Median User

[00:00:00] Ramsey: Hey everyone, and welcome back to another episode of Median User, the Humans Guide to the Age of AI.

[00:00:06] Ramsey: In this episode, we’re going to be covering something that is probably close to a lot of our hearts, and that is ChatGPT, but more so how can we get the most out of it? A lot of us use ChatGPT in our daily lives, and I know that me and AK definitely do, and we know a lot of other people who do as well, but people always seem to misunderstand the extent to which ChatGPT can actually help them and how they can really make the most out of the platform.

[00:00:31] AK: Yeah, I think it’s an incredible unlock to get good with the product of ChatGPT, all the features that it has. Things like memory, custom instructions, voice mode. I swear by this shortcut on the Mac, Option-Space, that you can set up with it.

[00:00:47] AK: Because what that lets me do is use it anywhere, anytime, one click away. And so these are the things that, for me, actually make a huge difference in my use of AI. And that took me from a maybe once a day, few times a week kind of user to, uh, 30 times a day type, you know, uh, ChatGPT user. But a lot of the conversation online, you know, and, uh, discourse, the news centers more around the models the product features to use them. And I think it’s of course, important to know which model you’re using, what it’s good at, and what we’ll cover that in a second. But I find it actually personally, the feature set like matters more to me than the models.

[00:01:30] AK: That’s why I, I don’t end up switching from ChatGPT to Claude or Grok or Gemini, or whichever the latest, hottest model is of the day. If it doesn’t have like the full feature set, the shortcuts that I’m used to, I’m not gonna make the switch unless it’s a very specific purpose.

[00:01:46] Ramsey: So let me ask you something as an like a regular user of ChatGPT or someone who might just be using it as a daily driver to get questions answered, does it really matter that much, which model that person is using, or does it matter more how that person uses, whichever model is the default setting for them?

[00:02:03] Ahmad: So there’s a little bit of a change for this section of this episode. We were originally going to go into model selection and talking about four oh versus oh three and reasoning models and how they work and why you choose different ones. But since the episode was first recorded, ChatGPT has now obviated all of those with ChatGPT.

[00:02:27] Ahmad: So now you almost never have to think about model selection anymore. GPT-5 is a great all around model and it automatically selects when it should turn on its thinking or reasoning mode based on the prompts that you give it. In my own experience, I’ve found this works really well 90% of the time, and I never need to select any specific model or mode anymore. Of course, you might think about Claude versus GPT-5 versus Gemini versus other models and what they each might be good at, but at least within ChatGPT model selection is not really needed anymore. Just go with whatever it does by default.

[00:03:10] Ahmad: So back to our episode now.

[00:03:13] Ramsey: So just to branch off a little bit, you mentioned reasoning and obviously a lot of the I guess hype around AI is because it can reason like a human or that’s the goal, right? When we get to AGI, it can think for itself. Now, there was some news fairly recently around how someone had debunked the reasoning side of ChatGPT, I believe, or was it something else.

[00:03:35] Ramsey: But as someone basically debunked the, the reasoning behind a model, how do you think that affects our use? You know, does it mean that we actually get worse answers than we think? Or is it not really affecting on our experience using these platforms?

[00:03:52] AK: I mean entering the realm of philosophy here of what is reasoning really, and are we ourselves truly reasoning or are we also just pattern matching? I’m not gonna go deep into that, but I’ll say for…

[00:04:03] Ramsey: Broke the fourth wall, man. How am I meant to think for the rest of the day knowing that my thoughts aren’t mine, shoot.

[00:04:11] AK: Yeah, I mean, I watch my kids grow up doing a lot of pattern matching and that seems to be the dominant way they think, so I wouldn’t be surprised if it goes all the way, all the way up.

[00:04:22] Ramsey: To be honest, I mean, I think about it a lot, like what is intelligence? And it comes to like, you know, in a general sense, and we’re not talking about if you’re smart at art or you know, whatever. But I think it, most of it boils down to pattern matching, right? It’s like you have this context and you understand certain things.

[00:04:39] Ramsey: Can we extrapolate that pattern recognition to other things? And that makes you seem smart. Like the smartest people I know wouldn’t call themselves smart. I think they would probably just say, look, we have like fairly good common sense and we know how to pattern match. And I think it makes sense, right?

[00:04:58] Ramsey: If the models are literally just pattern matching and calling it reasoning, I don’t think it’s that far of a stretch.

