The future of generative AI, a great Chrome extension, and using AI to examine an exam

AI Goes to College No. 7

Here's what to expect in this week's issue:

• My vision of the future of generative AI

• Harpa - a great AI Chrome extension

• Using Claude to examine an exam

• Should higher ed fear AI?

The highlights of this newsletter are available as a podcast, which is also called AI Goes to College. You can subscribe to the newsletter and the podcast at https://www.aigoestocollege.com/. The newsletter is also available on Substack: (https://aigoestocollege.substack.com/).

Rant of the week

Future of Generative AI - AI fades into the background and the rise of agents

What is the future of generative AI? That's a question of considerable interest to many people. The true answer is that nobody knows for sure. But we can make some guesses. Here's my guess; I'm pretty sure about my prognostication, but at the end of the day, it's just an informed guess.

Overview

• AI fades into the background - AI-enabled tools

• AI agents

• Agent to agent networks

AI fades into the background

The future of generative AI is in AI-enabled tools. Using AI chatbots is a weird mix of really simple and frustratingly complex. It's easy to get output and sometimes it's easy to get useful output. But, other times it's horribly difficult to get the chatbot do to what you want. That's why there are entire courses on "prompt engineering." Peruse AI sub-Reddits and other social media and you'll find many discussions on different prompting techniques and prompt templates. Some of us find the prompting puzzle interesting and fun, but most people only find it frustrating. That creates opportunities for AI-enabled tools.

The recently opened GPT store is a step in that direction. Custom GPT's can be very useful when they're well designed. One of my favorites is Cosmic Dream(https://chat.openai.com/g/g-FdMHL1sNo-cosmic-dream) which was developed by OpenAI. It creates colorful, kind of psychedelic images based on the prompt. It fun and sometimes useful. Here's an image from Cosmic Dream.

I created a simple GPT to help with this newsletter. It's not complicated; it just gives ChatGPT some context so that I don't have to enter it every time I want to do something related to this newsletter.

AI-enabled tools take this further (much further sometimes). There's a growing list of examples. I use CastMagic to quickly create all sorts of material for my podcasts. Otter.ai, which I've written about before, is awesome at transcribing audio. Grammarly uses AI for a variety of functions. New AI-enabled tools are coming out at a dizzying rate.

The great thing about these tools is that they are easy to use and tend to be good (or at least OK) at their tasks. One of my podcast hosts, Buzzsprout, has an AI tool called Cohost AI. All I do is upload my audio file and Cohost AI creates a transcript, show notes and suggests titles. All of these are easy to edit, so I just have to fine tune the output. It's incredibly useful and efficient. There are dozens, maybe hundreds, of examples from all sorts of fields.

Imagine a smoothly functioning AI test generator. Record your lectures, give it access to your notes or textbook chapters, set some parameters for the exam and click OK. You soon have an exam with solid questions, ready for printing or uploading into your learning management system. It can even grade the exams. How much time and effort would that save you?

Such tools could free higher ed professionals, staff and faculty, from considerably drudgery. It's a step towards moving humans "up the stack," which just means that AI can do the boring, semi-mindless work that so many of us hate, freeing us to do the more engaging, interesting parts of our jobs. Maybe you're different, but I HATE developing and grading exams. It's boring and whatever the opposite of enriching is. (I'm writing this at 3:30AM, so my vocabulary isn't at its finest.)

AI-enabled tools have almost endless possibilities to improve the lives of knowledge workers in almost any industry, including higher education. Although I can't predict the future with certainty, I'm pretty sure I'll end up being right about this. We'll all be using AI, but we won't think about it because the AI will be in the background, doing its thing to help us.

As I find interesting new tools, I'll let you know about them and you should do the same (please!). If you hear about something that looks useful, send me an email at craig@AIGoesToCollege.com. I would very much appreciate it.

