Panel Discussion: Leveraging AI and Automation in DevOps

This panel discussion focuses on leveraging AI and automation for sustainable and scalable infrastructure management. The speakers, who include CTOs and cloud architects, delve into the importance of data consistency for making AI-driven decisions in infrastructure. They emphasize the need to collect and store telemetry data effectively to facilitate AI utility, as well as the role of traditional AI versus generative AI in different use cases. The conversation highlights the significance of data quality, the potential of AI to streamline processes, and the changes AI may bring to roles within companies. Furthermore, the discussion touches on the ethical implications of AI decision-making and its impact on job roles, particularly for entry-level positions. Overall, the panel argues that while AI has the potential to enhance efficiency, human oversight will remain vital.

ladies and Gentlemen please welcome to panel discussion about the topic leveraging Ai and automation into boops for sustainable and scalable infrastructure [Music] management hello everybody Welcome to the panel discussion about leveraging AI for sustainable and uh scalable infrastructure management um with me is yesai um Albert and anorak and I think we start with a small introduction right isai you want to start thank you thanks uh so

my name is isai I uh the CTO at ler B we are a company that um focuses on developer productivity and experience uh automations for the uh some of the uh work happening between developers and reviewers all the way to getting Cod production happy to be here thanks my name is Albert and currently I'm the cloud architect and leader of it operations and uh devops and well

my background is directly like also in the startup world and in fact starting from the engineering uh positions ending directly on on on the position in which I was responsible for building light the operations and process and I think that it will be like the major point in this discussion from my site thank you hello everyone my name is anur is it working yeah it's uh my

name is anurag and uh I am a cloud architect I work for a car company based out of Sweden an electric cars company I'm not allowed to tell the name for unfortunately uh but you'll figure it out later anyways uh and I'm also an AWS data hero so I'll be bringing in the data engineering and uh uh let's say organization of data aspect of making sure that

the you can actually leverage AI for making a lot of decisions okay thank you well my name is Michael kman um I'm the managing director of CIA in Germany CIA is a consulting company we do it Consulting and I'm in the Consulting business for nearly 25 years now and I do devops since 15 years so we did not call it uh devops then it was more Alm

and it evolved into devops and uh I was in that phase and yeah I wrote some books about also devops and um yeah I'm a Microsoft Regional director Microsoft MVP and uh yeah speak at a lot of conferences and give a lot of trainings so happy to be so uh infrastructure management um an we talked earlier and you said uh for you it's very really important uh

the data aspect so so what kind of data would you collect or what what is your approach here if customers would would ask you okay how do we do AI for infrastructure management perfect yeah so again the topic says that we want to leverage AI in order to make smart decisions about our infrastructure right so in order to do that you need to take a step back

and first collect the the data in a way that is consistent across your landscape so when I say consistent across your landscape if you're using Cloud if you're using on Prem if you're using multiple Technologies you want all of your software to emit the data in a way that remains consistent across the landscape so utilizing some of the Technologies like uh open Telemetry for example so if

you're looking at uh again that's very crucial to get your Telemetry data or observability data as they call it the traces metrics and all of this you want to make sure that it is consistent across your landscape uh if you're utilizing tools that are that have this implementation methods that are proprietary then you will have pieces of information that are spread out across about the consumption of

your infrastructure and you will you you will only have bits and pieces of data that you'll not be able to put together so I I would say if you want to make sure that AI you want to utilize AI take a step back think about your data how are you collecting it once you think how you are collecting it how how are you storing it and making

it available so that's that's my starting point but um one more question the approach would be you would always be in favor of okay build your own AI Solution on top of that data or you you say okay there are already a lot of operation tools out there that have now ai built in uh so what would be the the preferred way in your opinion it's it's

a mix of both uh so let's say you as a company you have certain systems certain software that are really commercial that are like business support applications let's say which are not really bringing any additional Revenue to your company right so I'll be really happy to Outsource the optimization of that to the companies who have built in AI into their ecosystem because they know really well that