[00:05:03] AK: Yeah, I think there are some good definitions of what true reasoning outside of pattern matching is and what a good reasoning problem can look like. But also the study that you’re talking about that debunked that these models reason was a bit of a joke study. But I think people latched onto it because they want to negate what’s happening, you know, they want to say, oh no, no, nevermind. This isn’t really as far as you think it is. And for me, I don’t really care one way or another. What I care about is, is this useful for me? Can it do a data analysis task, which I myself only loosely know how to do and would’ve taken me an hour to do? Yes. And I can verify the results in other ways. Like if it does the data analysis and I can just check that whatever it came up with, if I just check a few of the data points I see, yeah, okay, it looks pretty correct to me. You know, I’ve done enough work with these reasoning models that I find like, wow, they’re really good.

[00:06:08] AK: They help me. They get my work done faster. They help me do things that I’m not so good at doing. I don’t care what’s like the philosophy underneath that surface. You know, if it helps, it helps.

[00:06:19] Ramsey: Yeah, no, I think that’s the right way to look at it, right? You can’t let these, the thinking behind why this might or might not work affect you actually using the thing and getting the most value out of it as possible.

[00:06:30] AK: Exactly.

[00:06:30] Ramsey: Like that. I guess that’s a good segue. Right? So we have these two models. You have the GPT-4 daily driver that you might use. You have another model, which might be good for certain tasks, but at the end of the day, how are we meant to make the most out of each one? Are there any ways that we’re meant to prompt these models or speak to them or any features that we might be missing that are actually leaving money on the table, per se, in terms of improving our productivity or opening insights that we might not have had before? What are your, starting with the prompting side, your kind of prompting hacks for anyone who’s looking to make the most out of ChatGPT?

[00:07:09] AK: Yeah, prompting is this fun game that you build an intuition for, and most of the time I think you don’t need to think about it too hard. You can just say whatever you would say in pretty natural language. But the more complex the problem, the longer the output—if you’re trying to come up with, for example, like a proposal to a major client—you definitely want to be a very good prompt with all of the context that the model needs, you know, formatted in a particular way. I’d say as a general principle, it’s learning to be very specific about your request, the content, your output format, your goals, and what you don’t want it to do. If you think about it like talking to an intern who only has a loose grasp of the area you’re interested in, the more specific you can be with your intern, the better they’re gonna perform. And so I think that’s like just principle number one. The harder the job is, the more specific you have to be.

[00:08:56] AK: I do a lot of context dumping. This is also where ChatGPT’s memory feature can come handy because over time it gets to know things about me, like where I live, where I work, my role at work, the team dynamics at work, in my family. It’s like, you know, my wife’s allergies and my kids’ dietary needs or the kind of food that I like to eat—like high protein, generally pretty low calorie. And so these types of things that accrue over time so that that context is building without me needing to put it in every single prompt. Then I can be a little looser. I can just say, hey, working on this thing for work and here’s what I need next. And I know it has enough to go on. But if you’re new to using ChatGPT and you haven’t given it all that context yet, I would spell it out like, here’s my role, here’s my team, here’s what the company does, here’s our website, here’s, you know, here’s a bunch of stuff to help you do the best job you can.

[00:09:10] Ramsey: Yeah, you mentioned two things there that are pretty interesting. I think the first one I want to talk about is negative prompting, which you said obviously, you know, tell it what you want it to do and also tell it what you don’t want it to do. Now I think a lot of people, you know, write this so they don’t understand how the model thinks, and no one’s expecting you to understand how it picks out information from its huge cloud of brains and provides you an answer—like, it’s a very complicated thing. But what you can do to help yourself get an answer that is most aligned with what you are looking for is to really give it negative guidelines.

[00:09:43] Ramsey: So for example, if I was to write a blog about the best dog food, right? That doesn’t know if I’m a vet or I’m a dog chef or whatever it might be, but if you give it guidelines and say, hey, look, I’m looking for an SEO-focused blog. I want to target these certain customers. I don’t want to target these certain customers, et cetera, then you’re providing not only a positive reinforcement, but a negative reinforcement for it to stay clear of certain things, and that will align it more with what you want from your end output.

[00:10:12] Ramsey: Important thing to do is to affirm things with it, right? Affirm what you’re really looking for and affirm what you’re really not looking for. And when it comes to the memory, one thing that I found really helpful recently is that if I have a specific task—for example, in my ChatGPT UI I have multiple tasks and chats open for different things—for example, creating YouTube video scripts. Nowadays I don’t have to tell it we’re creating a YouTube video script. I’ve just said, hey look, here’s an outline.

[00:10:38] Ramsey: Create the video script and it has all the context from our past scripts to work on and to align that script with the ones previous. But one thing that I found helpful is if you’re gonna create like a new chat for a new task, the first line of whatever you write in that chat should say, hey look, this is a chat for—and then insert task. So that way it always has a strict context that, okay, whatever we’re doing in this chat should be aligned with this.

