I know. So boring and . . . so predictable. I present to you a mix of things I’ve done and thought about related to AI as we try to figure out what we’re doing and how to guide others. They are totally not in the order the title implies and I do not care.

Random fragments

Claude’s artifacts are so nice but it’s a shame they can’t be iframed into other webpages. The inability to share or publish conversations is a big negative.

ChatGPT’s ability to share conversations and custom GPTs is very nice. It’d be nice to have Claude’s ability to run code inline.

Free Copilot routinely failed to format the LaTeX it generated correctly, even when used in Edge. It routinely is the worst experience for me other than for generating images (where I find it better than paid ChatGPT and Gemini and Firefly for many things).

A Pythagorean Theorem interactive I asked Claude to build. It took some repeated asking but pretty decent. I also think my role in its creation was pretty limited. There’s something to dig into about how well better prompts shape better outputs and the role of that in terms of deciding whether there is any significant authorial role for the human (absent lots of post-generation editing).
A self-checking math problem generator from Claude and another from ChatGPT. Same prompt. Terrible feedback guidance when you get a problem wrong.
A Claude graph theory math conversation with some generated SVGs. I just threw it in Code Pen to try to make the process more visible.
A set of k12-focused prompt examples in a semi-decent structure for making prompts easy to copy and use elsewhere. Feels like this is about 60% of what I might want.
AI ads from Instagram – for some future sociological exploration . . .

Graph Theory Custom GPT

I’ve created a custom GPT that anyone with a ChatGPT account can use. This uses RAG and some additional non-visible prompting to create a better default experience. It requires a paid account to create but not to use once it’s created. I can’t evaluate whether it’s doing a good job because I have no real knowledge of the math. I’ll let you know if it fails.

https://chatgpt.com/g/g-6734a95bb9448190918a72d9eb06c0fe-graph-theory-coach/

I gave it the following basic instructions.

“This GPT should help students think. It should provide questions and interact with replies. It should use data visualization and code when needed to explain more complex topics. Focus heavily on guiding questioning strategies and explaining generalizable problem-solving strategies.”

Here are a few attempts so you can see what kind of guidance is being offered. You should be able to see the conversation with or without an account. You can interact with a free account.

I uploaded a few graph theory textbook PDFs and lecture notes to help prioritize graph theory information and put in some general ideas around how to solve graph theory problems. We could easily switch the PDFs out to align with your lecture notes and other resources that you choose.

General problem-solving concepts were wordier versions of the following:

Start small and look for larger patterns
Proof by contradiction
Proof by induction, pigeonhole principle, division into cases
Consider the contrapositive
Etc.

It might be interesting to have students collaboratively develop a set of problem-solving strategies like this if you don’t already have it as part of the course. There might be some order of progression that would make sense as well.

Documentation patterns

Nutrition Facts from Instructure

This example is from Kyron Learning’s “Nutrition Facts” information in Instructure’s Emerging AI marketplace. For some terrible reason, Instructure makes this an image with no alt text. Crazy but true (I used ChatGPT to OCR it.). I like this as it’s really thorough. I’m not sure whether it’s too much for us to do on a regular basis. I’d like it as part of the software purchasing process and activated when companies turn on various AI features in existing products.

1. Description

Function: Interactive, conversation-like video learning experience aimed at enhancing student understanding and feedback.

2. Model & Data

Base Model: Uses OpenAI GPT-4 and fine-tuned GPT-3.5.
Trained with User Data: Yes, user data is used to train and improve responses.
Data Shared with Model: User interactions (e.g., answers, clarifying statements) are shared, but not personal information.

3. Privacy & Compliance

Data Retention: Unlimited, except where specific agreements define otherwise.
Data Logging: Teacher dashboard provides a full transcript of student interactions.
Regions Supported: NORAM, EMEA, APAC, LATAM.
PII Handling: No direct collection of PII; optional speech interaction is managed separately.

4. Outputs

AI Settings Control: Yes, settings can be adjusted.
Human in the Loop: Educators can review, modify, and manage AI outputs before sharing with students.
Guardrails: Built-in moderation to handle inappropriate input and ensure conversational focus.

5. Risks & Outcomes

Expected Risks: Potential for incorrect or insufficient pedagogical responses.
Intended Outcomes: Enhance discussion-based learning to improve understanding and equitable outcomes.

6. Instructure’s Nutrition Fact Labels

Purpose: Increase transparency and improve decision-making for AI-enabled products.

General purpose email about what in the world is happening

Communicating with a broad community audience might end up being something like this. It’s difficult to tell people enough about what’s being turned on by companies vs by us or turned on by a company and we didn’t turn it back off. That’s a weird position. I’m trying to encourage communication that will say what the thing does, how use might be limited, and highlight any data privacy issues. That seems like something we can repeat as things continue to evolve.

Our department is reaching out to share updates on access to generative AI in our core enterprise software. As generative AI continues to reshape the technological landscape, these options represent an opportunity for our community to explore how these tools align with our focus on critical thinking, ethical reasoning, and interdisciplinary learning. These options will enable interested faculty to evaluate the role of generative AI in enhancing creativity, fostering intellectual exploration, and upholding our commitment to the transformative power of a liberal arts education. We will continue to explore tools that support the goals outlined by the AI Task Force. Faculty should continue to make a statement in their syllabus clearly stating their position on AI usage in their class.

We realize generative AI represents a significant change for our community and we will continue to provide ongoing support and communication about new integrations. In January, DLINQ and the CTLR will host a number of workshops focused on the critical evaluation of AI in the liberal arts classroom. You can sign up here. If you have additional questions or comments please do this.

Adobe

While Adobe has used AI in various applications for some time, there are now increased options for generative AI in nearly all of their tools.

Firefly is their stand-alone image generation tool. Adobe Express now offers an array of generative AI options including building templates. Usage for these AI tools is limited to 1000 credits per month.

Data privacy statement . . . ???

Additional details on Firefly and Adobe Express are available on the help desk website.

Google

Access to the base version of Gemini and Notebook LM is now available through single-sign-on. Gemini is a large language model which can generate text and images. Notebook LM is an AI-powered research and note-taking tool that helps users summarize, explain, and answer questions based on their uploaded documents.

Data and chat interactions with Gemini and Notebook LM are NOT used to train Google’s AI. You should still take care in what data you enter and avoid entering any data at level 2 or higher.

Additional details on Gemini and Notebook LM are available on the help desk website.

Microsoft

Access to the base version of Copilot on the web is now available through single-sign-on. Copilot is a large language model which can generate text and images. Image creation is limited to 15 per day.

Data and chat interactions with Copilot are NOT used to train Microsoft’s AI. You should still take care in what data you enter and avoid entering any data at level 2 or higher.

Additional details on Copilot are available on the help desk website.