Welcome, everyone. For anyone I've yet to meet, I'm Deirdre Schreiber, associate director of learning and development in HUIT and this is our fifth AI Tech Talk series. So good afternoon and welcome. And today we're very excited to have a few folks from Harvard Business School who will be sharing some of the things that they've been doing in AI over there.
So I will let them go ahead and introduce themselves and share their roles, but I will kick it off with Kara Amana. Thanks so much, Deirdre. So good afternoon and thank you so much to the IT Academy for asking us to come and present, we were honored. My name is Kara Amana and I oversee a team supporting learning technologies at the Business School. And I'm here with a few of my colleagues to share how we are embracing IT within HPS's IT organization, how we're embracing IT. Oh, my goodness.
How are embracing AI. Well, there we go. It's going to go smooth for everyone else. How we're embracing AI in HPS's IT organization and with the community at large. So a little bit about our AI Alliance.
The AI Alliance is a collection of talented HBS IT staff who've been tasked with exploring all things AI to build excitement within our IT organization and the greater HBS IT community. We have three co-facilitators of the AI alliance, myself, Ryan Conwell, and Laurence Tardiani, both of whom you'll meet in a few moments. I want to take a moment to give you an idea of the structure for our presentation this afternoon and what you can expect. We're going to begin with some quick words from each of our areas of focus, which are listed here on the slide around the work that folks in these areas have been doing, along with some intriguing demos.
And after that, our hope is to have a guided discussion around AI topics with all of you. And so this is your prewarning that we expect you to participate later. More details on that. Let's dive right in.
So first I will say that we will share this deck after the session. I'm sure Deirdre and Justin can help us with getting that out to all of you. So don't feel like you need to be capturing any notes if the information is on the slide. On behalf of the community building and engagement group, I want to highlight a few resources and encourage you to visit them often because we find that they are ever evolving.
So the top two resources are only available to folks who have an HBS login. So apologies for that for folks here who don't have an HBS login. But still, I think worth mentioning, the first is a share site that we've developed in HBS IT designed as a central mechanism to share information out to our HBS community about generative AI tools and most importantly, has links to some fabulous AI training that our instructional designers are running. We've also highlighted here some of the wonderful work that our colleagues at Baker Library are doing around AI.
The first link, once you have the slides, will take you to the generative AI newsletter, which anyone can subscribe to and they also have an elearning that they developed for our MBA students that I actually found to be really informative and a great primer intro into how generative AI works. And lastly, I'm sure that most folks here are familiar with these Harvard resources, but we like to call them out just to make sure. First, the generative AI at Harvard site, and second, the AI community of practice. This link will take you to the Wiki for the AI community of practice where they are continuously updating with all great information.
The other engagement opportunity that our team has been working on in the AI Alliance is AI community at HBS Microsoft Teams site. So for those folks-- [BACKGROUND CHATTER] Oops, thought that was a question. So we've developed a team site where the whole HBS community can come together to talk about all things AI. So to learn what other folks are working on, or sharing news articles there, and just having some fun as well. So it's intended to be a social place to share and interact around AI.
And if there are HBS folks on the call today who are interested in joining, please do reach out and we can help you with that. Unfortunately, the delight of our expansive Harvard Universe, we are on a different instance of Microsoft Teams than the HBS folks. I'm going to pass things off to, if I can click my slide, to my colleague Ryan. Hi, everyone. My name is Ryan Conwell. I'm the manager of our business analyst group.
We reside in the strategy, planning, and governance team within HBS IT. And with regards to the AI Alliance, the subteam that I head up is really evaluating our subscription-based tools. And there's a couple of different reasons why we decided to break things up in this way. So on Kara's earlier slide, you saw there were three different little subteams that we have going and one of them is specific to the subscription tools because we really want to see if within the IT community and our administrative use cases, is there enough of a justification to either expand on subscription-based tools or can we really accomplish what we need to with the widely available tools, whether it's free things over the open internet, whether it's something that we can make available to the entire community within HBS systems and environments? And where can we draw that line based on what we're able to do? So there's really three things that we try to do to that end. One is play around with the subscription-based tools that we have.
