Zohar 

That's a good question. That's a good, very good question, and it's a pretty good point. I think at its core, if someone doesn't want to actually learn something, there's not much you can do about it. They're going to be able to use AI to just completely mitigate it because at its core, all of our AI methods are, at the moment, reactionary. They're not proactive. You're not predicting, oh, so this new AI is going to come out, therefore, we can take these steps to counteract it before it ever comes out. This thing has already happened. Here's how we can circumvent it. And reactionary kinds of defensive mechanisms never work. So like you're saying, you can't really do anything about it, but what you can do is make it significantly more appealing to slow down, to take the time, to add friction into the process. Also, it's possible to just force it. Say from an educator standpoint, it's possible to just create assignments that explicitly have friction that can't really be solved with AI like we’ve been talking about.  

 

Then the second part of the AI conundrum is something that was very well established, but foundational knowledge needs to exist before you use AI. There's a paper that kind of builds on this. It's called The Memory Paradox (A/N: Zohar meant to cite The Extended Hollow Mind), and it argues AI can only really effectively build on knowledge if you have preexisting knowledge for the AI to build off of, which seems like a very obvious statement until you think about it for a moment. What this basically requires is for a user, before they even get introduced to the concept of AI, to be applicable to a certain situation, to have the basic foundations. It's like-- for a five-year-old, you're introducing them to addition before you introduce them to multiplication. So if they want to use AI to help with multiplication, they can because they have the basic steps of addition down. You need the kind of-- Essentially, you need the kind of foundational framework but for everything to build off of before you can use AI. And an integration of those two kinds of mechanisms is probably the ideal way to go for future development in education. Questions? 

 

Ali 

Yeah. So Zohar, you mentioned about first building out our foundational knowledge, which is really essential as we grow in our career or even in our education, in our school, high school, or be it university  

 

Zohar 

Mm-hmm.  

 

Ali 

... et cetera. But then if I-- if in this age of AI, I, I can just ask a question to AI that could build my knowledge, even if it's not really building it, it's at least giving me the answer to it in less than three seconds or even less.  

 

Zohar 

Yeah. [laughs]  

 

Ali 

Then what is the point of even gaining that foundation knowledge without AI in the first place when you can just ask it? And wouldn't it be more important to learn how to use that AI rather than building that knowledge by yourself?  

 

Zohar 

That's a good question, and I do-- Don't get me wrong. I do think it's important to know how to use AI effectively. I think that's one of the most important things that you can do if you want to use AI. But at the same time-You're not always going to have the time to use AI. A lot of times in a professional context, you're going to be in a little bit of a rush.  

 

Ali 

Mm. 

 

Zohar 

Maybe, I don't know, you have a project that came in two hours ago. It's gonna be six hours of work, and it's due in two hours. You only have four hours to do it. It's six hours of work. You're going to have to go up to an AI, ask it, "Okay, can you do these six hours of work in four hours?" Because a lot of the times when it comes to AI, effective AI use specifically, most of it revolves around asking it for revisions, giving it revisions, making it take what it takes, and then working on it. And oftentimes that might take even longer. It's fine in the normal circumstance because if you have the foundational framework, you know what's wrong. You can look at it, see, oh, this is wrong, fix it really quickly. If you don't have that foundational framework, you can look at an AI. Maybe it's giving you something completely right. Maybe it's giving you something that is abhorrently wrong, and anyone with a sane understanding of the groundwork would know it. But because you don't have that groundwork, you can't fix it. So I guess what I'm trying to say is the groundwork's important because you need to know when the AI is wrong [laughs], and it's wrong a lot. There's a recent study that came out that shows a lot of more advanced models of AI are significantly more prone to hallucinations, so you have to work around it, I guess.  

 

Ali 

Mm-hmm. But for example, if you ask a really simple question to AI, then obviously it's not gonna hallucinate. It's gonna give you the correct answer in most cases. Because, for example, I saw a sixth grader asking a multiplication question to AI, and as you all know, multiplication is quite easy.  

 

Zohar 

[laughs]  

 

Ali 

AI won't give a wrong answer to that.  

 

Zohar 

Yeah.  

 

Ali 

Then what happened in that case? Wouldn't those sixth graders or even seventh graders just keep on asking AI to do their homeworks?  

 

Zohar 

Oh, yeah. No. Oh, yeah, no. But the thing is, the sixth grader knows multiplication. By the time you're in sixth grade, you have a good understanding of what multiplication is. You're... Isn't sixth grade when you start the introduction, the very basic rudiments of algebra and stuff like that? So--  

 

Alfonso 

Yeah.  

