Episode Transcript
Gary
Okay. Hi, I’m Gary Huberman, CEO, founder Unqork.
Dave
I’m Dave Ferrucci, Chief AI Officer and CTO at Unqork.
Gary
So, now this is gonna be one time I am going to be so out of my league in the room here today. I’ve never felt less technical between Dave and our guest, Kefir Gawrich, who’s CEO of Mindfulness, spelled like mindfulness with the extra A for AI, which is very creative. But Kefir and I go back all the way through about 2013, 2012, when, you know,
I was a CIO at MetLife. He was head of technology at BlackRock. He then went on to become the chief innovation officer at BlackRock. He has been behind the scenes with us since Unqork was launched, helping us from every aspect, from the technical side, which has been amazing. I have never met, I’ve never been in the room with two more technical individuals and my intelligence level is down here. They’re definitely up here. I did though get come armed with a
brand new panel that our chief evangelist Rich Army with, which means I could play the role of the buttons and all these little gadgets here, which I’m gonna do by the way. with Jeopardy there for Dave. Kofir, just give, you have an amazing background. I remember all the way from the Europe to Israel to US to back, just give us as much as you can. What could you share that’s not national secrets here or anything?
Kfir
First of all, thank you very much for inviting me and it’s a pleasure to be here. I’ve been at Unqork indeed over the years. think my first interaction with Unqork that was not you was a forum of your developers in the firm where I did an open mic session. And it’s kind of an AMA and I managed to talk with people about software development, about the future of
automation in software and all kinds of other things. And it was fantastic. I think it was with Tony Kim here. That was when he did the investment and I came over to do the technical. for those who don’t know Tony Kim, let out series C, he participated in B. I have never met an investor like Tony. So he he will literally email me and say, Hey, Gary, I’m coming from the West Coast to East Coast. When do you have time? I’ll sit with you and talk strategy what I’m seeing.
Gary
That’s, I can’t tell you how few investors actually do that reach out.
Kfir
Tony is amazing and I managed to do with Tony some amazing things in the AI world, believe it or not, the last couple of years that they’re quite pioneering if you want. So that’s a very good connection into Unqork.
Gary
And Tony was one of the ones that basically said years ago, SaaS is dead. What’s your, give me your feeling. Let’s jump right in and then we’ll come back to you.
Kfir
SaaS is dead, I’m not sure about it. I think majority of example of SaaS is dead, and again, I’m not, you want to jump directly in, that’s a jump of two trillion dollar loss for the market of software that happened over two days in February. And it’s not a joke, it needs to be taken very seriously. With that said, I think it’s kind of, you know, it’s an 80-20 game, like most of the problems. If you got the piece of SaaS that was doing for you all kinds of things that you don’t need and just a few things that you needed, I assume right now when you’re going to go into these new directions of AI working with your prompt from a line, and later on when we’re going to go into, you know,
native AI companies, this kind of software will be dead. But if you are running a piece of software that is entrenched into process, that is doing a lot of things that AI will not be getting there because regulations, because you are uncomfortable sharing certain type of data sets.
because all kind of other reasons. So that’s kind of the 20 % of the map. They have a great future. They they’re buying themselves time and they will be able to reinvent themselves either with what’s coming up into the agents journey and we’ll discuss about it I’m sure whenever you want. But otherwise with this kind of native AI approach.
They will be saved.
I think that people, need to ask themselves also how this is happening. We had a lot of things when people say things are dead and they are still alive with us, very much alive. I remember when I built at HP a practice called data center transformation with a lot of friends around me, even prior to HP. And people say that data centers are dead because of cloud and data center transformation came to kind of make this transition into it.
Right now, not only data centers are not dead, they are the core of the topic in the world nowadays. You call them AI data center, hyperscale, whatever you want to call them. But at the end of the day, by the way, you know, things they do not live up there in the cloud. They live in data centers. People still don’t figure it out till nowadays.
You are charging these tokens for things that they are happening at the end of the day in data centers. And if you did not pay attention to that, Jensen Huang is charging a lot of money for what he’s putting into these data centers. So data center transformation was something. till nowadays, all over the world, people are talking about the digital transformation.