There's a few of them out there, we do have the Microsoft Copilot 365 that is fully integrated with just about all of your Office 365 tools. We do have a ChatGPT instance that is a little bit of a higher level that allows file uploads and some of the more detailed custom GPTs that we can experiment and play around with. And another big one is GitHub Copilot. And so I'll talk in the next slide about what are some of the use cases we do within those. But within each of those tools we do some experimentation and then we try to see what has some legs to it and try to drive that to a more detailed use case to be more practical within the administrative environment. So we're not trying to jump into any classroom solutions.
This is really trying to help the IT staff more effectively accomplish whatever it is they're trying to do. So we have a large enough IT organization where we have developers through training, through project support, and tier 1 support, all those things go on within IT. So there's lots of different areas where we can apply AI tools. And so we're just really trying to see where we can most effectively deploy those resources.
And once we do develop those detailed use cases, let's compare those to what we can accomplish with the widely available tools. So sometimes we're able to play around with these subscription tools and come up with some great prompts that can probably help people out. And then it turns out those work just fine in the widely available tool set. So we can share those out and let other people start taking advantage of things we find. So those are really what we're trying to do as a subgroup.
So with this next slide, I'll be walking you through what some of these things are. Like I said, one of our main tools is the Microsoft Copilot, and so this is what's in the Microsoft 365 environment. And there's a few things that it has helped us with, one of those is the meeting notes template creation as well as meeting note summarization.
And that takes a couple different forms, one of which is that Copilot meeting transcription, which lets you turn on a Microsoft Teams transcript and it will automatically let you know with 5 to 10 minutes left in your meeting. Hey, do you want to know what the notes are and what your action items are? And it's a little proactive with doing that thing where you don't need to go in and ask for it afterwards. You can also take the summarization from a Zoom transcript.
You can take something from a team's transcript or anything else and dump that into Copilot and say, hey, what are the top six action items coming out of this meeting? And it'll go through and analyze those and let you know who said what and when and who's taking responsibility for different items. So that is something that the free open internet version doesn't allow us to do because it doesn't have those same tie-ins, it doesn't necessarily know who everyone is. But you can get a lighter version of that done with the widely available tools, but it's not quite as robust and detailed as what the Copilot tool has offered us. Another interesting thing that I've played around with is email inbox summarization. And like I said before, the Copilot tool is really tied into everything in that Microsoft environment.
So if anyone's ever come back to that mountain of emails from a vacation, let's say you're off for a week and unread count is a little bit higher than you ever want to deal with, you can just ask really quickly. Hey, what are some of the most important emails that were addressed to me that are unread right now? And it would spit you out a whole list of what the key things are, and it would even summarize what those emails are. You could craft replies that you could just paste in and start taking quick action items without having to go through and read every single one and make those decisions yourself right away. So again, nice little key shortcuts that maybe it's not going to save you a day's worth of work, but every little bit here or there can definitely matter to make life easier and help deal with that inbox wrangling that I know we all tend to struggle with.
The next type of tool that we've been playing around with aside from Copilot is the ChatGPT. So this is the ChatGPT 4.0 enterprise, and the subscription based model is really within our own environment.
So we can deal with the higher data level within the Harvard classification, and so therefore you can upload slightly more sensitive things. And what we've done with that is we have a couple team members that have done some really interesting work in the knowledge management area. So I'm not sure how widely known the knowledge management concept is to everyone and where you are on your knowledge journey. But just a quick background there, we've been working on our knowledge base for a couple of years now and it's still something that we're focusing on and growing and trying to evolve. And part of that is making sure that we have good quality articles in our knowledge base for our support teams to reference as well as our general IT staff, and now our entire community is able to view our knowledge base.
So having good quality articles is very key to making sure that, that resource is best utilized by everyone. So a couple of the custom GPTs that have been made by our team is the first one, knowledge management, article evaluation, and grading. So this is still being worked on, but a really interesting thing that we did was we took our knowledge standard and so we have a really well-documented list of criteria that a good knowledge article needs to meet in order to pass and say that this is something that we want to be running with. And this GPT is trained on that standard and we can start pointing it to articles and saying how well does it meet the standard? And it will start highlighting where there are differences and it's working on an actual grading of, OK, this is an 85% match, this is a 95% match. Those of you who are very familiar with some of these AI models know that grading and numerical type of thing is not straightforward, so the grading definitely needs some work, but just a heads up on that. The other thing is knowledge management creation.