 

Zohar 

... you have a good understanding of what the-- you're asking the AI to do. So when you're-- a sixth grader's asking an AI for multiplication, it knows what it's talking about. It looks at this and it's like, "Okay." The sixth grader is like, "Okay, I know what this is. The AI's given me this right result. It's fine." Then when the AI gives them a wrong result, they’re like, "Wait, that's not right. Maybe I should think about this more." So when the sixth grader moves on to stuff that they’re not 100% sure about- 

 

Ali 

Mm-hm?  

 

Zohar 

... it has the... The sixth grader has that kind of awareness that AI has the potential to be wrong. It has the potential to be wrong, and so the sixth grader is going to build up their own fundamental knowledge alongside using AI, which allows them to, like I said, build up the knowledge to expand their network of information.  

 

Alfonso 

That point just made me think about a question. Would you think there's a good idea to break the use of AI into levels? So depending on your level of education, you should approach AI in different paths. And that suggests that we should as a global community, create different frameworks to approach AI based on our levels of knowledge.  

 

Zohar  

That's a really good idea [laughs]. Holy shit. That's a really good idea.  

 

Alfonso  

Yeah, just a thought that came up out of my mind.  

 

Zohar  

Yeah, no.  

 

Alfonso  

And you mentioned the point of the, like  

 

Zohar  

I can see the  

 

Alfonso  

... the multiplication  

 

Zohar  

... that's a really good idea.  

 

Alfonso  

Yeah.  

 

Zohar  

Like, the,... I'm just thinking about that. That's hu- that's huge because, you can have no AI sections for introductions, like kindergarten. There's grade one, grade two, grade three, grade four. Elementary school, basically. Once you get into middle school stuff, you can have the very basics concept explaining version. You can have it... You can have AI for kids [laughs] .  

 

Alfonso  

Yeah.  

 

Zohar  

You know, you, you have, like, YouTube kids, ChatGPT kids [laughs] . 

 

Alfonso 

Oh, that's amazing. Imagine Google integrates this into Gemini and before you start using the chat, you can select the level of  

 

Zohar  

What, yeah, what age range  

 

Alfonso  

... of grasp you have on this topic, and--  

 

Zohar 

Mm.  

 

Ali 

Yeah. 

 

Alfonso  

... the AI will break the--  

 

Zohar  

Adapt to it. That's actually a really good idea.  

 

Ali  

Yeah.  

 

Alfonso  

Yeah, that's a good idea.  

 

Zohar  

And then I guess the issue there is that some people might say they might overestimate or underestimate their own abilities, but you can't really prevent that. But that's a human error thing.  

Alfonso  

Okay. Before going to the depth of my topic, I want to start with the-- going back to the question Zohar asked me a couple of minutes ago about how we ensure the correct use of AI into students going to the approach of the Architect. So a lot of students get into trouble [laughs] when using AI. If we just give students access to ChatGPT and say "Good luck," it will actually cause serious harm to their education. So a major 2024 study or research named [unclear] points this topic out. They found that because AI chatbots give us answers so instantly, they completely disrupt our normal learning process. If we use it as an Oracle our independent skills and our ability to struggle through a problem and find the answer will actually deteriorate over time, so we are going to lose our academic muscle. So to safely make AI an Architect, we have to use it for something called metacognitive support. I know that sounds like a very heavy academic psychology term, but it really just means think about how you think [laughs]. So it’s a great 2025 study, and the research team goes deep into this concept. They found that if an AI is programmed to ask the student questions rather than giving answers, everything changes. Imagine you have a huge essay due. Instead of asking AI to write it, you tell the AI your topic, and you prompt the AI to act as your coach. The AI asks you, "What's your main argument? What are your three main points? What is your timeline to finish this by Friday?" Something like that. And this study points to metacognitive support. This system of AI will force students to reflect on their own strategy, so the AI becomes a sounding board. It does not do the work for you. It forces you to organize your own thoughts. And this builds up on another study. In this context of problem-based learning, they found that AI is incredibly powerful. When it helps you build blueprints, let's say you have a giant confusing math problem or a complex research question, you use AI to break it, that one massive problem, down into five small, easy-to-manage steps. The literature calls this a temporary bridge. The AI builds the roadmap, but you still have to drive the car. You are doing the actual learning, but the AI removes the initial panic, that cognitive overload I mentioned earlier, so you can actually get into the work with confidence.  