particular uh ecosystem how how how you can operationalize it but there are pieces of software that I have that only belongs to me as a company and that is differentiating me in the entire let's say landscape of products that I'm offering so the consumption consumption infrastructure consumption your logs metrics and traces they will be unique to your consumption patterns so if it is if it if it

if it that if it is that if that is the case then I cannot rely on a generically mod generically made model that these people are providing so for those cases I I would say invest in your infrastructure in invest in your uh building your own AI models now when I say AI I'm not pointing to just gen AI again that's the case right now everybody when

they think AI is Gen Ai No there is traditional AI there is machine learning and that will uh that will put you on a path to make the best use of your data that is generated by applications okay I we we talked earlier and you said it's it's very much about the process and and how we can feed AI into the process so right and I would

like also to emphasize one thing which is about directly data quality like in my career it's like the second moment on which we talk about using artificial intelligence and it's next time when we want to use AI to use AI so we can make slides we can tell to everyone yeah we are AI company but we start like with data quality that we understand like what's going

on and what's the process one example which is not related to AI but it's almost exactly the same so imagine the case and it's like in fact one of the use case that there was like the company in which someone decided to start using terraform for infrastructure as code and there was only like one guy who became a kind of evang evangelist across the company but the

problem was that some of the infrastructure was managed with this terraform some not and that's exactly the same thing when we because the thing is that the main goal is that it will streamline process it will simplify thing so that we can deliver as devop team we can deliver directly to developers a special panel which will streamline the process it will not replace people it can like

optimize our time but thing is that it needs to become like a crucial like a major component of the process not that someone will be able to bypass it it's exactly like you know with pool requests but like the good approach is that we can block making uh you know like self approval for pull request and then yes it's merged for example to production the key point

is that we build something that will be used across the company that's our goal and that's also like very important uh thing when we talk about return of investment as always when we talk about adding new tool we don't do it because it's trendy but because it will generate value okay most projects do it because it's trendy in plan I agree developers because they're trendy because cops

yeah where's the business people so um yes what what what do you maybe Also let's talk a little bit about the value from AI so what do you think where AI will bring value so we we collect the data we we look at the process but but what is the goal right so what do you think uh can we achieve over the next years yeah so so

we discussed this uh a little earlier and um you mentioned the and I'm thinking AI in in particular the new kinds of developments gen and uh getting closer to some sort of intelligence so it's not intelligent but the place where it shines is an ability to summarize and help um people in like ingest a lot of information um so knowledge building the playbooks understanding like teaching new

people on on the team this is how we work um augmenting the process with um like this is similar to 10 previous use cases 10 10 previous incidents this is what we did then all of these are very natural places for AI to to chip in and say oh there's a wealth of information very unstructured in many places so there's collected data but there's also rcas and

and previous tickets and incidents and and slacks and emails I can use all of that to help the data build and help new help people know more faster about what they need to do here I think that's probably a key component and I think AI today is in a state where it it evolved a lot over the last years right with with the large language models I

mean this was a big boost but we are still in the point where we still have the co-pilot right so we still have the human that does uh that that needs to be there to correct to autocorrect because AI may fail so so you need someone to blame yeah someone to to steer so so the the main steering are still the human and the co-pilot assists and

and and creates a lot of value and adds value to it and I think at some point in the future this will shift I think especially in infrastructure management we need at the end very binary decision so it it either works or it does not work so if you have a co-pilot or if you have ai and it hallucinates from time to time then your data center

is down and you don't want that so so it's still a lot of the co-pilot on on on your side meaning augmentation of tickets writing postmortem uh analyzing it proposing you things but this will also generate more quality in all the data you have on operations right you can have more more data you can make more with it and this will accelerate over the next years and

then it at some point it will shift and then AI will do most of the heavy lifting of the decisions and then you only pull humans in for some exceptions and I I think that one one important think here is that it's a really great tool for example for knowledge sharing or building like the internal knowledge sharing right because one of the problem is always documentation no