So again, using that standard to create new knowledge articles that meet that standard is another good thing. So just starting right off the bat, rather than grading existing things, it's creating new high quality articles. And another interesting thing has been using the analytics tools within ChatGPT to analyze our Live Online Classroom data. So taking that user interaction information from an online classroom, how many hand raises people have, how many questions they've asked, how many polls they've answered. And it can take a lot of that information and let our professors know how engaged their student is. And then finally, the GitHub Copilot.
So this is really for our developers to be watching alongside as they work within GitHub and are working on code itself to either help with troubleshooting syntax problems and help guide them. So they're still evaluating how much of a value add that is versus some other tools that are out there. But it is one thing to know. I see Steven's note here in the chat. It is very important to know where your own school's licenses are with regards to some of these tools, which is why we've been conscious about making that decision of subscription-based versus widely available is when you are in these subscription tools, it's also making sure you know what data levels are allowed and what type of information can be entered.
And if you're not in those, also being very cognizant of what can be used and where especially when you start talking file uploads and beating transcripts and we can have all sensitive data in there. So thank you Steven for entering that note and I think everyone should definitely take a reference to that. So with that, I'm going to turn things over to Michelle. Thank you.
That's actually a perfect segue to talk about our subgroup which is focused on the widely available tools at HBS and thinking about the types of data that each of these tools is eligible for, I suppose. That's one of the things we're keeping at the forefront of our subgroup of how do we use these tools in ways that are productive and effective without putting out any of our data at risk? So I am part of this subgroup. Our mission statement is to examine the widely available tools, experiment with them, identify use cases, compare them to the subscription-based tools, and create matrices of what tool, when. And really try and in one sense push those widely available tools to their limits, again, keeping that data protection in mind.
There are three that we are most focused on, which is the Copilot browser-based chatbot or formerly Bing Chat, the ChatGPT, just the free version, and then the AI Sandbox, which we know that is fine for level 3 data and not everybody in the community has access to it, but all faculty and FTE staff do. So if we could move to the next slide. Perfect. So keeping the conversation on data, we have this handy chart here that separates out the tools that are, I'm going to say, at least within HBS, largely supported by us that are widely available. And on the side there you can see the data that is allowed for each one.
And then in black are the tools that aren't technically supported, but the community can use. There's nothing to stop them from using. I would say ChatGPT probably should be in the red because we largely do support that one. But we've been trying to collect up this list of what's out there, what's available, and then thinking about what are the use cases, what data is protected if we use these, and trying to get really creative.
So when we think about how to protect our data, I run a training, for example, on the browser-based chatbot Copilot and I get this question that asked to me a lot, how do I keep it to level 1? And I'm thinking about, well, what are you doing with the chatbot? You're engaging it in a dialogue, you're maybe asking it to produce some content for you, maybe a template, maybe helping coaching you on writing an email or some of the things I like to use it for, making sure that I've covered all pieces of a research project, for example. When we're thinking about it in that way, there isn't really a situation where I'm inputting any level 3 or up data into it. I can keep it vague because I'm engaging it in a dialogue. If, for example, I was asking it for assistance coaching me on an email, I'm not going to tell it. Anything that I would be concerned constitutes level 3 and up.
But for the most part, because you're not uploading documents into that, you're pretty safe. And so that's part of what we're trying to do here is really think about, well, how are we interacting with these tools? What's the outcome of them while still keeping that data protection in mind and also, what support is available? With that in mind, if we could go to the next slide, we can talk a little bit about some of those creative uses. Kara are you driving the slideshow? Oh, perfect, thank you. Sorry, my computer just decided right at that moment to lag. But I think you're good now.
Thank you. Yes, it's perfect. So we do want to talk about some of those uses that we've been exploring with the widely available tools.
So for me, I'm an instructional designer and I do a lot of training development, documentation, a lot of executive summaries and use case capture and that type of thing. And so what the eye has been really useful for me for and full disclosure, my tool of choice has been the Copilot chatbot is really coaching me on my ideas, helping me think through topics. I don't like to give it a paragraph prompt.
I like to engage it in a conversation where I'll start off with a really broad, you're going to be my assistant, help me think through these ideas. You ask me questions and then I engage back and forth with it. And I found that to be a really productive way to get something out of it that really works for me.
