Curious about the latest in AI and how it’s shaping work at Middlebury? This session highlights new developments in AI tools for teaching and research, on campus and beyond. Join us for a fast-paced, practical look at what’s happening now and what’s coming next. Co-facilitated with Middlebury College Research & Instruction Librarians.
I don’t think I wrote this description, but I was asked to help give this presentation. It’s a lot of stuff to try to talk about.
Trend: AI is integrating into every layer of software, often without full visibility, without institutional choice, and at additional cost
The ability to say “don’t use AI” is pretty much already really difficult and likely a statement that isn’t going to be understood by either party. Remember the quaint days of worrying about Grammarly? Now we’ve got the OS doing local models with handoffs to cloud models of various origins. That will only increase. I feel like I can’t even get to Microsoft things without having to fight through Copilot layers that I have no interest in.
If you haven’t seen Speechify lately, go see how that’s turned from a text-to-speech tool to trying to be a centralized AI productivity home. I really did accidentally create a podcast on unsolved heists. Absolutely nuts.
Trend: AI access is increasingly complex, with rapidly shifting aspects that influence privacy, quality, and cost
Middlebury Edu protection: Gemini and Gemini Notebook
“Your Middlebury chats aren’t used to improve our models. Gemini is AI and can make mistakes.”
We have the free Google edu tier. It has our expected extra protections in Gemini Notebook and Gemini chat . . . but not in Google AI Studio, Google AI Labs, and Gemini in Collab, etc.
Same login. Same company. Tool-specific privacy changes. What a mess.
Google has also decided (I think) that edu users shouldn’t be able to see their token usage.
Then we’ve got a student experience that might have 3 different experiences/tools/capabilities just within Google Sheets. Middlebury offers the free edu version. No Gemini in the Sheets toolbar. I don’t know for sure if that’s a setting choice or a license choice. Students with a free Google account will have Gemini in Sheets . . . They can also get a free year of Gemini Pro (Sheets integration plus other stuff) after proving they are a Middlebury student BUTTTTT that will be in their personal account not their Middlebury account. Sheets Canvas fits in there somewhere.
Keep in mind this is just stuff to worry about in Google.
Trend: AI is increasingly central to search and discovery, inside licensed repositories, on the open web, and within individual items
Lots of library stuff in the slides. Do note that there are contractual statements about how that content is allowed to interact with AI tools. Now try to convey that to students/faculty and contextualize it with different Midd-provided and personal AI services.
Trend: AI is increasingly effective at converting sources into machine-readable text
I think the Maidu text example is interesting – I am adding context with the Maidu alphabet and also shaping how the OCR occurs to better align to my cut/paste needs in a way I couldn’t do with traditional OCR. So it’s kind of a mix of easily tweakable OCR and tool building to facilitate a workflow. That kind of stuff is hard to resist.
The Georgian script example is also pretty strong.
Trend: AI is increasingly effective at converting text to natural, human-sounding speech
I stumbled across VoiceBox and made a few simple clones in a variety of languages. I think it sounds pretty much like me. I can say it did not like Hebrew. I do want to stress this is minimal effort voice cloning on a free tool.
Trend: AI tools are increasingly grounding interactions in source material, reducing hallucinations, and making answers easier to verify
Example tool: Gemini Notebook (formerly NotebookLM and Gemini is formerly Bard . . .1).
You probably can’t access these, but Notebook is pretty solid for this kind of thing and faculty and students are using it extensively.
Trend: AI is letting coders and non-coders build increasingly sophisticated tools, visualizations, and interactives
Both vibe coding (where the person doesn’t really know or understand the underlying code – or doesn’t care to look at it) and AI-assisted coding (where AI helps with speed, complexity, and problem-solving) are enabling a whole new range of custom options. Maybe these things are just functional prototypes. It has helped me dramatically in terms of bash scripts, GitHub actions, and some other things I just wouldn’t have had the time to do otherwise. The line between an IDE that autocompletes and more aggressive AI integration is only going to get blurrier.
This is just a bunch of random examples in various subjects.
experiments.middcreate.net/extras/vibe-examples/
The idea that faculty or students themselves can build explorable explanations for various topics is wild. The fact that I built these math and comp sci examples is cool, but I also don’t know enough to evaluate them. That’s what the last slide is about. We can make so many things, so quickly. It’s easy to outpace our ability to evaluate what we’re doing and to create ongoing obligations that we can’t easily maintain.
Maybe Jim Groom and I need to revive the zombification theme but for projects, AI code, etc.
Trend: AI now interacts with tools like Lean and TorchPhysics to create verification loops in math and science research
Example: the Fourier neural operator, and TorchPhysics.
AI research is increasingly self-verifying: Lean checks proofs against a formal kernel, TorchPhysics constrains outputs to governing equations, and AI-designed proteins get tested against real lab data — so claims are checked against an external standard, not just a plausible explanation.
This is another world where I don’t live and I’m looking through a glass darkly.
Links
- Mathematical experiments becoming abundant through human-machine teaming
- TorchPhysics
- Claude accelerates protein design (Anthropic)
- From trust to verification: Lean’s impact on mathematics (Simons Foundation)
- AlphaProof coverage (phys.org)
- ClimODE (GitHub)
Trend: AI is increasingly capable of chained, sophisticated actions, across data sources, using tools, and producing finished products with minimal human intervention
I feel like this is something that works well for me. I struggles to talk about “agent” with tool language (Gemini or Gems or whatever), but really it’s less about model or brand name and more about your participation and when it occurs. As you give AI more info, more tools, more actions to take without your approval . . . you move from a chat interaction to an agentic one.
Guiding question: “What am I comfortable allowing AI to do?” This is less about prompt engineering and more about context and the harness: what data/tool access does the model have, and which choices and actions happen with or without human review?
Some faculty are already letting AI churn on tasks for 90+ minutes at a time or update aspects of their course on their behalf.
Trend: AI has dramatically lowered the cost of first drafts and raw output, shifting the bottleneck from creation to verification and understanding
Verification now happens on two fronts: human review and mechanical (AI) verification.
If someone sends me an AI product, I’m generally angry. Right now it’s the equivalent to sending me a Google search URL or a LMGTFY2. To make it worth my time, I need the human to make some sort of vow that they reviewed, tweaked, approve of the content, and will suffer any consequences resulting from it. Otherwise, we get stuck in a loop where people just bury you in AI content and you have two choices:
- suffocate trying to parse all of it with your human brain as it consumes all waking hours
- use AI defensively in a way that makes both the humans relatively useless
Establishing and agreeing to new workplace norms for this is going to be just awful.