AI agents

My next prediction is that we will see a significant increase in the use of AI agents. An AI agent is just AI-enabled software that performs specific tasks on behalf of a user. The language is a little messy here. AI virtual assistants such as Siri and Alexa are sometimes called agents. I'm talking about something different here. The agents I'm talking about are intended to be part of workflows. For example, an AI agent might scan your email inbox and classify messages into different projects. Or an agent might scan student admission applications and extract key data and automatically enter it into an admission system. I'd love a highly capable AI agent that could grade homework assignments.

Here are some key characteristics of AI agents:

• AI agents mostly operate autonomously, with limited human intervention

• AI agents are task oriented and limited in scope

• AI agents are embedded in workflows

• AI agents are often embedded in applications

• Using AI agents doesn't require specific expertise in AI (e.g. prompt engineering)

• AI agents can be combined (more on this later)

The interesting thing about these agents is that they'll be operating behind the scenes. In many cases, you may not even be aware they're running, although you may trigger the agents use or use its output without realizing it. The agents will just be in the background, doing their thing and making your life easier ... hopefully.

When refined, these agents will be massively valuable, but building and refining them will take a lot of effort and considerable resources. That's one reason I see most of these coming from large vendors. For example, a learning management system vendor might have the resources and the data available to create useful AI agents. However, there will be room for smaller organizations to develop highly specialized agents. Keep in mind that AI agents will be subject to the same limitations and problems as any other AI system, such as dependence on training data quality and the very real potential for bias. That's why human oversight will remain important.

Agent to agent networks

The real power of AI agents will come from networks of agents working together. Just as networks of people can accomplish more together than they can separately, networks of AI agents will be capable of amazing things.

Imagine if you had an AI agent network that continuously scanned your inbox, sending messages off to other agents for further processing. Your email processing agent recognizes an email from an student who needs advising, so it hands off the message to your advising agent who coordinates with the universities class scheduling agent to find out what relevant classes are being offered in future terms. The scheduling agent passes this information back to the advising agent who figures out what the student should do and sends this information to your email composition agent who writes a response and sends it to the student. All of this happens without you lifting a finger. Pretty sweet, right? Of course there are risks here. You're putting a lot of trust in the agent and there are places where this could go horribly wrong. But a smoothly working agent network could be a wonderful thing.

AI ecosystems will be critical. To reach their true potential, agents will need to be able to communicate and coordinate with each other. This means that there will have to be standard ways of communicating and interacting. This is where I see the big tech companies, IBM, Apple, Google, etc. taking control. It will take the big organizations to build and gain traction for a particular ecosystems. The interaction may also be standards-based, much like the Internet operates on a specific set of standard protocols.

It also seems to me that we'll see the emergence of what might be called super-agents, or maybe supervisory agents. As the name implies, the job of these agents will be to manage and coordinate the activities of individual agents. In fact, I think that most of us will interact primarily with the super-agents. More advanced users may be able to create super-agents to carry out specialized workflows.

CONCLUSION

So, that's my vision of the future:

• A shift from AI chatbots to AI-embedded tools

• The rise of AI agents, and

• Networks of AI agents taking on complex workflows

Of course, I may be wrong about all of this. AI is, to say the least, a dynamic area. Nobody really knows what the AI landscape will be even a couple years from how. I think we're in for an exciting, interesting future, one that has the potential to free higher ed professionals from some dreaded drudgery. That's the future I'm hoping to see, but time will tell, I suppose.

Tips of the week

Using Claude to examine an exam

Last evening, I went into a mind panic when I realized I needed to have a midterm exam ready to print and copy the next morning. Due to conference travel, I was going to be off campus on the day that I would normally have made the copies.

Fortunately, I had a draft ready to go and just needed to make a few changes, which I did. Then came the interesting part. Normally, it takes me quite a bit of time to scan for duplicate questions and questions that are hard to understand. I was reasonably confident of the quality of the questions, but needed to double-check. Given the time constraints, I decided to use Claude 3 Opus to do the check.

