- AI Sidequest: How-To Tips and News
- Posts
- Harvard's AI tutor experiment: What worked and what didn’t
Harvard's AI tutor experiment: What worked and what didn’t
Students used a custom chatbot for accounting class. The results are in.
Mignon Fogarty
August 07, 2025
Issue 74
On today’s quest:
— Open AI releases two free open-weight models
— How Harvard Business School students are using an AI tutor
— Eleven Labs launches music generation tool
— Google begins to curb AI workflow during power demand spikes
— Word watch: clanker
— Word watch: AI vegan
— AI in medical research
— But I don’t want to emotionally connect with a customer service agent
OpenAI just released two free models you can run on your own computer
NOTE: This is the biggest news in AI right now, but you won’t care about it if you aren’t interested in running AI on your home computer, so if that’s you, jump to the next section.
The new OpenAI model, gpt-oss, comes in two sizes: 20b and 120b. The larger size is too big to run on most home computers, but I can run 20b on my Mac Mini that has 24 GB of RAM. I can even run it with Chrome open in two windows with 40 tabs running (although it’s noticeably slower than when Chrome is closed).
Early testers say gpt-oss:20b is about equivalent to the OpenAI o3-mini model, although it has a high hallucination rate if you use it “out of the box.”
Christopher Penn explains that although it’s a powerful model, it doesn’t have as much built-in knowledge as the online models you’re used to using, such as ChatGPT, Claude, or Gemini. But it’s great for tasks where you give it information to process — such as summarizing, rewriting, or analyzing text — just don’t use it as a search engine.
The benefits of running a model on your own computer include privacy, consistency, the ability to do work without an internet connection, and the ability to know what kind of power is running your system — a benefit that is especially important to me because I run my home on solar power.1
How to use gpt-oss-20b on YOUR computer
Getting this running is now surprisingly easy! This is how I did it:
Make sure your computer has at least 16 GB RAM.
Open the app and choose gpt-oss:20b from the models pull down menu. (If you don’t see it, start typing the name.)
Send your first prompt, and the app will download the model and then process your prompt.
I did this on a Mac, but Ollama is supposed to work on Windows and Linux too. LM Studio is a similar app I’ve seen other people use and which may give you more options.
If it runs slowly, try closing other programs on your computer. On a Mac, you can see which programs are using the most memory by opening the Activity Monitor, which you’ll find in the Utilities folder in your Applications folder.
If you want to delete the model later (it takes up 13 GB of hard drive space), open Terminal (in the Utilities menu on a Mac), type ollama list and press Enter to see the list of models on your computer and how much memory they are taking, copy the name of the model you want to delete, and then type ollama rm gpt-oss:20b and press Enter.
How Harvard Business School students are using an AI tutor
A new case study from the Harvard Business School describes how 930 first-year students taking an introductory accounting class used a chatbot built on their class materials. Of all prompts entered into the system:
45% asked for explanations of accounting concepts (example: What is time-driven activity-based costing?)
42% asked for practice questions, an activity recommended by the instructor (example: Can you give me a practice question on bad debt expense and allowance for doubtful accounts?)
6% asked for calculations, the small number reflecting a general concern about the inability of LLMs to do accurate calculations
Bot use rose significantly before midterms and finals.
The professor who led the pilot study said that “his accounting students were noticeably more prepared for class, enabling professors to spend less time reviewing basic accounting mechanics and more time engaging students in deeper learning.”
ElevenLabs launches a music generation tool
ElevenLabs has launched a music generation tool that looks likely to compete with Suno. The company says, “Created in collaboration with labels, publishers, and artists, Eleven Music is cleared for nearly all commercial uses, from film and television to podcasts and social media videos, and from advertisements to gaming.”
Interestingly, the terms of service ban a few industries from using the tool. Some of the more surprising are tobacco companies, pharmaceutical companies, religious organizations, and political organizations.
Google begins to curb AI workload during power demand spikes
Google is expanding a program to shift or pause AI tasks during times of high energy demand, for example, by shifting AI training to other locations during heat waves, according to an announcement by the company.
The machine learning program is currently limited to Indiana Michigan Power and the Tennessee Valley Authority and follows a trial program with Omaha Public Power District, but Google says it also already shifts other non-urgent tasks — like processing YouTube videos — during times of high demand in Belgium and Taiwan.
Ad: Syllaby
Burned Out From Content Creation?
Syllaby.io helps creators turn content ideas into faceless videos in minutes. No editing. No filming. Just results.
Build your brand and stay consistent—without burning out.
