The Student Who Was Afraid to Ask
A professor of engineering walks through custom GPT tutors and AI projects built at Northeastern and Humanitarians AI, including the student who used a chatbot instead of office hours because she was afraid to ask directly.
A student had a math question. She did not ask her professor, because she was afraid the question would make her feel stupid in front of someone she respected. She asked a chatbot instead. That single detail, buried partway through a professor's show-and-tell of Northeastern and Humanitarians AI's educational AI projects, says more about why these tools matter than any list of features could.
Why the talk is a show-and-tell, not a syllabus
Nik Bear Brown, a professor of engineering at Northeastern and founder of the nonprofit Humanitarians AI, opens by admitting the obvious limitation of the format: there are hundreds of AI projects across the two organizations, and a short talk can only cover a handful. Rather than try to summarize everything, the approach is demonstration, actually showing a few of the tools in action, because the underlying philosophy of the work is "build it and see what happens." Learning by doing is not just the subject of the talk; it is the method behind the projects themselves.
Custom GPTs solve a real institutional problem
The most immediately practical thread in the talk is about funding, or the lack of it. In most educational institutions, getting money approved for anything can take a year. Custom GPTs sidestep that bottleneck entirely: if one person has a GPT Plus account, they can build a custom GPT that then serves hundreds of students in a class, without needing a budget line or an approval process. Each of these bots takes roughly half an hour to build using carefully written prompts, and hundreds have been created this way.
Three examples anchor the discussion. Ada is an introductory calculus bot used by Abby Williams in her class. Nuke is a physics bot used in introductory physics classes at Northeastern. Grace is an algorithms chatbot used in a Coursera algorithms course. All three were built from prompts that are shared openly alongside the bots themselves, so the exact instructions behind each one are visible rather than hidden.
What happens to office hours
Grace's impact is described in concrete, observable terms: after introducing the bot into the Coursera algorithms course, office hours attendance dropped off almost entirely. A couple of students might show up in the first week out of habit, but after that, questions get answered through the chatbot instead. That shift shows up both in usage statistics and in the simple fact that people stop coming to office hours, which is treated as evidence the tool is genuinely doing the job it was built for, not just supplementing it.
Scaffolding, tone, and the student who used it instead of asking
Ada's story is where the emotional core of the talk lands. Abby Williams took over prompting the bot after it was initially built, and one design choice she insisted on was tone: the bot needed to be encouraging even when a student gets something wrong, pointing out what was done correctly before addressing the mistake, rather than simply returning a correct answer. That scaffolding approach, telling a student why something is right or wrong rather than just supplying the answer, reflects what teachers generally want from a tutoring tool.
A live demonstration shows this in action: a handwritten calculus problem, photographed and uploaded, checked by Ada. The response identifies what was done correctly, flags a specific error, and explains what change would make the work correct, all in an encouraging tone that mirrors how Abby herself talks to students. Abby relayed that a student who had struggled with math came to her specifically to say she had been afraid to ask questions in person because she worried it would make her look stupid, but that she used the bot regularly instead. The bots have also collected survey feedback from students, averaging around 4.8 out of 5.
Beyond chatbots: writing a book with AI
The talk shifts to a different kind of project: using language models to draft a full-length book on cancer medicine and nanomedicine, developed with a distinguished professor of physics and nanomedicine at Northeastern for a program that teaches the subject to people worldwide, including in places where access to books and educational materials is limited. The scale of the book is substantial enough that, without AI assistance drafting chapters, the project likely would not have been written at all within a reasonable timeframe, a task that could otherwise take a decade is being completed in a matter of months. A related tool, referred to in the talk as the Popper program, addresses a different problem within the same project: each chapter contains hundreds of factual assertions, and the tool uses language models to search the web and gather evidence supporting or checking those claims automatically.
A companion project called Medhavi turns the resulting book into an adaptive intelligent textbook. Rather than presenting the same content to every reader, Medhavi lets someone log in, identify who they are, a physician versus a high school student, for instance, and then presents the material along a path suited to that reader, rather than forcing everyone through the same linear table of contents.
The larger ecosystem, briefly
Running out of time, the talk closes with a rapid list of other projects underway: Dayhoff, working on protein prediction and agents for computational biology; Popper's broader role in computational skepticism, automating verification of whether AI-generated claims are accurate; Madison, extending standard reinforcement learning into agentic AI systems for branding and marketing; and a wastewater and public health project involving sensor and spectroscopy work. None of these get more than a sentence, but their presence signals that the calculus and physics bots are a small slice of a much larger portfolio.
