Summary
A short tutorial on how to get started with generative AI tools at the University of Michigan, and tips for avoiding common pitfalls. Focus is on U-M GPT and language models GPT 3.5, GPT 4 Turbo, DALL·E 3, and Llama 2.
Body
Summary
The University of Michigan's Information and Technology Services released a campus-wide generative AI tool in August 2023. Codenamed U-M GPT - part of GenAI at UM -, this tool provides popular language-processing AI models for use with moderate-sensitivity and public information: GPT 3.5, GPT 4 Turbo, DALL·E 3, and Llama 2. Llama 2 has the ability to query public information on the web. DALL·E is able to generate images. This article is a short tutorial on how to get started, and tips for avoiding common pitfalls.
Steps
Getting started with U-M GPT
- Open https://umgpt.umich.edu/ in a web browser.
- Login using your U-M level-1 credentials (@umich.edu email and password).
- If prompted, review and accept the information sharing agreement.
- Select the appropriate artificial intelligence model:
- GPT 3.5. A large language model from OpenAI which leverages public information, commonly known as ChatGPT.
- GPT 4. A large language model from OpenAI which leverages public information, has some human checking, and is designed to be faster and safer than version 3.5.
- Llama 2. A large language model from Meta which leverages publicly accessible information and data sets.
- DALL·E 3. Text-to-images models to generate digital images from natural language descriptions.
- Type your question in the chat box, then press Enter.
- Chat history is saved as a tab. You can create new chats by clicking the + New chat button on the left panel.
What can I do with it?
U-M GPT offers access to large "language models" which shine in particular in writing tasks such as:
- Summarizing documents
- Proofreading
- Criticizing writing
- Writing articles
- Providing pro/con arguments
- Translating, either into human languages or computer (programming) languages
- Providing ideas and suggestions (e.g. business plans, writing prompts)
Language models in U-M GPT are not great for:
- Web searches
- Fact-checking
- Accessing live websites
- Analyzing research data
General usage tips
- Accuracy can be improved by adding to your prompt: "don't make up answers" (or don't make up URLs, etc.). Note that even so, results may not be correct.
- These AI tools tend to add additional context in their answers and include your question as part of their response. Simply ask them to "be concise" to give you shorter answers.
- When asking for computer code generation, be very specific with the requirements, parameters and exclusions. Then, iteratively ask the tool to improve, fix or change aspects of the code until a final product is acceptable. Do not expect the code to be perfect - it may still need human revisions.
- When asking for public information from specific websites, use Llama 2 instead of GPT as GPT had limited access to websites.
- Always fact-check responses. In current AI models, there is always a possibility that information maybe incorrect. Ask the tool to correct information as needed but always fact-check the final response.
- For example, when Llama 2 is asked: "What are the main keywords used in publication synopsis, titles of summaries, for articles and other publications authored or co-authored by Cathy Goldstein of the University of Michigan?" It responds with incorrect information stating that Dr. Cathy Goldstein is an "associate professor" (she is a professor) and that she has done research in mindfulness and psychotherapy (her research is on sleep medicine).
- To be able to merge results from different language models, you can specify the output format to be the same in your queries. For example, GPT by default displays tables as HTML, while Llama displays them as Markdown code. You can ask both to "display as HTML" or "display as Markdown", or ask to generate in CSV format.
Usage ideas
- You can use the GPT and Llama models for spell checking, grammar correction, translating into other languages, transforming into a different tone of voice, and even for using the writing style of a specific person (after providing text to analyze from that person).
- Prepare for public speaking events and debates by pasting your speech text, and asking the tools to give you arguments pro and against, issues that others may raise, rebuttals, or both sides of the debate.
- Examples database: https://github.com/DepressionCenter/AI-prompt-database
- See the Resources section below for listings of examples and ideas.
Notes
- U-M guidance states that human-subject research, whether coded or de-identified, cannot be used in third party AI tools (e.g. ChatGPT embedded into an app). It is however unclear whether the U-M version of these AI tools can be used for human-subject research data. Further IA guidance is needed.
- Note that the Eisenberg Family Depression Center is not affiliated with U-M GPT or U-M GenAI. This article is for informational purposes only.
Resources
- Questions? Contact the University of Michigan's GenAI team (genai.umich.edu)
- AI Tools at U-M
- Getting Started Guides
- Examples, Templates and Ideas
- Policies
- Other Resources
About the Author
 |
Gabriel Mongefranco is a Mobile Data Architect at the University of Michigan Eisenberg Family Depression Center. Gabriel has over a decade of experience in data analytics, dashboard design, automation, back end software development, database design, middleware and API architecture, and technical writing.
| |
|