258 – An AI Assistant for STACK

[00:00:07] Santiago: Hi, and welcome to the IDEMS podcast. I am Santiago Borio, an Impact Activation Fellow, and I’m here with David Stern, one of the founding directors of IDEMS. 

Hi, David. 

[00:00:17] David: Hi, Santiago. You’re wanting to talk about the STACK AI agent today. I’m looking forward to this.

[00:00:25] Santiago: Well, we recently had an episode about our work with several academic institutions on STACK authoring, and I mentioned briefly that AI helps in this work. I wanted to focus in on how AI helps. And we’ve mentioned it before, and I believe that we even had an episode on the STACK AI assistant that we set up. It’s a well-known company that has its AI models and you can train a specific chatbot with specific content to help specialise in an area.

And what we’ve done is we got example questions and we’ve got parts of the STACK documentation and we fed it into a chatbot to train it on how to author STACK questions. And since we’ve done that, I have to say, I have to admit, I haven’t authored a single question without using the assistant.

[00:01:41] David: This doesn’t surprise me. But let me just clarify, when you say you haven’t authored a single question without using the assistant, you are still putting in almost as much time as if you were authoring from scratch, but the quality of the question you’re getting at the end is potentially higher. Or is it, is it about efficiency? It is certainly not removing you from the loop.

[00:02:03] Santiago: Oh, no, it’s certainly not removing me from the loop. One of the things that we mentioned in that episode is that one of our contracts requires us, which we’re very happy to do, to train local members of the academic institution we’re working with to author questions. And I think that, even with the assistant, my expertise of STACK and how STACK works is what allows me to work more efficiently to get better quality questions spending the same time I was spending before.

[00:02:41] David: Yeah. So what’s interesting about this is this is not even a time saver for you, this is a quality enhancer. Is that how you would frame it at this point in time?

[00:02:53] Santiago: I think it’s a bit of both.

[00:02:55] David: Okay.

[00:02:56] Santiago: It is a time saver because I can, in the same time, achieve higher quality questions. So if I was aiming for the higher quality questions, it is reducing the time I spend. 

[00:03:06] David: Okay. Yeah, I’m happy with that as one framing of it.

[00:03:10] Santiago: It’s the chicken and egg I think that we’re getting.

[00:03:13] David: It is. And I guess it doesn’t really matter but I think, I do think it is an important distinction about whether we are seeing this as being used to increase efficiency and to be a time saver or whether we are seeing this to increase quality and therefore actually achieve more than we could have achieved otherwise.

And I think there’s an interesting distinction on how we frame that, which I think is essential. It’s not that one is right and one is wrong, it’s that both are valid. And as you say, it could be a bit of both, and there may be elements which are a bit of both. But one of the things I really appreciate about how you and Michele and others are using this agent, is that it’s never replacing your understanding. You know, your understanding is paramount to everything, it is just actually assisting with the process.

[00:04:11] Santiago: Yes, and let me give a very concrete example. I was authoring a question the other day, it was a multiple choice question, and I wanted to assess a question using the function that is defined, which is MCQ correct, which stands for multiple choice question correct, and it detects what I defined in the variables for the question as a correct answer in a set of multiple potential answers, and it’s more robust than just putting the correct answer in the assessment tree.

And I didn’t remember the syntax that needed to be used. And what the agent gave me was incorrect syntax, and it wasn’t working. And I had to say, okay, this is not working, specifically check the documentation of multiple choice questions, for the correct usage of this. Sometimes AI hallucinates, it was hallucinating and it was giving me the incorrect syntax and things weren’t working. I could have looked at the documentation myself and gone through and checked that, but I asked the agent, just look at the documentation, tell me what it is, and within seconds I had the correct answer. 

[00:05:37] David: I like that as an example.

[00:05:38] Santiago: But I had to be there to interact with the agent, and I had to tell the agent, okay, this is not working, and I think this is a possible reason, can you check this specific thing? Which is why the agent itself is not going to replace my expertise. 

[00:05:58] David: But this is exactly where the agent could now have this much wider knowledge of the syntax in a way that you don’t need to hold that information. It is something which can just help. And over time these more specific agents, I believe, can be trained better and better to be able to use the syntax, to be able to take in the syntax, and build from it in ways which support the process you’re defining.

