In this episode, Gabriel (Founder & CEO) speaks with Nirzar (Duck.ai lead) about how we’re making AI more useful by letting users choose the tone and length of AI responses.
Disclaimers: (1) The audio, video (above), and transcript (below) are unedited and may contain minor inaccuracies or transcription errors. (2) This website is operated by Substack. This is their privacy policy.
Gabriel Hello again, welcome to Duck Tales, where we go behind the scenes at DuckDuckGo and talk about things going on inside the company and features for building, cetera, et We’re gonna talk AI again. Today I have Nirzar with me. Nirzar, you wanna introduce yourself?
NirzarYeah, sure. the designer for Duck.ai. I kind of lead the product. And yeah, we’ve been working together. Yeah, mean, yeah, it’s actually been super nice to have like that product here. And yeah, I’ve been kind of working on Duck.ai for last two years. Actually, we started like doing the MVP together, Gabe, you and I, we were kind of playing around.
Gabriel You’re more than a designer. You’re more than a designer at this point.Yes, indeed. So Duck.ai is our chat equivalent, AI chat equivalent. ⁓ It’s private chat. You can access popular models from within it. What are the model providers we have now? Here’s our.
Nirzar Right now we are offering GPT4 Mini which is our default. We also started offering GPT5 Mini which is the newest model from OpenAI. Actually a lot of people are using it a lot more than we expected. But we also like focusing on open source models obviously. So the OpenAI is open source model, Llama and Mistral as well and Cloud Antropiq. So yeah, it kind of fits into our idea about just giving a lot of choice to the users, yeah.
Gabriel Model choice, all the major providers. Yeah, okay. So my quick spiel on our AI approach that I gave last time, but for anyone new. Approach to AI, private, useful, optional, private. In this case, know, it applies to all our AI features, but in this case, you know, we anonymize chats. We don’t train on data. We have bunch of other privacy features in there. Do you want to hit on a couple?
Nirzar Hmm.Yeah mainly I think the storing the chats on your local devices, I think a big one. Yeah. Yeah. And I think the like the biggest sort of thing that we’ve been pursuing is also having like retention of chats and like not having any retention in most cases, which is actually like most industry standard. ⁓
Gabriel Yes, if you have recent chats, they’re gonna be stored on your local device, not on DuckDuckGo servers.
Nirzar And we’re also looking into some private inference stuff, but we’ll get to that later.
Gabriel Great, mainly anonymous and not training. Useful, we’re gonna get to that in a second, because we’re gonna talk about customization, which I think is a super useful feature that we added recently. And then optional, just a couple words on that. ⁓ All our AI features are optional, including Duck.ai. I mean, obviously you can choose not to use it if you don’t wanna use it in general. In our integrations into our search engine and browser, we have settings that will turn it off.
Gabriel So there’s no integration of the search engine if you don’t want it. No integration of the browser if you don’t want the entry points. Although we do think you should check it out because we do think it’s useful in private. ⁓ But we understand people who don’t want to do that for various reasons.
Nirzar Yep. If you’re gonna use it somewhere else, it’s better to use it here if you care about privacy.
Gabriel That’s a good way to put it. Yeah. Okay. So back to useful customization. We’re here today to talk about feature that we’ve been working on. You want to introduce it, maybe share screen.
Nirzar Yeah, I can go through it. Yeah, no, I just like remember where it came from when you asked me to talk about it and I remember you were kind of annoyed at like use of emoji and responses and also like how big the responses are. I sometimes like it but I gotta get that and I think what you mentioned was like if like I don’t like it, I’m pretty sure a bunch of people don’t like it as well. So we kind of talked about like, hey.
Gabriel Yes, I’m very Gen X and I don’t like emojis. ⁓Yes.
Nirzar Like I think what kind of we concluded was just the idea that like there is not like a single personality that we can land on that will like kind of suit everyone. And we always try to give choice to users. ⁓ So I think this kind of fits into the choice and control obviously. And this kind of fits into that ⁓ category. Let me just give you a demo quickly of how it works.
Gabriel Yeah, while you’re doing that,
where I was coming from with that too, was like chatting, this whole feature of chatting is very obviously conversational. like you’re talking to somebody, know, that we’re personifying the AI in this case. I mean, there are people you like to talk to and there are people you don’t like to talk to, there are people you like their texts and there are people you don’t like their texts. Here you can control that. And that would be the idea is to give users control about like what kind of responses. If you’re going to be chatting with this thing a lot.
NirzarYeah.
Yeah
Gabriel Like, what do you want it to sound like, you know? I think that’s kind of the idea for me. And not everybody would choose the same thing.