Then template creation is another great one. So again, it's keeping it really broad. So I'm not worried at all about if I'm inputting any of the wrong type of data because I'm asking it to give me a template filled with helper text that myself and my colleagues can use for a lot of those things like needs assessment that we do over and over again, and the same questions that we need to ask. And so having it create those templates is incredibly helpful.
And with Copilot it's even better because it has a little word icon, and it can save it to my OneDrive. Decorative image generation, and then something I think we all need is that grammatical writing assistant. So if I'm struggling with a sentence, I just cannot get it right, I can throw that into Copilot and have it help me with fixing that or offering these suggestions. And so with that in mind Kara, remind me, are you playing the role of Christen today or is Lawrence? I am playing the role of Christen. Yes. So I'm playing the role right now of Christen Goguen, our associate director of AV operations and media services.
And so these are her slides and her use cases. So if you have questions, I will do my best, but I may have to put you in touch with someone more experienced in the area than I am if I don't know the answer. So first, so Christen has some use cases in the realm of video production, motion graphics, and lecture capture. And first we are looking at an example using Adobe Premiere Pro and Elements. So the AI that is built into the Adobe suite analyzes your video transcripts and allows you to do a number of things, remove filler words like, uh, and, um, and ah, which hugely helpful.
I'm always saying those things you, also then have the ability to find a certain text within the transcript of a video. You can cut the text and paste it into another area of the transcript. And the video then follows the text, which is just mind blowing to me.
And so it allows you to, and these are Christen's words, it allows you to assemble a rough cut very quickly. So you have a big long video and you need to very quickly produce a rough cut of what it's going to look like, you can do that very quickly. And similarly, it can help you to identify and remove retakes. So when you're filming, often you will have to retake and have folks repeat things multiple times and the AI can help you to find those retakes and eliminate them. And in Premiere Elements, you can create highlight reels. So the AI analyzes the footage and creates tags automatically based on the content and it detects scene and sound changes and then can help assemble a highlight reel for you then you can refine as opposed to you having to start from the beginning there.
And at HBS we use both Panopto and Kaltura, two different lecture capture programs and highlight reels similar to what I was just talking about are both on their product roadmaps as they work on further integrating AI into their platforms. The next use case that Christen was kind enough to share with us. This is about using ChatGPT Enterprise and Adobe After Effects, so you would be using Adobe After Effects to do production editing.
And so what Christen is demonstrating here in these images is that you can go to ChatGPT and ask it to make after effects code for object rotation. The object should speed up by the end of the 30 second composition. It then generates the code for you, you simply copy and paste the code from ChatGPT into After Effects and into the rotation property in After Effects. And After Effects then just goes ahead and ingests that code and the image rotates. And as Christen would say here, you'll just have to trust her that this image did actually rotate.
We should record a little video with the rotating star I think. So a benefit here is this type of editing usually requires painstaking keyframe edits. So you can imagine a stop motion situation in order to make that rotation happen. And with AI, it happens very, very quickly.
I'm going to pass it over to Lawrence. Great. Thank you Kara.
So I am going to embody another one of my colleagues here today is Greg Porretta So he's actually a senior user interface and instructional designer and animator with the multimedia development team. Very long title, really smart guy. And similar to what Kara was talking about with the Adobe products, he primarily works in two different ones as far as AI goes at least, and that's Adobe Photoshop and Firefly. Now both of these can be text to image-based AI tools. And so the cool thing about the Adobe suite at least is that it's trained on Adobe Stock, openly licensed content, and public domain materials. So if you're using any one of these tools, it's freely available for commercial use.
You don't have to worry about any copyright strikes or any of that stuff. So the Adobe Photoshop tool specifically has generative fill, which means that you can add and subtract areas within your image just via text prompts to create something unique. Whereas Adobe Firefly is more similar to DALL-E, if you've played with that, where you're creating something just brand new from scratch just using text and it will generate that image for you.
Side note for what it's worth, the copilot.microsoft.com, which we've been talking a little bit in chat backwards and forth, that does have DALL-E built in. So if you don't want to play with Photoshop or Firefly and just want a quick image of something, again, as long as you're typing in some level 1 stuff, it does have DALL-E.
So you can type in, generate an image of insert thing here, cowboys, and it will spit out an image for you. So feel free to play with that as well. But anyway, back to the-- oh, yes.