 

Zohar 

So, Alfonso, let's kind of put this into practice for a moment. Say we have a pretty complex project, like a super heavy Python script, like Assignment 2, or a really difficult math proof. How exactly would I use AI as an architect in that kind of scenario?  

 

Alfonso 

Good question. So there are actually two great papers to address this. One is called Leveraging Large Language Models to Enhance Self-Regulated Learning in Programming Education. They found that instead of asking AI to write the final code, students use it to outline the logical structure of the algorithm before they type a single line. Another study of Self-Regulated Learning with AI in Prompt-Based Learning showed the exact same thing for general problem-solving. You use the AI to build a blueprint, but you have to build the house yourself because our brain wants a clear map of the things we want to do. And if we just read the first sentence of the assignment and say, "Build a complex system that decrypts information and helps to build security in the application." And you will think, "Where do I start?" So the AI will just tell you, "Take this problem, separate it into small steps," and then you by yourself need to go through each step. And this will even increase the learning outcomes and also the quality of the final product because imagine the difference between when you start a project and you already know what's your final goal. Because sometimes we start assignments, and we are like forty percent of the [way through the] assignment, and then we notice that, "Oh, I should not go through this path. Maybe I need to go back some steps and then take this other way." So if you already know the final output… So using AI for this makes you think about that, you will take each step of the assignment focusing on the final output, and that will increase the quality of the final output, let's say.  

 

Ali  

Alfonso, if we accept the idea that you just presented, then how do our university, University of Toronto Mississauga, how can this university's professors change their assignments or their quizzes tomorrow to require this kind of AI use?  

 

Alfonso  

Nice question. So to answer this, I just wanna first say that my answer is not the final solution to this problem. I think that there are other solutions, but I would just like to bring to the table an idea I had. So professors need to grade the process, not just the product. So there's a brilliant paper called Teaching Project Management with Generative AI: A Pedagogical Model for Responsive and Sustainable Practice. So professors can literally assign a project where step one is to submit an AI-generated workflow and timeline for how you will complete this assignment. It will explicitly teach students project management and force them to use the AI to organize the tool. So I have seen this even in some of my courses in university in these semesters. So the professor will post the assignment, and we will be required to not directly use AI, but to build a blueprint of how we're going to break the assignment into steps. And in this step, some students may use AI, some will not use AI, but it helps to break the assignment into steps and start learning from the basics. For example, I think that in our computer science class, they do this by the worksheets of the assignments they give us in the classes, the lectures. And even though it's not saying directly to break the assignment into steps, they are teaching us a small part of the assignment, even the final output, like how we're going to... how the system will work even before we start the system coding. Because it comes down to the same idea of having a clear vision and view and image of the final product. It will help you in the building process of the actual assignment.  

 

Ali  

But here's the thing. Do you think that the profs should mandate the use of AI to do- to complete these assignments? Because there are a lot of other students who for certain legitimate or illegitimate reasons don't want to use AI even if it's recommended in those assignments.  

 

Alfonso  

So when it comes to recommending the use of AI or not, I always point to the framework of the learning path or the learning skills that each student has. So I think that the most important thing that we need to impose on students is the motivation, discipline, and willingness to learn.When we have that point clear, then we will start analyzing the different path the students want to take to achieve the assignment. So I will not say that we need to impose the use of AI. Like in the CS class, they are not telling us to use AI. They are just giving off the worksheet, and we'll need to think about the final process. And in our classes, they like to tell us to break the assignment into smaller steps without telling us to use AI. So then it comes for the student on how to use AI. But then, as we were mentioning earlier, [there’s] the different levels of knowledge that each person has. So I know some students that have come to university with a huge math background. For example, they have gone to some tournaments, Olympiads of mathematics, and they normally like [laughs] ... I don't know if this is correct to say here, but they skip some readings or they don't pay attention to this specific thing.  

 

Zohar  

Or they go take 157. [laughs]  

 

Alfonso  

Or they go to take harder courses. And you may say, "But what-- is this correct?" And then you see that they have a higher score than the student, they read all the readings and didn't use ChatGPT. And it comes down to the background of the student. So we cannot really have a specific workflow of, okay, each student is going to use AI like this and if not like this, and they need to break into steps and then make this ex- th-this exact prompt. So I think that we need to teach based on the level of understanding of the student, but there's also a baseline of if you don't want to use AI, at least break the assignment into steps by yourself. And if you want to use AI, you can go for it.  

