202 – Building Custom AI Assistants

[00:00:06] Michele: Hi, and welcome to the IDEMS podcast. I’m Michele, an Impact Activation Fellow, and I’m here with David Stern, one of the founding directors of IDEMS. Hi David.

[00:00:18] David: Hi, Michele, your first time to be leading a podcast. What are we discussing? 

[00:00:22] Michele: Yeah, I am surprised and excited. And one of the reasons why I’m excited is that we are discussing a project that I really care about, which is related to STACK, but also to AI. Before starting to work for IDEMS I started developing for myself a tool, an AI tool that would help me author STACK questions.

And surprisingly, even though I knew nothing at the time about AI, it became incredibly useful. I was learning how to author STACK questions, so part of the usefulness was having some kind of interactive documentation that would help me just learn more.

But then also another side of it was the efficiency, I was able to produce questions, quality questions, I don’t know, maybe five times as fast as I would have normally, maybe 10 times, I don’t know. And also the quality of my question became better because everything was more tidy and I had all of the comments, and I could make questions multilingual very easily and so on.

So I realised, again before working for IDEMS, that this was solving a real issue that STACK has, and the real issue in my opinion is potential authors are scared by how difficult it is to learn how to produce a STACK question, or maybe even mainly by how time consuming it is to produce STACK questions.

Teachers are just used to go to the textbook and give exercises to the students. So for a newcomer, this seems so inconvenient that you can keep on talking about all of the advantages, all of the learning advantages for the students, and all of the statistics you can have at the end of it, and all of the feedback, and so on and so forth. But if it’s too inconvenient, nobody’s gonna do it.

[00:02:42] David: Let me just interrupt briefly there, it’s not that nobody’s going to do it because there’s 2000 institutions across the world who are doing it. It’s just that that’s 2000 institutions where you could have 2 million institutions where this could be used much more widely. So it’s not that no one is, that you are limiting the set of people who are gonna use it.

[00:03:00] Michele: You are very right. It was a good approximation as a percentage of total amount of possible authors.

[00:03:08] David: Exactly, and I think maybe to tie this in to IDEMS more generally, this is something we believe, on almost all the things we’re working on, this idea that you need people to be expert authors in different ways and to actually build these systems, which are complex, but require expertise to author well, this is central to what we’re doing, and therefore that same problem applies almost everywhere.

And this is why in the past, the sort of solutions that we are proposing, they would’ve had niche audiences where only small numbers of people could use them compared to the total need. And I think what you are articulating is we live in a new world. The power of AI agents to change this and to make that expert authoring accessible, there’s something really powerful there.

[00:04:05] Michele: Yeah, I agree. I suppose part of what we are trying to do, part of it is just to make technology more accessible. 

[00:04:13] David: When you say make technology more accessible, the thing you’re missing is it’s make technology development more accessible. Building technology that’s for experts, whereas what we actually want is we want everyone to be able to build and own the technologies they built. That’s what we can change.

[00:04:35] Michele: Yeah, I do agree. I do agree. And this was and is, generally speaking, my vision for AI, to just break barriers. But specifically for this STACK assistant, I’m very surprised by how well it works and how useful it is. And so I really wanted to make it into something that people would just use.

And when I started working at IDEMS, knowing that in this company we do have people who use STACK professionally for eight hours a day, my initial thought is, let’s help them, let’s make them 10 times more efficient, let’s just give them what I have at the moment. But then you and I started having some follow up conversation about maybe developing this assistant more, not only in order to make it better at what it does, but also to add functionalities.

For example, an idea that you gave me that it’s super valuable is to eventually make the assistant interact with GitHub, so that in GitHub we would have our question banks and that would enable teams to work on question banks at the same time, maybe by only, or almost only interacting with the AI assistant without having to actually go to GitHub and so on.

[00:06:16] David: Let me come into this because this is something where the way you are framing that is exactly where, this is not just about STACK. This is about what I would imagine AI agents and the role responsible AI agents can play to enable us to make it easier for teams to work on these bigger projects together.

I actually think the tools for individuals to build different stuff themselves, actually, they’re pretty good already. And the problem is putting teams together becomes extremely expensive and inefficient, and you need really high level of expertise to enter into that collaborative team.

And this is where I believe with the right AI agents, we can reduce those barriers so those teams can be more diverse, they can have people who have technical expertise working with people who have context expertise or subject expertise, so there are less technical barriers to contribution because of the AI agents. And this is something that is part of, if you want, the broader AI vision for IDEMS.

This is where we really believe responsible AI agents can enable us to be more collaborative. Not to take power away from humans, but to actually enable us to collaborate better as humans. And that approach of actually thinking about, well, what does that collaboration look like and how do these AI agents help that collaboration on these bigger projects so that we can get better things, which are used more widely, this is central to what Kate and I have been discussing for years, and with Danny as well, of course, about how we actually build the, if you want, the technologies of the future.

