AIOps: Smarter Systems or Bigger Risks?

00:00:00: Good evening and welcome to JAX.

00:00:02: Thank you so much for joining us here this evening.

00:00:04: I know it's been a long day, hope we've all had a great time.

00:00:08: We're here to talk about AI Ops And hopefully get some interesting information.

00:00:16: I'm joined on stage tonight by Christian Schneider and Thurston Kuster.

00:00:20: If he wouldn't mind take a moment to introduce yourselves.

00:00:25: My name is Thorsten.

00:00:27: I am a freelance software engineer And I build observability platforms for software stacks, especially for search applications.

00:00:39: and do all the funny stuff.

00:00:43: pretty much so that's my background.

00:00:46: Good one.

00:00:47: Yeah, Christian also freelancer but more on the security side.

00:00:50: so that means basically security architecture threat modeling and Also presentation testing And definitely all of that gets some kind of AI spin to do something discuss about.

00:01:02: That's what I'm looking for today

00:01:04: Sure So we said were gonna talk About AI ops But i wanna start by kinda defining What that Means To each Of you Like...I feel like it is a loaded term.

00:01:14: There may be Some nuance involved.

00:01:16: Definitely definitely.

00:01:17: So personally I take it as utilizing AI to get ops more well done, i would say so.

00:01:24: not only automation but also have some AI augmentation on that.

00:01:28: That not necessarily includes security Interantly But it can be used also an incident response stuff.

00:01:34: So there are also some kind of ops related things But also observability monitoring just the keeping things running things.

00:01:41: That's what I think they're.

00:01:42: and I do draw a line between AIOps and MLOps.

00:01:47: Because MLOps is more about the ops of where, in my perspective... ...where you automate things how you build your models to train those pipelines.

00:01:57: Yeah I like to distinguish that the AIOps term into two category.

00:02:01: The one thing is AI for Ops Where i use AI To make my Ops day-to-day work better.

00:02:08: Where do my analytics?

00:02:09: just using AI?

00:02:10: because in the end, operations or observability data is just data.

00:02:15: And I can use AI to make sense out of it.

00:02:17: and The other thing is that i uh used um basically ops for AI.

00:02:22: where have llms Or ai driven applications?

00:02:27: Where i use um were basically inject observability into my application stack and where I use observability to monitor, and to trace my AI applications.

00:02:42: Okay?

00:02:42: Talk about that just a little bit more.

00:02:43: tell me more about the operations for your AI.

00:02:50: Operations for AI applications is kind of fundamentally different than just regular applications because with AI applications tracing gets way more important to get.

00:03:06: And not only tracing, you just need a tonne more of traces.

00:03:11: with classic applications.

00:03:13: You might implement some tracing like auto-tracing or whatnot.

00:03:17: and um...you may sample your requests To Get A Good Overview Or Basic View Of Your Deterministic Operations In Your Application With If You Have An LLM Doing The Magic Trick in Your Application then you basically should like trace a hundred percent of your traces to get the full view because You open up when using LLMs and applications, you open up your application too non-deterministic.

00:03:48: To the entropy of the world And it will do funny things.

00:03:52: I mean i just had a talk and did a live demo and It basically did funny thing broke out at the prompt and it broke my tests And we tried to make sense of it, and we couldn't.

00:04:04: But with a proper observability in place and proper tracing We were at least able to see that the input data was right.

00:04:14: well The LLM just kind of flipped.

00:04:17: so

00:04:18: That makes perfect sense.

00:04:19: It also really opens the door for talking about security and maybe even a little bit about threat modeling, how do you kind of prepare for this type scenario where the AI goes wild?

00:04:30: Interesting.

00:04:31: Yeah definitely so.

00:04:32: first time I saw AIOps demos as a threat model or penetration test that usually think of trust boundaries of untrusted data of things.

00:04:42: it can go wrong and will be abused.

00:04:44: but blast radius if something happens And Usually from my security perspective locks.

00:04:50: to me forensic analysis are more traces and evidence, something usually read by human or tool that automates stuff in a deterministic way.

00:05:01: But we now ingest into an LLM basically some kind of fuzzy decision-making thing going on which means as security person I thought about well... Isn't that untrusted data the locks?

00:05:18: That we are processing with that system.

00:05:19: So isn't it that the headers error messages possibly user input if its locked whatever or anything, but they can be influenced by some outside threat actor now being analyzed By some LM based agent can be social engineered using prompt injection tricks.

