# mk-notes-on-hist-ai-seminar-20231214

## presentations

### Goel and Levy (HKS)

- "StatGPT"
- [Stats API-201](https://www.hks.harvard.edu/courses/quantitative-analysis-and-empirical-methods) at HKS.
- key elements/goals
- guiding through problems rather than giving answers
- providing always-on resource
- consistently supportive and non-judgmental
- interacting in preferred language
- also Slack DMs
- logs a key feature
- Dan asks
- easy to use for faculty and teaching team
- easy to monitor student use and learn from it
- easy to tweak in response to student use
- Dan used the custom GPT builder
- Q:
- student-by-student customization mechanics?
- QC on problems sent to students?
- how did you handle the logs (quantity esp)?
- 3.5 triage and then on to gpt-4 if needed
- Chris S: interview profs (or fill in form) and then generate prompt from this
- Dan: if the student wants the answer, they can go to ChatGPT, so a lot depends on social capital building with students
-

### Greg (FAS)

- TeachGPT
- Chris:
- comparing to one of our best instructors... if this were me... etc
-
- Q:
- one thing that comes across is the way certain assumptions about personhood determine one's thinking here (like when Dustin asks why 3 bots and Greg is saying "well, you could call them three or 1, just different system prompts")
- so the Q => when to present one vs many personae to students? when is the mental model of personhood valuable to the learner? and is this a question for an engineer or someone used to crafting characters for theatre or film?
- next steps:
- student tracking
- query the data, look for students that are struggling, etc
- simulations
- build in vision
- multimodal interactions
- upload student handwritten work
-

### Bill and Christian (HGSE)

<iframe width="560" height="315" src="https://www.youtube.com/embed/Gqf9euf7TuI?si=cCWDn5OoUzmaNCgp" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" allowfullscreen></iframe>

- How People Learn
- students overwhelmed by amount of content
- need for more personalized feedback
- LLaMa inference engine, eith
- 78 LLMs including
- mistral 7b
- openchat_3.5
- llama-2-7b
- sqlcoder-7b
- langchain
- autogen
- streamlit for the interface
- Personae
- talk in different ways for different audiences
-

### Malan (SEAS)

- goal: provide with virtual office hrs 24/7
- text as ephemeral delivery system for logical structures
- answering questions without ego and/or judgment
-

## next steps

- send updates to Nick etc
-