[00:11:03] Ramsey: Because sometimes, as you get much further down the line and there’s a lot of context, sometimes it loses core things that’s what you’re looking for in the chat, and might not fully understand the true nature of that chat. They might just kind of think you are asking for random things. So if you can align off the bat and say, hey look, in this chat we’re going to be focusing on doing X, Y, Z, then you can just provide bits of context and it already has the overarching theme of what this chat is about, to create outputs that are more aligned with what you are looking for anyway. So again, nowadays in my YouTube video scripting, I don’t have to say, hey, this is meant to be YouTube scripts. I just say, look, hey, here’s some information. Okay, great. YouTube video script is done, and it’s very, very simple.

[00:11:35] Ramsey: So one thing that this also leads on to is leading the model. And what I mean by this is, let’s say I, again, I’m writing an article or I’m writing a script for a YouTube video, and I’m unsure based on the context of the video that ChatGPT will get it right the first time—maybe because there’s a lot of information out there, maybe because there’s not a lot of information out there on a certain topic—what I might do is lead it to action something else beforehand. For example, hey, can you please write a blog on the best food to make for a dog in Siberia? Okay? Can you please research X, Y, Z different articles or X, Y, Z different search queries before you go ahead and write the article. And normally I’ll actually say, before you write anything, show me your research so then I can vet it. So then you’re leading the model just like you would with an intern that you have. You’re just making sure that it’s in line with, again, the output that you are looking for versus having the model jump ahead and just do something that you don’t really want. And I’ve had this many times where it’s like, hey, we wanna do this task and you just tell it what task you wanna do and then suddenly—I’m sure, AK, you’ve had the same—it just does it. And you’re like, no, no, wait, we haven’t started yet. Just let me know you understand and then we can move ahead. So, leading the model and really making sure that it understands when to the task or what task it you really wanted to execute in that moment is really, really important. It will save you a ton of time and many a headache when you are using these platforms.

[00:13:14] AK: Yeah. I have such an interesting example of having done this recently. I was working on, appropriately enough, an LLM evaluation setup for a healthcare scribe feature that we use in our platform. So the feature itself is recording encounters with patients and doctors and then coming up with the chart notes for the doctor afterwards. And so it takes a good bit of intelligence that we need to also be very accurate on—like no room for error, obviously medical context. And so we were setting up an LLM evaluation, so you have one LLM evaluating the output of the actual feature LLM. And in order to do that, I needed to learn a lot about this general topic of LLM evaluation.

[00:14:02] AK: And so I started reading and finding all these big, like deep, well-researched articles, you know, like 10-page articles, full breakdowns of how to do this well today. And it’s an evolving topic. It’s probably changing every week. There’s very much a feeling for me of just overwhelm with the amount of information I had to process to come up with a good plan. And so then what I did is I found like the top two, three articles, long, well-researched articles, into ChatGPT, and said, read these articles and knowing the goal of like what I’m trying to do at work, come up with the plan that I can share with my engineering team. And it was just an incredible result because then you’re getting a human-written, well-researched, like, you know, up-to-date article fed into ChatGPT as its input. Then I don’t—I mean, I read the article too—but I don’t then need to do all the mental processing of how does that translate to us? Which tool does that mean we need to use? It’s just like processing all of that, coming up with, okay, in your case, use this tool, use that thing. This is what your engineers need to know. This is what you need to know. And I was just like, wow, this is amazing. I passed it to them. I do this thing a lot where I don’t say that this was an AI-generated output, just to kind of see does this feel right to you? And they’re like, yeah, this looks great.

[00:15:24] AK: We’ll get started. I was like, hey, just so you know, like this was ChatGPT-generated. They’re like, oh, cool. Like, it’s good, you know? So yeah, priming it with these things—and I find that topic of priming is generally how to get great results. So feeding it articles, feeding it research is one way. Have you ever played much with role assignment or telling it you are a world-class marketer who’s an expert at writing blog posts about dogs and then starting the rest of your task from there?

[00:15:59] Ramsey: Yeah, I used to do that early days when it seemed to find it hard to find the context to which I was asking for a task. So I would provide it a role so it would know, okay, based on all these trillions of parameters of information I have, I should only listen to those pertaining to this task or this role.

[00:16:20] Ramsey: So I would say, hey, look, you are a creative writer, or you are an SEO specialist who’s very well-versed in helping companies rank for Google search terms. And then I would go ahead and write my prompt. I don’t think I ever did it in the sense of, hey, you are a [insert industry professional] just because maybe like the nuance of where I work—like a professional expert in blockchain and crypto—like there’s not very many like real experts in this technology.

[00:16:52] Ramsey: And a lot of people who say they are are just absolutely full of shit. No offense, sorry for swearing, excuse my French, you know, whatever. And so like every time I did, because I did try it at the beginning and I was like, this is utterly rubbish. Like this is not any sort of information I would wanna have in an article. So I would just like fact check it afterwards. But the role playing in the sense of, hey, you are a writer or you are X, Y, Z, or you are very good at insert whatever you are looking for—that did help a lot.