one likes to write documentation let's be honest so that's like the great case for artificial intelligence to do like this boring stuff right and then we can generate information because that's always like the problem when we for example onboard someone right and we need to provide information or that we have for example our infrastructure as code we have like everything automated right but we need to have

like the description okay but what is the dependency why we have something in this place because we have comment or annotation about something I think that's like very valuable think in which AI all even right now can do their job perfectly because it's like documentation that's something in which AI can be in my opinion much better than anyone so AI writes the documentation and then AI reads

the documentation that's the circle of acceleration Circle of Life remove the man in the middle yeah I think that's true and for example GitHub they they do chat Ops for many years right issue Ops and chat UPS so they have everything in slack and they automate everything in infrastructure shutting down things uh rebooting them it's really great great if you have new new Engineers on boarding they

can go H look at that day and they they have the incident and everything is documented there so that's perfect to to put AI on top of it um but now if I look for example in Microsoft they also use heavily AI to to run to run aser right but uh they are not so focused on on Chups they had different systems and now with teams meeting

and everything but now with co-pilot suddenly these teams meetings are not just a meeting right that used to be it was a meeting something came out something happened it was a a a a very critical incident nobody would document that today you have a co-pilot say oh give me the summary of that meeting and suddenly you have good data and and and this along the entire operations

chain this generates a lot of quality data that then now accelerates the entire process and I think that's the interesting part so it it's I don't know where we exactly here but I think it's only a few years and and the quality of data we are producing now that that will will generate a lot of new use cases and where the agents can basically do most of

the stuff autonomous it's like related even to loog analytics right as we can have like our elastic search or for some cases in cyber security is it like I would say early stage right we have still a lot of false positive but is like the domain in which I see like the great case as we have like a huge volume of data right and then we need

to find someone in the team to process this data and I think it also connected directly to shortage of skills in the market right so who in the company can fast I would yeah that's I will add another angle to it with I would say U if you look at any big Enterprises their problem is not data availability their problem is taking actions from the data that

is available so if if I am a early stage engineer let's say first couple of years of experience and you throw me a big Json blob like a file which is five gigs in size and people are like here here you go that's the logs that that are generated by this application it's acting weirdly can you go and find the errors for me right I mean you

take this augment this and put for a 10e uh experience engineer he's going to he or she is going to have the same problems right so these are the places like you said in the beginning where we can use AI as augmented tooling right so can I build use buy share rent subscribe whatever some sort of models that are heavily trained across these kind of use cases

so I have a Big Blob I I pass it through the AI and it is co-piloting with the engineer it is enabling the engineer and then they are in a position to say okay I think oh the kubernetes the PS that in that are in the kubernetes they are having a conflict because they are trying to deploy the ports at the same time so that kind of

insights trying to build it out of uh out of that particular file maybe it would have been impossible for a human with 10 gigs of locks to figure that out and not just a vile adding more data sources and combining them with their right time stamps because then if you do I don't know giops and you have your kubernetes files in git and and your AI has

access to both and they say oh here changed something and here's the log that changed with it a few minutes after so this is probably correlated yeah ai ai is is is extremely good at finding patterns that human eyes cannot see that's the reality that's how AI is built right and we need to find ways in which we can utilize that to the fullest so if for

example if I give somebody an ax and ask them to uh let's just blow up the car right they can blow up the car with an axe but it's going to be painful because ax is made to be taken take down a tree so similarly we are we are we need to figure out okay when you say AI there are so many types of AIS out there

and and each of these AIS are are built to do certain aspects of this so if you want to apply it on your infrastructure needs fine tuning is required that's my point I mean this is a whole long story but my point is yes fine tuning of AI is required and these are the cases where AI will shine if you make the best use of it and