The first step was to create a pdf version of the exam and upload it to Claude. Yes, there are some privacy concerns with this, but I'm really not worried about Claude having my exam questions and there was no personal data involved, so I proceeded.

Then I put in a pretty simple prompt:

Please analyze the attached exam for two areas: 1) unclear questions, 2) duplicate questions.

Here's a screenshot of Claude's response. I've blurred some of the details for obvious reasons. There were no duplicate questions, but Claude did point out some concepts that were covered in multiple questions (which was intentional.)

I checked the unclear questions, but didn't make any modifications. This was kind of interesting. I didn't give Claude any context, so it flagged some questions as being unclear, but in the context of the course, the questions actually are clear. For example, Claude thought a question about the differences between requirements and goals in decision-making was unclear. However, I spent A LOT of time on this in class, so I'm sure students will understand the questions ... at least the ones who were paying attention will.

This went so well that I decided to get Claude to do another check, this time checking the exam against the study guide. I give extensive study guides and promise my students that if a concept isn't included in the study guide, it won't be on the exam. Now the QUESTIONS aren't on the study guide, but the topics are. For example, the study guide might say, "Compare and contrast requirements and goals in the context of decision-making," but the exam includes a multiple-choice question on the difference between requirements and goals.

Checking the exam against the study guide isn't hard, but it is tedious. So, I asked Claude to do it for me. I uploaded a pdf of the study guide and gave Claude this prompt:

Thanks! Please compare the exam to the attached study guide. Are there any questions on the exam that relate to topics not on the study guide? Note that the study guide does NOT contain specific questions for the exam, rather it covers topics that students should know.

Claude did its thing and responded:

Claude raised a few potential mismatches, but after reviewing these concerns, I decided that the questions were fine. Claude's concerns were reasonable, but human judgment was required to make the final determination about whether the questions were covered in the study guide.

Since I was already pretty confident about the exam, I felt entirely comfortable using Claude to do the checking. But, I suggest running a couple of experiments before fully trusting AI with this sort of task.

The big message here is that Claude made a couple of tedious tasks much more palatable. It also saved me about an hour of time, which was nice at the end of a long day. So, if you have some tasks that you find mundane and tedious, see if you can find a way to get AI to help free you from the drudgery. But remember, when all is said and done, you're still responsible for your work. (By the way, if you want some help in figuring out how AI can reduce some pain points, contact me at craig@AIGoesToCollege.com.)

Resource of the week

HARPA AI - A worthwhile Chrome extension

Harpa.ai (https://harpa.ai/) is an interesting new (to me) Chrome extension that integrates AI in your web browser. Harpa can use many different AI models, but the normal chat-base interface seems to use GPT. You can also connect Harpa to Claude and Gemini.

Harpa's main use is to summarize web pages, but it has many other functions as well ... maybe too many. Harpa can be a little overwhelming. Here are just a few of its functions:

• Summarizing web pages

• Generating content, including email messages, based on a web page

• Monitoring websites for changes

• Extract YouTube transcripts

• Extract data from websites,

• and much more

It does all of this from within a Chrome browser window, which is extremely useful. More on that later.

For me, the ability to summarize a web page is Harpa's killer app; it's what got me to start using it. This could be insanely useful. Here's an example. I asked Harpa to give me a long summary of this article: https://www.nytimes.com/2024/04/01/business/ai-tech-economy.html?unlocked_article_code=1.hU0.2t9e.-an9u3fVwioM&ugrp=m

Here's what Harpa generated:

Key Takeaway

David Autor, an M.I.T. economist known for his skepticism towards technology's impact on the workforce, now argues that artificial intelligence (A.I.), particularly generative A.I., could benefit the middle class by creating new opportunities for employment and increasing wages.

Summary

• David Autor, an economist at MIT, traditionally known for his research highlighting how technology and trade have negatively affected American workers' incomes, now contends that A.I., specifically generative A.I., could reverse this trend.