Nirzar Yeah, it’s very personal. think this personality thing is very kind of difficult to nail down on anything. In those cases, it’s just better to sort of give that ⁓ to each user. They can decide what they want. So I’ll just show one quick thing. By the way, I think a lot of this was like, it’s reusing a lot of these tools. like ChatGPT also has something similar, but I think what you and like what we wanted to do is just like putting it like friend and center. So basically like this is a very small example. Don’t take greetings to chatbots are any popular anymore, but let’s say you are, no?
Gabriel I’m not seeing it yet, Nirzar. I got a black screen, so, nah, weird. Try to reshare. It worked before. We tested this. Try it one more time. it doesn’t work, I’ll try mine. there you go. Yeah, it works now. Go for it.
Nirzar Yeah, we tried that. I would be surprised. All right, perfect. ⁓ Yeah, so we also wanted to like kind of put it front and center. And this is one example of just like saying.
Gabriel I think that’s super important because it’s like it’s a front and center thing. you’re going to, this is a very important thing to change the personality. So it’s not hidden behind settings. That’s a big design departure.
NirzarYeah, yeah, yeah, I don’t think like anybody else had this like this prominent and it kind of like incentives to use it as well. Try it out, see how it works. If you don’t like it, that’s fine. ⁓ So yeah, I think the personality thing that we were talking about was just like, like the tonality, like use of emojis, for example, but like, as you can see, this is like a base model. It’s like Claud, Sonnet 4 and if you ask it,
How’s it going? It’s like, it’s nice to connect with you. How are things on your end? This is like not a very good way to like, these words just like make me a little bit irk. But yeah, like I mean, this is like if you just customize it and this is what we came up with to begin with, which is like just the idea of like customizing the tone or like how long the response should be or like naming is also pretty good.
Nirzar Like if I just ask it to call me by name. By the way, all of this is kind of stored on your local device to protect your privacy again.
Gabriel Yeah, actually, one thing we didn’t say is you don’t even need an account to use Duck.ai. You can just show up at Duck.ai and start doing this without anything. And you can use pseudonyms too, you know.
Nirzaryeah. Actually, you know what?
Yeah, but you know what like when we didn’t think about it when we like made Duck.ai without logins But like when we released it like the biggest positive thing was like people were like, I don’t have to sign up It’s like a huge deal apparently but like I don’t think we thought about it that much but it’s really good I think that that’s why it worked really well. Anyway, sorry coming back to this I digress ⁓
Gabriel That’s what we would call a strategy credit for being a privacy company since we don’t have accounts. We didn’t even have an ability to log in.
Nirzar Hahaha Sorry, we..Sorry, I meant we thoroughly thought about it and it was a great decision. ⁓ Okay, so I’m gonna ask it to call me by my new real name and I’m gonna say my tone, I want it to be like a little bit more playful, as I said connect with you. This is pretty simple. If I apply it, it’s going to store it on your thing. there is a much great I like it. ⁓ Anyways, but instantly just like such like just a very different response to the same model. And this is what we mean by like, like even a little bit of instructions can like make the most out of it. ⁓ This just made me like want to talk to it. I don’t know. Just calling it by your name or your nickname.
GabrielAnd we, and like you mainly, but you like built in these options, right? Like you, like the ones that you suggested there.
Nirzar Yeah. Yeah, so I think we worked a lot on these. I really like the ducky one that you wrote, these instructions which are pretty cool. Kind of pretends to be a duck, which is fun. And actually I use it a lot in like work stuff, because it kind of adds a little bit of these like...
Gabriel Yeah.It was more like a throwaway idea, but yeah.
Nirzar⁓ But yeah, just battling around digital port. It’s kind of stupid, but I like it. But yeah, I think we worked on this. The thing that we worked on the most was like this AI roles. ⁓ I think like it was actually kind of fun to write a bunch of these.
I particularly like Chef one, cause I cook a lot and I think I used all my cooking knowledge to kind of inform to like all write all the instructions for Chef, like with the templates and stuff. like if you ask for a recipe, it’s gonna like give you a very specific template. The other thing I really like here is the instructions part. And this is kind of like goes to transparency thing. ⁓ We show like whatever we tell to. what instructions we provide to Duck.ai. All of this is transparent. You can see exactly what instructions are getting passed onto the model, which is kind of nice. But yeah, I think it seems very simple on the front, but there’s a lot of complexity behind it running in the background. And I think that was kind of the goal for designers as well, to make it super simple and doing all this work in the background.
Gabriel Sweet. All right, you can stop sharing your screen if you want. Let’s drill down to a couple things though there. So we were talking about the tone and response, which I think is critical if you’re gonna talk to something all day and like it. But then the AI roles that you mentioned briefly, like the chef, that is kind of a separate thing. Because now we’re saying, hey, independent of how you wanna talk to me, could you act like a teacher or a chef or a weightlifting assistant or we have like 20 of them or something. And as I understand it, like, you you especially with the team went and, you know, found some good instructions to help the AI become that role. And if you actually want to do that kind of thing and jump into that role, it’s just like, you get a lot better response.