And the digital transformation might include moving to cloud, doing all kinds of things, touching AI. So that was also a program that a lot of companies, they managed to do. I think the next thing maybe should be something like a SaaS transformation that will basically take SasS for where it needs to be, replaced, augmented, or re-engineered with AI.
And that’s also connected with a big problem that we have nowadays with AI that might be also related to the suspocalypse, right? The story there is people look at these things nowadays in a very simplistic way. And they are asking, you know, how would you look at this; how would you invest in this; how would you approach this from all kinds of ways? And what people forgot.. the basic things of anything about technology at the end of the day, it’s a system. It has block diagrams. It has a way to be analyzed. And people, they just lose it.
Even Jensen Huang, when he’s trying to simplify it for the public to understand that, in particular for investors, right? He’s presenting this, know, layers of a cake and he’s giving an explanation with that. And I think that’s very important for people not to forget about it.
So when you’re talking about SaaS and what this, before it’s dead, it needs to be replaced by something. Spoiler alert, it’s not going to be replaced by one prompt. And the examples of people that they replace it with one prompt like the example of, I’ve been able to replace a workflow into project management with, I don’t know, a few prompts and maybe some base 44, or I don’t know what people did there and all kinds of examples that came into the newspapers. It’s very simple functions. These are not entrenched functions into an enterprise.
When you look at some of these SaaS products, you look into implementation times that they go between months to more than a year. And this is not happening because the companies that are providing these kinds of services are incapable. This is happening because the processes are difficult, because there is governance, because there are things to go through, like data sets, because a lot of reasons.
You know, for me, I had the biggest frustration where basically I was thinking that I can do any kind of corporate application running with AI in three to four months. That was my target
Gary
Wait, just to make sure, Vibe coding, this is you were going to Vibe code anything internal?
Kfir
Before even Vibe coding, before even Vibe coding, just, you know, just do something with AI, ML AI, let’s call it, before even Vibe. Vibe coding came later on.
Right now when we say later, it’s months, right? This is crazy, but that’s kind of how it is. So basically, I was totally disappointed. It took me sometimes six to eight months. And then I started looking around and talking with people that they were as advanced. months to get to the place you needed to Yeah. And so what was the problem?
The problem is just to touch a data set to be kosherized by everybody, including teams that were part of my organization—tree to four months. In order to get access. And then you build the product.
Building it’s easy. Building it’s easy, but right now the data is the cart in front of the horse that is the AI. That’s the problem.
So basically, you know, you’re having this situation and I was very frustrated but when you look at other bodies that they are even more regulated, even more complicated, more bureaucratic, I’m aware about corporations that it takes them 12 to 18 months to put the data set in front of people so they can touch it.
Nowadays, 12 to 18 months, you’re doomed, if you’re a small business. If you’re a big business, if you’re something that running some mega or scale kind of processes, maybe you’re going to be delaying that, but not for long time.
So that’s the complexity of the suspocalypse story. I think it is requiring rethinking. It’s requiring a transformation. people that they sit there, and if you’re even looking into successes that you had with AI, you know that the successes were very fast for things that are like a process, like to replicate something that you can easily explain. That’s promptable.
Vibing them, vibing these things. But in the moment, things are getting complex, touching a lot of points. We spoke before here, before we started the podcast about neuro-symbolics. Imagine that you start running in all kinds of places, in all kinds of neuro-symbolic points that you cannot touch because of compliance, because of all kinds of reasons.
You’re going to be in some problems there. yeah.
Gary
Dave, what do you think?
Dave
About which part of it, that was amazing. I mean, I think I agree with many of the points that being raised. I think we can underestimate – you think about SAS 00 we can underestimate some of the complexity, the touch points, the complex governance and workflows that these things do.
You pick one function and you could vibe code that function, but when you think about all the knowledge that went into the construction of those systems with regard to how much they reflect about that business, you start really needing people who are experts at understanding that stuff.
