Baruch Sadogursky and Leonid Igolnik on Reliable AI — Interview with Alexy

This is Leon Golink sitt This is Leon Golink sitt and this is Baragarski head of developer and this is Baragarski head of developer relations at taxare. So before we talk relations at taxare. So before we talk about the highlights uh we should talk about the highlights uh we should talk about how it feels right now working about how it feels right now working with AI with AI before even that something is wrong in before even that something is wrong in your customer. your customer. That's right. It's like back to the That's right. It's like back to the future and you have my goggles. That's future and you have my goggles. That's the problem. I knew something was the problem. I knew something was missing in this. missing in this. Now we can talk about AI. Now we can talk about AI. Uh as we just talked in our Uh as we just talked in our presentation, working with AI is presentation, working with AI is tremendously rewarding but tremendously rewarding but frustrating frustrating and frustrating at the same time. and frustrating at the same time. Frust. So it starts with rewarding. Wow, Frust. So it starts with rewarding. Wow, this thing is cool and it becomes very this thing is cool and it becomes very frustrating very fast. frustrating very fast. So uh that's why in our talk we talk a So uh that's why in our talk we talk a lot about how do you take the output lot about how do you take the output from those modern AI tools and turn it from those modern AI tools and turn it into a reliable result. And that's been into a reliable result. And that's been an area of interest for us for the last an area of interest for us for the last almost a year. almost a year. And in the end of the day, the highlight And in the end of the day, the highlight is when this happens. When you not only is when this happens. When you not only feel good about the outcomes of AI, but feel good about the outcomes of AI, but can reason can reason that the end result is exactly what you that the end result is exactly what you wanted it to do. This is this is the wanted it to do. This is this is the rewarding part. It's much harder than we rewarding part. It's much harder than we would expect when we just saw this would expect when we just saw this amazing LLM thing for the first time and amazing LLM thing for the first time and we really thought that hey we ask for we really thought that hey we ask for it. It actually does that. The first it. It actually does that. The first vibe coding experiments looked like wow vibe coding experiments looked like wow but then you look closer but then you look closer and it says you're absolutely right. The and it says you're absolutely right. The funny thing it reminds me of the early funny thing it reminds me of the early days of getting into coding like that days of getting into coding like that every engineer gets into coding because every engineer gets into coding because you go I have an idea and I can turn it you go I have an idea and I can turn it into the reality and I think AI done into the reality and I think AI done right can short circuit that cycle and right can short circuit that cycle and allow you to build bigger and more allow you to build bigger and more complex realities. But that that complex realities. But that that endorphin shot when it works and it endorphin shot when it works and it works reliably that's I think what's works reliably that's I think what's rewarding about working with AI at a rewarding about working with AI at a bigger scale than a single human could bigger scale than a single human could do. do. Yeah. So it's kind of a cycle you Wow, Yeah. So it's kind of a cycle you Wow, this is exciting. It does what we need. this is exciting. It does what we need. Oh, actually not really. This is very Oh, actually not really. This is very frustrating. And then we fix it. We add frustrating. And then we fix it. We add some guardrails. We add some shared some guardrails. We add some shared understanding, shared context. Oh, this understanding, shared context. Oh, this works. It does exactly what we want. And works. It does exactly what we want. And this is where the real doment kicks. this is where the real doment kicks. You you quoted every single keyword from You you quoted every single keyword from our talk. our talk. That was by design. That was by design. Ah, available on the internet. Ah, available on the internet. Exactly. You go to show Exactly. You go to show notes.taxcar.com, taxcare.com there is a notes.taxcar.com, taxcare.com there is a recording like a video slides all the recording like a video slides all the links everything is right there links everything is right there you can watch it as many times as you you can watch it as many times as you want want and in different versions we have 30 and in different versions we have 30 minutes we have 45 we have three hours minutes we have 45 we have three hours depends on your availability you can depends on your availability you can pick and choose when we think about pick and choose when we think about reliability and AI uh I think it's reliability and AI uh I think it's important to understand that AI by important to understand that AI by definition is a stochastic system it's a definition is a stochastic system it's a non-deterministic system every time you non-deterministic system every time you put the same input you get different put the same input you get different outputs and that's desired in that outputs and that's desired in that system so for us reliable able is how do system so for us reliable able is how do you make it repeatable, you make it repeatable, right? So because stochastic right? So because stochastic nondeterministic systems means no nondeterministic systems means no reliability ever. This is what it sounds reliability ever. This is what it sounds like. You will get different results. So like. You will get different results. So what are we even talking about? The what are we even talking about? The answer is yes. Let's see how we can answer is yes. Let's see how we can wrangle it into doing the right thing. wrangle it into doing the right thing. Now right thing and same thing are not Now right thing and same thing are not the same things. You might have