And I think what you’ve done with the way you’ve built this AI agent for STACK, demonstrated that the barriers to building those agents is now low enough that, you know, you said you had no expertise in AI, but you did have a strong mathematical background, you were an expert in other ways, so a non-specialist expert is able to quickly build things which are useful.

And that’s really where I would argue the technologies are at the moment. They are not yet, in my mind, efficient, they’re not yet ecologically friendly, there’s other issues around them, but they are at a stage where a non-specialist expert can easily use them to build something which is useful.

[00:08:55] Michele: That was surprising to me as well, how easy it was to have a product that is maybe non-professional, maybe non-complete, but useful. And again, it was just a matter of understanding what the available tools were and set everything up. That was the main struggle and it was not a big struggle.

And also, you mentioned this idea or fear that AI would take away something important from humans. This is very widespread as you know, there’s a lot of hype surrounding AI, but also a lot of fear. And one of the fears is what is our role at some point.

But I’m very sure the aim should be that AI takes away all of the repetitive, boring parts of a job so that you can actually focus on the creative and interesting part, and you can spend more time on what actually brings value.

[00:10:07] David: Well, the problem with framing it like that, and this is a genuine problem, is those boring, repetitive jobs in big companies, they’re the jobs given to entry level people. And so the thing which we are already seeing is that there are now a shortage of entry level jobs because a lot of those boring, repetitive tasks are no longer being given to entry level people, they can be taken over by AI.

And so, I want to sort of frame, while I agree that that’s part of what I think AI can do, I believe responsible AI is about changing the roles that people play in the team and enabling an individual to do more. You sort of said, if you can be five times as effective, 10 times as effective, and other studies showed this can be up to 20 times as effective, then one person can do the job of 5, 10, 20 people. And so now there’s less jobs. So that is a fear.

My claim is that actually, if we look from a societal perspective, it means that we are going to need to change where the jobs are, but there’s plenty of work to be done. We just need to reallocate, and rethink about how those jobs happen. If people are more effective at certain things, then it shouldn’t mean that there are less jobs, it means that certain tasks which need more attention can now get more attention.

That’s what we need to try and do, but our societal structures will have to shift possibly to fit that. So I do think this is going to be a period of transformation for societies, but I absolutely agree with you, done right, this doesn’t remove the need for humans. On the contrary, it enhances what humans can achieve. Well, that’s exciting. Imagine as a society we are now 10 times more efficient or effective.

Now I’m not saying that’s what AI is going to let us get to, but I do think it will shift. You know, I think there’s certain roles where therefore we will need more people and there’s other roles where we will have less. Coming back to teaching, maybe the life of teachers will get better as they can spend more time focusing on teaching. Maybe it’ll even get to the stage where we can invest more of our wealth into education more generally. These are the shifts which might happen, they might not happen, but there are interesting discussions. 

[00:12:40] Michele: Talking about the shift that I’m sure is going to happen, it’s happening at the moment already. I think the issue here is the speed of the change, because any change for the better, from a societal point of view is good. But then what about the individuals who now have some jobs or are experts in some fields, even though society, or the jobs may change, the single individuals may struggle, and that is given by the speed of change.

That does worry me, at least in the short term. But I’m still an optimist and I do believe that the end state may very well be a good state or a better state.

[00:13:29] David: I’ve been called an eternal optimist. My current version of this, is I don’t know. I don’t know whether the changes we are going through now will lead us to somewhere better or somewhere worse. I genuinely, however optimistic I am, I think it’s in the balance. I think it’s really down to decisions which are going to be made in, you know, probably the next, certainly in my lifetime, I think those decisions are going to be crucial in terms of how we as societies move forward.

And I don’t know what the right or the wrong answers are, but I do believe that things are in the balance and those decisions do matter. If we make certain decisions, we could get to a state which is better than our current state, and if other decisions are made, it could lead to states which are worse than our current state.

[00:14:26] Michele: Yeah, I completely agree. A field that I would like to explore more is the field of alignment of AI, which is more or less, at least one of the topics that is necessary to explore if you are worried about the future of AI or the future of humanity given AI. And more generally, as you are mentioning, responsible AI, you said that the important decisions are going to be made in your lifetime, I’m sure you’re right. I actually think in the next few years, maybe five.

[00:15:00] David: I think important decisions have been made, important decisions will be made over the next five years. I don’t believe the tipping points are gonna happen within the next five years. I think important decisions are being made all the time related to this. But with societies in different ways, if we talk about this from a global perspective, I don’t think we are getting to the tipping points yet on this.

