The Case for Less

We have spent most of our time, most of our decades, beautiful decades, teaching in computers how to give us more, how to give us more recommendations, more suggestions, more... We tried to train it in such a way that it has to provide us with more importance, more things to do, right?

How about not talk about more today, let's talk about less.

what if more isn't always helpful because sometimes when you are when you have given someone more in abundance that might just give them a confusion less clarity in their mind it's just like you giving them another decision and another and another until they reach to a point where they don't know exactly where to begin. They're stuck in this loop and where to start, how to do it.

So most of the time, I've always thought about this. I've come across a lot of people who have a lot of things in their plate.

They've taken a lot of things, yet they want to start, they want to work on something and they want to begin somewhere.

It always comes to this choice paralysis. We call it when a lot of people have it, it's just that we don't accept it. So why is this happening?

This is happening because we have made AI to give us more. Like can it generate more images? Great. Can it give us more recommendations? Better.

Can it also think about what I can do next? that's even better

but imagine what if we gave it imagine we give them the true values into constraints

Kindness as a Technical Constraint

this is where I was talking about what happens when we have kindness as technical constraint basically, imagine we hard coded kindness and compassion into a guardrails to a system

that helps us reduce cognitive load onto our head, onto our mind. It gives us better, easy ways to move forward rather than pouring all of this information onto our heads.

So I've tried to build a system that helps people who are suffering, suffering with, who have little bit of neurodivergency, to give them a direction of how you can reduce this stress that is onto your mind when you have too much of things, too many things on your plate.

So what happens when a kindness is a technical constraint? Instead of saying make the experience kinder, what if we had said that system must not overwhelm people with a lot of choices yeah we have something we can actually design

here I think I believe we should when people are we have right now AI can generate and I can create anything you want you just need to know your how how and why but you can create anything you want the thing is it's not about what we are creating, it's about how we are creating it.

So next time when we are building something, why are we not, so we should be talking about is this going to overwhelm someone? Is this going to make them go into like choice paralysis? Is this going to help them cope up fast?

Too Too much information sometimes actually makes people, like you know, it overwhelms them to a point where they don't, they can't take decisions quickly.

Designing for Choice Paralysis

So it's just that I had a friend who I was, he was my classmate in my master's. I had, this idea stuck to me because he has hyper ADHD. So, he had a lot of things in his plate.

By the end of semester, we had to submit about five games, we had to build five games. He had so many things in his plate, he was not able to even move with one, just like, we just had 10 -15 days to submit our projects and he was not even able to move one step forward.

The reason was, when I asked him what was the reason, then he said that it's just that I am choice paralyzed. I have so much of confusion because of the systems that have been built and the information that is readily available for us. Too much information, too many next steps.

And when we can see that behavior, we can design against it. And then we shift our focus to designing for the weakest moment, not the happy path.

But what happens What happens when the person can't choose? What happens when they make too many mistakes or when they want to give up? Does the system make it harder for them or it helps them to cope up from this?

Janus: A Calmer Way Forward

So this is where I built Janus. Janus is a system which follows three steps process. It's a calm three steps process. I will not bore you guys, I will just show you the demo, maybe you will like it.

Hello Shiva, good evening, just after half past six on Tuesday, what can I do for you? Janice, I'm stuck with my inbox. Hey Janice, I'm stuck with my inbox. Let's make this smaller together, just three steps, no rush.

Open your email and pick just one message to read, doesn't matter which one. one. Start here. This one's tiny.

Decide what that one message needs. Reply, delete, or move to a folder. Take a breath and look at the next message whenever you're ready.

When you've done one, say next, and I'll be right here. You can stop any time. Any? Thank you, Janice.

So what you have heard here is this system, what it does is it tries to put less load on you, whenever you head to Janus, it's not going to judge you saying that you have got so many works to do, you've got so many to -do lists, your to -do list is full of a lot of things. No, it's not going to do that.

What it's going to do is, it's going to give you three steps that you can follow by doing one at a time. What it does is it removes this pressure onto your mind that you need to do a lot of things and you have so many things piled up and you are answerable to people.