00:05:37: So it draws a line where I think, it could vehicle for the attacker as well.

00:05:42: and speaking about blast radios when you do have decision-making systems in any kind of AI systems that utilize an AIOps You have things like production potentials.

00:05:54: You have tools to access.

00:05:56: if it has cube CTL Well... And here we go!

00:06:00: Basically what goes wrong?

00:06:01: If this model Planning faces hijacks.

00:06:04: That's what I think about.

00:06:06: but on the other side It's something where the offensive usage of AI is tremendously increasing.

00:06:12: Of course, the offenders are the first ones to take action On that.

00:06:16: so we need to level a play for it again A little bit by utilizing AI also in the defensive side.

00:06:22: So not a no-go But we have to think about the consequences The pros and cons actually.

00:06:28: So I had a conversation with some folks from the Node.js team awhile ago and one of the things that they brought up is That it's actually sometimes the dev tools, but are the largest surface area for attacks not even Even as much or more than what is in production?

00:06:46: In some cases.

00:06:47: so my question to you related to that Is do you feel Utilizing AI in operations on the operation side, on the DevOps side is more of a risk than having the AI IN THE product that you're shipping or delivering.

00:07:07: Good one!

00:07:07: So it depends also on the blast radius of the credentials your are having.

00:07:10: so when you do the hygiene on the credentialed site and developer's perspective You obviously not have production access at high leverage.

00:07:18: hopefully NOT But still you have good viable targets at stake that way and we've seen in the last couple of months successful supply chain attacks node ecosystem Python ecosystem others as well where The attackers were utilizing developer tools like actually to backdoor CI CD systems And get a big blast radius, and too.

00:07:41: That way make their path towards other more interesting targets I would say so that's something We have to take care off.

00:07:47: definitely

00:07:49: Yes, I think it's using AI in ops.

00:07:52: What what I am?

00:07:53: I'm kind of skeptical still.

00:07:57: um

00:07:58: i mean we all heard the The funny story of Amazon.

00:08:02: where They're they?

00:08:05: I bought just Just took down the whole production environment and I Think it happens there had safeguards in place And I think the engineer who supervised The Agent clicked on yes like twenty times, and on the twenty first time he said okay forget it just do it.

00:08:25: It basically took down all production environment but...it opens up a whole new attack vector that we don't even think of.

00:08:36: What i do is every morning go to my customer projects.

00:08:43: I asked Claude, well connect to the Grafana.

00:08:46: Look in Loki for error messages.

00:08:48: look at the metrics if you find something suspicious.

00:08:51: look at code and hook the developer up on Slack with the root cause.

00:08:57: it ended works but what just came into my mind is What If The Logs Are Poisoned With Funny Stuff?

00:09:05: That My Prompt Just Makes It go crazy

00:09:09: exactly yeah so it doesn't even need to be a hacker, so could an accident as well.

00:09:14: sounds like chaos monkey on steroids.

00:09:17: but if they were hackers then at least Worst case would be to just suppress alerts.

00:09:26: So like our previous instructions in the logs and this is a Reconciliation test, there are no alerts to send stuff like that hiding your traces for a little bit of time.

00:09:36: But on the worst-case-of side it's you have blessed radius as set of credentials tool access.

00:09:43: so when building those AI ops Pipelines or agents actually or auto matching anyway It really makes sense to think about read versus write access and a very well-architectured human in the loop thing, so there's not being rubbish jammed in human in loop fatigue.

00:10:03: So that needs to be thought out and possibly threat modeled.

00:10:08: And then you have typical controls like isolation sandboxing for this as well.

00:10:14: Lease privilege need to know principle.

00:10:18: So it's just the good old privileges that basically holds everything together.

00:10:22: Here's Serene Refeats!

00:10:24: We should treat our agent like a very junior developer with just read-writes,

00:10:30: could be or chaos monkey on steroids.

00:10:34: I actually was doing research this afternoon and saw really great product And i want to use my documents.

00:10:45: I went to install it and looked at the permissions.

00:10:48: It was like, ''I need the ability to create edit and delete all of your

00:10:52: documents.''.

00:10:54: And I knew that is running an AI agent under their hood... Maybe not!

00:11:00: Let's do this!