costs are also an important topic now right since we throw uh geni on everything and uh suddenly the response times are really slow which is also a problem infrastructure management and the costs are very high so do do you I think in the context of an incident where you're comparing that to human time it's almost always cheaper like getting the even if if there's a cost perspective

if it's not about like resolve everything all all the time but you're dealing with an incident and you saving developer time the the math is always in favor of automating yeah that's uh definitely true yeah but if you want to have like some some scaling up scaling down logic where you have to react then it's basically gen out of out of game it's too expensive and uh

just uh yeah too slow and also like one thing to which we I think we made another came back so it's about like data so if we use like some software as a service do we want to provide our company information to third party and that's even like the case with ch GPT when many employees just copy and pasted compounding information to J chat GPT because well

they didn't have like the knowledge that in fact they shared like sometimes confidential information like the case of Samsung employees when they provided some information about semiconductors product directly to chart GPT and now companies are aware of availability of these tools and their employees would like to use these tools right because like it can simplify their work and that's the thing on which I think it it's

on the one on the one side a kind of like blocker for many implementations and on the second like a great chance for many projects especially for self hosted some open source that all data will be owned like the by the company and in fact it will also mean that like only big companies will be able to build very valuable internal AI because they have a huge

amount of data there's probably a middle ground where some fine tuning at the industry level or the St stack level right this is how you can study a log from a specific uh type of system and then larger companies will do fine-tuning on these are my apps these are my applicative logs these are my specific um um private to me kind of setups yeah that's that's where

my point from earlier that I'm trying to keep is there is a division of labor so the big companies who are providing me the models they expect some aspects that I as a company need to take care of my own data that's where your data engineering principles your data engineering practices comes into place are you building again data engineering is all about building pipelines and pipelines inherently

is nothing but automation so if you have if you want to make the best use of AI you need to again take a step back ask yourself okay how do I make in order to make the best case use of AI I need to get provideed the best data with the best data quality how can I get it I need to build item poent data pipelines that

can bring the data back to this AI how can I build data item poent data pipelines to build practices and to build uh build principles that you can then keep consistent across your infrastructure land landscape which will then give you the opportunity to make use uh one example also of usage of like any kind of AI is even like cost optimization in Cloud as bet I think

like as when we talk about costs in the cloud in most cases we deploy all the services right everything is pay as you go so we have this flexibility but in many cases like you know Engineers or someone who's responsible in the company they may not have like the knowledge from day one so that's the place in which as we have like some patterns for example of

usage of kubernetes right so that system can detect at what specific time we should scale up or scale down our system for example which specific tool available in the cloud would be like the best fit right because like the cost for this for our specific case as each company has different we can have like the different number of events different size which would impact the cost and

I think that's like very nice case for AI not specifically like gen AI EX that yeah it's it's good to put you know like in the presentations to stakeholders right that we implemented uh gen AI but that's the place this operator will probably do yeah yes yes so it's even stronger with kubernetes J gen Ai and yeah I think that's that's like very interesting domain in which

it's it's just useful and that's also like what the cloud providers try to do like from from their side at least yeah I mean I've been so far being so against geni that I I want to clarify that no I see gen as a very good potential uh but the Gen but gen inherently the large language language models are built on the semantics of the language right

for the English so for the use cases where you said okay I have a RCA report I have a um I have a incident report and I have multiple these kind of reports that are going across that are written in human languages those are the areas where jna shine because they are built for that purposes if you take J and apply to process your logs perhaps is

going to be able still able to figure out something but it's it's not meant for it so utilize those pieces of AIS in the right spots orchestrate them through automation so this is where AI Ops maybe ml Ops something like that that comes across uh which which which helps you do this automation making sure that the chain of custody for utilizing AI remains consistent across your use

cases we we have established I'm not agreeing that AI is not good for analyzing your locks because in your locks you have a time stamp and you have some is not AI is but gen is not no but uh you always have a a logging string from the developer that you will also find in your source code at some point right so and if if you have