Lawrence, sorry. The part that I should have said on the previous slides and it applies here as well, all of these Adobe products are available to all of us as Harvard employees through the Adobe Creative Cloud licenses. Great. Thank you Kara.
Great segue, I should have said that as well. Thank you. Yes. So you can use your Adobe Creative Cloud to find these as well. They might not be installed by default on your machines, but if you get Creative Cloud, you should be able to install them. So going to Greg's specific use case, he has two super interesting ones.
So the first one was for a course called data science for managers. And part of this course was all about this historical water pump. Actually the image where you see after, on the right hand side there. And this case was all about John Snow, who had actually removed the handle and as a way to combat a cholera outbreak back in the 1800s. So Greg had this after photo.
And as part of this case, he wanted to represent what the pump may have looked like before this John Snow protagonist had removed it. So by using Generative Fill, he was able to generate a number of different handles that may have been attached to this pump to create a before and after image for use in this case. So as you can see, using this Generative Fill, there are several different designs all along the bottom and I think they ended up going with the one in the top left that's blown up with the word before in it. But it's just really incredible and amazing to see all the different handle types, but also how accurate the Generative Fill was able to get the texture and color and tone and lighting so that it really look like these handles were supposed to be there. So really, really interesting use case for that course.
The second use case, these images are created brand new. So unlike Generative Fill where you started with something and added something in, these were created from nothing. So this was a case where it was a study on 5G networks and he actually used Adobe Firefly to generate these images just purely from text prompts. So he experimented with depicting a 5G tower in various locations, giving it varying levels of detail to get exactly what he wants. But what's interesting when generating this purely from scratch is when you look really closely, you start to see some of the little details get missed, like maybe some of the spokes on some of the towers don't quite line up or go in impossible patterns or something like that.
But nevertheless, it was a really easy way for him and this was one of the main benefits that he wanted to make sure was shared with you is just the speed at which he was able to create something that he couldn't find answers for online, Google image search or what have you. He was able to generate something extremely specific in a very, very quick way. And yes, I had that note. Thank you Kara, both are available through the Creative Cloud accounts.
Thank you. So those were his two use cases. So if we can go to the next slide, please.
Great. This is the fun part. Not that the rest wasn't fun, but this is the extra fun part. So we have been doing a lot of presenting over the last 35 minutes and now we want to open a conversation.
So we're going to leverage the power of Poll Everywhere. And we have a couple questions to kick things off. But of course, if you prefer just to throw your questions in chat like you have been before or you want to just raise your hand and come off mute and talk then, that's fine too. But this is really a conversational piece. This isn't just the Lawrence talking head show for the next 15, 20 minutes.
So Kara, would you like to open our Poll Everywhere please? Yeah, I'm going to do that. Stop for a second, switch over because you're probably getting a message at Poll Everywhere but it's not open yet. I apologize for that.
Oh, thank you, Deirdre. Throwing that in chat. Wonderful.
Thank you. Thank you. About the Creative Cloud, I wasn't aware we had that as a perk. And as I discovered it, there's just a whole range of possibilities in there, so definitely check it out. Between the Adobe Pro stuff in After Effects that Kara was mentioning earlier on behalf of Christen or if you're interested in the Photoshop and Firefly stuff, it's all in there. So definitely check it out.
And one theme that I like to point out with some of these AI tools is in our slide deck when you do see that come out and you can go through slide by slide, the image on my team's intro slide of the three people standing there with the blue was created in Adobe Firefly as a very uninformed, not graphic designer person created that. And then you can compare that to what someone like Greg can do when it is in the hands of someone that does have that skill set. And you can see how much of an aid it can be when you combine that with your own knowledge and skill, which is something that we try to emphasize as we talk about the use of these tools within HBS IT. They're really good job aids, they're not necessarily just solutions.
So I see a few folks have successfully gotten into Poll Everywhere. The URL is still up at the top or should be up at the top of the screen and the QR code. This first quadrant we're just asking you to plot yourself where you fall. On the left axis is the vertical axis, I should say, is personal interest in AI technology, so how interested are you in AI? And then finding the intersection of that and across the bottom, the horizontal is your current knowledge of AI tools.