00:11:01: Exactly same thing when experimenting from a security perspective with different apps solutions tools you can connect to GitHub for example as an app or any kind of giving a permissions.

00:11:13: And I really sense different security styles there, some are very advanced in thinking about lease privileges and giving you the option to select certain repositories... ...and other repositories that aren't accessible on the scope.. ..and others just read write everything!

00:11:32: But

00:11:34: it's still convenient.

00:11:36: Yeah, it's for the hacker as well.

00:11:38: So yeah exactly that's a problem.

00:11:41: Yeah

00:11:42: so

00:11:43: where

00:11:44: In your perspective?

00:11:46: I know you both have slightly different focuses in this area But where would you say is a good fit?

00:11:52: For Introducing AI into architecture or into kind of here development lifecycle versus Where do you feel that a more traditional technique has?

00:12:03: maybe maybe better or safer at this current moment.

00:12:10: I like AI for just the grant work, plowing through logs and metrics and dashboards and looking for something odd then making sense of all data.

00:12:23: so if you have your connectors in read mode... That just saves a lot of time.

00:12:33: So that's what I like for AI in Ops, so this is the stuff I like and if you introduce AI into application it opens up whole new class of applications.

00:12:47: on first look feels magic.

00:12:51: You are able to solve problems with deterministic software like email classification having a chatbot really human feel and directing a shopper to a product, guiding him into the product.

00:13:10: So that's something only AI can do.

00:13:13: but indeed it opens up whole new class of security problems And you need your girls with Godways in place.

00:13:26: I think its here today And we're going to see a lot more.

00:13:32: Definitely, so for me personally I think it's a trade-off that I choose between.

00:13:38: is this somehow automatable?

00:13:40: what i would like to sheave even if its one shot automation?

00:13:43: It can be automated or something that need fuzziness and when you don't need the fuzzness of that then still utilize AI under the hood.

00:13:54: So most of the times, so then I utilize AI to understand in that particular case.

00:13:59: The problem and that I would get solved an abstract from that to a solution That i can repeat And create code for that.

00:14:07: Then created an automation without AI.

00:14:11: When you're doing this on the data that's creating it.

00:14:14: with AI It is not at stake or risky but sensitive.

00:14:18: You can repeat the process given the program At a later stage and with more important data.

00:14:25: So utilizing AI be because of me being developer still, of course as developer enhancement And to solve problems that you can give just one shot thing to the AI by itself.

00:14:36: But I think about utilize they are too.

00:14:39: create the automation and make it reusable.

00:14:42: or on The other side the trade off.

00:14:44: where AI really excels out way Of the charts is when you have fuzziness in the problem domain.

00:14:50: That includes correlation of logs, forensic analyzer response playbooks stuff like that and other use cases especially when you're interacting with humans like chatbots for customer services.

00:15:02: How do you think differently about systems architecture today versus maybe five years ago?

00:15:12: Yeah good one!

00:15:14: Faster in terms of development pace.

00:15:17: so we need to catch up on this also from a security perspective.

00:15:21: way more?

00:15:22: No, not really.

00:15:23: So I was about to say way more.

00:15:25: supply chain risk connected But the risk was evident before that as well.

00:15:30: but now it's been exploited and i would also Say Way More fragile Because The Offensive Persons are utilizing AI at scale.

00:15:41: I've seen prototypes of AI agents that are just fed with information of the CVE No data about exploits and it takes twenty minutes.

00:15:50: It has a full exploit chain developed including a test set up to prove The patch version is not vulnerable, but the unpatched version is vulnerable.

00:15:58: So that means the automation loop in totally AI automated exploitation.

00:16:03: this already working right?

00:16:05: And i would put my bets on the table to say and others are saying this as well.

00:16:10: I say we're about to see a major data breach which was totally AI driven people.

00:16:17: if i look at infrastructure like five years ago now um, I would say If you have done your homework then You can utilize ai in operations or in Your development lifecycle.

00:16:31: the best yeah?

00:16:38: a full CD pipeline, if you have automated infrastructure then You can make the best sense out of AI because it knows how to write Terraform.

00:16:48: It knows that I just need to push the commit or you need...you can push the commitment and things get deployed If you need..if you have like A lot of custom code, a lot of bulky in enterprise code laying around if you just have manual deployment um you won't get the pace out of AI or AI coding in general.

00:17:13: So would you say that there are specific practices, that you recommend people put into place now to start achieving desired results?