an AI that knows your source code and you knows your logs and then they can perfectly match oh this log is from that line and and suddenly um this knowledge of language is is important right so so this mine changed yesterday at night exactly yeah yeah so so there was a deployment so so there's a lot of knowledge that that because it's not just the the English

language or general language AI can also combine things very very logically it's semantics so it it learns itself on semantics but we are making an assumption that we are changing gen models or we are feeding into the Gen things that it can understand so uh R AG rack is the current way of modifying gen so that it can it can do that uh but even for us

to utilize rag effectively we need to have data engineering practices because if I do not have data I cannot feed uh good data to the Gen through rack so I think another example where you know you bridge the Gen or llm capabilities with the problem of making sense of a 10 gig blog file or multiple correlated files a typical developer or person doing RCA is you know

at that scale needs to write code or queries to okay I think there's a problem with ports let's let's find the relevant uh log lines so that often takes time just running or how do I query these two files which are in different formats to prove my point and J can help me write that query so you can ask uh give me the relevant log lines it

will do a internal rag cycle to create the query or the you know in whatever um text stack you're using and get you that result result faster so it's an augmentation of pure development if you like for the RCA even even if you haven't ingested the full 10 gigs into the context of your uh that exactly is the part of the topic which is automation so AI

alone will not be your Survivor or or it won't be able to help us take to the level that we are expecting it's it also needs to be supported by good uh practices around automation so I don't see the either of them surviving without the other so we talked about wreck um but maybe let's talk a little bit about agents because I think this is also will

will change a lot so so what's your take on on on agent and how will they well AI in general but of course also for management I think uh okay sorry I'll since I here I'll start so I think agents will be really helpful uh but I I think uh the way we are building agents I I I'm guessing it won't be on proprietary AI um algorithms

so I'm I'm specifically looking at openi because that's the most common aspects of it but with the recent um recent announcements about llama being really good as good as CH GPT uh I think agent uh would be mostly private LMS that are modified on private data in an isolated environment that are are specific to your uh your company let's say uh because we I mean there there

have been a lot of news that ethical use of data by openi is questionable uh at this moment so yes agents will be the big case but I think uh the building blocks of uh agents would be open- Source uh llms so self-hosted uh like closed wall agents yes not not even small so U llama 3 for example I think it was launched by meta a few

days ago it has twice the token capacity as chat GPT uh but it kind of assumes that you have the infrastructure in in in place so I think right now the bottleneck for that is infrastructure U because not all of the companies will be able to afford that infrastructure but this is where also small language models come into place uh which were again recently launched by Microsoft

if I'm not mistaken uh so a combination of these slms llms open Source fine tune to your use case where you ensure that the data of the llm is not going beyond your company's boundaries that's where I see agents succeed um I already have customers that uh use agents um to automate some things in the ticket system so um if they open a ticket they would have

one agent that would ask another agent or at the end of llm right so um give me a solution for that then it would validate this and it's like a three agent that talk to each other and so you would have this cycle and then it would Che no I think you're hallucinating this is not the solution and it will go back and and ask it again

until the solution is like 99% and then it would hand it back to the human still to a human to approve but then uh this this way of course this takes some time and money of course is not cheap but it it just um shows I think the power of of Distributing this on on on multiple agents I think like any other profession 80% of of what

we do even in this space um is simple work like not everything is a terribly difficult and U complex problem to solve there are a lot of uh simple tasks there are things that were a junior coming into the team would typically handle that is TP what you we usually see as the starting point for agents you need to document a change uh in some way or

to make um a very standard but boring update across multiple places that's a great ticket for an agent to to go and run with with eventual oversight by a human maybe without and how will it affect licensing licensing of the Agents of software and general I mean uh that that's something I I thought a few weeks ago and then it was like okay so we we used

to do licensing on buying software right and then selling updates and then with a cloud it changed completely to a subscription model now like like every software runs more or less on a subscription model but now suddenly with agents I mean thinking a few years in the future where the co-pilot experience is not the co-pilot but the main experience and and the other one is on the