 

Ali 

Scaffolding. 

 

Zohar 

Assume there's a base level of competence and work there.  

 

Alfonso 

Yeah, of course.  

 

Ali 

Exactly. That's a great answer.  

Zohar 

Yeah. No, it's a really good answer. I feel like, kinda makes me think that there's an idea that I have that if you don't want to use AI, you don't have to, but you should be willing to use any other tool that you can get your hands on to succeed. It's a hard world out there. You gotta try your hardest to get somewhere in it, I guess is one of, is one of my thoughts.  

​

Alfonso 

Yeah, of course. AI is a technology that has come to the world to stay. I don't think that AI will disappear in the next decades.  

 

Zohar  

It's like Pa- it's like Pandora's box. You can't put everything back in the box.  

 

Alfonso 

Yeah. Yeah. Like all the revolutions we have, internet, the computers, transportation are like--  

 

Ali 

Mm-hmm. 

 

Alfonso 

... we enter a new environment, a new society, a new reality. So we need to learn how to adapt AI to our today's learning AI outcomes. 

 

Ali 

Exactly, exactly.  

 

Alfonso 

So I think that the correct path is to have these conversations and discuss these types of ideas.  

 

Zohar 

Mm-hmm.  

 

Alfonso 

So a-actually my, my point was built up on the idea that I think about the last revolutions we have through history, a lot like the printers, then we have the internet, and then we have globalization of all the knowledge. So many thousands of years ago, the problem was that you didn't have access to the information. Nowadays, on the internet, you have access to a lot of information, but you need to select the correct information. And now the problem is you can select the information because AI can filter all the information, but then you need to ensure that the AI has the correct information, and then you need to also ensure that the AI is not making you not think about other possible solutions or limiting your creativity process to come up with some solutions. So with this problem, trying to make connections between each revolution through history, I came up with the idea of how we can take a different approach to AI, not just the Oracle, but like an Architect that can help us increase our learning outcomes.  

 

Ali 

Let's expand on what Alfonso talked about on using AI as an architect rather than an oracle, while also taking the risks into account that Zohar mentioned. So on one hand, it's like having the AI as a specialist at your own fingertips. Those AIs use a process called natural language processing, NLP in short form, that analyzes the user's previous writing or coding styles and tries to imitate that style in their specific voice or tone, and using these same types of vocabulary in the outputs of the AI. And it allows the AI to act like a ghostwriter, a really fast ghostwriter, not like the traditional ghostwriter. It can literally replicate your style within not seconds, but literally milliseconds. But here's the thing, the research that I did emphasizes one of the biggest cons of this really fast ghostwriter. They show that even the AI-generated responses from models like Text-DaVinci-002, as mentioned by Becker et al. Those AI often, those AI often admit the lack of understanding of their AI-generated outputs by the end users as those users just simply take in the output without learning even a single thing. Now, if the AI just gives us the answer blindly, we feel like we've learned the concept. But in reality, the thing is that we've actually taken out the real mental heavy lifting out. We feel like we have done the work, but it's the AI who has done it. It's an illusion of achievement… It's what Michael Kölling calls the illusion of achievement. He states that the learning only happens when you actually struggle with the problem. In the case of most undisciplined students, they literally just put in their prompts into AI and get the output without learning anything, and it stops them. It literally kills their struggle and stops them from building a long-term retention.  

 

Now, before I brought this into computer science, I had to ask myself, are we really building a bridge to understanding or just a crutch that we'll eventually trip over? In computer science, and even in many other fields and courses, this is really prominent. But in computer science, the problem is even more intense. I mentioned it before that we have all had the Sunday night problem where the-- we are stuck on a complex logic error, and we're stuck on it for the past couple of hours. We don't have any support available, so we only have two options, either to try-- either to cry or ask ChatGPT. Most of us just end up using ChatGPT or any other AI. But what starts as a fix for a single bug, that, that thing, that action triggers a cascade of automated help by the AI for all of our other questions of the assignment, for example. And it erodes our own critical thinking. It prevents us from using our own brainpower. It lets the AI do the entire work. It lets the AI act like a ghostwriter. If those students who use AI at that point, at that moment, then as Wermlinger speculated, those students may spend a massive amount of time on actually customizing their prompts so that they get the perfect solution, the perfect output from AI, rather than spending their own time efficiently in producing the solution to the problem all by themselves. Hence, this is where the idea of which I call the Socratic AI Tutor comes in. Socratic refers to the Greek philosopher, Socrates, in which the Socratic term means that it involves a sort of discussion between one person and the other. In this case, it's the AI and the end user.  