[00:06:26] Santiago: Yeah, and I have, in my list of things to do, two questions to include in the training of the bot, of the assistant, so that it remembers these things a lot more readily, all the materials that we put for the training, it has at hand every time that you put a query in.

[00:06:50] David: Yeah.

[00:06:51] Santiago: And the, I think, most time-effective thing that I found in one of our contracts that we have is to review and improve questions. And the agent is able to read the exported XML for the question, a file that contains all the question information, it can read it and interpret it, and I can then point it to specific parts of the question where it can focus in, narrow in, and I don’t have to be sending screenshots or copying and pasting too much code, I can just upload the question file.

[00:07:39] David: Yeah. And this is the thing, this is really why it’s so powerful and why this whole process of taking specialist agents is, I think, what this is. This is an agent which is specifically trained to be a specialist. And it’s a beautiful application, which I think as an organisation, we’re gonna lean in too much more. But it really is the work that’s happening on STACK, which is helping to identify how easy this is to do, and how powerful they can be.

[00:08:12] Santiago: Yeah. And another example is I uploaded an XML of the original question, I uploaded what I thought was the final XML of the question once I made the changes that I thought were necessary, and I copied and pasted the feedback that I was given originally by the lecturer on what they wanted improved, and I said, have I addressed all the feedback? And it identified a specific bit of feedback that I hadn’t addressed.

[00:08:47] David: Oh, great. Yeah, beautiful. And that’s exactly the sort of thing it can do very well.

[00:08:56] Santiago: And it is currently a large language model, and you were saying that we want to specialise a bit more and potentially work towards small language models. However, one thing that I’m not sure it’s going to be able to achieve through a small language model – you are the expert on this – is the wealth of knowledge of mathematics.

One of the things that I mentioned in the episode that I referred to earlier is that we’re not subject specialists, and sometimes we have to author questions on topics that we don’t fully understand. Having the knowledge that a large language model has hugely helps framing the question, understanding the background, helps with the randomization. “Is this randomization going to lead to different variants that are of different difficulty?” That’s a very common question that I pose. Because I don’t understand mathematics, I cannot fully assess if a particular randomization will lead to variants that are of varying difficulty, and we want all the variants to have the same difficulty.

[00:10:05] David: Yeah. Well, in this case, we want the variant to have the same difficulty. There are other cases where having variants of varying difficulty is also desirable. This is what they found in SmileyTutor, which used to be called tutor-web, so there are advantages to that as well, but that’s a whole other story, we shouldn’t get distracted by that.

Let’s come back to these AI agents and the way you’re using them, because what is so appropriate about this is the fact that this sort of assistance that you’re describing is low risk. The way I like to think of this is, well, what happens, how robust is our system to the agent getting it wrong, whether it’s you, the human agent, or the AI agent.

What I’m loving about how you are describing this is the iteration cycles between you and the AI agents, because at the heart of this, the systems are powerful when they’re robust to actually, any agent, human or AI, getting them wrong. You said you had done it and then you asked the AI agent if you had covered all the bits of feedback, it found a mistake, it found something which you hadn’t covered.

You are then getting things back from the AI agent and you are finding the things that the agent is getting wrong. It’s those cycles, you are reducing the probability of there being remaining issues considerably, over and above what an AI agent could do or a human agent could do, because you’ve got these cycles. And that’s really what’s so powerful and effective, I think, in the way this is being used.

[00:11:45] Santiago: And the particular model that we’re using has improved over time, and it now remembers my preferences. So, when I interact with the agent, errors that it gave me before, or not necessarily errors, preferences in terms of how I define variables, whether I put a space before a colon when I’m defining a variable or not, it’s remembering those sort of things, and it’s giving me the output in the way I prefer to have it. 

[00:12:17] David: This exactly is “continually evolving”. It’s just one of our principles as an organisation, and that’s what you’re describing now that your relationship to the AI agents is continually evolving, which is great.

[00:12:31] Santiago: Yes, and another area where I found it extremely useful, some of the work that we’re doing is reviewing old questions that were used beforehand, in previous iterations of courses, and they being authored by other people, the question variables for the question itself are very hard to decipher. And I just feed them in and say, give them back to me with comments.