It makes sure that we are, so that person who is moving forward is not stuck there. We are trying to, why I built this is people don't get stuck in that place where they are, they would rather move from there, movement is everything is what I believe and that's

Building Systems That Know When to Stop

how the Janus works but there are, so what I have actually, why I wanted to do this was I wanted to tell people that whenever we are building something, we are trying, we should be able to answer, instead of building it around a problem or around the capability of what a model can do we should be able to ask questions like who could this

overwhelm what could the system do to feel unkind that feels unkind and what rules prevents it can we test the rules when we do all when we ask these four questions it makes it eventually gradually you have metrics that you have built for yourself, when you're building a product, you're trying to put kindness as their constraint.

That's what happens when... So when we do that, the system itself becomes kind, and the kind system actually knows when to stop, not to say more.

Thank you. Thank you very much, Shiva.

Questions: Implementing Kindness Constraints

Okay, we have about five minutes for questions. Does anyone anyone have any questions for Shiva?

Yes. Sorry, how did you add the constraints? Did you train a model, or how did you add those constraints to this?

So this, I built it, so I have, there's a code I wanted to show, but then I thought that people would get bored of it. So there is a certain level of code that I've hard hard -coded in this.

That's where the kindness constraint comes in. So I've added certain levels of guardrails, certain hard -coded kindness compassion points which it follows regardless of what.

Let's say you went and used prompt into a model or trained it by creating projects or adding empty files, it's still not going to be as, it's some point, at some point it's going to hallucinate. It's going to step away from where it is.

But when you're hard coding, like, you know, kindness as a guardrail, what happens is it's going to be kind even when you are, even if your context is missing.

Hi, Shiva. Thank you for the talk. It was really informative.

Why Prompts Alone Are Not Enough

I have a question about how does your, I guess, model kind of differ to, for example, if I put prompt into one of the other models, like OpenAI, and said, oh, I want to solve this problem, but don't overwhelm me with too many options. Give me one or two options. So how does your model differentiate from that? Thank you.

Yeah, so I've tried it. I've tried ChatGPT. I've I've tried Claude. I've tried to talk to the voice agents that they have. But guess what? Except for ChatGPT, I think the rest of the models, they don't even want to listen to you. They just want to tell whatever they have in their mind.

It's just like you ask him, hey, ChatGPT, do you know where I can get a burger nearby? It's going to tell you 50 restaurants around you. It's going to be like, okay, you're in Farringdon. That's a beautiful place. You should go there, that, do this, that. My God, there's so many things that it tells you. It just floods your mind.

And Claude has not been that good communicator, I believe. I think even if you give it a prompt, I tried that as well. I'm actually, I've given some, I have a takeaway for that. You guys can try it at home. There's a PDF there, a free PDF. You can download it. it, and my Janus is open source, you can actually try it yourself, but you can see the difference.

So the thing is, when you are giving prompts to this system, they don't actually understand you well, because they don't know your context, where you're coming from, right?

Context, Memory, and Personalization

So let's Let's say, since I built it for myself, I've tried to teach it as much as I have, I know about myself, so it works around, so there is something called Hermes agents, if you guys have heard of, there is something called agent harnessing, so in that there is something called memory .md, it's called a memory .md file, which has 2500 characters, that's a a limit for it.

So what you can do is, this memory file, you can actually, so whenever I talk to Janice, I'm talking to it everything. I'm just putting everything out to it. I'm just telling it, you know what, I have to talk to this guy, I have to call him. It catches it up. And then it makes a calendar for myself.

It says that, you know, remember, you need to call Barry. He was going to, like, we had a conversation about it. So it just keeps keeps you that reminder when it's needed.

So the thing is, that's just limited to 2 ,500 characters. I mean, that characters, if you can build that memory that you can use, it's really powerful. I think it's gonna go big.

So yeah, I've tried my hands on OpenAI and other models. They haven't have been that good. Yeah.