Arun Joseph on Reliable AI — Interview with Alexy

Yeah, I'm Arun Joseph. Uh I'm the Yeah, I'm Arun Joseph. Uh I'm the co-founder of Mosaic Agentic Systems. co-founder of Mosaic Agentic Systems. Role of course uh is building in a Role of course uh is building in a startup right along with my co-founders startup right along with my co-founders Jasper and Amad. Yeah, that's what my Jasper and Amad. Yeah, that's what my and my startup is all about large scale and my startup is all about large scale decisioning engines uh for operational decisioning engines uh for operational intelligence as we call it. Yeah, that's intelligence as we call it. Yeah, that's pretty much what we do. Previously the pretty much what we do. Previously the head of AI engineering for Dodge Telecom head of AI engineering for Dodge Telecom in Europe. in Europe. I would I would say that the I would I would say that the accomplishment was we developed a few accomplishment was we developed a few things which might have been something things which might have been something like a small mini Xerox spark moment. We like a small mini Xerox spark moment. We built potentially one of the first built potentially one of the first agentic platforms and put it live early agentic platforms and put it live early late 2023 late 2023 and uh it it powers it used to power and uh it it powers it used to power multiple countries in Europe open source multiple countries in Europe open source mood and Eclipse Foundation early mood and Eclipse Foundation early or mid 2024 we gave a talk about it in or mid 2024 we gave a talk about it in our production agents when most of the our production agents when most of the enterprise a AI was failing. So we gave enterprise a AI was failing. So we gave a talk and it was only a week later a talk and it was only a week later OpenAI even released swamp. So we OpenAI even released swamp. So we started building our own protocols. So started building our own protocols. So we had a small great team um who who we had a small great team um who who aspired to build the foundations of what aspired to build the foundations of what we call as a H&D computing and we put it we call as a H&D computing and we put it in production and there were so many in production and there were so many learnings. So both the business outcomes learnings. So both the business outcomes as well as having had the privilege of as well as having had the privilege of leading such an elite team reliable AI. leading such an elite team reliable AI. Yeah, this is excellent. So I'm one of Yeah, this is excellent. So I'm one of the original contributors to the small the original contributors to the small group that u wrote reliable AI. Reliable group that u wrote reliable AI. Reliable AI is something I would define as um if AI is something I would define as um if the outputs from an AI system or any the outputs from an AI system or any system, right? U if they are verifiable system, right? U if they are verifiable and reproducible and consistent under and reproducible and consistent under real operational constraints. I would I real operational constraints. I would I would say that's the definition. would say that's the definition. So this is interesting. So of late I So this is interesting. So of late I have started to notice a shift in also have started to notice a shift in also treating this as a true engineering treating this as a true engineering discipline. discipline. and not only and not only uh model research. So reliability uh model research. So reliability engineering loves constraints. So engineering loves constraints. So engineering shines in constraints, engineering shines in constraints, right? So AI the pure model, right? So AI the pure model, how do you constrain it to produce how do you constrain it to produce verifiable verifiable uh consistent results is an engineering uh consistent results is an engineering discipline which has started to pick up discipline which has started to pick up and more effort should also start to go and more effort should also start to go in there. uh and more such use cases in there. uh and more such use cases will allow push the boundaries of will allow push the boundaries of computing as we know it and right now a computing as we know it and right now a lot of this u is definitely there on the lot of this u is definitely there on the model side but I'm absolutely imagining model side but I'm absolutely imagining new operating systems new ways to have new operating systems new ways to have makes deterministic and nondeterministic makes deterministic and nondeterministic computing as a foundational paradigm computing as a foundational paradigm there are new line and straw walls to there are new line and straw walls to emerge I'm looking for where they are emerge I'm looking for where they are going to come from that's all I would going to come from that's all I would say yeah the next five years, say yeah the next five years, five years is uh light here in in the AI five years is uh light here in in the AI space. Um I would say that there is a space. Um I would say that there is a lot of lot of hype in the AI world. We started to see hype in the AI world. We started to see at least AI systems as computational at least AI systems as computational primitives which is different from primitives which is different from expecting unicorns and rainbows, right? expecting unicorns and rainbows, right? So the next five years I would imagine So the next five years I would imagine at least personal computing would have at least personal computing would have definitely definitely uh definitely changed into a mix of uh definitely changed into a mix of traditional deterministic and uh and traditional deterministic and uh and agentic computing as I would call it in agentic computing as I would call it in every personal device I would imagine every personal device I would imagine models also to be running such that the models also to be running such that the ease with which you are interacting with ease with which you are interacting with the computer to do things for you is the computer to do things for you is radically going to change. You can call radically going to change. You can call it AI, you can call it whatever, but it AI, you can call it whatever, but this is the fundamental construct. Other this is the fundamental construct. Other than that, I would imagine than that, I would imagine the large scale enterprises to the large scale enterprises to completely shift in how they build their completely shift in how they build their IT information systems are very prone to IT information systems are very prone to change. change. I would imagine companies to be run I would imagine companies to be run purely on agentic systems at least some purely on agentic systems at least some divisions which is going to produce divisions which is going to produce massive significant results. This is massive significant results. This is essentially why I quit my job to build essentially why I quit my job to build such systems such systems thinking that this is a future that it thinking that this is a future that it holds large scale operational holds large scale operational intelligence companies or divisions intelligence companies or divisions which can self-optimize on information which can self-optimize on information systems right and not AGI I would rather systems right and not AGI I would rather say how do you selfoptimize a system to say how do you selfoptimize a system to increase the revenue or reduce cost increase the revenue or reduce cost leakage because information system can leakage because information system can self- adapt with a very small nimble self- adapt with a very small nimble team of experts