 

So rather than a cheat code that we use to write the code, for example, we are looking at a form of AI, a customized AI that acts like a human teaching assistant. It's like-- It's more human. In a study conducted by Deng et al, [clears throat] Deng et al. in twenty twenty-five, they assessed the learning efficiency of several students, and they tested their learning efficiency by giving them three types of AI, namely proactive AI, which suggests [...] code or strats without being asked to do so. The second one is passive, which only responds when the user asks the questions. And the third one, the last one, is collaborative AI. Collaborative AI is a kind that lets the user interact with it in a more natural back-and-forth conversation. This is where the bot asks the question at the right moment to let-- to guide the user towards the solution rather than blindly giving out the solution in one go. It-- This happens in several outputs, not in one single output. And this is where the learning happens. But the thing is, this requires a form of customizing our prompts that we give to AI. We have to change our old way of talking with AI. We have to change our prompts. Now, if the students that are using their old prompts, if they customize their prompts now, then they can prevent what Lin et al observed in their own study. They call it the copy, run, error, fix loop in the context of computer science, in the context of programming. What this loop means is that when a student gets an assignment question, they put it into ChatGPT, for example. They get an output, they copy the entire code and paste it into their interpreter or where the code runs, the console. They run it. They encounter any so- any, any form of errors. They go to ChatGPT again. They try to fix it. They copy the new generated code. They run it. They get an error. They fix it. So this loop happens again and again. Copy, run, error, fix. Copy, run, error, fix. It goes on and on and on in an endless loop. Now, if the AI intervenes at just the right time, for example, during the error phase, and it prompts with a question rather than a solution, then it would keep the students in what the researchers call the zone of proximal development. It gives these students or the users just enough of a small push or just enough [of a] nudge to actually think about the problem and let the s- the students themselves do the actual heavy, heavy lifting of the critical thinking, for example.Consequently, the AI that gives the formative feedback of the answers of the questions-- of the AI's questions, that formative feedback can form a loop which could eliminate any misunderstandings or misconceptions of the students of the particular concept, and which would allow to, [clears throat] allow to help build the foundational knowledge, as what Zohar mentioned. And this type of Architect way of using AI, this repetitive process of AI that you-- that acts like a feed-- that gives a formative feedback in a loop, it could guide the student towards the final solution in several outputs, just like I mentioned, instead of just one.  

 

Another thing I want to highlight is a really telling visual I found in Lin et al's study. In Figure 1, they map out certain conflict levels between student practices and instructor norms and how the instructors respond to the use of AI by students in different areas. For example, assignment questions, quiz solving, writing ideation, project generation, or project management, et cetera. When you look at the figure, you see clear conflict, [clears throat] clear conflict in things like quiz solving or solving an assignment question. This is where almost all the professors rejected the idea of using AI, as it eliminates all of the students' own input in answering the questions all by themselves. It undermines their foundational capabilities. And that makes total sense. The profs don't just want us to copy the code just for a grade. The next category is called the partial conflict area, which happens in project implementation, for example, where you're implementing a project. Instructors in that area are actually more tolerant here because they realize that LLMs cannot really solve complex coding tasks which could be related to machine learning that involves training the AI, for example, or handling complex files yet. AI is still developing, it's-- it cannot yet solve those problems. Comparatively, the third category is called the broad agreement. This is the most tolerant area of student intents and practices. This con-- this area consists of AI usage for, for example, writing, writing revisions or writing ideations or even just creating summaries using AI. All instructors in this area tolerate this type of AI assistance as-- and they also encourage the use of AI for these types of works. In this moment, they can eliminate the-- they can save the time that it takes to summarize those things or even brainstorm before writing. But they distinguish this from the actual writing generation using AI. And even if they tolerate this generating AI summaries or brainstorming, they emphasize a lot on proofreading and selectively taking out stuff from the AI's output rather than just copying word for word or verbatim.  

 

But this brings us to the ethics of bias. It's not just about cheating. The ACM Code of Ethics reminds us that we have to avoid harm. And I found that AI tutors can actually create socioeconomic gaps. If you look at the best AI tutors or AI chatbots out there, the ones that don't hallucinate as much, the ones who are really confident and the ones who give out correct, accurate outputs. If these types of chatbots have a version of GPT-4, for example, then would their students only get the high quality support? Or would other types of people or students be able to get that type of support?  