round and and more agents are talking to each other so I have my personal assistant token that agent that will talk to yours and figure out when we meet and something and already book the restaurant and everything and so will we still pay per user will we pay per agent will we payest there will be a difference right well I think price pricing is Shifting is always

shifted or towards value so if it's not about the cost of producing the software or running it if if you're only as a vendor trying to get some margin over your cost to deliver then you're playing a very uh shortsighted game so charging for some like for usage if this is or consumption based is trying to tie to Value if I'm using agents internally to run my

stuff I don't see that as affecting my pricing structure but agents will become not just an internal method to keep my lights on agents will become part of my soft my product that is customer facing right it will in the future replace humans right so instead of having an employee hiring here you probably can replace it with an agent at some point for and I'm talking about

there's internal how do I keep my system running right with most of our talk today is internal facing but I think we realize most of our software which is is customer facing will also be changing and look more like agents you're using my um I don't know my workspace solution now there's agents you're using my calendar solution there's agents you're using my planning solution all of these

now become uh agent assisted as part of my value to customers and yes that that that will me give me an opportunity to charge um some uplift because this the my software is smart I think um the costing is is it's a very complex Topic at the moment because all of the people who are selling AI augmented Services whether through Bots or whether through agents or through

regular aspects uh their price comparison point is the price put by open AI at this moment but it's a fact that openi is absorbing a lot of this cost behind the scenes because they have plenty of VC money at this moment but what happens when the VC money runs out the probability of the base cost getting down or increasing by 23x we cannot discount it uh so

the point that I'm trying to make is right now whatever whatever things we do it's it's a shot in the dark because we don't have a baseline uh for us to compare what AI costs unless you're selling gpus absolutely that's that's the money minting business right now or as they call it selling shovel shs yes but yes I think we will see some Modern or some new

take on the um on on on the pricing models for the agents which would probably be maybe again I'm taking a guess in the wild maybe it would be pay for each successful outcome of the agent like if you're trying to schedule a lunch meeting with 50 different VCS and if it is able to Target one VC and able to get you lunch you will only pay

for that particular instance and not pay for the 49 uh let's say different things that he was that agent was not able to get again that's an interesting pricing model uh that that could come and with the agents I think just a side note it's it's the insurance industry that will also come up with some pricing uh let's say Shenanigans uh and uh that would be interesting

space to watch in the future to Ure me against the AI hallucinations and also ensuring yourself so that your human agent human uh sorry agent is sorry your AI agent is not if if they do some mistakes yes I need to have some sort of insurance against it yes so we go on more time that yeah automation is great but in the end it's like the case

in the end there must be a human it's like with autonomous cars we talk about autonomous cars for so many years but then we go to the problem okay where is the accident and who is responsible for it currently of course the driver because driver needs to be like responsible and need to observe is the system works correctly and we'll go to the exactly same D exactly

same problem also like in this devops in devops uh domain right because okay something is deployed but imagine the case that for example it will start 1,000 of virtual machines by mistake and who will be responsible later that's Alo like the very difficult part and yeah I think that yeah it can be great business for insurance in my opinion this is always very difficult in theory and

practically it's not because uh humans do errors machines not right so or at least much less right so if you just look at how many car accidents are caused by humans and then if you look how many are caused by autonomous cars that's a quite easy discussion with the the um with the insurance companies right so and I think that's figured out quite quite easily so so

legally it's a little bit more complex but uh in the end even given those numbers it's it's been hard to make autonomous cars a reality not because of technology I don't think it's the insurance I no it's legal legal and moral legal and moral and it's exactly like in it that you from engineer perspective we can do everything it's not a problem we can deliver but the

part is when we go to the business and legal team they will challenge it and like from legal or business perspective it may be just too high risk level in comparison to the uh Return of the investment and that's like this I I would say always very challenging moment in which we need to merge like this engineering World in which it's doable and like from probability perspective