[00:17:20] Ramsey: I haven’t used it as much. I just haven’t felt like I had to. And I think obviously with what we mentioned before around the ChatGPT—like the platform understanding and having memory of our previous history—it’s, oh, I sound like I’m dating it. But having its memory, like, you know, being full of what we’ve already done, it makes it a lot easier. It already has the context. It doesn’t need to really dig too deep to understand what I’m looking for because it has so much already.

[00:17:46] AK: So there’s one technique that I’ve recently come up with that I find to be like the next-gen version of this role prompting. It’s something I’m calling like the expert insights primer prompt. And it’s a simple thing. I kind of almost discovered it by accident, but it’s so powerful for getting a better result when you need it to think through an expert’s lens on a specific topic, especially if you are not an expert in that topic. It’s hard for you to prompt it as if you were the expert.

[00:18:15] AK: For me, for example, like product design, I can describe exactly what I’m looking for, use the words that help clue it into what kind of analysis I’m trying to get or what kind of design ideation I’m trying to get. But if I was to try to say, write a marketing blog post, I wouldn’t know how to write that prompt in a way that it’s gonna get a great result. I would just be like, make it really good. You’re an expert, but I don’t know what good looks like. I don’t know how to describe good. And so one thing I’ve come up with is that I first start the conversation—the chat—with “What does every great marketer know about what they do that is not obvious to an outsider?” You can fill in whatever job title or whatever sort of action expert role instead of marketer there.

[00:19:20] AK: So what does every great consultant know? What does every great chef know? What does every great real estate agent know about what they do that is not obvious to an outsider? And then what this prompt does is ChatGPT usually returns a list of about eight to ten expert insights from that person’s role. And so the first time I used this was for a consultant. I was interested in doing some AI consulting. I spoke to a friend of mine who’s been a consultant his whole career at some of the major companies. So I spent like an hour on the phone with him just getting his full ten years of experience distilled into one hour. What is the real heart of consulting?

[00:19:59] AK: And there were insights in there I would’ve never thought of—things like, you know, 50% of the job is you’re actually just giving validation to the client to do what they want to do anyway. Like if they need to lay off a team or they need to restructure the company, they’ve already had those ideas, but they didn’t want to be the one responsible for them. And if they say, “McKinsey told me to do it,” that’s actually what they’re paying for. It’s not the idea, it’s the validation. I was like, that’s crazy. Like they really just pay you that much just to have someone to blame basically.

[00:20:15] AK: And so after this conversation, I wondered, would ChatGPT have come up with the same insights? And so I used this prompt for the first time: “What does every great consultant know about what they do that is not obvious to an outsider?” And it was like almost all of the same insights that my friend gave me from ten years of work. It was just an amazing list. And I was like, wow, like this is really a great way to get it to think as an expert from this area. Then where I took this one step further, that’s been amazing, is if you start the chat with that question—you don’t even care about the answer—anything you want to do afterwards, write a proposal, edit this email, draft this, change that, anything you want to do from there, it’ll now do it with those expert insights clued in, primed for that expertise. And I think it works so much better than if you just say, write this proposal as if you are an amazing consultant. It doesn’t do the same thing. But if you first sort of ask it for those specific insights, it writes it with those—it’s incredible from there.

[00:21:29] Ramsey: It’s really weird, you know, like it’s almost like you’ve told it to go to the library and pick out a book. And then once it’s read the book, it can never do anything without information from that book. And I think that’s a nice way to think about how you talk to any model—is that the way that the information is stored and structured doesn’t have to be understood in a deep level by the end user.

[00:21:53] Ramsey: But if you just think about it like a huge library of information and prompts you give it, it’s gonna go and find the right book on that prompt and read it. Then it will use that book to continually iterate on whatever you are going to provide it next. So as I mentioned, if you are doing, again, YouTube video scripts, if I primed my message just by saying, hey, we’re creating a YouTube video script, it’s just gonna go and find the most basic book or piece of information on how to create a YouTube video, and then it’ll create your output from that.

[00:22:24] Ramsey: But if I said, what are the most non-obvious things that the top 100 YouTubers do in their videos? Then it will go and look for information around that topic, which will be a lot deeper, a lot more precise, and something that, again, much more non-obvious and something that’ll be likely higher quality.

[00:22:42] AK: Definitely.

[00:22:42] Ramsey: It will page that into every output it provides after that. So I think that’s a nice way to look at it—it’s like it has all the information, but you also need to guide it to help it pick out which information it should use for your outputs.

