Opening: Unknown Knowns, Roles, and Why This Matters

The nice questions posed by you in his talk, the unknown knowns, the things that are already existing but we don't know them about ourselves for example and as well as the time spent on the roles is a time well spent, the time lost on the roles. So I will pick up from those parts directly.

Introduction: Teaching AI Who You Are (and What You Want)

So hey guys, I'm Tavish and today I will be like we will be teaching AI who we are and rather what what we want the AI to be for us.

So just before we start, hands up. So Andrea's already asked who all has used AI, but a bit more specific, who has used AI for, and given it context about its project

or something before asking. Specific context to yourself. Perfect.

And the next one is whether you have actually made an automation of sorts where you don't need to give it context anymore and the context keeps getting updated by itself.

Okay, perfect. So, oh, it's cut a bit, but it's fine.

Why AI Doesn’t “Lie”: Prediction, Not Truth

Last time I was here, Andreas asked a beautiful question. Why is the AI lying to me? So let's set it straight.

AI does not lie to anyone. AI does not tell the truth to anyone. It is simply predicting.

As you already mentioned earlier, there is no ground truth. It just has a memory.

And initially, out of the box, it doesn't even have a memory of who you are.

So, as you rightly mentioned, statistical patterns, it is just making probabilistic guesses.

Now, the exact gap this falls into is a lot of people still use AI like a search engine. They ask it one question and they stop. They ask another question and they move on.

Building a Useful AI System: Four Foundational Layers

Well, you can use it like that, sure, but the actual people who are getting value out of AI are building a whole system around it. No need for coding, it's a system. It can be a system without any code and they have four basic layers and those are the four layers we will be talking about.

These are the four base layers and I mention three because most people who are non -technical they can build three, so the identity, the memory as well as the constraints. Tool is a bit more difficult to build for the non -technical people but we will move.

Layer 1 — Identity: Values, Style, and the Kind of Assistant You Want

So the identity, as we already know, how you think, what you value and this plugs into the exact part of how you want your AI to be. Do you want it to be a mirror of you? Do you want it to complement you? So the things that you don't know per se and the AI fills that out.

Layer 2 — Memory: Supplying and Updating Context

Next, the memory. So this is where the context part comes in. AI out of the box has absolutely no memory, you need to give it context to make it actually perform the way you want it to.

Layer 3 — Tools: Giving the Model “Arms” to Act

Now the tools. This is the part which you can't build yourself per se, maybe up to some levels, but think of the language model itself like a brain. If you ask it a question it can reason, it can write you the textual outputs, but it has no arms, it cannot act. This is where the tools come in.

So what are the tools? It can be a simple Google search. It does not know information, it connects to Chrome, makes its search

query gets the answer back boom and finally the constraints things which you never want the model to do again everything fully is dependent on you what you want to do it is fully customizable there is no right or wrong answers now we will be looking at the absolute bare basics today so the

identity can be as simple as a plain text file now when this thing goes in you can tell it how you think what you value and what decisions you want the AI to make in a way that you want it to be so some specific examples let's say you want the sentences to be short no buzzwords casual precise it's up to you what you value so tangible things over abstract decisions and all those things

now as the systems get more and more complex it can go from a simple readme file or a text file into a fully like full -fledged system again fully up to you. We will look into a bit more later on.

The memory. So this video here, this is something which I use myself. As I said, it's a pretty fleshed out one.

It's a knowledge graph, fully useful for my coding projects. You can see everything that is mapped. So the exact files, which functions are in which files, which are being called by which other files, but you don't need to make it that complex.

Again, it can be a bunch of different readme files or text files, which mentions that, okay, this is the context. This is the working memory, which you need to keep to know what it's about, the whole thing.

Now, memory does get stale, like humans. Maybe I know you, after a while you change some of your opinions, so I need to continuously update my memory and that should be done as well.

Tools, as I said, is a bit more on the technical part, so we can't really build them out ourselves, the non -technical people.

So, some simpler things which you can do is instead of making tools, you can make like session start.

So why I mentioned session start is let's say you start chatting with your agent. Every time you start chatting, you want a few specific actions. Let's say, okay, I'm working on project A. What exactly were the things which I worked on in project A yesterday? Give it to me.

It can quickly check your history and give you back the entire thing that you worked on yesterday and the exact action points that you want to work on today. Very simple tool called per se.

Layer 4 — Constraints: Guardrails and Plan Mode

And lastly the constraints. So this is again fully up to you. Do not execute before the plan is clear.

The last point which I put in is because I use a lot of this thing called the plan mode which is present in cloud code, codecs and all those harnesses.

People don't use plan mode often enough. It is beautiful like the plan mode.

You start chatting with your AI. It keeps asking you questions.

It does not execute the plan it keeps asking you questions it keeps asking you questions until it actually knows what you want and then it gives you the plan and you still have the ability to change the plan

yourself and this live demo it is on my laptop if i get time i'll boot it up and i'll show you guys and now late so earlier people had this friction of executing the plans the friction is still there but it has shifted.

From User to Director: Harness Engineering

Once you implement those four basic systems, you are not the plain user anymore. You enter into the mode which is called the director and a bit more technical, this is where harness engineering comes in.

I'm assuming most people have heard about wipe coding, prompt engineering, context engineering. How many of you have heard about harness engineering? Nobody.

It's a pretty new thing because these agents, they are being used as something called a harness And the entire system you build around it, that particular thing is called the harness engineering. So it changes your role from a simple user who is executing things to the actual director. And then you start asking the right questions.

And what are the right questions per se? What is worth building? What does good actually look like? And if this is what I actually mean versus maybe not. So when you start asking those questions, that is when the entire system gets built and you start getting more value out of these AIs.

Wrapping Up: Summary and Takeaways

Well, that was my very short presentation about the entire thing.

QR Code, Templates, and Where to Go Deeper

Since you and my talk, they are very similar, we actually created this particular QR code. And if you scan it, you actually get to this particular website, which has, let me scroll down a bit, the opening frames by you. So the things that she mentioned about the known unknowns, the AI that works with CIDRs.

And if you scroll down more, this actually covers the entire thing about what the identities are, the memory, the tools, the constraints. And if you want to execute those things pretty much yourself later on in the evening, you can scroll down and I have a template here as well. You can pretty much copy paste it, modify it based off of your need.

And there's a bunch of things here like this is the tip of the tip of the iceberg. bug so a bunch of things done you can switch to the developer track as well if you want to look into the more coding heavy aspects so these are significantly more developer friendly things but yeah those are the things i wanted to talk about today thank you