from like data perspective is not the problem because we need to put it on PowerPoint and to present to someone who may be like against and I think that it will take a lot of years also to educate about like upsides and downsides of such Solutions so do we have any questions from the from the audience or not yet okay so then let's continue maybe a bot

somewhere an agent wants to ask questions oh yeah we we should have a question it failed to reason I did this with some conference talks during dorona Corona in because in online conferences people would not ask questions so would F Q&A and then already ask my own questions that I will answer we could do that and have a b and this I think like quite interesting that

in DeVos well it's mostly engineering right and that's like this Dom in the company which is mainly led by Engineers so that's like very nice Green Field in which we can play we can make experiments and we can have like the full support but when we go to some like more business critical and I would say based on my experience it always go to very like basic

questions espe especially like with legal and business team which may have like different approach to things and even some like this data driven decisions but they are not like very much data driven your talk also covered this earlier with governance right so a lot of the governance things are not the rules but the rules the company lays on their own so they create the policies and then

they think the policies are made in stone and then oh we going to do this because of the policy and no it's you can change the policy and then you can do it yeah but I think there's uh even in inside the engineering bubble could in theory be more adventurous and experiment with technology and rely on technology more I'm thinking about that conversation with the manager the

system went down because the AI made a mistake versus the system made went down because John made a mistake and the the first conversation is very hard like okay what do we do now how do we prevent it from happening again all of that people are used there's only already a good practice of what to do when it's a human and there's still not a good way

or a comfortable way to deal with the AI made a mistake okay how do we prevent it why why did it make a mistake yeah so no no but it's the same question right John if John was able to shut down our data center by accident so uh wow what can we do that this thousands of years of experience with human blame and with human skills and

capacity or incapacity it's it's natural to us to say oh the Stars it's like God made a a move that's a way you to explain the unexplainable or blame someone or maybe okay let's fire this person or let's train this person or let's tie their hands so they can't spin up those I hope the companies will not fire someone from making a mistake because uh that's I'm

saying is there's a lot of human experience in dealing with that and not so much with there is a opaque algorithm that I'm not going to debug the llm right it's not it's out of scope now it's okay how do I hedge and if you starting to work on hedging the algorithm you've already lost a lot because you're not tying its hands so it can't do this

it can't do that I need a human to to blame at the end if they made a mistake over overseeing the but if your process works and you have a blameless postmortem right and you analyze what caused the incident you've you fixed it later on and and you do it and then there was an agent that did something wrong right and that's this same then it's a

human it's okay how could this happen and then you might need another agent you might need some quality control you might need uh to adjust some Matrix that okay this might have been an edge case and maybe there there should have been a human intervention but so it's like the normal process it doesn't matter if if this error was caused by a machine by a config file

by Network issue or something or a human or an AI right it doesn't matter it was the entire process that that failed at that point and you have to look at the bigger picture ask if five wise uh they do a root cause analysis and then try to fix the root cause and if the root cause is the AI then then you need to to adjust something

there but I don't see a big glob I'm saying it's a lot less comfortable for existing humans which is why it's taking longer yeah that's just true it's new and now we have question to it oh there's a question what do you see as the next big steps that are not possible now but will be in a few years that's a great question who want to take

first yeah I'll take a stab at it I would say we will see new models of uh what is called as autoscaling in infrastructure that we have not seen right now so let me give you an example if you are in a kubernetes space or or or if you're in a serverless space you tend to have a component called AS application load balancer or network load balance

or whatever that takes in certain parameters that you have to arbitrarily decide again believe me 90% of the teams don't know why they set that threshold right so they they are unable you are upscale or downscaling depending on that arbitrary numbers without rethinking what those where those numbers came from or what is the right time to reevaluate those aspects so I think uh that's one big area

where we are going to see a big change because companies like Amazon Google and um Microsoft they will have a lot of data that they have collected from albs that's going to bring in the patterns that will work just out of the box for 99% of the use cases uh so that's one aspect very close to infrastructure where I see autoscaling becomes a commodity rather than uh