And the way it comments and puts context into the question variables is remarkable. It’s remarkable, and I can then read through the variables and understand what each variable is doing.

[00:13:13] David: Yeah. And this is exactly what is good practice anyway, but it is time consuming. And so this is now making good practice to be effortless.

[00:13:24] Santiago: But it’s not just the time, it’s the quality, because the quality of the comments that the AI agent is adding to the question variables is better than what I would have written myself in the first place.

[00:13:48] David: To me, this is understandable because broadly what it’s doing is that it’s just summarising information that it has in a way which is exactly what it’s trained to do.

[00:14:00] Santiago: Yes, and it is frustrating sometimes, it has gone round in circles many times.

[00:14:09] David: Yeah.

[00:14:10] Santiago: I have an example of a Parsons question that I had to edit recently, and the original question was with the first code that was designed for Parsons questions, for proof understanding, and it went round in circles and the question wasn’t working, it was giving me things that just didn’t work.

And it got to a point where it started suggesting the first thing it suggested that was wrong until I realised, “is this actually the right function to use, can you check the documentation?” And it checked the documentation again, and it said, “oh, no, actually, you’re using legacy code, we should do this differently”. And it’s something that I don’t think I would have realised just by reading the code. It has access to the documentation and it can get details so much quicker than any human could, it’s been a great help in so many different ways.

[00:15:16] David: I’m glad this is going so well and you’re enjoying using it because that’s also a part of it. Improving the quality of work is part of it as well. Even if it doesn’t improve efficiency or improve the actual quality, if it’s improving quality of work, that’s good enough.

[00:15:35] Santiago: Yes, and it is most certainly doing that. And I think that it’s really important to reiterate how it’s not just the assistant, it’s me working with the assistant and Michele working with the assistant and other people working with the assistant. If you don’t have that basic understanding or that depth of understanding, I should say, of how a STACK question should be authored, it just doesn’t output the results that it should, and authoring a question from scratch, if you don’t have that layer of understanding of STACK as a system, you wouldn’t be able to use it well enough.

[00:16:24] David: But I think there is more to it than that. And I was just having a discussion today about the fact that I don’t actually use AI much, or hardly at all myself. I would only ever use them with people like yourself, with other members of the team. And the reason for that is that I’m not spending all my day, I’m not spending large chunks of my time on a problem which means that I actually feel competent enough to know where the models might be going wrong and how I need to interact with them.

Whereas what you’re describing is exactly that symbiosis between how you’re working with them and the value they’re bringing, and how you’re using that value. And that’s what I’m observing in a lot of the team who are now using AI for so much. The big failure points recently, and we had one of these where there was a task which members of our team hoped to do through AI, and it was a huge failure point, but it was exactly that point where they were trying to use it to avoid doing the work themselves, rather than to enhance the work they were doing. And it just didn’t work, it couldn’t do that.

[00:17:38] Santiago: Oh, you’re being very cryptic. I’m interested to know the details.

[00:17:44] David: It was related to the website. There’s been a lot of work updating our website, and there are some really interesting tools, AI-based tools, to do this now. And it was interesting how good they looked without actually being good at producing the quality that was needed. So they could produce something that looked good quickly, but they couldn’t produce something which served our needs.

[00:18:13] Santiago: I’d like to finish by reflecting on how using this assistant has inspired me to start creating my own assistants for other tasks that I do. It’s really interesting how assistants like this can make our work more effective in general. But it’s, again, the relationship between me and the assistant. Because it will not get it right, it will make mistakes, and it’s those interactions, having trained an assistant and knowing its limitations and knowing where I need to look into in more detail and where I can rely on it a bit more, what makes this work.

[00:19:11] David: Absolutely. And it is exciting that these different pieces of work are moving along very nicely within different members of the team at the moment, because we’ve known for a number of years now the importance of these AI assistants within our broader work, but to actually have the STACK assistant as the first one which is getting built out concretely, and which is opening the doors to having the other assistants getting built out when they’re needed is exciting. Watch this space. This is the responsible AI ideas that we were talking about three, four years ago, this is part of what these are leading to, which is great.

[00:19:56] Santiago: Yeah. We should leave it there, thank you, David.

[00:20:00] David: Thank you.