There’s a whole nother set of philosophies around that, of when does a decision then become irreversible? When do you get into a state you can’t get out of? I actually don’t believe that’s going to be in the next 10 years. I think there’s a lot of noise happening now, there’s a lot of movement, there’s a lot of decisions happening, but I don’t think we’re actually on the edge of tipping points. I think that we’re getting to sort of crisis points, but I don’t think from a societal, a global societal point, we are quite at those tipping points yet.

But then again, I’m an optimist, so I’m hoping we’re not at those tipping points yet, because the way a lot of societies are going at the moment, I’m not sure we’re making good decisions yet. But there are big decisions being made, and so there’s this mixture, there’s good decisions, there’s decisions I consider dangerous, there’s this balance between different decisions being made.

I hope, and I believe, we are not in that urgent ‘things are gonna have to happen in the next five years’. I think it’s more like a 20 year period where these big decisions are gonna be coming out and playing out in ways where we will reach tipping points, where there will be points of no return.

[00:16:32] Michele: Well, as you mentioned, nobody really knows. I hope you’re right, because one of my worries is the speed of change. And also some of the worries around AI are, you know, geopolitical and economical and so on. But again, given all of this, all of the reasonable worries that we are having, still, at the moment at least, AI is such an amazing tool that can be used for applications that are not yet fully explored, by far.

[00:17:09] David: But let me come back to a little bit of why I’m actually pretty confident in my position.

[00:17:15] Michele: Sure.

[00:17:16] David: You get down to the mathematics behind it, actually, we are only scratching the surface, it’s in its infancy still. You know, the work which has happened on large language models is incredible, the advances in the last few years have been absolutely phenomenal. But if I get down to the underlying mathematics behind this, we’re scratching the surface of what’s possible.

I believe in the next few years we’re gonna go from these large language models to actually having really powerful small language models. That’s been some of the advances which have sort of been made. When we get to really powerful small language models, which could compete with the current large language models, well that changes the dynamics, all the environmental concerns basically disappear because the small language models don’t have the same environmental impact.

I think we’re still in our infancy in terms of what’s possible mathematically for us to do. And so the advances that are gonna happen over the next 10, 15 years I think are gonna be fantastic and move things forwards in leaps and bounds again, but behind the scenes, my guess is, people’s vision of AI was the Turing test.

Passing the Turing test took society about 80 years. This was a momentous shift forward, which happened very quickly, which is where that urgency has come from, because that was what people imagined. Actually a lot of the big powerful things happened 40 years ago, which have actually been built into our systems and we’re only just now getting on top of those advances from 40 years ago. It’s gonna take another 20 years before this actually irons itself out into technologies which are permeated through society in stable ways, is my expectation.

[00:19:05] Michele: By the way, I was so wrong about the Turing test. I remember having discussions with friends, and I was supporting the idea that once the AI would pass the Turing test, that would be the tipping point. And I was so sure about it. And now that that’s done, we passed by it almost without noticing that it happened.

[00:19:28] David: Sorry, have you heard the discussions around Turing Test 2.0? But anyway, that’s a whole different discussion, we shouldn’t get distracted by that now.

[00:19:35] Michele: Right. But I do agree there are ways in which we can make the Turing test even more interesting. But, on the other point you were making about small language models, there are two at the moment, two races that are happening in AI.

One is physical, just who buys more, mainly from Nvidia, and who buys more has an advantage. But of course, there’s also the software part, and I’m convinced that the state of the art is not the most efficient state of the art that we could have, so that, as you were mentioning, we could have potentially smaller models, but more clever, built in a more intelligent way that then become more efficient with all of the advantages that you were describing. 

[00:20:30] David: This is exactly, there’s so much potential for that latter. And the problem is, is it fast enough compared to the former? At the moment, the speed of development, everybody’s rush, is just, well, let’s throw more compute at it. If we throw more compute at it, things happen faster, and we don’t need to worry about the efficiencies. Now, I believe that switch, that’s part of the battle. If we go down a world where we just have more and more compute, I think that’s going to lead to a dystopian future.

If at some point over the next 15, 20 years you want efficiency in your models, I think that’s where we have hope because actually the sort of things that are needed, if we could build not just your AI assistant for STACK, but an AI assistant for STACK which is just what’s needed and has the capability that’s needed, at the moment your AI assistant for STACK uses basically the ability to talk about anything, to just do a single task.

Whereas, once we actually are able to build these smaller language models, which can just do specific tasks and they become more efficient and that becomes effective, that would be a whole nother leap forward. And there are people working on this. These are some of the things which are really progressing.