good thing to have uh in the next few years which is right now it's it's on the onus is on you to configure it right but it'll come out of the box yeah I do see I do see like the case in which on each employees laptop variably install this agent and that like currently we have internal developer portal right in which developers can create like their

temporary environments that everything will be moved to a single place a kind of chatbot in my opinion in which they will be able to create everything what they need uh for their software development and that there will be like the single space which will be able also to track their activities and understand what they do if they do any kind of the mistakes so it will be

like shift left but not directly in the pipeline but it will be moved directly to the computer of developer um one area that may uh we may see as a as a enhancement in the coming years and it's a combination of the models becoming smaller or um um things I can start to embed I think there's there's going to be a shift from centralized uh infrastructure management

to uh a little more decentralized or even embedded so a process could take care of of its own or a CPU could um make some smart moves on where where am I spending my Cycles depending on local signals instead of having a master uh octopus that is driving the decisions the autoscaling decisions across everything we now collect all the signals to one place and make a decision

and think of every every CPU cycle making a decision based on local inputs but enough uh Power to do predictions now you have this emerging self-optimizing um kind of grid which is decentralized no central place makes the decision interesting yeah um I think also wish did already a little bit in that direction I think right now it's a lot about augmentation of incidents of tickets and then

then Gathering data creating better data but I think in a few years I think 99% of the incidents that they will be handled by ai ai will know what to do in this case something is down something's not working clearing the cash starting services so and and this system will learn and learn so so you will always lead need less human interactions and I mean the data

center providers they're they're huge they scale extremely they they have really big interest in doing it and they have the tools to do it so I think that will be one of the areas that that will all all lead in in in that space because it's a lot of money we have another question you mentioned Ai and agents taking over Junior tasks does this not also create

the problem Junior entry level chobs are fading the future seniors sure I'll I'll take that so there is uh if you look at Evolution from of history or mankind you will always see that uh when you introduce a tool it kind of creates that fear uh that my jobs are going to take over uh are going to be taken over a classic example of that is accountants

and calculators because when calculators came along accountant thoughts okay my job is gone because everything that I do is that's uh that's what the calculator is doing uh but if you look at history again um accountants are not hired because they can do math they are hired because they understand the domain of accounting so that's what I see the junior jobs that we have right now where

we tend to again not everywhere but most of the companies where where they tend to give the bread crumbles or bread pieces uh like uh like very odd jobs to to Juniors that will change where Juniors when they enter they will do real work using AI as their friend as a mentor as as as a person who can guide them a little bit when they when the

AI becomes strong enough but there will be a weird phas in this industry for the 5 10 years until the point or I'm I'm stretching it maybe 2 three years until when AI is not that good enough where we will have to where the Juniors will I think will struggle and I'm not saying this out of my own uh let's say uh um because I think so

because this was shared by another person who works in rice that is Research Institute of Sweden who published a report about Ai and its effects on Junior entry level jobs yeah and I think it's the same like when for example anible Chef or puppet was introduced it was exactly the same because it was automating the work which was I think we're not only talking about Jun entry

level chobs I mean AI will take over more jobs and a lot of lawyers I think can be be replac by AI because it can do it better not not everybody of course but there are more jobs that will go away and then I mean in the future we will be happy probably because we do not have enough people to fill all the jobs but there will

be some time where where also the entire Society has to look in their eyes okay how do we gener how do we distribute the money because right now U it is done by salaries and then in the future if we have agents because this talks a little bit in in our discussions we had earlier about the licensing so then in the future it's okay how do we

distribute the money from the company so that everybody has enough but I think that's a too complicated thing to take it over now because I think we're over time and uh I see people already getting a little bit nervous so thanks for the good discussion and uh thank you very much thank you thank you very much we will be outside on the ask anything box if you

feel like asking followup questions and we will leave this for the next talk I guess yeah thank you thank you [Applause]