[00:22:58] AK: So one more technique to throw on top of that, and this is getting a little advanced, but some people might find it really useful. And that’s meta prompting, which is when if you need to write a prompt for something quite complicated. Let’s say it is that create a project proposal for this major client for this specific project. And the meta prompt in this case is instead of asking for that directly, you ask for the prompt that you would then use with the model. So you say, hey, ChatGPT, I need the best possible prompt for a proposal for this major client for this. So what this does, like, especially for non-experts who aren’t using AI every day, all day for work and everything, but if you’re just sort of arriving at ChatGPT, you’re not very good at prompting.

[00:23:49] AK: Or sometimes for me it’s like I just don’t have the time to think of the best prompt. With all of those guidelines we just shared, one thing that works really well is the primer prompt. So what does every great job title know about what they do that is not obvious to an outsider? By now using that, using those insights you just came up with, write me the best possible prompt to create a project proposal for a consulting firm or for a client. And one thing that can really help you do is quickly, and you don’t have to read any of the answers. Your ultimate goal is the proposal. You don’t need to read the steps along the way, but what you’ve done is you’ve kind of just made ChatGPT do the work for you of adding the context, being very specific, doing those do’s and don’ts. It knows that at this point it knows how to write a great prompt because the amount of in, you know, information about prompting is now widely available on the internet, so you’re kind of like cheating through everything we just described without having to actually know how to do it yourself.

[00:24:56] Ramsey: Yeah, you get like really good advice like that. And there’s other ways of making the LLM do what you want by threatening it, which also works. So you can pick poison, really meta prompt, or threaten, which.

[00:25:09] AK: Oh yeah. I mean, emotional manipulation works really well.

[00:25:13] Ramsey: You know, I think I was using Cursor to code something and I’m basically, I’m coding a game. It’s just a game that me and my friends play, but it’s like a pretty unknown game and there’s no online version for it. So I was trying to build it and it was getting something wrong. And I was like, if you don’t fix this now I’m gonna use something else.

[00:25:33] Ramsey: And they’re like, oh, okay. And then they fixed it. And I was like, there’s like no way. There’s no way. This is real. Like we’ve done this 10 times, you probably waited so many of my credits, but only until I threatened you with the jumping ship. Are you gonna fix it? By the way, that is actually a real way that you can prompt ChatGPT if you threaten the model and say, hey look, I don’t like this output.

[00:25:56] Ramsey: If you do another bad output, I’m gonna go away and use another model. It actually should provide you, yeah, a weirdly good output.

[00:26:06] AK: Do I need to get Claude or Gemini involved here? Come on, man.

[00:26:09] Ramsey: But yeah, I may next one, like, don’t let me tell Sam about this. You know, get the boss involved.

[00:26:17] AK: Yeah, yeah. Being encouraging also works. I think there’s actually some hard research to show that being very encouraging, like being like, you’re really great. Thank you so much. That was amazing. Being all kinds of like buttering up, buttering it up seems to work. So you can go either way.

[00:26:37] AK: You can either sweeten the prompt or go very threatening and either way, I guess it just raises its stakes or mimics human behavior in some sense.

[00:26:47] Ramsey: You gotta play good cop and bad cop. I mean, I would say thank you to GPT for many other reasons. I mean J-G-B-T, I love you. If you can hear me through my Siri.

[00:26:55] AK: I love you, but if you don’t get this right, I’m leaving you.

[00:26:58] Ramsey: Yeah, exactly. It’s like dating, you know, you gotta keep them on their… That’s also not dating advice. So, yeah. But I think even with the positive affirmations, something that is nice and that can lean from that is actually asking it to ask you questions, which is something that a lot of people don’t do.

[00:27:16] Ramsey: So, for example, if you have a more complicated task and you feel that, okay, this might be quite nuanced and specific and maybe a bit complicated, just ask it, hey, before we start anything, do you have any questions about this task for me? It’ll ask you and then you answer it, and then it will go ahead and I, that really yields much better results.

[00:27:33] AK: Yep, definitely. I mean, it’s so important for me that I put it in my custom instructions, which is one of those killer features that I think everyone needs to know about. For me there’s this sort of like trinity of special features I described at the beginning, it’s custom instructions, memory, the option space shortcut on Mac. I don’t know what it is on Windows, but having those three things set up for me has like 10x’d my usage of ChatGPT, and just my productivity with it overall.

[00:28:06] AK: So for those who don’t know, I want to run through this really quickly, ‘cause I think if you’re not using this today, you absolutely need to be, no matter what you use ChatGPT for.

[00:28:15] AK: So custom instructions is a field in the settings where you can tell ChatGPT how you want it to behave, how you want it to answer in general, in every chat. And so for me, I started having the inspiration to do this just by following what people on Twitter were doing. I came across something that Eigenlayer on Twitter wrote—a custom prompt which is now called the EigenPrompt. It is a really good starting point for anyone, I think. And what it does is tell it things like, be very terse, don’t use many formalities, don’t qualify your statements. If there’s a policy being broken that you can’t answer or be specific, what policy is broken and why, however smart you’re being right now, act as if you’re two standard deviations smarter. Millennial slang, not boomer slang. Throw in some Gen Z slang occasionally, if these kind of things that just change the tone that makes, by itself, just a huge difference for me.