• He suggests that A.I., if implemented effectively, could revitalize the middle-skill, middle-class segment of the U.S. labor market, which has been hollowed out by automation and globalization.

• Modern A.I. is viewed by Autor as fundamentally different from previous technologies, offering new possibilities by changing the economics of high-stakes decision-making and enabling a broader range of people to undertake valuable work.

• He posits that this could lead to increased pay for workers, including those without college degrees, thereby lifting more individuals into the middle class.

• While some economists view Autor's perspective as speculative, others consider it a stimulating thought exercise.

• Autor acknowledges the uncertainty surrounding the impact of A.I. on employment and wages, but he argues for envisioning positive outcomes and preparing for a better future.

• His vision of A.I. as a "worker complementary technology" suggests that it could enhance the productivity of less-experienced workers, potentially benefiting them the most.

• However, the realization of these benefits depends on various factors such as corporate behavior, worker bargaining power, and policy incentives.

• Despite challenges, Autor emphasizes the importance of imagining positive outcomes, fostering debate, and making informed decisions about the use of A.I. technology.

That's a pretty solid summary and it took a few seconds to create. One potential use for this is to quickly help you decide whether you need to read an article. Or, if you trust Harpa enough, the summary might be sufficient for your needs.

You can even ask Harpa to repurpose the summary, for example by turning it into a blog post. This might be a good way to quickly create content, although you should keep the usual warnings about ethics and the need for human review in mind.

It's pretty easy to switch between models, so if you're not getting good results, you can just try a different model. Overall, I'm pretty impressed with Harpa.

I'm considering a full review of Harpa, but for now, I recommend checking it out for yourself. The free version seems pretty useful, but I'm considering their lifetime deal, which is about $140. Harpa's pricing structure is a little odd. From what I can tell, most people would probably be fine with their X (lifetime deal) or S1 (monthly) membership. You can get the full pricing rundown here: https://harpa.ai/pricing.

AI in higher ed news

Should higher ed fear or embrace AI

Generative AI is a hot topic in higher ed (that's why I produce this newsletter!). So, many in higher ed are wondering whether AI will ultimately be a good thing or a bad thing for education. A recent article in Government Technology (https://www.govtech.com/education/higher-ed/opinion-higher-eds-reasons-to-both-embrace-and-fear-ai) laid out some arguments regarding the pros and cons of generative AI for higher ed.

The article is well done overall, although it doesn't break any new ground. Here's the gist:

Pros

• Personalization - AI has the potential to provide individualized instruction, which could be a boon in an era of ever-rising class sizes.

• Efficiency - It's going to take some time, but we will eventually see efficiency gains from AI; some are already seeing significant boosts in productivity.

• Versatility - Frankly, I'm not sure what point the author was trying to make here. He wrote about multimodal AI, which combines different types of data (e.g., text, video, images, and audio) to produce interesting output and claimed that this will allow for more versatile agents, which is probably true.

Cons

• Job displacement - There will be shifts in the workforce, which is true of any technological disruption. There will be winners and losers, with some jobs being eliminated or vastly reduced, and new jobs emerging to take their place.

• Bias and inequity - As I've written about before, AI models are largely dependent on the underlying data on which they're trained. That means that biased data goes in and biased results come out, unless steps are taken to reduce the bias. Just like any form of bias, AI-based bias can lead to serious inequities.

• Over-reliance on AI - This issue deserves more attention, so I'm glad the author brought it up. It's easy to become dependent on AI, which isn't necessarily a terrible thing. We're dependent on a lot of technology, but that's the price we pay for technology's benefits. However, there's a balance between taking advantage of technology and technology being a replacement for thinking and judgment. This problem is especially important in higher ed since we have to guide students on finding the right balance between using AI as a learning aid and as a replacement for learning.

Overall, the article is worth checking out, especially if some of these ideas are new to you.