And the coding is one part, the embodiment of all that business knowledge is another point. So I think it, I sort of agree with Kefir that it’s not going to disappear. These things are not, they’re going to be around for a while. I think prices may be affected. I think that where the, you it’s interesting you raise the point of where they’re going to go.
I think they kind of get cannibalized from a bunch of different vectors, right?
One of them is just people realizing they’re probably overpaying for stuff they don’t really use. And they can now have the opportunity to build that themselves. Another place they might get cannibalized is startups that say, hey, I know a lot about this. I can ground-up, rebuild a lot of this with AI.
So it’s kind of interesting because if I had a better idea, let’s say for a Salesforce or some big SaaS product. And I went to invest it and I said, if I spend $2 billion, I think I can grab 5 % of the market share because I have this differentiation or 10 % of the market share, whatever it is. I’m making these numbers up. They might not work out, but you get the idea.
And he says, you know what, given the risk and everything else and the upfront investment is not worth it. But now if I come and I say I can grab five or 10% of the market share and it’s only going to cost me $20 million, that’s a different investment now.
Kfir
Total different story, but you need to be careful with that. One thing, I will make an observation that AI uses a service, by the way. So maybe even as a service model is not yet dead.
Dave
Well, yeah, I’m sorry. just wanted to, yeah, you’re making, I just want to jump in there because that’s why I don’t like this question, Gary, because SaaS is too broad of a term. Everything is like SaaS. It’s all running on the web. I remember when SaaS was the big idea was you’re not going to send me a CD. I’m going to get a web app. mean, that was the big idea. I think when we say SaaS, we’re referring to these big companies like Salesforce, not to the technology of SaaS.
Kfir
Yeah, but at the end of the day, you cannot run away from the fact that you are, you know, it’s the software and it’s the model. So you need to approach it one way or the other. So the question is legitimate question, you know. I want to give you an example also to be careful with these things. Some people jump on these things and, you know, they started using, you know, one European company I heard about, they basically, they were using all kinds of SaaS products in order to run their call centers, and they fired everybody in the call center, and they stopped using all the SaaS, and they put AI agents to run everything. It took six months till they collapsed.
After six months, they hired everybody and paid more expensive for the SaaS product that they had before here. Real story.
So therefore, you know, it’s, you know… people need to do these steps, you know, fast and slow.
Gary
So I gotta ask, so, you know, there’s very few people I know could answer these questions. So in my mind, when we talk AI in today’s world, there’s the AI agents related to code generation. It’s code gens.
How long have you been doing code generators and like seeing code generators in your life? Like you’ve been doing it forever. Yes. So there’s code generation. We use Claude. We use internally.
Then there’s the AI agents. And these are like the fairy tales. These are like, that’s the example of like, we’re going to put in an agent and it’s going to take over all the back office functions and do all these. But in a regulated business,
How risky is that in your mind?
Kfir
You just can’t do it because first of all, you will not have the level of explainability and compliance and atomicity that you need to have in such kind of a process when you’re running this.
Gary
Explainability, compliance, good.
Kfir
You will need to basically, in order to do that, it will be done. It will be done. I did it in all kinds of ways and all kinds of applications.
For compliance. Been there, done that, works. It was, again, 80-20. 80 % was easy, 20%. 80%, six months. 20%, more than a year. It’s not so easy, but you’re getting there.
When things are starting getting complicated and entrenched, the story will be there, you know, layers of agents that they will need to be, you know, under, you know, a workflow or some kind of hyperautomation, whatever you want to call it.
Look, it’s the models of developing and the models of charge will change because of that. So nowadays we’re talking already about, you know, workers, right?
And the workers are basically a bunch of layered agents that are put together into a process automation that it’s visible, it’s audible. You can audit that, you can understand what it’s doing.
And then things are getting complicated when they are getting too smart. So when people start right now, in some applications they can do it. Financial world or even healthcare, I don’t think that you can do that. When these things are started learning and starting doing things autonomously, this is when the smart becomes a problem. You can do that.