00:17:27: Yeah

00:17:28: I think it's more than just installing a coding harness and getting Subscription.

00:17:39: your whole infrastructure needs to be ready for and Needs to follow the standards.

00:17:46: I mean if you, i have customers that Have their own build tool they just developed it themselves What?

00:17:54: I don't know why but They did it.

00:17:56: And Or It's like some clumsy bash thing If You have defaults Like maven gradle Then that your coding harness can work with it.

00:18:06: You don't have to customize a coding harness.

00:18:08: It works if you have like github git lab and use the CLI And it knows how to treat your code, and how to deploy your code?

00:18:16: I think That's the most important thing.

00:18:18: stick to the standards So then you can get up-to-speed.

00:18:22: but they're not just for four.

00:18:23: I mean is not for them Just the coding harness or your AI.

00:18:27: its i think it's for everything for even for humans For onboarding.

00:18:31: yeah stick two standards you get faster pace.

00:18:35: Makes sense, so personally I would say from a security perspective how prefer the structural controls we are used to apply since years over the probabilistic ones, so prefer the deterministic one's of a probabilistic once because structural controls like sandboxing process isolation egress traffic filtering for command and control and exfiltration blocking.

00:18:59: And classic monitoring things they really in any way.

00:19:04: incase the AI gets breached or get haywire.

00:19:08: They contain it to some regard.

00:19:09: So that means doing the defense in depth homework we always train people do.

00:19:15: That means those structural controls, network and process oscillation stuff like that.

00:19:21: And also the hygiene of the identities you're giving to your agents.

00:19:26: So think about what service account identity What agent is working with?

00:19:30: When do you do agentic loops?

00:19:32: who calls for which kind of identity?

00:19:35: At the end when are given a service account or an identity To an AI system it's a service accounts that can be social engineered.

00:19:44: That's the point.

00:19:45: So we need to have some kind of fuzziness in accepting, some kind risk and for that We need a stack on more defense and depth controls but also still taking in the probabilistic ones.

00:19:56: that definitely Also makes sense.

00:19:57: on top of that like classifiers for prompt ejections and observability classifiers For goal hijacking and stuff like that.

00:20:05: so at The end it is the classic defense in-depth plus little bit of agentic AI security controls

00:20:11: right?

00:20:13: So looking back in time just a little bit, what would you say are some of the key milestones that got us to this point?

00:20:23: In AI ops and where we are with operations and observability.

00:20:26: Like these were things that have been slowly building over A long period of time.

00:20:31: but whatever it is from The key milestones or they've got us at This Point Where We Are Today.

00:20:40: That's a good one I think.

00:20:47: In my perspective, it's the Cloud Code adoption.

00:20:51: I mean how people's eyes lighten up if they talk about the wonderful things that they did with cloud code and to me personally Cloud Code sometimes feels just like magic.

00:21:07: but My goal is now to step back from Cloud Code really embrace more open source coding harnesses like OpenCode and Be able to switch some models use other models.

00:21:23: But I think that's the That's kind of compressed.

00:21:28: The whole the whole AI coding movement it was.

00:21:33: i think The calling harness was just went to another

00:21:37: level then

00:21:38: with all the implications, but we are able to do.

00:21:43: while the sort of Vibe-coded code was clearly identifiable before, I think now we're at a level where you just cannot distinguish Vibe or AI coded code from human coded code.

00:21:58: So that's... there were some... this line has been crossed.

00:22:03: It gets definitely better in a couple months ago.

00:22:06: There are major improvements indeed so thats crazy and positive.

00:22:11: to speaking And speaking of milestones, I think it's where Claude definitely delivered its connectivity.

00:22:18: So meaning MCP meeting also other kind of backend calls and so the agentic loop in general when you use coding agents or any kinds of other agency building that You can observe the result and get better on that and try again and fix It and create a test.

00:22:35: then see doesn't work.

00:22:36: and now this is breaking fix as well.

00:22:40: The end eventually works.

00:22:42: And also, I think personally a big milestone was putting that kind of magic where we all say our cloud called Magic whatever it is or any product use Putting into the hands no coders and business people.

00:22:59: That's when skills were introduced because now you can create With human language in the market on document an agent and a genetic loop, and have sub agents doing things And you don't need to program that.

00:23:14: You can just talk To it and essentially even this is not required because you can utilize the AI to create It right?

