- NYU Tandon Unveils a New Strategy and Playbook for AI in the Classroom — NYU Tandon School of Engineering released a framework for governing generative AI in the classroom that rejects both blanket bans and unrestricted use, asking instructors to decide assignment by assignment whether AI advances learning or replaces the work students must do themselves. The guide, Protecting Learning: A Practical Guide to AI-Use Policy for Instructors, Departments, and Schools, sorts assignments into three categories — red, where AI is prohibited because students must show they can do the work independently; yellow, where AI plays a limited supporting role; and green, where using AI well is part of the skill being taught — and most courses will use more than one. It grew out of a school-wide review in spring and summer 2026 in which about 80 percent of full-time faculty took part and 13 departments and programs submitted written plans, and it was co-authored by faculty from every academic department plus four undergraduate and graduate students. The guide is open for public comment through November, with a first revision planned for January 2027; associate dean of AI initiatives Julia Stoyanovich said the biggest concern faculty raised was not cheating but learning loss, and that students chose stricter rules than she would have set.
📌 Key takeaways:
- Assignment-level policy, not an institution-wide rule, is the shape that survives faculty governance: three tiers, a default per course, and an explicit statement of when AI use is the learning objective itself.
- The process is the deliverable — near-universal departmental participation, student co-authorship, a public comment window and a dated revision cycle buy legitimacy that a central policy memo never gets.
- Because learning loss outranked cheating as the faculty's top concern, the IT ask shifts from detection and proctoring toward assessment-redesign support and instructor-facing tooling.
🏛️ UCSD angle: TritonAI's Canvas integration already lets instructors choose which course materials an assistant ingests and set the default model per course — the plumbing course-level AI policy depends on, and where assignment-level permissions would have to be enforced.
- Texas Lawmakers Will Consider Creating a Department of AI — A proposal to establish a Texas Department of Artificial Intelligence is among three bills filed ahead of the 90th legislative session that would add oversight requirements and restrictions on state agencies' use of AI. House Bill 605, filed by Representative Salman Bhojani, would create an independent agency — separate from the Department of Information Resources — governed by a five-member board appointed by the governor, charged with monitoring compliance, evaluating government AI systems and maintaining a public registry of covered systems. Agencies would have to inventory covered AI systems with their use, purpose, data categories and safeguards, submit to risk-based reviews, publish impact summaries for systems under heightened scrutiny, and disclose in a searchable public registry whether human review is available for certain decisions. The department would also shape procurement: a model checklist covering performance testing, data restrictions, vendor accountability, incident reporting and subcontractor responsibility, plus a requirement that agencies summarize intended use before signing a contract for a heightened-scrutiny system.
📌 Key takeaways:
- The inventory-plus-public-registry model is the one spreading across states: public institutions should assume they will need a system-level AI inventory with data categories and human-review status, producible on request rather than assembled after the fact.
- Procurement is where this lands in practice — performance testing, data restrictions, vendor accountability and incident reporting become standard contract terms, so IT should put them in boilerplate now instead of negotiating them per deal.
- Putting the oversight body outside the central IT agency is the structural choice to watch, because it separates the people who run AI from the people who audit it.
🏛️ UCSD angle: UC San Diego's LiteLLM gateway already logs model usage by identity, the raw material any inventory requires, and ITS has scoped an AI governance audit and inventory agent that would auto-discover systems, risk-classify them and export audit evidence.
- CMU, Pitt embrace AI as region debates data centers — Carnegie Mellon and the University of Pittsburgh are going all-in on AI while their region argues over planned commercial data centers, and the two universities are themselves adding capacity. CMU's $3 billion gift from Citadel's Ken Griffin funds a new Miami campus and what the university describes as state-of-the-art compute on the Pittsburgh campus, and CMU, Pitt and West Virginia University share a $160 million NSF grant on regional electrical grid infrastructure. The Pittsburgh Supercomputing Center, owned jointly by the two universities since the 1980s, is developing a new data center at an undisclosed location to house the NSF-funded Bridges-3 system and TangleLab, with construction expected in 2027. A CMU spokesperson said local needs are met from on-campus resources, PSC, national research facilities and external partners, with partnerships prioritizing sustainable and energy-efficient operations; University of Idaho research computing director Luke Sheneman framed the obligation as providing capacity while minimizing downsides, and a University of Florida energy economist noted universities chase AI infrastructure partly to attract faculty and students.
📌 Key takeaways:
- Campus AI capacity is now an energy and community-relations problem as much as a procurement one: facilities, sustainability and local government belong in the compute plan from the first draft, not the permit stage.
- The jointly owned regional center is the pragmatic middle path — shared ownership lets institutions add capacity without each building and powering its own hall, and it keeps the political heat at arm's length.