As we’ve discussed before, I think it’s absolutely sensible to progress using the large language models ’cause that’s what we can now do efficiently, effectively, with relatively little manpower, we can build things which are useful. But I would be really worried if that was the long-term goal. The long-term goal is once you have such an efficient large language model, which is useful for something, in five years time, how can we make this smaller to just do this task and actually make this more efficient so that this could get down and run on a laptop or a mobile phone?

Could we get down to that and actually have, you know, these really powerful systems not needing to be done in big environments which are energy consuming and all the rest of it, but could actually be brought down to size and still be effective at what we need them for?

My expectation is yes, that will happen over the next 10 to 15 years. This is why I don’t think it’s a five year time span. I think some of these other advances are gonna take a decade. But if we are ready for them and we line it up, then that’s where then you’ll outcompete because you’ll be able to do things that the large language models can’t do. And that’s part of the competition.

And all of this is good, this is sort of genuine progress happening in a competitive environment, but with different visions about how the future might look, and what you are building towards.

[00:23:15] Michele: Yeah, more and more we have to plan not for what exists at the moment, but for what is likely going to exist in the future. Since the change is so fast, we have to plan for the future. And I’m very sure we cannot generalise, but I remember when Deep Blue for the first time, beat Kasparov, who was the chess world champion at the time. Deep Blue was a huge computer only specialised in playing chess, and it managed to beat Kasparov, but it was not a given that it was going to win.

And now any mobile phone can beat the world champion without any possibility of comeback from the world champion. So, again, I’m not sure we can generalise, but it’s very hard to imagine how things are going to be 5, 10, 20 years from now.

[00:24:14] David: Well, there are elements of it which are hard to imagine. But the underlying mathematics, well actually the pathways are pretty clear. This is what is known, this is what’s going to be the next sort of leves. This is how we’re gonna be able to improve this. I don’t think from the actual technological advance perspective, I don’t think there’s that much uncertainty about what progress can and will happen.

I think the uncertainty is how is society going to engage in this? What regulations are gonna come in? What are we actually gonna see in terms of the societal impact of this? That’s really hard and really challenging.

But from the underlying technological advances, it’s pretty obvious. Just as in many ways, for those who were behind the scenes working away at it, the progress that has been made over the last, over 40 years since the chess match. Well, these advances, they are just slow and steady.

It isn’t something where suddenly out of nowhere, the large language models emerged. No, this was slow and steady progress by lots of people making progress on the underlying mathematics and systems on this, and that is ongoing.

What will the next levels of advance lead to there? Well, I am confident that with the right research and with the right areas behind this, we’re gonna be able to take the large language model areas and get them down into small language models, which is more specialists, and going back into solving specific problems, but in really powerful ways, building on what’s happened already and what’s already there.

Those advances are happening behind the scenes. And so I’m confident that’s going to be there as technologies in the next 10 years, next decade or so. So planning for that, this is just good forward planning. From a technology creator perspective, this is going to exist and how we’re going to have access, what’s going to exist and how, we don’t know exactly but we know enough to be able to build for it.

[00:26:20] Michele: Yeah. I would like at some point, I’m not sure if today we have the time, but I would like to ask you what you think about the consequences that AI may have, positive or negative, for low resource environments. 

[00:26:35] David: This is something where I feel very passionately about this, and it is part of the reasons that I’m really so optimistic about what could happen once we get to these small language models and so on, because the impact for low resource environments, this could be devastatingly bad, where inequality just continues to grow.

Or done right with the right technology serving them, I think this could be an incredible equalizer and force for development in really low resource environments. You know, giving access to information, to services. The only parallel I have for this within education, education technologies, people often sort of state these are going to be great equalizers.

The last great equalizer educational technology I’m aware of is the printing press. I wonder whether we could see educational technologies having an equivalent impact as the printing press on how widespread high quality education could become. If educational quality becomes that widespread in different ways, which is part of the Sustainable Development Goals.

I believe if we could get that widespread educational quality, then actually what we are leading to is a world which will look very different. So this is where I hope equality, not total equality, but the lack of extreme inequalities might become a realistic possibility. An equal society is actually not effective, but extreme inequality is shown to be negative for society. A sensible range of inequality within societies and across societies is what will lead to a much more stable world and global environment. That’s part of what I think these technologies could bring. It’s not to say they will.

[00:28:39] Michele: Yeah.

[00:28:40] David: Anyway. That’s a good place for us to stop because I think that is a longer discussion if we get sucked into that.

[00:28:47] Michele: Yeah. And as always with podcasts, we are just scratching the surface. But this has been a very interesting conversation. We could have follow ups if we wanted. Thank you David again for the conversation and for the opportunity, for supporting me with the STACK assistant, sharing the ideas, it’s been a blessing.

[00:29:07] David: No, it is great, and it is really great to see what you’re doing on that. I look forward to seeing where it goes. Thank you.