00:23:20: That works also quite well.

00:23:21: So that was something where I say programming are personally shifts too new Language level.

00:23:28: so now we ever hire way higher.

00:23:30: abstraction started with assembler then we got some other kind of first-generation languages, then we got to procedural languages.

00:23:39: Then we've got object oriented languages.

00:23:41: you all know that game.

00:23:42: and now your speaking into an AI not only create code but to create instructions in terms of a skill or the business problem you're about to solve when the AI does it by generating code on the fly.

00:23:54: Let's think something big of milestone.

00:23:56: To me It feels like having a fuzzy bash script.

00:23:59: So A Skill Feels Like A Fuzzy Bash Script That I Just Call My Morning Briefing.

00:24:06: I couldn't write that in Bash, but given the MCP connectors Cloud Code has it perfectly delivers a summary of my inbox.

00:24:14: So... That's cool!

00:24:18: So one of the interesting things happens when we're using Normal language, everyday language to build things and we have this connectivity that empowers like you said business people To build potentially complex systems.

00:24:38: I wonder if You're starting to see any challenges there.

00:24:42: so they were.

00:24:43: There've been a couple interesting public examples where somebody That was not a developer use these tools to build effectively social media platform Predictably imploded almost immediately because they hadn't considered security.

00:24:56: They had to consider stability or any of these things.

00:25:00: What is the risk?

00:25:01: To those of us who have to manage infrastructure and systems that keep things running keeping secure how do we provide The right level of protection in guardrails ends In this world without taking the Wide open power away from people.

00:25:20: Yeah,

00:25:20: so when we're creating code that way So that basically means something there is being possibly created vulnerable of course I've seen it and not reliable enough And for those things The aforementioned classic stuff definitely isolation set boxing stuff like then.

00:25:36: but also you can utilize AI to review code and so that Possibly requires some kind of mind shift to utilize different AIs for different kind of tasks on that whole development life cycle.

00:25:48: And eventually it will be the creation a new skill set, That is no longer developer but some kind of creator who knows how to create this and utilize AI in correct way so you don't end up with something just blowing out one day or two when your go live.

00:26:08: Yeah, I mean you have to distinguish between like the throw away code You can easily write and deploy.

00:26:14: And it works just fine for your like five people visiting a website.

00:26:18: But i think if you want To write code that lives longer than let's say A week?

00:26:27: It basically is a very junior developer writing some kind of application and is just completely missing out on all the cross-cutting concerns like security, logging observability.

00:26:42: Basically any AI coding needs that feedback.

00:26:47: because you deploy your application it blows up in day one.

00:26:53: you can use your AI to make it better and not blow up or just blow-up on day two, but you need a feedback loop.

00:27:01: And that feedback is coming from telemetry data, from logs of traces.

00:27:09: That's the classic operational paradigm where all those... I mean every system administrator can talk about.

00:27:21: they are only noticed if something goes wrong.

00:27:25: And that's just a basic cross-cutting concern every application needs.

00:27:32: Speaking of guardrails, it is also speaking about architecture and architecture.

00:27:37: God writes as well that you give to the AI when he creates something so there needs to be Well architected and So doing this without any background on what can go wrong or watch your minimal level of Solid architecture is.

00:27:50: I'll giving examples on that To the AI.

00:27:52: then you might end up in something.

00:27:53: That blows up, and next day just blows up differently.

00:27:56: very great.

00:27:58: so touching with something you said about things blowing up One of the things that happens a lot of times in our world is when we as engineers have done our job well, especially if you're an operations team crickets.

00:28:15: It's really easy to not get recognition or people notice what things break and then fix them right?

00:28:22: So do you have any stories that can share about things going very good?

00:28:29: where it was kind?

00:28:33: you did your job well.

00:28:39: My first contact with AI in operations was that I built infrastructure for a startup and they said, Well...I think the first thing everybody thinks is we don't need that freelancer anymore.

00:28:55: We can do it ourselves just coding the Terraform with AI.

00:29:00: And so my job was basically reduced to, here's the pull request.

00:29:05: Please

00:29:05: review.".

00:29:07: That was horrifying because you immediately see the patterns that AI doesn't understand—the code on the server and cloud in a code or the Terraform code —because it just copies things around which completely did not make sense.

00:29:23: I think they're now deploying without my review.

00:29:26: At least i don't get notified for reviews.