- Expect the anchor-tenant dynamic to reach institutional campus: institutional AI demand is already being used to justify commercial data center siting nearby, and universities will be asked to take a public position.
🏛️ UCSD angle: UC San Diego sites its shared AI capacity at SDSC rather than in standalone campus data centers, and SDSC now leads the NSF's NAIRR Operations Center — the same shared-facility logic operating at national scale.
- University of Alabama Builds Statewide High Performance Computing Center with Dell Technologies — The University of Alabama is deploying Dell high performance computing, AI and data infrastructure at the core of its new High Performance Computing Center, a facility funded in part by the National Institute of Standards and Technology. The center will serve the University of Alabama System and create collaboration opportunities with other academic institutions, government agencies and industry partners across the state, giving Alabama researchers access to capabilities that have historically been available mainly through national laboratories and major federal computing facilities. The environment combines Dell PowerEdge R770 and XE7740 servers with NVIDIA RTX PRO 6000 Blackwell and H200 NVL GPUs over Quantum-2 InfiniBand, with PowerScale storage for research datasets, and it supports traditional scientific simulation alongside AI, machine learning and data-intensive research plus UA's new School of Data Science.
📌 Key takeaways:
- Statewide shared compute is becoming the funding and governance model: design for multi-institution access, chargeback and federated identity from day one, because retrofitting them after commissioning is where these projects stall.
- Mixed workloads are the real design constraint — simulation, model training and coursework on one cluster means scheduling, quota and fair-share policy decide whether faculty trust the service.
- Federal co-funding is a live path for campus AI infrastructure, so the CIO and the grants office should be planning submissions together rather than meeting at the award announcement.
🏛️ UCSD angle: UC San Diego runs the same pattern through SDSC, which hosts TritonAI's on-premises inference and anchors campus research computing while a single gateway brokers access for the campus.
- Gemini Decisions API: Fast, calibrated multimodal decisions — Google introduced the Gemini Decisions API with a new Gemini Decision 1 model, an endpoint built for closed-set decisions: up to 1,000 yes/no, multiple-choice or rubric-scoring questions answered in a single pass over text, images, audio, video and PDFs up to 256K tokens. Rather than generating answer text token by token, it scores candidate options in one forward pass and returns probabilities and confidence scores at 150–250 ms end-to-end latency for $0.10 per 1M input tokens with no output tokens billed. Google says the model leads its benchmark suite for agent routing, reasoning, ranking and multimodal understanding with the lowest calibration error, and that it is nearly five times cheaper than Gemini 3.5 Flash-Lite through the Interactions API on the shared suite; the stated uses are tool selection, policy checks, model routing and context grading.
📌 Key takeaways:
- Classification is turning into a first-class model primitive, which changes gateway economics: routing, policy checks and guardrails can be priced per decision instead of per generated token, so cost models built on token throughput need revisiting.
- Calibration is the number to demand in evaluation — confidence scores, not accuracy alone, are what human-review thresholds and auto-approve rules are actually built on.
- Low-latency multimodal closed-set scoring makes real-time document triage and moderation viable inside existing workflows, which is where most campuses have stalled.
🏛️ UCSD angle: TritonAI already routes models and enforces access through its LiteLLM gateway at campus scale, and calibrated decision endpoints are the kind of primitive that would let per-course and per-department AI policy be enforced at that layer rather than in prose.
- llama.cpp v1.0.0: A stronger foundation for the local AI ecosystem — The ggml-org project released llama.cpp and ggml at version 1.0.0, committing both libraries to semantic versioning in which a major version bump is the signal of an API or ABI break — a stability promise aimed at the many downstream applications that already depend on the project's interfaces. llama.cpp runs multimodal LLMs across hardware ranging from consumer machines to high-end enterprise servers through quantization, optimized backends and a portable C/C++ implementation, and the maintainers frame the milestone around local AI and privacy becoming more important as adoption grows. Going forward, version boundaries will be made strictly visible to downstream consumers so applications can depend on the library without chasing breaking changes.
📌 Key takeaways:
- Local inference is graduating from experiment to pinnable dependency: a stable API with declared breakage points is the precondition for putting an open runtime on a production path with a patching and upgrade policy.
- Once the runtime is versioned and stable, private-by-default deployment — data never leaving institutional infrastructure — stops being a bespoke engineering project and becomes an ordinary service to operate.
- Ask vendors which inference runtime they embed and whether it can be swapped; a versioned open runtime is one of the few real exit ramps from a proprietary platform.
🏛️ UCSD angle: TritonAI runs on-premises at SDSC and served 90.6 percent of the 527 billion tokens it handled between January and September 2026 through self-hosted or internal routes, so stability on the open-source inference stack is directly operational at UC San Diego.