So we like to do this one at the beginning of sessions just to get a feel for where everybody stands currently. Great. Thank you Kara. I do always find these interesting to see. I mean, if someone doesn't have, for interest sake, any interest in AI, maybe we can spark some of that today. But I can see that there is certainly a lot of interest and even a lot of knowledge in this group, which is fantastic.
I'm sure we will learn some things from you in a second because we are going to talk about your use cases in just a moment, as Kara teased earlier. But this is really cool to see, see where everyone lands as far as interest goes, or are you just totally AI'ed out. Have you heard that phrase way too many times over the last, few months and you have no more interest? Well, maybe we'll see that A little bit. Too much AI. Sometimes.
Wow, this is great. Thank you Kara. I can certainly see that we have a lot of interest in the room and definitely also a lot of knowledge, which is fantastic. Maybe let's jump to the next one. Yeah, sure.
The next one I find super interesting as a conversational piece. So again, feel free to type it in here, but also feel free to just come off mute. This really is that conversational time.
So we had been sharing a few use cases that we've been playing around with in HBS IT, but we're really interested in also hearing what are some use cases you've been up to? Have you been playing with this tool or these varying tools? Have you saved any time or is it just being a fun experiment? All answers are welcome. So I'm curious where you've been playing. Paige go ahead. Hi.
I can share what we've been doing. So I work in HUIT and we develop software and we work with faculty affairs at the FAS and finance. And one of the things we've been trying to do is we match professorship funds to faculty salaries so that we can support their salaries and costs in an unrestricted way. And we do that through an application that does the matching process. And it gets to the level of if you're a faculty and say you teach chemistry, so it knows institutional data, it knows that you teach chemistry, and it also knows that the fund supports chemistry and so it will match and the money will go there.
There are funds where we have money left over or we have money where we didn't match faculty because it gets to a particular level like organic chemistry, which isn't an institutional value of a department. And so what we used, we used an API and ChatGPT to take all the data about the funds for the ChatGPT to come back and say, what is the academic subject you would really associate with this fund. So a donor gave a lot of money and said the terms are you have to use it for organic chemistry. And so when we got that list of our funds that really had money left over, we then use ChatGPT too also and we took all the course descriptions of all of FAS's courses to have it pull out which one, for example, touches upon organic chemistry. And then when we had that data, we said, oh, this is the faculty or teaching fellow and we can move money to them. So AI's been helping us a lot with that and it's allowed us to find money for the institution.
That is super interesting. So that API, ChatGPT integration, is that something that your team built yourself? Tell me more about how you got that going. So I have no idea to do it, I signed it to a staff member, so I can tell a little bit. So he spoke to a group within HUIT that was able to give us a license for the API, and then he wrote some Python scripts to help do some of the process for it. And then the ChatGPT did the rest of it. Oh, that's super exciting.
So I probably don't have enough. No, that's great. I can imagine that saved you quite a lot of time too, compared to trying to decipher and process this all yourself. Oh, huge amount of time because when you think about FAS having just 800 faculty alone, besides all the teaching staff too, just thousands of people. And we first asked the question, anyone know who teaches or we could move that money to? And it's like, no, but if you take the course description, it will say they teach in that. And of course, that's just one small example, but we did it across 78 funds that had leftover money across so many different academic areas.
Well, that's so incredible. That's so great. I haven't personally played with the API integration of the GPU. I've used them out of the box, if you will. But I know there are people within the HBS IT community who are definitely using the API integrations of ChatGPT and they're definitely getting a lot of learning and value out of it. So what a really cool example.
Thank you Paige. You're welcome. We're trying to learn.
If you don't mind me asking, how much time did that take to set up initially? You know what, it didn't once like we figured out who to connect with. It didn't take much time at all, and we had the data in our database. So it was just a matter of getting our schemas set up, writing the exact scripts.
And then the hard part was, we had to send it to client to say, some of the results that it gave us were so off, there was no reference. It might come back and say Greek mythology and we're like, this course is completely on economics, like there's no reference. So there had to be some work, but it saved our client tons of time.
We don't have the exact numbers yet, but we believe it probably identified $1 million worth of data that can be moved to free up unrestricted dollars. So significant financial benefit. I'm talking very little time, like maybe in total 12 hours of trying to figure it all out. And this was a learning exercise. Wow.
That's incredible. Thank you Paige. What a great example. Thank you for sharing. You're welcome. So I see a few others that have been posting in chat and I forgot to mention earlier and you all are jumping on it, which is fantastic is that the thumbs up feature.