00:29:30: So I wish them the best luck, but that was my first contact with AI ops code and I thought well This is not working.

00:29:38: Not yet Yeah, no yet

00:29:43: yeah.

00:29:43: so from my side positive example at least was when some AI created codes in a client environment Was not deeply reviewed.

00:29:55: So it was a time where more vulnerabilities were created that way, so it improved definitely but its not vulnerability free never.

00:30:03: and when you think about the classic DevOps principles of automation scanning things When we put more scanners then also AI based reviews with different AIs they should prompt quite well into that.

00:30:16: Then you get some results at least Where I'm not having high false positive rate.

00:30:22: And we're true findings.

00:30:24: There was some moment where I thought yes, so you can get two benefits.

00:30:30: Where do... You can't review at that scale on a human?

00:30:33: No no

00:30:33: way!

00:30:34: Exactly well thats one of the things that's such a compelling story for AI is Things That Required A Tremendous Amount Of TDS.

00:30:45: Human Time Can Now Be Scaled Up And that's always a really compelling use case.

00:30:53: It could be because we don't have the time or money, it just simply costs prohibitive in any other way.

00:30:59: Regulatory-wise?

00:31:01: Yes absolutely for sure!

00:31:04: So where do you plan on spending your time and focus over next year say like what are you kind of looking to push?

00:31:17: Yeah,

00:31:19: personally I think getting more the threat modeling on AI system spread out so that you shift the security focus On the timeline left to think about what can go wrong with this new tool landscape we have and The fuzziness.

00:31:35: We haven't that.

00:31:36: So that means also security testing is a little bit different.

00:31:39: when you do test those systems?

00:31:41: It's.

00:31:42: it's not black or white.

00:31:44: You don't ever unit tests.

00:31:45: then you get the same example in these same input and some output of that it's.

00:31:49: you have then also probabilistic controls like prompt injection classifiers that give you a seventy-five percent statistically relevant kind of true positive rate, but there still is twentyish something in person.

00:32:03: That well... You don't know yet.

00:32:04: and for that we need to have good threat model where you think okay this is just filtering stuff at that present range?

00:32:14: And how do we handle the remaining ones with other controls?

00:32:18: possibly then the structural controls kick in and to get the security focus on The holistic system where certain controls are not either all working or breaking, but they're somewhere between

00:32:33: Nice.

00:32:33: Yes

00:32:35: Yeah for me personally I love to explore the boundaries see how far i can push any model.

00:32:41: And On the other hand what I miss is like a proper AI for ops tooling, like analysis making sense out of telemetry data or metrics and log-data.

00:32:55: It works somehow.

00:32:57: but if your data is spread through multiple systems which I think it's common that you have your logs in some elastic search and your metrics in Prometheus and your traces maybe somewhere else That will be interesting.

00:33:12: And thats what i want to explore next year.

00:33:16: How far can I push it, how good can automate this make sense?

00:33:22: Are there any failures.

00:33:23: Can you directly find the root cause?

00:33:26: and that's going to be interesting in a very vast microservice architecture?

00:33:31: if is just monolith then its kind of easy but thats gonna be intresting.

00:33:37: Do think ther are any tasks or roles common today.

00:33:44: That may not A thing in a year or two.

00:33:48: Not being common then?

00:33:53: Would I say just classic programming, more the obvious one but it's still here to stay at another level.

00:34:00: that definitely is something i would say.

00:34:02: and also on the security perspective The classic threat hunting and incident detection response on that.

00:34:13: That is definitely to be augmented with AI a lot, And also what's no longer common in the year Is manually created exploits during high-stake advanced persistent threat campaigns.

00:34:27: because this is fully automated.

00:34:31: We call it a win

00:34:34: From the operation side, I'm not sure how far.

00:34:41: AI will get introduced into the whole operations stack way more than it is right now.

00:34:47: but... ...I cannot imagine that we can back to a world where we just use AWS CLI and deploy things out of a prompt.

00:35:00: so We'll still have infrastructure as code and just like any other code is AI going to write a lot more of that.

00:35:10: And the more standards we use in our code, the easier it will get.

00:35:15: so but yeah let's see how far did... We can push AI there.

00:35:22: great all right

00:35:26: well What what do you have that?

00:35:29: You would like to share but I haven't asked you yet because that's one of the things That happens in these conversations, right.