[00:29:17] AK: And then what I’ve done is I’ve started to add in my own custom instructions on top of that. And one of them is like ask follow-up questions wherever appropriate. And that can be either follow-up questions because you need more context to answer my request as best you can, or follow-up question that I could then use to ask you like where to go from here.

[00:29:39] AK: So if we start by saying, hey, write me this email, or actually, sorry, let me say, tell me about how the Coliseum was built, and it might say, okay, here’s a bunch of information. And then instead of me having to think about what to ask next, it could say, do you want to know more about the time that it was ransacked and then rebuilt? Or do you want to know more about the other buildings in Rome? Or do you want to know about the stone quarrying method they used? And so it kind of gives you a few directions to go on whatever it is. And for me, that’s now built into my custom instructions. So are you a custom instructions fan like me? ‘Cause I could go on all day, but what’s in yours?

[00:30:22] Ramsey: There’s one feature that I just love and that’s voice mode. I just sit there and talk to ChatGPT. It sounds ridiculous, but sometimes when I have a conundrum that I can’t articulate very well in words or I just don’t, I can’t be typing because I’m doing something else.

[00:30:41] Ramsey: It’s just so easy to click that feature and have a conversation with ChatGPT. The only thing I have to remember is telling it to write down what we’re talking about, ‘cause sometimes it’ll forget to do that.

[00:30:52] Ramsey: It’s just so easy. I can just say like, hey look, I’ve got an issue. I’m trying to come up with a new offer. I’m not sure how to structure the pricing. Can we do some research to figure out what the best type of pricing model would be?

[00:31:02] Ramsey: Boom, I’ve said it in 15 seconds. I don’t have to type any words. I just sit there and wait for an answer and then I obviously ask it to write it down for me so I don’t have to go ahead and remember it or recall it later.

[00:31:14] Ramsey: And I think it’s just so powerful. Like this is, I think, where the future of it all heads when you’re thinking about wearable AI merch or accessories.

[00:31:26] Ramsey: For example, there’s companies out there who are building this hardware. I forget their names. I should probably get them for the next one. Might be interesting to cover, but, you know, stuff like earrings or headphones or necklaces where you’re gonna be talking to the AI through those devices, and it’s gonna be picking up on what’s going on around you.

[00:31:43] Ramsey: You’re just gonna have a copilot with you the whole time. And at the moment, in the phone is perfectly fine, right? I can just open the ChatGPT app, I can click go, and I have the whole model ready for me to converse with, which is really, really powerful. That’s probably the one I use the absolute most.

[00:31:58] Ramsey: I’m not too much into the, I would call it more like finicky prompts and hacks. I think this is like really more than enough for what I need.

[00:32:12] AK: Yeah, I like voice mode a lot as well, but I actually, I use the voice input and then I get the written info back ‘cause I don’t like to wait for the full voice response.

[00:32:24] Ramsey: Yeah, I think there’s something in the conversation though, because you can have real conversations with it. It’s not just a case of question, answer, question, answer. Like, I really like that aspect and I think how it’s developing because I use it a fair bit, so I notice when things change and it’s coming along very nicely and I think it’s really nice when you have a situation that you might not be able to talk to many people about, maybe because there’s no one who really understands the situation or it’s not something a little bit more, I’m not gonna say intimate or personal, but just something that you don’t want to deal with other people and you just wanna kind of deal with it yourself. It’s great.

[00:32:57] Ramsey: And I’m talking mainly about work related things, right? For example, maybe research tasks that I don’t have time to sit there and write out, and I just need to plan on how to research something or when I was developing this game for the first time. I just told it the rules. I was like, hey look, this is how it’s played. This happens, this happens, this happens, and it’s a lot faster than me typing out, alright, so when you put down a queen, the next person gets like, just let me say it, it takes three seconds as opposed to the many more that takes to type it out. And then it will ask questions back like, oh, okay, so when this happens, what happens after that?

[00:33:32] Ramsey: And I can just kind of spit through it and it just gobbles up all the context and we can then work with that information afterwards. It’s just so much easier. And I think down the line, obviously there’ll be ways where this transcribes into your thoughts and your thought waves as well.

[00:33:47] Ramsey: Like that’s something that I think a lot of people are working towards is using like electrical pulses to figure out what people are thinking and then transcribing that into, because that’s already a thing, there’s people who are developing like, mice, like computer mice that were governed by what you were thinking about on the screen and stuff like that.