Key takeaways
- Custom GPTs solve a real funding bottleneck: one person with a GPT Plus account can build a tool that serves an entire class, without institutional budget approval.
- Grace, an algorithms chatbot, effectively replaced office hours in a Coursera course, visible both in usage statistics and in students no longer showing up in person.
- Ada, a calculus bot, is deliberately prompted to scaffold answers, explaining what is right and wrong, rather than just returning a correct solution, and to maintain an encouraging tone.
- A student used Ada specifically because she was afraid to ask her professor questions directly, illustrating a use case beyond convenience: reducing the social cost of asking for help.
- Language models are being used to draft a full-length book on cancer and nanomedicine, a project that could otherwise take up to a decade to complete.
- Medhavi presents that book adaptively, tailoring the reading path to whether the reader is a physician, a student, or another kind of learner.
Who this is for
This talk is aimed at educators and technologists curious about how AI tools can be built cheaply and quickly inside a university or nonprofit setting without waiting on institutional funding. It doubles as an invitation: Humanitarians AI describes itself as ready to help anyone who wants to start building similar tools, with resources available on its website, YouTube channel, and GitHub.
Full transcript(auto-generated, with timestamps)
[0:00]Hey, uh my name is Nick Bear Brown. Uh Open AI asked me to talk about uh the educational AI we do here at Nor Eastern plus my nonprofit humanitarians AI. I'm a professor of engineering at Northeastern. Um what Open Eye asked me to talk about is what we teach, but we have hundreds of AI projects. So I'm just going to do a showand tell for a couple, not hundreds. But this is why I made this web page here. Um, I'm going to quickly go over what we built mostly just show and tell for the how we built it. I made this web page here so you can come to
[0:39]This web page. We have other tutorials. We have GitHub. We have a nonprofit humanitarian AI that can help you build it. But this focuses on what we build. We build a lot of things and we do it by doing it with AI. You basically build it and see what happens. So, we're going to talk about some angic AI systems we built. Uh, we're going to talk a little bit about an AI fluency framework called Botspeak that we have. We're going to talk about some things we've done for Nor Eastern, just a hod hodgepodge of uh educational tools that we built. We have a wonderful project called lyrical
[1:14]Literacy, which is using AI to create songs for people who learn in different ways because everybody's brain is different. We have a lot of people just building things for good causes because you learn by doing. So we'll talk a little bit about that. We have the computational skepticism framework. What this means is uh we have tools which automatically look and check and give evidence for whether something is true or not. It's a big deal. AI is what it's returning right. Uh we do a lot of the tutorials including tutorials on how to do everything that I'm talking about here. And we do research and publications. Uh
[1:50]I've done a ton of courses so I won't go over them. Uh again we have a humanitarian AI which is a nonprofit which does this as well. We believe learning doesn't stop when you graduate. So until a recent graduate gets a great job they continue learning and doing even a graduate who graduated two years ago a year ago there's AI courses which simply did not exist when they graduated two years ago. So they'll come back and upskill and learn as will mid-career people with a humanitarian AI. We also have a group of AI skunk works. We believe you should start building the first day you're on campus. Just build
[2:27]Things, do things, learn by doing. And I've won a bunch of awards on this. So, let's get going. I'm just going to do a show and tell, show some of our projects. We've created hundreds of custom GPTs. They're super easy to create and they're really effective. What's beautiful about the custom GPTs is if you've ever worked in an educational institution, getting the money to fund anything is virtually impossible. it can take you a year to just get a little bit of money to fund something. But with the uh custom GPTs, all you have to do is create them. If one person has a uh a GPT plus account, they can create a
[3:10]Custom GPT for hundreds of students in the class. So it takes the custom GPS are super si simple to make. Basically, you give it prompts and for all of these GPTs here, uh, we put all of the actual exact prompts that we use. So, these are the prompts using ADA. So, this is ADA, an introductory calculus bot that's used by Abby Williams in her class. All of these bots I'm going to talk about, I'm just going to talk about three. We have hundreds of them. I'm going to talk about um Nuke which is a physics bot intro introductory physics bot that's Tuesday at North Eastern for the
[3:47]Introductory physics classes with Shri and Tom Kelly um Grace which is for my uh Corsera algorithms course after using Grace in my Corsera algorithms course I just don't have people come to office hours anymore everything gets answered through the bo I mean the first week maybe a couple people come but After that people just using the the chatbot to answer everything and we can see that from the usage statistics and from you know the fact that people stop coming to office hours and ADA is for introductory calculus. So here are the prompts super easy. It takes about half an hour to make these things. We have some data uploaded these
