AI decision implementable across a campus instead of announced at it.
- Self-paced faculty development plus visible course-level redesign travels further than mandates, and doing both produces the internal evidence — engagement in a 130-student general education course, a microcredential in a 600-student writing sequence — that budget conversations require.
- The real classroom change was assessment redesign toward work AI cannot complete on its own; tool access by itself did not move teaching practice.
🏛️ UCSD angle: TritonAI Connect runs a comparable structure at UC San Diego — TritonAI Champions, an AI Use-Case Group, department-level groups and AI Consulting Hours — with Discovery Series training and course redesign sequenced the same way, from community sessions to assignment-level practice.
- As AI agents take control, Nvidia rethinks the security playbook — Nvidia added Scale-In, a new pillar in its AI networking line aimed at securing, managing and operating agentic AI factories, with secure access, high-performance data movement, tenant isolation and more predictable performance where many agents run on shared infrastructure. It is built on BlueField-4 data processing units and DOCA software over Spectrum-X Ethernet; networking senior vice president Gilad Shainer describes BlueField-4 as carrying in-silicon security engines that work out of band and collect telemetry from memory, storage, GPUs and CPUs. The design extends Nvidia's earlier Open Agent Safety Platform and open-source OpenShell runtime, with Sentry monitoring agent behavior and able to quarantine agents that move outside their defined software boundaries. The architectural argument is that the monitor must sit on a device the agents cannot influence, rather than inside the environment being monitored.
📌 Key takeaways:
- Containment is becoming an architectural question: if the enforcement layer runs inside the same environment as the agent, the agent can reach it, so campus architects evaluating agent platforms should ask where enforcement lives relative to the workload.
- Agent isolation and multi-tenant separation belong in procurement language for any shared research-computing or AI-factory investment, because "which agents can touch which data" now has hardware-level answers that differ by vendor.
- Open-source agent runtimes that govern execution, access and inference routing are emerging as the integration layer; piloting one is materially cheaper than building custom governance around every agent an institution deploys.
🏛️ UCSD angle: The TritonAI Harness already implements this lifecycle in software — intake, access, build, sandbox, deploy, operate — with the LiteLLM gateway attributing model traffic by identity, which is the same out-of-band control principle applied at the platform layer rather than the network layer.
- BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model — Tether AI Research released BotPsy, a 0.8B-parameter model built on Qwen3.5-0.8B for function calling and agentic workflows on device, covering tool use across search, travel, commerce, device control, files, calendar and messaging. The team reports state-of-the-art tool-use performance for its size, leading five of six agentic benchmarks tested and outperforming models up to 1.2B parameters, including a variant 2.5 times larger on most tasks. The stated motivation is operational: frontier cloud models carry fluctuating costs and provider risk, since vendors can modify, deny or deprecate access, and not every task needs frontier-level intelligence. The authors are explicit that a 0.8B model still makes mistakes and is not intended for high-stakes use.
📌 Key takeaways:
- The case for small self-hosted or on-device models is shifting from cost to control: benchmark parity in narrow tool use means routine assistant tasks can move off frontier APIs, leaving frontier spend for work that genuinely needs it.
- Privacy-by-architecture removes a common blocker — assistant workloads that never leave the device avoid the data-sharing review that stalls cloud pilots in advising, health and HR-adjacent services.
- Match model size to consequence: leadership on a 0.8B tool-use benchmark says nothing about reliability for grading, admissions or benefits decisions, where human review stays mandatory.
🏛️ UCSD angle: UC San Diego already runs this arbitrage at campus scale — 90.6% of the 527B tokens served through TritonAI between January and September 2026 went to self-hosted or internal routes, with frontier models kept available through the same gateway for tasks that need them.