[00:34:03] Ramsey: So I think that’s where it will go. And I think it’s a good practice to kind of talk to the AI instead at this point, especially if you are someone who is using AI to maximize productivity. And I mean, I’m a bit scared that people will lose their critical thinking muscle.

[00:34:22] Ramsey: But if you’re there to just like max out productivity, I don’t understand why you’re not talking to the AI. Like you can be typing and doing other things on your laptop and talking at the same time. You can get twice as much done realistically. And yeah, okay, it’s hard to think about two things at the same time, but if you just consistently switch every second, it’s fine.

[00:34:40] Ramsey: Trust me, it works.

[00:34:41] AK: I’m awful at that.

[00:34:42] Ramsey: You get used to it. I think, I mean, I think I was just doing it in school since I was young ‘cause I just didn’t wanna work. So I would pretend to listen and then do something else and pretend to listen and do something else, like the whole lesson. So I’m just used to it. But if you put it into practice and you really put the time into just getting used to talking to it, it frees up a lot of time for your, you know, your hands and, uh.

[00:35:04] AK: Your hands and something else. So I know my wife uses voice mode for kind of more personal conversations, almost like a diary where that talks back. And I think this, she finds it really useful just to like process. If she’s got this big wave of emotion about something, something really tough’s coming up for her. Do you use it like that at all yourself? Like personal, just outside of work?

[00:35:32] Ramsey: Nah. Be honest, I’m not the best person when it comes to like modeling how to deal with like mental health stuff on. Not that I’m like bad in my brain, I just like, I’m very good at just compartmentalizing and just moving on and dealing with it. So I don’t think I need it for that.

[00:35:45] Ramsey: It’s more for if I need to maybe, I don’t wanna say outsource brains or outsource computational power, but more so like, okay, I’ve got a problem and I’ve thought about it and I dunno how else to think about it.

[00:35:59] Ramsey: Let me see if ChatGPT can offer a different perspective or extrapolate my thinking into another way or just provide like another door that I can walk through, and then I can go down that path myself. That’s kind of how I see it, because at the end of the day, again, if we’re talking about reasoning and wisdom, there’s many ways to think about things and ultimately the models have been trained on so much data that they should have different ways of thinking about different problems.

[00:36:25] Ramsey: And so I think it’s beneficial to see how they think about things and if there’s anything that you can learn from their way of thinking about it, because I only know one way to think about things, and that’s the way that I think about things. And it could be a, and it sounds like actually kind of stupid when I say it out loud, like the way that anyone thinks about anything could be a mix of loads of different things, but you don’t know what you don’t know.

[00:36:44] Ramsey: And if you can get another perspective that might fill that gap or maybe color in another part of the picture for you, I think it’s useful to have that information. And that’s basically why I would use it.

[00:36:57] AK: Interesting. I should really try it more ‘cause I think I got a little impatient with it, and I just like to kind of have my answers brewing in the background while I’m doing other things, but I think I would like to give it another go. One thing I’ve held off from that I’ve been scared about is it with my children, who they’re two and a half years old, twins, and I was like really thinking sometimes, like it’d be interesting to try them on this to be able to just talk to an AI, but I don’t know if I think this is a good age yet for it. I mean, they don’t use much technology otherwise. Like they watch a few kids shows here and there, but like, what if they could just ask questions? What if when they’re really talkative? I was just like, here’s ChatGPT. But I don’t, I…

[00:37:42] Ramsey: Uh, it’s dangerous. I think that’s like, you know, there’s, there’s one-way doors and there’s two-way doors. I think that is most definitely a one-way door. Like, oh man. I think that’s like my biggest fear is that I don’t want to instigate behavior, whether it’s my future kids or my friends or people I work with.

[00:38:01] Ramsey: Like I’m all for using AI. Don’t get me wrong, like we’re doing this for a reason, right. But I’m very conscious that if you rely on it too much, there is a level of degradation of your mental capacity that happens. And this is not a case of theory, like we’ve seen it in other aspects of technology where you have attention spans have shrunk so much in the last 10 years because of.

[00:38:23] AK: Definitely.

[00:38:23] Ramsey: Instant gratification through content. Doom scrolling, and whoa, whoa, whoa. Like, we’re victims of it. Even my dad, I remember when I was like 12 and I got my first BlackBerry and he was like, this phone’s gonna ruin your life. And then I caught him last week sitting on the sofa, like scrolling through TikTok and I’m like, that phone’s gonna ruin your life, man. He just kind of laughed, but it’s true. I think this stuff rewires you, and I just don’t want that to happen too early to kids, especially before they really develop their own sense of how to think.

[00:38:53] AK: Yeah, agreed. I think I’ll wait a little while. It’s just tempting because they’re so curious right now. Like, everything’s a “why.” Why is the moon up? Why is that cloud small? And I’m like, there’s only so many whys I can handle in a day. It’d be nice to offload some of those, but yeah, I don’t want them learning to rely on it before they learn how to reason for themselves.