So if this is a use case you see in the Poll Everywhere poll and it's something that you yourself have been doing or just think it's really interesting. Definitely give it a thumbs up, it'll bubble to the top for us to talk about. And I think the first one is so timely too, like a draft of my annual selfevaluation based on a list of bullet points and trainings I've been completing.
We're all coming up to that point in the year where we have to do our selfreflections and annual reviews and sometimes it can be difficult to reflect on those or know what to say. And that is such a great example. I know I even used it to help me write smart goals for the year before. I had a goal and I just couldn't think of a really great way to make them smart. And I know I should be making smart-based goals. So definitely think about it in terms of next year's goal setting too if you're struggling like I was.
Not to call out necessarily the person who put this up there, but do you want to talk about or anyone else done something similar? We've got some four thumbs up there on how you used it as part of your annual goal setting or evaluation. So I entered that. Hi. Kelsey Walsh from AA&D. So I used it.
I just started in my role about six months ago, so I had some metrics that I wanted to meet this year, so I included those. I took data points from other projects I did and then I also just went into Harvard Training portal and copied and pasted a bunch of the trainings I completed and sent it through AI and just edited what it put out. And my boss is actually in this session with me right now. And so hopefully it's all good.
That's great. Thank you Kelsey. I definitely find that some of these tools strongest use cases generally lean back to text-based things. And obviously, a lot of these, they're large language models.
And we are seeing more adoption of more of a multimodal approach where it doesn't just respond with text, it can also respond with images as we're seeing like it can generate logos and stuff for you. So I am seeing more and more of that, but I totally hear you where my default reaction is usually to just think about things that are in writing, like the second one, create an initial draft of an email. I certainly do that a ton as well, particularly if I have to write something that's a little bit wordy. Maybe it's more of an announcement or something like that, and I want to make sure it's coming across and I'm hitting all the bullet points that I want to hit without it being too wordy or too short where it's not clear.
I definitely use it for things like that. So I think that's great. Any other use cases people want to raise their hands and share? Lawrence, just an evolution on your drafting emails. We're having a little conversation in the chat about how sometimes it feels like it's more effort to get the AI to do what you want it to do, that there's this cost benefit analysis. Could I have just done this myself? And I found myself battling with that in the beginning and I still often do battle with that.
I will say some of the things that have been helpful to me in that frame is, if I have already written something in the frame of having it write for me, if I have already written something and I actually need it to be combined with something else, or I need it to be simplified or summarized, make this bullets, make this more brief, make this longer, et cetera. I've found that, that can be a helpful way to look at as opposed to trying to figure out how to, though I have to say Michelle here has become quite the pro with prompting. But I'm not as good as Michelle with making sure I'm telling it all the things to get exactly the output I'm looking for sometimes for like initial generation of things. But pasting, texting, and sometimes having it make changes for me has been helpful. I'll be totally frank with you as well Kara. I'm in the exact same boat, so I'm a person who has the M365 Copilot one, so it is baked into Outlook.
And I don't find myself using it to create emails for me because I can often just type in what I need to respond and hit Send faster than it would be for me to tell Copilot to write the email, then me proofread, and then hit Send. So I totally hear you. There is that cost benefit analysis piece, but I do like the examples you shared because I'm always trying to think of this myself too, is like, when's the best time to use this? When am I going to actually save time versus just playing essentially. So I hear you and those summarize or put this in bullets or make it more succinct or make it more friendly, things like that, I think are great examples. Thanks. A cold call quickly to Michelle.
Michelle, there's a question in chat learning about improving your prompting skills. Do we have any good resources there that we might be able to share? Yes. So I just posted an OpenAI prompt engineering guide, but there's actually quite a few resources.
I know you had some good one. I have so many on here, I'm trying to see. I think I'm going to post in HUIT's top level prompt engineering because then that links out to others. But I do want to just quickly do a note for what I found to be really helpful is less about getting the prompt right the first time and more about having the chatbots help you with your prompt. So you can-- That's right. --again, engaging it in that dialogue.
What am I missing? Can you ask me questions to get me where I need to go? I found that to be the best way to get those tailored responses. And then you don't have to really overthink what to put in the prompt as well. That's right Michelle, I had forgotten that you gave that good guidance before. Thank you. Love that. Alison.