Dave
Well, I, yeah, I was going to say that, you know, you bring up an interesting point, sort of an escalation of the notion of hallucination, right? I mean, hallucination is those two sides of the coin, right? On one hand, when you give an LLM a problem to solve, it’s typically under-specified. A lot of the hard work is specifying exactly what you want the solution to be. But the beauty of using an AI is it fills in all the gaps for you.
So you could be under specified. When it agrees with you, you’re happy. When it doesn’t agree with you, it gives you an unexpected outcome because you didn’t completely specify it. You’re upset and you call it hallucination.
But everything at Philadom was an hallucination. When you have agents and you’re amplifying that across a workflow and they’re making a lot of decisions that you’re you didn’t necessarily specify that they should make or respect responding conditions that you didn’t provide a deterministic program for, you’re going to be surprised, and that’s and that surprise is going to amplify through that entire workflow.
So this becomes extremely unpredictable and it’s extremely interesting because one of the things we thought were really beneficial about computers was that they were deterministic data exactly what you told them over and over again exactly the way you expected them to do it. It’s the advantage over using humans sometimes.
Now we’re kind of building these agents and we’re saying we don’t want the trouble of programming it because that’s too hard. So we’re going to let the agent do all the work, make some of these decisions. Sometimes it’s right, sometimes it’s wrong. This now gets amplified. We don’t understand the risks. They might be tolerable, they might be intolerable. What’s your baseline?
Is your baseline a deterministic process? Because if it is, it’s going to suck. Is your baseline a bunch of humans to make the same number of mistakes? Well, maybe that’s a different baseline. But do they make the same type of mistakes?
It’s kind of a really uncharted territory. We don’t know the answer. And it’s going to evolve. And we’re going to change things. But part of that whole ecosystem of the way it’s happening right now presents this other really interesting phenomenon, which is that you have people who are non-engineers and non-programmers saying, I’m just going to use an agent. So you can tell the agent what to do and plug that into the workflow.
Half the time, those individuals, my claim, my hypothesis, my thesis, is half the time those individuals don’t even understand the difference between needing a probabilistic process because you can’t specify it and a process that’s completely specifiable which you can create a deterministic, error-free thing to do.
So you’re introducing unnecessary risk because of how easy it is for non-engineers to create those things who don’t even understand the difference.
Kfir
You know, you raise a couple of points there that make me smile. One is I heard, you know, lately I read about some people that they developers, they got fired. So they told some stories that suddenly right now companies where they got fired, they basically let business people do vibe coding and the code is going directly into production. Forget about four eyes, forget about everything.
So speak about hallucination. What hallucination? I do not see good things coming out of it.
And to the point that you mentioned related to specification, you know, I would, you know, I want people to pay attention to the latest trend into prompting, which, you know, spoiler alert, it’s look a little bit like coding. If you want to do real good specification. Did you see that lately?
Dave
It’s getting so specific you might as well write the code and this is my point for half the time
Kfir
It looks like it looks exactly like code
Gary
I haven’t seen this.
Kfir
It’s like oh yeah. You basically write the code. But right now you know you’re more vertical because you’re trying to see everything for for basically in order for for these prompts to be understood and to be able to be to run you know a responsible AI. You need to write this different that you will write the problem because the problem runs you know the prompt runs doesn’t matter how you write it. You know you really LLM it You hit, you know, whatever you’re hitting there.
But now, you know, if you write that in the, you know, in the way that, you know, Dave is basically hinting, you need to write it vertically, such basically that all these things, you can see them real time how they are running audit. Yeah.
Dave
Well, this is hilarious. It’s hilarious, but it’s not surprising, Because you can’t get around the fact that if you want to control a formal system, you need a formal language. What’s the difference between a formal language and a natural language? A formal language is unambiguous. You can’t hallucinate. It is what it is.
And so now what happens is we’re realizing, gee, when we use natural language, the AI makes stuff up. No kidding. Because it’s ambiguous. You left things out. Who knows what you really mean. It made a thousand micro decisions when you weren’t looking. And if you want that complete control, you’re going to have to give a complete specification that’s in an unambiguous, meaning formal language. And guess what you’ve done? You’ve reinvented code. You go through pseudo code.