[00:39:15] Ramsey: Yeah, exactly. Because even the curiosity part, it’s good when you don’t get an answer sometimes. It makes you dig deeper, right? If you just get answers instantly, you kind of lose that drive to investigate or to connect dots.

[00:39:29] AK: Exactly.

[00:39:31] Ramsey: Speaking of digging, one other thing I want to talk about is the projects feature in ChatGPT. Because this is another one that, like, no one really talks about enough.

[00:39:42] AK: Oh, that’s one of my favorites, man. Projects and Docs. And the secret chats, to a smaller extent.

[00:39:49] AK: Projects is how I keep my different workstreams alive. So I’ve got one for my job, one for the podcast, one for meal planning, and so on. Each one’s like a little persistent space. The coolest part is when you go back later and say, “Process this episode,” and it already knows what that means. It knows we transcribe, clean, write show notes, make the cover art, all of that. So it’s not like starting over every time. It just keeps that workflow memory.

[00:40:22] Ramsey: Yeah, it’s super nice. And I think once you’ve set them up properly, you can actually treat it like an extension of your brain. Like, I’ll be out somewhere, think of something for a client or an idea for an episode, open the app, drop a note in the project chat, and it’s just there waiting for me when I sit down later.

[00:40:43] AK: Yeah, totally. And then Docs takes that even further. Because instead of like losing track of what’s in the chat or scrolling up a hundred messages, you can just have a live document that we both edit while the model’s still in the loop. That’s been huge for like proposals or longer scripts.

[00:41:03] Ramsey: Yeah, I’ve been using Docs a little bit too. Sometimes it gives me like weird formatting at first, but overall it’s so much more useful than trying to edit in the chat window.

[00:41:14] AK: Yeah. And I like that you can move seamlessly between chat and doc—brainstorm here, polish there. I think this is the direction everything’s going. ChatGPT’s product is maturing fast.

[00:41:29] Ramsey: For sure. And the whole GPT Store and Agents stuff too.

[00:41:35] AK: Yeah, like prompts becoming apps. I’m convinced that’s where it’s going. For most jobs, a well-designed prompt inside ChatGPT beats building a standalone UI. Like, why have ten apps when one agent can do them all?

[00:41:53] Ramsey: Yeah, especially when you think about the friction cost of switching tools all the time. If I can just say “open my design GPT” and it knows my style, my fonts, my client preferences, that’s it. That’s my app.

[00:42:08] AK: Exactly. And the multi-format support just keeps growing. Like, soon I want to be able to hand it a spreadsheet and say, “Make this chart, clean the headers, export as CSV.” No syntax, no formulas, just natural language.

[00:42:23] Ramsey: Yeah, 100 percent. Because right now, even with things like Excel Copilot or Google’s Duet AI, you’re still kind of chatting about the data instead of editing it. I just want to say, “Hey, use column A to generate this,” and it does it.

[00:42:41] AK: Yeah, exactly. I want the manipulation, not the commentary.

[00:42:45] Ramsey: It’s funny how that loops back to what you said at the start—that features matter more than models.

[00:42:52] AK: Yeah, exactly. You can swap models all day, but if you don’t have that workflow and those affordances—custom instructions, memory, projects—you’re starting from zero every time.

[00:43:03] Ramsey: So I guess for anyone listening who’s like, “Okay, this all sounds cool, but where do I start?” what would you say are the first two or three things to turn on or try?

[00:43:14] AK: Custom Instructions, number one. Voice Mode, number two. Those alone will 10x your experience. Then, when you’re ready, add Projects so that you stop having to remind it who you are every time. That’s really all you need to feel the shift.

[00:43:32] Ramsey: Yeah, and then after that, start experimenting with things like image generation, diagrams, or custom GPTs.

[00:43:40] AK: Totally. Those are fun and powerful too. I’d say start small—get ChatGPT embedded into your workflow little by little, and before long, you won’t want to work without it.

[00:43:52] Ramsey: Yeah, exactly.

[00:43:54] AK: And if all else fails, threaten it.

[00:43:56] Ramsey: Yeah, yeah. “Do this or I’m switching to Claude.”

[00:44:00] AK: That’s right. And then say, “Just kidding,” at the end so it doesn’t hold a grudge.

[00:44:04] Ramsey: Alright, that’s it for today’s episode.

[00:44:06] AK: Hope you enjoyed this one.

[00:44:08] Ramsey: Make sure you follow, rate us on Spotify or Apple, and we’ll see you in the next one.

[00:44:13] AK: Thanks everyone.

[00:44:15] Ramsey: Later.

[00:44:17] AK: Bye.