[4:30]Abby has taken over writing this bot. So basically we wrote this uh Dev Shaw and I wrote this initially and then uh it took maybe an hour to write the bot. Students love it. Uh we've put some uh surveys on these. They average around 48 out of five for all free bots. And what's nice about it is Abby told me that one of the students came up to her who had math there. She was afraid to ask Abby questions because she would feel stupid. but she uses the bunt. Okay, so let's try Atta with a handwritten problem. I'm told by Abby that often in calculus people write
[5:10]Equations on pieces of paper and then upload you know what they take from their camera uh because math is sometimes hard to write in a proper way. You don't know latte or other things. So let's go ahead and try that. So, we're going to upload just a little handwritten um so this may or may not be right and we're going to ask added to check it. What every teacher wants is they want scaffolding. They don't want the answer to be spit out. They they want if it's right, why is it right? If it's wrong, why it's wrong. And what's also beautiful about this is uh tone. Aby's very friendly and so she
[5:55]She took over prompting it. She wants it to be encouraging even when you get things wrong. What you did right, what you did wrong. So this this has this, you know, this is not correct, but you got this part right. So good job on getting this part right, but you made a sort of a trivial little error here and it would be correct only if you made this change. This is very much happening. is very much the tone of yeah you got it partly not you are wrong but you're sort of right but you know you you made a mistake here and so not only
[6:26]Is it very useful these tools but it even has the sort of a tone of Abby in the GPT where it's encouraging and friendly even when you make a mistake so that's it for this and we'll go on to the next show and more show and tell. Uh we use uh language models to write first drafts of books. This is an extensive book. Uh this is working with Shri and Evan and Shri is a distinguished professor of physics here at nanom medicine at northeastern and uh for his canare program. Cancare program teaches cancer medicine and nanom medicine to people all around the world even very sort of places where
[7:14]It's really hard to access books and things like that. So Aan primarily is leading this project. We wrote a lot of tools to help write this book with prompts to get a first draft and then once we have a draft of a book a book like this is huge. I don't think this book gets written if it's not for AI helping us write this book. It's a very extensive book and just getting a first draft of this book is taking a few months using the AI tools but it's an important book I think and um the language models help us write drafts of chapters further the popper uh program
[7:53]What it's doing is each one of these chapters has hundreds of assertions you know where did you get evidence for this where do you have evidence for this where do you have evidence for this and so we also have tools which go search the web, use the language models to help us, things like the the GPT 40 model to gather what is the evidence for, you know, this being a tumor suppressor, what is the evidence for that and automate that whole process. I don't think a book like this gets written if not without the help of AI. It'll take us months to write this. It's an
[8:25]Extensive book. This could take a decade easily to write pre AI. um in order to actually view the book just reading the table of contents is a chore. We've also created uh Madavi which is an intelligent textbook and what it allows you to do is it allows you to log in tell who you are and it'll present the book to you according to who you are. So if you are a physician, it might lead you in the book in a different path than you would be if you're a high school student wanting to learn about cancer, cancer nominal medicine. But I don't think this book gets written if it's not
[9:04]For the help of tools like JBT and other language bonds. Okay, my timer says I'm out of time, which kind of makes me want to cry because we have so many more things to show you related to what we do with AI. We have day off which is we're doing a lot of stuff with uh protein prediction and agents for uh doing computational biology and acting as research labs. We have popper which is computational skepticism which is automated tools to verify whether AI is right or not. We have Madison which are some extended uh reinforcement learning engines which extend a standard reinforcement learning model to angenic
[9:42]AI with branding and marketing. Um we have uh ramen effect which is uh agents to wastewater um Sears um Ron and spectroscopy for uh public health. But I'm out of time so I'll just leave you with do it. If you want to do AI just do it. Humanitarians AI here will help you do it. Everything I've talked about we have on the website YouTube and GitHub for code. Just do it.
More videos
2:08Bridging the Pixel Gap in Browser Automation.
2:23How One Narrow Safety Rule Can Make an AI Less Safe Everywhere Else.
2:04Why splitting a chunk from its document makes it retrieve for the wrong question
4:20Three You Can Take Back. One You Can't.
2:21Why a 50-turn agent pays for the same screenshot 35 times unless it caches the pixels
1:53