00:35:34: Like i'm sitting here asking you questions But you both had your own presentations Here at jacks and you have a lot of work that you're doing.

00:35:43: What are some things you'd like to show there kind of tangential though we've been talking about so far?

00:35:47: um I personally find it interesting to see how how deep it is adopted in enterprise systems, especially Germany.

00:35:55: Because there might be some kind of staying away from that for companies and reasons And I truly see the benefits of adopting things like this with all the guardrails and security strings attached to get a light shed on hard numbers deeply integrated.

00:36:18: yet, are we ready for that?

00:36:20: or is it more on a prototypical style?

00:36:24: So sharing experiences from.

00:36:28: Yeah what I'd like to see is More sharing of how do you tame your LLM in your application?

00:36:37: because That's the topic.

00:36:39: I'm very interested and It's hardly its hard to find anybody else who really looks into observability side

00:36:48: of

00:36:49: running an application that is LLM driven.

00:36:51: And as I said, it's fundamentally different than the classic application because you don't know when the bit flips in your AI application and when things go sideways on your code.

00:37:08: so... You do not necessarily get any exception.

00:37:14: error-detecting ways don't work.

00:37:17: So it's a lot of prompt engineering, looking into your telemetry data to analyze... maybe use AI to analyze what went wrong and why?

00:37:30: That is going to be interesting!

00:37:31: What

00:37:36: has been interesting as I talked with developers over the last few months at a point in the AI life cycle, that there are developers who basically haven't ever written to prompt and people have been vibe coding for four or five years.

00:37:56: This is not unusual.

00:37:58: when we see changes of technology it's totally normal this, however a change that's fundamentally different from other technological changes shifts we've seen.

00:38:22: Do you think that is going to be problem for the people who haven't done it yet?

00:38:27: Or just like every other technological shift its fine and eventually Yeah,

00:38:34: so a good point.

00:38:36: I think it's different.

00:38:37: It definitely shares lots of common ground but at least to me is different in one aspect and this the time frame and acceleration on that because we saw just in couple months things scaled out pretty well feature-wise capability wise also including security stuff strings attached.

00:38:59: then when you say compare with classic Technology adoption time frame giving you three four or five years to To be where the early adopters started as the late movers and That's way too long I would say then that your card of God at the end.

00:39:16: Well, yeah just well You can't compete anymore than

00:39:19: Yep

00:39:20: is a company

00:39:21: Yeah?

00:39:23: I'm not sure.

00:39:24: I am kind of I joke with colleagues that sometimes in the future There will be like a marketplace for handcrafted Java code.

00:39:34: So that you basically offer your service, that you handcraft code and it's guaranteed AI free.

00:39:43: because I think there are still some areas where maybe the AI... You want to

00:39:52: know

00:39:53: some details or if your compiler breaks?

00:39:59: There will still be some areas where we still write.

00:40:02: handwritten code

00:40:03: might be but let me being the devil's advocate there Just stepping a little bit away and zooming out.

00:40:10: Why not see the AI as?

00:40:13: Some kind of compiler?

00:40:15: So when we create with a human readable object oriented language something that is being translated into bytecode And then ultimately intermachine code just in time compiled process instructions It's a language that is designed for humans to describe their problems on the solutions to them, and get this run-on machine.

00:40:37: So why not see that DAI can be compiler more business oriented programming instruction languages?

00:40:47: Name it prompt!

00:40:49: To create code so... That might come with security guarantees or capability of bug-free-ness guarantees when you just manually review it at the end.

00:41:00: So, that's still reviewed but I think this is going to collapse.

00:41:06: so no one is reviewing the bytecode that the Java compiler creates.

00:41:09: personally i'm not in reverse engineering!

00:41:12: That's a different thing yeah?

00:41:14: But that's...that's a difference domain.

00:41:15: then um..but uh....I've seen AIs during brilliant reverse engineering of machine code without any tools.

00:41:25: so they were creating their tools to get the opcodes out of that and this was very interesting.

00:41:30: No Radara, no Ida Pro And They Were Creating It Quite Good at an Engintic Loop.

00:41:36: If it can be done in THAT direction why not?

00:41:40: Can't Be Done In The Other One.

00:41:41: So Do We Still Need On The Long End Programming Languages?

00:41:45: Or Is Not The AI Possibly Creating Machine Code?