Lawrence. I actually started using the Sandbox to help me do some coding. So I have a background in Python, but it's not my preferred language, so I feel I'm like mostly fluent.
But sometimes if I'm trying to solve something specific, I'll end up going through all these Google search of like Stack Overflow pages of how do I solve this thing? And so using the Sandbox, I'll say, write me some Python code that does this and it just gets me started. And if I chunk it out like that, the things I've hit up against are sometimes it's completely wrong, it's actually just making up functions that don't even exist in Python. And so sometimes it goes off the rails even though you've turned the temperature way down. So I think it's like getting you started and it's giving you the building blocks of work from things which I found really helpful actually. And then also, I've just been exploring using the ChatGPT API, so we actually have a license through HUIT to explore that and write a Python code to cycle through some data in a folder structure, do some analysis on that data, and then spit it back out.
So I feel like we're at the precipice of starting to use it for more things for sure. Absolutely. And I will say as part of my role Alison, I do a lot of PowerShell programming and I've ran into similar experiences where I actually do like to use it a lot and it can often give me the starting point for how to write the rest of it or edit it to my use case and it saved me a ton of time. And again to your point, to save digging through Stack Overflow and a whole bunch of other places, on occasion it has given me switches that don't exist.
So it's like, oh, man, if it did, it would have been fantastic. So sometimes I got to modify it a little bit, but for the most part, it gives me a starting point. It gives me that 80% of jumping off and getting going. Lawrence, we've got about five minutes left. Yes, perfect.
Deirdre, did you want to have the floor for the last five minutes to wrap us up? I mean, don't need anything particular. Just a huge thank you to Kara, Lawrence, Michelle, and Ryan for coming and sharing. And Thank you to everyone who participated in the conversation. But if there are any questions, feel free to jump in.
I've been throwing some LinkedIn learning resources into the chat as well for folks. Oh, perfect. Because I think Paige asked earlier about IT Academy courses around AI.
And so at this point just as everyone knows, AI has been evolving rapidly, and Harvard is still in more of an experimental stage. We haven't created our own customized Harvard IT Academy class, we have information security or digital accessibility. But that is something we're exploring, what is the right balance, what are the needs.
And I have flagged writing effective prompts as something, whether we create or partner with others like CWD, that's definitely an area where we're looking into. And Liz asked in chat, I don't know if anyone on this call, I don't know that we have anybody specifically on the call who's doing this at HBS IT, but anybody who's doing AI data cleaning who might be willing to share a little or might be just willing to raise their hand so that Liz can connect with you? We don't have anyone with us today who can share. I would also be curious about that if anybody does want to raise their hand about it. The closest I've done that I've seen within our group was translating requirements from ability statements to user stories.
We did have one of our business analysts use ChatGPT to do that. So you can take existing information and just change it into a different format, so it's more consumable by the project team. So it's a form of data cleansing, but probably not the exact same thing that somebody had put in the chat. That was me. So I do a bit of data cleaning with ChatGPT, the enterprise one, so I can throw in confidential data. And when I just want to do quick things like delete these columns or I only want x, y, and z, pull that out and make another CSV for me.
It's pretty quick and easy, but anything too complex, it's just not too wieldy. I have a couple examples, but they get a little bit into the weeds that I don't think we can cover in two minutes. Probably not in two minutes. I was going to say, yeah, I've done just some very basic Excel stuff but I don't know that I would call it data cleaning. But Liz, I bet that Justin would be happy to talk with you later Though also I know many are-- Or maybe a future TikTok.
And I was just going to say that Kara. So we're not doing one in June because we have the IT summit and there's a lot of breakout sessions around AI, but we are looking to continue this topic in Tech Talk series. So Liz, thank you for the suggestion. Justin, you might be presenting, but we can get you some other folks as well. So we can definitely flag that as a topic as we're planning out our FY25 calendar. Great.
And Deirdre, we can probably find some folks at HBS IT, we don't have anybody with us today. Great. Great. Well, thank you, everyone. Almost at time. Thanks so much.
It's been a pleasure. Thank you all. Thank you again. Have a wonderful afternoon.
Thank you from HBS. Sure. Thank you, everyone.
Have a great afternoon.
2025-01-16 06:05