WN #5

Prompting Unchained! … er, Chain Prompting

Much has been written about the usefulness of prompt engineering, with some believing it’s the skill of the future and others thinking it’s quite overrated. I linked several articles on this topic a couple of weeks ago. That said, there are a few advanced prompting concepts out there that can help you play to a model’s strengths and mitigate its limitations. One of these is chain prompting, which helps AI break down its processing into more defined steps, thus improving output. For example:

• “Chain of Thought” prompting involves encouraging AI to explain its reasoning by simply asking it to take things step-by-step, or even better, including an example of how to process the prompt along with the prompt itself.

• “Chain of Verification” prompting involves asking AI to reflect on its output to verify information and “fix” its own hallucinations, resulting in a higher rate of accuracy.

Recently, a new paper put out by Salesforce, MIT, and Columbia explores another offshoot, “Chain of Density” prompting. While the prompt itself may seem technical and complex, the concept is relatively simple. While ChatGPT can summarize text (such as an article), its summaries are often more vague than most people would prefer. “Chain of Density” prompting asks ChatGPT to review its own summary and add new pieces of information to it while maintaining the same overall summary length. By asking it to repeat this process multiple times, you get an output that includes several iterations of the same summary that progressively get more and more dense, and thus, more and more optimal (up to a point where the summary becomes too dense):

It’s important to note that this style of prompt currently only works on GPT-4 (the premium version of ChatGPT), but it has several potential uses in education. For example, instead of using the “Chain of Density” concept for summarizing articles, you can use it for giving multiple iterations of feedback on student work, with each version increasing in density. In the following example, which I explain further in a thread on Twitter/X, I have GPT-4 generate 5 rounds of feedback on a student essay. Here are rounds 1 and 2, with the 2nd version adding additional strenghts, limitations, and suggestions so the feedback is slightly more dense than the 1st:

By round 5, the feedback, while still the same overall length, is packed with ideas and suggestions:

Having multiple iterations of feedback can allow teachers to match the density with a student’s ability level. You may want to give a student less dense feedback (rounds 1 or 2), or they may be ready to handle more dense feedback (rounds 3-5). You could also use these iterations to provide for student choice, or you can alter the prompt so each round of feedback is independent, allowing for a more step-by-step approach.

There is plenty of opportunity to riff on this prompt to optimize it for you and your students (Here are the text versions if you want to copy and paste. Remember, this will only work on GPT-4), or play around with the chain prompting concept to create iterations of something that becomes increasingly more creative, complex, diverse…etc.

Increasing Calls for Regulation

UNESCO is urging governments to regulate generative AI in schools, citing a recent survey showing that less than 10% of schools and universities have institutional policies and/or formal guidance regarding the use of generative AI. They’ve put out their own guidance to help with this process. If the size of the UNESCO report seems a bit daunting, here’s a brief and helpful PDF put out by ISTE and other ed organizations designed to help school leaders successfully bring AI to their buildings.

Building Students’ AI Skills

Furthering the theme of policies and guidance around AI, this critique of Harvard’s AI policy places emphasis on helping students navigate the process of using AI rather than just being explicit on if and when they can use it, a sentiment echoed by the director of the school’s writing center. The School Library Journal thinks librarians are well-suited to help students build these skills.

How are Schools and Teachers Using AI?

For the districts that have already begun embracing AI, how are they going about it? This article gives a comprehensive overview. What about at the classroom level? How are teachers using it? Lots of great ideas from teachers here, here, and here.

📌 GPT-4’s “Premium” Features

How is GPT-4 better than the free version of ChatGPT (GPT-3.5) besides better outputs, better memory, and fewer hallucinations? For one, it has access to plug-ins that connect it with other websites and tools. Here are just a few of the many plug-ins that could be helpful to teachers. GPT-4 also has an “Advanced Data Analysis” feature that can read files and analyze/visualize data, among other things.

📌 AI Integrations for Meet and Zoom

Google Meet is rolling out a host of new AI features including note-taking, summaries, and even the ability to have AI attend a meeting for you. Not to be outdone, Zoom has announced several similar features for its platform.

📌 Quizlet’s New AI Tools

I’ve used Quizlet for years to find, create, and gamify study sets for students. If you haven’t used Quizlet yet this school year, there are a bunch of new AI features to explore (in addition to the chatbot tutor, Q-Chat, that they launched last year).

Keeping Families in the Loop

One of my goals at the beginning of every school year is to increase parent/family outreach and engagement. Yet, inevitably, it slides down my priority list as my time gets split among 36,000 other tasks (sound familiar?). With AI, it can now be a quick and easy process. See my thread on Twitter/X to see what this looks like.

Seeing Things from Another Perspective

One of my favorite uses of AI is leveraging its ability to generate different viewpoints and counterarguments on the spot. Helping students see things from another perspective has never been easier.

• Coke is releasing a new AI-generated mystery flavor. It tastes like the future. Yum!

• Warner Music has signed its first AI artist. She actually made her debut five years ago as a digital model, but now she sings.

• Hearing aids are set to get a big AI upgrade. The hope is that they become more attractive to people who need them but won’t wear traditional hearing aids.

• Amazon wants to help you pay for things with the palm of your hand.

That’s all for this week! If you appreciate the content in this newsletter, consider subscribing for free, or sharing with people in your network who may find value in it.