the same things. You might have something that slightly different but something that slightly different but still right. still right. That's right. When it comes to the That's right. When it comes to the reliability, I think it's important to reliability, I think it's important to look at the history of our industry look at the history of our industry developing from the very 50s when uh developing from the very 50s when uh Grace Hopper who built the first Grace Hopper who built the first compiler heard that uh I will never compiler heard that uh I will never trust the code I didn't write myself and trust the code I didn't write myself and yet here we are. We've been continuously yet here we are. We've been continuously increasing levels of abstractions. Now increasing levels of abstractions. Now in order for us to maintain those levels in order for us to maintain those levels of abstractions successfully, we also of abstractions successfully, we also built a set of tools. For example, we built a set of tools. For example, we went to the cloud, we changed our went to the cloud, we changed our observability, we built better observability, we built better automation, right? automation, right? similar uh solution should be applying similar uh solution should be applying to this new level of abstraction where to this new level of abstraction where English is your new coding language and English is your new coding language and stochastic nondeterministic things are stochastic nondeterministic things are your compiler. We need a set of tools your compiler. We need a set of tools and set of approaches to making this and set of approaches to making this abstraction work and and the industry is abstraction work and and the industry is getting there. We see a lot of tools getting there. We see a lot of tools which are AI native which were designed which are AI native which were designed first with AI in mind and especially first with AI in mind and especially when it comes to reliability. The stuff when it comes to reliability. The stuff like specdriven development is now older like specdriven development is now older age and there are tools emerging. There age and there are tools emerging. There is a kirao from Amazon that was GA is a kirao from Amazon that was GA yesterday. There is spec kit by GitHub yesterday. There is spec kit by GitHub that was released couple of months ago that was released couple of months ago and get new features every day. Um the and get new features every day. Um the industry is is starting to realize that industry is is starting to realize that what we need on top of LLMs are those what we need on top of LLMs are those guardrails that will make specdriven guardrails that will make specdriven development, reliable development, development, reliable development, intent in integrity chain and we see intent in integrity chain and we see those tools emerging. This is very those tools emerging. This is very exciting exciting and it's also predictable. Uh you know and it's also predictable. Uh you know for better or for worse Gardner cur for better or for worse Gardner cur coined the the curve and we we moving coined the the curve and we we moving from a peak of excitement and eventually from a peak of excitement and eventually we'll end up in the valley of despair or we'll end up in the valley of despair or this is where a lot of industry right this is where a lot of industry right now now or the tr of disillusionment and soon or the tr of disillusionment and soon we'll like any other cycle in our we'll like any other cycle in our industry we'll reach that plateau of industry we'll reach that plateau of productivity. productivity. Yep. Yeah. And this is what's starting Yep. Yeah. And this is what's starting to emerge. to emerge. Yep. Yep. Simple answer. No idea. like as as Simple answer. No idea. like as as unpredictable as it can be. We can look unpredictable as it can be. We can look at the changes in the last 6 months, 9 at the changes in the last 6 months, 9 months, couple of years and none of us I months, couple of years and none of us I think can honestly say I saw that think can honestly say I saw that coming. coming. But I think we should talk we can talk But I think we should talk we can talk about the elements that the stack will about the elements that the stack will contain. contain. Oh yeah, Oh yeah, predictability and reliability because predictability and reliability because at the end of the day we're trying to at the end of the day we're trying to drive business processes and business drive business processes and business outcomes with this and unreliable outcomes with this and unreliable unpredictable businesses don't survive. unpredictable businesses don't survive. A better integration between different A better integration between different teams in the organization like for teams in the organization like for example LLM makes the communication example LLM makes the communication between developers and business people between developers and business people and product people much more effective. and product people much more effective. We will see a lot of that because that We will see a lot of that because that benefits everybody and at the end of the benefits everybody and at the end of the day like with any other abstraction day like with any other abstraction layer we added in our industry more layer we added in our industry more software gets written time and time software gets written time and time again. So we don't know how the stack is again. So we don't know how the stack is going to look but we believe that those going to look but we believe that those are the key elements that the stack are the key elements that the stack would have to satisfy to stay successful would have to satisfy to stay successful in this space. in this space. Will we see uh stuff like uh um what is Will we see uh stuff like uh um what is called IGI and universal income for called IGI and universal income for everybody and everybody not working everybody and everybody not working because AI does all the job. Uh probably because AI does all the job. Uh probably who knows probably not but who knows we who knows probably not but who knows we will be pleasantly surprised if that will be pleasantly surprised if that will happen.