 

Zohar 

So did you find any LLMs that happen to reduce the socioeconomic gap you mentioned?  

 

Ali 

Yeah, most certainly. There are two open source LLMs that I found, namely Llama and Mistral, that run on particularly modest hardwares, [and] they are open source. So open source means that the developer can customize the models based on their own alternative creative methods of solving the exact same problem, and it mitigates the problem of the socioeconomic part of the AIs.  

 

So as I was saying, would the AI that is based on GPT-4 create a two-tiered education system where only the wealthier students get the high quality support? There's another point that Becker et al. state that OpenAI's Codex is trained on more than fifty million GitHub repositories. They are trained on more than fifty million codes. That includes the vast majority of GitHub Python code, for example, and it totals a whopping one fifty-nine gigabytes, [clears throat] gigabytes of data.  

 

Overall, to emphasize once again, one solution for the problem I mentioned before is creating certain prompts to stop the AI from giving away the entire solution in just one go. Mollick and Mollick describe this as a recipe for an AI tutor. They suggest that we must customize our prompts so that the AI never gives away the answer, but instead breaks the problem down into more simpler parts and gives-- and asks the questions to the users, and guides them to the solution at the end.  

 

There was another similar clever solution I found by the CodeAid study at the University of Toronto, our university. They designed a tool that only outputs the pseudocode. Pseudocode is a code that only uses English statements to describe the logic of a coding problem. It only gives you the logic, but it actually forces you, the end user, to write the actual syntax. This technique of using the pseudocode moves the student from being just a consumer of AI to actually developing the code mostly by themselves, which is a skill that Denny et al. mention and argue is more important now than ever before.  

 

Alfonso  

So here I have a point. You mentioned that CodeAid study where the AI only gives pseudocode. But honestly, if I'm lazy, can’t I just copy the pseudocode into another AI and say, "Write this into Python"? Isn't your solution just a two-step cheat code?  

 

Ali 

Well, that's a great question. So the thing is, at the end of the day, it's all up to the student to be honest in their work. If they really want to take something from what they are learning in university, for example, then they need to be honest in the way they use AI. And if they have a purpose other than just receiving that piece of paper at the end of their degree, then they must actually put in the right amount of effort at the right amount-- at the right time in learning those concepts rather than just simply using AI. Just like how Zohar mentioned, unethical or continuous use of AI may cause us to lose our brain cells or even stop us from building the foundational knowledge that is required in our courses. And there's also another thing that taking help right now may seem okay for those people, for those students, even us. But at some point in the semester, every single student would have to take the final exam where they are alone with a piece of paper and a pen, and there's no, literally no support available. There's no chatbots. It's literally you and the exam paper. At that moment, you realize the consequences of use-- of overusing AI.  

 

Now, ultimately, the research that I conducted suggested that AI tutors should be scaffolded, meaning that the help should fade away as soon as we get better. And if we do it right, AI becomes what Frank Vahid calls the teaching assistant that never judges you. This is because the students are more comfortable to ask stupid or embarrassing questions to the AI. And the word teacher has literally now become more prevalent among students (to describe AI), according to a post-study survey conducted by Ma et al.  

 

Zohar 

You-- Yeah, that's a good point, but isn't judgment or at the very least accountability part of why we learn? If there's no social pressure, do we just lose discipline to finish the work?  

 

Ali 

Well, that's true that we learn when we are criticized, but AI is not as it was before. ChatGPT's default settings used to be that it's really respectable. It literally appreciates every single thing that you do, every single answer that you do, even if you're entirely wrong in what you're saying. You're saying that, for example, elephants grow on a tree, and ChatGPT is like, "Oh yeah, that's right.” It literally just supported you in whatever you're saying. But nowadays, there's a feature that they introduced in ChatGPT, for example, where you can customize the behavior of the AI. You can put in a customized prompt, to adjust the behavior. For example, “Be a ruthless mentor, admit my trash ideas whenever possible, or criticize where necessary”, so that the AI doesn't always support you in everything and can actually act like a human teaching assistant by criticizing you more.  

 

Zohar 

All right. [gentle music] Thank you so much for listening to our little podcast. Hope you guys enjoyed it, hope you learned something, and hope you all came out with a little bit of a new perspective on generative AI. Thank you very much, and we'll see you next time. 

 

Alfonso 

See you next time. 

 

Ali 

See you next time.