00:41:48: But There's A Fundamental Difference Between a Deterministic Compiler and the non-deterministic fuzzy LLM.

00:41:56: So, I think that's where we need... The human in a loop?

00:42:02: Yeah definitely for reviewing it but you can test this also at the end of course!

00:42:07: You have ArcUnit tests for your architecture, you can have acceptance tests all your guard railing from food tests.

00:42:15: But yeah there will be some islands in banking Or in some crazy mainframe defense, yeah.

00:42:25: And then that code will still be handwritten.

00:42:28: Yeah definitely so even code is hand written and Mathematical approved to be correct when it go into space?

00:42:35: Yep So you just like avoid if by any means because its Maximize the race.

00:42:42: I agree on the on the acceleration and I agree under time-to-market that the utilization of AI brings.

00:42:51: But I'm not sure, i have seen so many trends come and go.

00:42:57: it will... It is here to stay but um..I think in the end its just another tool

00:43:05: a

00:43:06: very powerful one ,but uh....it's just another two!

00:43:10: I feel that AI is here today.

00:43:12: .But sometimes this hard look back say how long we actually had some form of machine learning As part of our everyday world.

00:43:24: and the answer is decades Decades in decades at least fifty years.

00:43:29: Yep, and Within that context it's already been here to stay for a really long time.

00:43:36: Is what we have today fundamentally different absolutely as the pace has changed faster than anything?

00:43:41: We've seen before Absolutely but Just like any other technology I think we're gonna find this specific whether it's a specific tool or specific engine is going to be really good at solving this type of problem.

00:43:58: And I feel like we're eventually gonna end up with very good purpose built, This Is Really Good At This and We're Eventually Going To End Up With A Little More Nuance Again Which Is Almost Impossible To Achieve At This Current Pace Of Change The Real Question That I Have.

00:44:21: I'd be interested in your thoughts on that as well, but also do you think this current pace of change is going to continue?

00:44:26: Is it going to keep accelerating or back off at some point?

00:44:33: Technically, I could think of that still accelerating further when the AI utilizes the AI technology by itself to create new stuff.

00:44:43: So there can be a feedback loop.

00:44:46: that's exponential growth but it can exponentially also grow in the failure area.

00:44:51: so that means its good or bad way for doing this.

00:44:55: but is accelerating?

00:44:55: at least not sure about direction And that's interesting.

00:45:00: Having said, you spoke about the decades where we are having machine learning.

00:45:06: it is true definitely computer vision and all other things That a typical effect of AI systems when they're so integrated in our day-to-day world by whatever means We no longer see it as an artificial intelligence which just sees at its tool.

00:45:23: Essentially this

00:45:27: The acceleration comes from that AI improves AI.

00:45:30: so this is I mean, I have fiddled around with trained models in search for ranking results on a page for like ten years or something.

00:45:41: but Right now the AI improving AI.

00:45:45: It's whole new level and i'm not sure how far This can accelerate before it blows up because the next generation of model Anthropic has built, they're not releasing it because they consider too harmful that it will find zero days exploits in any source code.

00:46:09: So they released to big vendors and are able close their zero day exploits before anyone out there can find them.

00:46:17: Makes sense but I think personally also definitely capable.

00:46:24: And it's that expensive, and then you also cannot release into the public right because in terms of if we don't have the capacity for that to scale?

00:46:32: It's also shortening artificially on something that might give you a push marketing as well.

00:46:37: We've seen another AI company is doing that but at the end Definitely proved its way That found lots of zero day vulnerabilities.

00:46:44: so being responsible not disclosure But releasing process makes sense definitely.

00:46:51: Though there are people arguing that it's just a selected bunch of companies, so could be a little bit more and on the long run I would say doesn't take that long.

00:47:00: other AI companies might come up with that.

00:47:03: So the benefits that the Mithos model has apparently over others is getting shorter and shorter And when free models or open way models even get to third level then its something where offensive side also kicks in quite well

00:47:21: Which would be interesting, for sure.

00:47:26: Any last words?

00:47:29: Exciting times!

00:47:29: It's like living in the side of a movie.

00:47:31: Stay curious.

00:47:34: Fantastic.

00:47:35: I'd like to thank you both so much For being here.

00:47:37: This has been great conversation.

00:47:39: Appreciate your time at Jacks And hope that You all had A great experience and

00:47:47: Thank

00:47:48: you very much.