Word watch: clanker
“Clanker” is a derogatory term for a robot that goes back at least as far as the 2005 “Star Wars” universe, but is getting more play lately as robot technology heats up. It also seems to be expanding (at least a bit) to describe AI. For example, a recent NPR piece cites a post on X by Sen. Ruben Gallego, D-Ariz., promoting his latest bill: "Sick of yelling 'REPRESENTATIVE' into the phone 10 times just to talk to a human being? My new bill makes sure you don't have to talk to a clanker if you don't want to."
Social media users are also using “robot racism” and “robophobia” to describe these sentiments.
Word watch: AI vegan
An “AI vegan” is someone who avoids using AI for environmental, ethical, or personal reasons. Unlike people who are hesitant about new technology or skeptical that it will be useful, the reasons for some people’s resistance to AI are built-in and unlikely to go away, which is why David Joyner from Georgia Institute of Technology, suggests veganism is a more appropriate parallel than skepticism or hesitancy. He writes, “Unlike many other technologies, it’s important not to assume that skeptics and laggards will eventually become adopters.”
However, I would note that in the future, as progress is made on environmental concerns and models are more likely to be trained without violating copyright, some of the objections he cites could disappear or be lessened.
AI in medical research
I spent four years in a biology PhD program at Stanford followed by some time at a pharmaceutical consulting firm, and my husband founded a small biotech company, so I have an idea of what goes into lab research and drug development — and I have been skeptical of reports about AI dramatically speeding up drug discovery.
Don’t get me wrong, AI shows great promise in some areas. It may be able to generate better predictions about which drugs to prioritize, it can pull together promising results from disparate fields to find new ideas to pursue, it can speed up data analysis, and so on. It also shows great promise in the delivery of medical treatment.
But human cells are messy, squishy things that act in their own time and behave differently in different people. And if you want to know how a drug will affect people in the real world — good and bad — it just takes time. I do not believe AI will cure all disease in 10 years, as Demis Hassabis, co‑founder and CEO of Google DeepMind said in an April 2025 interview with 60 Minutes.
“AI will not suddenly lead to an Alzheimer’s cure” and the companion piece, “AI could help lead to an Alzheimer’s cure,” by Jacob Trefethen do an excellent job of laying out drug development areas where AI is unlikely to help and where it is especially promising.
But I don’t want to emotionally connect with a customer service agent
You may recall I recently had a wonderful experience with my first AI customer service agent on the phone — it was fast, smooth, and accurate.
But I still wasn’t thrilled as I listened to someone from a voice AI company on the Superhuman podcast say that whereas today, companies want to get you off customer service calls as fast as possible (or get you to not call at all), in the future, companies will try to get you to feel a connection with AI customer service agents and stay on the line to talk longer.
Quick Hits
Psychology
Illinois Bans AI From Providing Therapy. Utah and Nevada have also passed laws regulating AI use in therapy. — Gizmodo
Legal
Job market
“Palantir Technologies just achieved a milestone that would have seemed outlandish even to its boldest promoters a year ago: a first-ever billion-dollar quarter, propelled by a runaway boom in artificial intelligence that is now fundamentally transforming how the company operates — and how many employees it believes it needs.” (The CEO’s quotes about AI in this story are some of the most hyped-up I’ve seen outside of the model builders themselves.) — Fortune
Education
Custom AI Tutor Bots Are Transforming Learning at Harvard Business School — Harvard Business Impact
Other
OpenAI is providing ChatGPT Enterprise free to the entire U.S.federal workforce for the next year — OpenAI
You Can Now Have a Conversation With the Statues at Versailles Using Artificial Intelligence —Smithsonian Magazine
Google releases Genie 3, a model that builds a world as you explore it (I’m not that into video AI, but this seems like a big conceptional leap forward.) — Google
What is AI Sidequest?
Are you interested in the intersection of AI with language, writing, and culture? With maybe a little consumer business thrown in? Then you’re in the right place!
I’m Mignon Fogarty: I’ve been writing about language for almost 20 years and was the chair of media entrepreneurship in the School of Journalism at the University of Nevada, Reno. I became interested in AI back in 2022 when articles about large language models started flooding my Google alerts. AI Sidequest is where I write about stories I find interesting. I hope you find them interesting too.
If you loved the newsletter, share your favorite part on social media and tag me so I can engage! [LinkedIn — Facebook — Mastodon]
Apologies for saying this all the time! I’m not trying to virtue signal. The newsletter is just growing quickly, and there are a lot of new subscribers who may not understand why the source of power matters so much to me.
Written by a human