Nirzar Yeah. Yeah. Yeah.
You get a lot more. I mean, we did a lot of testing on like how much it differs and how good it is at following the instructions. ⁓ I think one thing that I would say is I think the hardest part was to like make, because we offer so many models, like it’s so difficult. Like if you ask the same instruction to Mistral versus to Claude, it’s like Mistral won’t do something certain times and Claude will do it. And then you just have to like figure out how to like do the prompt engineering.
So that was kind of challenging, but I think we got, I don’t know, we got like pretty good set of instruction that are like very, like we’re trying to get to like more predictable sort of responses where we can, you know, assume it’s gonna follow these instructions.
GabrielGot it. And then you also showed, I think I wanted to highlight is the system prompt. the, you, using an industry term here, but the, user prompt is what you’re typing in. The system prompt is what we’re talking about changing here, which is like the instructions that were given to the AI that apply to all the responses. But we’re actually showing you what the system prompt looks like in terms of these changes, how it applies. And you could, there was another box there. You can.
Gabriel You can change, you can add your own to, the end of it. So like if you, to your point, if you’re switching models and there’s something about the model you like or you don’t like, that particular model, you could add to it.
Nirzar Yep. Yeah, I mean, it’s already such a black box about like how things are working behind the screen. Like, do you want to add one more layer of like, okay, there is something else happening on top of what the model is doing. like, think giving that much at least like transparency is helpful in my opinion. Like, you just don’t know what’s happening behind the screens.
Gabriel Agreed. Okay, so what is next? What are we working on next with this, if we can say?
Nirzar Yeah, so I’ve been kind of noticing how I’m using it. Actually, I haven’t talked about this a bit with you before this conversation, but the way I’ve been using it is like, I think I’m like trying to figure out like there are kind of two different ways I’ve been using it. One is sort of this tone and like things that I kind of generally looking for in any of the conversations I’m doing versus these like tasks.
So like I do a lot of like these repeatable things. So like an example is like I’m learning how to code better. I’m a designer, but like I’m doing that. And like, I have this like set of instruction that I always kind of append to a blog of code so it can explain what the code is doing to me as a designer. And this is like a repeatable task. So like there are some of these customizations that are like good for particular conversations. And some of the instructions are like particular for overall sort of like always on your entire Duck.ai experience. And I think we’ve got to find out like how we can like make that distinction better and make the interface better so that we can kind of support both of these like doing the task like over and over again every single time versus ⁓ something that kind of like exists in everything. ⁓ Can I share my screen? I can show you.
Gabriel Yeah, I was kind of seeing it semi blurry. So it’s like, it’s possible Adam’s going to make us record this whole thing again. But so we may just want to like talk about it briefly and then share after, but instead, but I would say that I haven’t talked to you about this yet, but I completely agree with everything you said. in fact, I was coming out from a different angle, which is
GabrielActually two different angles one is I have prompts that I want to use all the time remember repeatedly and right now I have to copy and paste them and There really seems like there should be a way to like have that bank saved, you know and then independently the roles like I find myself not using them as much as I actually want to or should because It seems like it’s mixing the what I want to be constant as my system prompt for tone and this one-off task for a role. So it does seem like I know I advocate for combining them into one interface, but we probably should split them at some point, maybe in some of this and why don’t we talk after and then we’ll maybe we’ll have another episode about it. Any, yeah, cause we’re at like, I don’t know how long we’re supposed to go, but I think we’re, we exceeded the 15 minute mark on this.
Nirzar Yeah. Hahaha ⁓ Yeah? Yeah, sure sure.Okay.
Gabriel Okay, closing thoughts, Nirzar. ⁓ What about usage response feedback? Has it been positive?
Nirzar Feedback has been good. mean, you know, a bunch of people already like were asking for it but like tech savvy people obviously like ask for it more so they’re gonna be like, can...is like I don’t like emojis and like you have to kind of figure out where that balance is so people who are asking for it obviously happy about it.
And others I think like yeah like I just pulled up some numbers. 20 % people use tone which is like the highest usage of any of the customizations. The next one is the 16 % people use the role of the chatbot and then everything else is kind of like long tail after that. But yeah overall it’s been nice like 6 to 7 % of like all the Duck.ai chats right now happening are using some sort of customization which is like again a bit higher than I expected because we kind of did this yeah we kind of did this like throw away thing to begin with to kind of see the appetite and I think a lot of people are using it so I think it’s it’s worth sort of like spending more time on this to make it better
Gabriel Yeah, it’s higher than I expected too. Yeah. Cool, okay, so I realize I didn’t construct a closing line for these episodes yet. That I’m gonna have to do. So, here’s our, yeah. Thank you for coming. Thank you everyone for listening and until next time.
Nirzar Let’s chop. Yeah.
This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit insideduckduckgo.substack.com
