Roughing out some AI thoughts

I’m starting to rough out some things we’ll need to try to get right about AI . . .

Obviously these are just notes and aren’t meant to be anything other than stuff I’m trying to shape into coherent thoughts. I do intend to get the AI initiatives by various edu institutions into something more coherent. If you know of some I should look at, please let me know.


There is no feasible way to avoid increased use of AI at Middlebury. Depending on your definition, there’s a fair amount that’s already here in various products.

Existing software is integrating AI rapidly. This will require ongoing evaluation and communication to the community. These changes will impact both instruction and organizational workflows. It will bring up concerns around privacy as well as what is allowed by who and in what scenarios.

Faculty, staff, and students will be pursuing new tools and discipline-specific possibilities (in addition to trying to detect and prevent them). This will impact everyday teaching and learning as well as research and learning objectives. This will increase the need for guidance, expertise, policy, and infrastructure to support new use cases and to deal with repercussions.

Given the pace of change, evaluation, documentation, and communication will need to be rapid and ongoing. Security and privacy evaluation will be a challenge. We will have to decide to what degree we wish to develop internal expertise and technological capacity and to what degree we wish to outsource it.

It will be helpful to have clearly articulated support statements for each level.

Big questions

Given the Energy 2028 goals, we will want to think through through energy usage and environmental impact.

There are big questions about intellectual property, ethics, and bias that will remain contentious for the foreseeable future.


Existing Hardware

Most, if not all, of our existing labs and employee laptops more than a year or two old will be poorly provisioned to run AI on the machine. That will increase the need to run AI as a service or drive renewal of personal and lab computers. It may be that the pace of change in AI ends up overtaking this issue with more efficient models that don’t rely as much on GPUs or particular processors.

It’s unclear how AI would work in our virtualized spaces or on the high performance computing cluster.

Existing software

We will need to be aware of how AI is being integrated into the centrally-supported software that we already use.

That integration will need to be evaluated, documented, and communicated.

We will need a process for choosing whether to enable AI when it is optional and a mechanism for informing the community when AI becomes available (by choice or by fiat).

Microsoft, Google, Zoom, and Team Dynamics will be likely initial sources for consideration.

AI as a service

Primary vendors (existing)

We have existing relationships with a number of AI SaaS vendors. It will be important to understand what access we have, to what functionality, and at what cost.

While AI is an inclusive term, it’s important to realize that it’s composed of many very specific functions. Large language models are one aspect, but there are entire services dedicated to specific types of AI that meet much more specific academic and organizational needs.

There will be challenges around account management and the complexity of usage-based and process-specific charge models.



Likely new vendors


Computing cycles/High performance computing clusterInternal hardware

Assess current infrastructure. What kinds of processing do we have and how does it related to the needs for AI (GPU vs CPU)? Look at Google’s MLOps pipeline for possible guidance.


What, if anything, do we provide through Apporto? Are other ways we provide virtualized spaces for AI work?


Michigan AI services

Partner: Microsoft



Three tiers.

  1. U-M GPT
    1. Access to popular AI models (Azure Open AI and internally hosted models)
    2. Cost: free but limited
      1. “U-M GPT initially comes with prompt limits set at 25 prompts per hour. Each query asked is considered one prompt.”
    3. Infrastructure: cloud or internal hardware
      1. “Unlike some commercially available AI chat tools, ITS Generative AI tools are hosted by the University of Michigan or controlled in a U-M Cloud environment, secure, and initially provided at no cost for faculty, staff, and students. Interactions with ITS Generative AI tools are only being used for the purposes outlined in the Privacy Notice.”
  1. U-M Maizey
    1. Tweak AI with custom datasets (new tech, working with you to improve it)
    2. Cost: model dependent, per prompt length, billed monthly
      1. A one sentence prompt (10-12 words) with a paragraph reply (80-100 words), costs about $0.02.
      2. A three to four sentence prompt (75 words) with a one page reply (250-300 words), costs about $0.10
      3. The current maximum cost for a single prompt and reply is $0.48, which is about 25 pages of text.
    3. Data
      1. U-M Canvas (Ann Arbor campus only. Dearborn and Flint campuses coming soon)
        1. Course Announcements, Assignments, Files, Lecture Recordings, unlocked Modules, and Pages can be indexed.
  2. U-M GPT Toolkit
    1. Most advanced and flexible, full control over AI environments
    2. Available upon request
    3. Cost: token driven, model dependent

ASU AI Services

Partner: OpenAI

“multi-cloud, model-agnostic strategy, leveraging the capabilities of major cloud providers. It allows us to adapt components as new technologies and innovations emerge while providing the flexibility to incorporate emerging providers.”

VITRA process

AI tools

Support infrastructure

Contest – AI innovation challenge

  • enhancing student success
  • forging new avenues for innovative research
  • streamlining organizational processes.

Groups, grants, etc.

  • Intelligent Agents for Next-generation Cybersecurity – UC, Santa Barbara
  • Climate smart agriculture and forestry – Univ. MN Twin Cities
    • Neural and Cognitive foundations of artificial intelligence – Columbia
    • AI for decision making- Carnegie Mellon Univ.
  • biomedical research with AI “The overall goal of Bridge2AI is to generate flagship datasets and best practices for the collection and preparation of AI/ML-ready data to address biomedical and behavioral research grand challenges. Bridge2AI will require multiple scientific domains to come together with data science, data management and analytic experts to unlock the potential of AI/ML for the scientific community. “

2 thoughts on “Roughing out some AI thoughts

Leave a Reply