But I mean, it’s…I mean, it’s funny, but it’s ironic.
Kfir
But let’s see what this is doing next. And it’s bringing us back to the point of cost.
Gary
Yes.
Kfir
Suddenly, all these efforts, they are making an AI agent slash future worker, because we’re going to soon change the model of charge, be more expensive than the humans.
That’s what we’re seeing right now already.
Gary
It’s the question we ask is, is your cost going up or down? Because we expect it to go down.
Kfir
You expected, but so what? You know, we expected all the technology, you know, world to be to do the same things that we’ve seen going from agricultural world to industrialized world. which was quite significant. You jumped, don’t know, 50 % to 80%, depending what you’re looking at.
We did not see these kinds of things with technology. We hoped again that maybe right now we’re going to see into this S curve exactly the former, you know, improvement with AI.
Currently, you know, you see when people are going and they’re giving a hug to one of these large AI companies and say, I’ve seen a couple of financial companies around the world that they did that. And when they do that, you see the outcomes of this work. And you see, we managed to save X amounts of tens of hours per month of productivity. That’s why you came to work with AI. You’re in the business of making money. So you basically did not open your data sets of investments and all your investments story to the the AI surface. You just let them look into how your people run operation and deal with productivity.
So this is again, selective AI, selectivity in the stories. And it’s not about the AI companies. They are amazing in what they are doing, and very successful. It’s just when they are tasked, they are tasked with their hands behind.
Gary
It’s interesting because I’ve seen so many times of, and for those who don’t know, when a company is about to IPO, the investment banks make a lot of money during those IPOs. And those investment banks, Musk was the first person who came out openly in the past month or two and said, hey, if you want a piece of our SpaceX IPO, you’re gonna license Grok. But the reality was, that’s what’s always happened.
It always happened was if I’m gonna IPO a company, I’m gonna get the investment banks to be a client first. The problem is a lot of the statements you’re reading about are the productivity gains of these AI solutions are not real. It’s related to an IPO that’s gonna come in the future and I wanna be my seat at the table, which sucks. That’s the reality.
I do like, by the way, that Claude is telling people to take a nap during the day. I don’t know if you read about that, but apparently it’s been telling people it’s time to go to sleep. It’s like you’re working too hard, Graphe. You should go to sleep and take a little nap now.
Dave
Can you imagine? That’s hilarious.
Kfir
Yeah. You know, when they are telling you that you passed your amount of tokens for the session, why wouldn’t it? Another way of telling you that is why don’t you go and take a nap? you’ve so much. It’s awesome.
Dave
I think the prices are going to go up and then I think it’s going to create demand for a much cheaper way to deploy inference and it’s going to start, you’re going to see competition and it’s going to go back down.
So I think we’re going to see, we’ve steadily been seeing a decrease for sure. I think that as the demand goes up and the value starts to be realized, at least in some fields, that it’s price is going to go up and I think that’s going to spark competition.
We already know that you can run inference on small clusters of like M4 M5 chips and apples, apples on five chips and you can do you could do it much more efficiently. We know that you can fine-tune smaller models to be effective in many cases. You don’t always have to be paying the giant, you know, frontier models.
But I think people have to be pushed a little bit. think as the prices go up, they can go down. In other words, the technology is there for this to be done more cheaply. The big cost is in training these massive models. In the last year, think, you know, they evaluate these things. They have an ELO score. And the newest frontier models, you know, there are sometimes double or triple the price of the cheaper ones are technically by the ELO score like 5% better
Kfir
Yeah, the reason for that is, Dave, you did what’s on, you dealt with massive amounts of data. But when I build… you know, financial AI models. You know, just like SaaS that is basically SaaS was sold, you know, you need 20 % of the future, but you’re buying the whole SaaS. And these are the problem with these companies.
The same thing happens with the foundational models right now, because when you build right now something that you need for your application to attack per se, suddenly you switch around four models and average in order to get optimum for what you’re trying to do, because none of them are, you know, singularly killing what you need to do there.
So that’s an interesting direction. Now, when when these kind of things are happening, I will tell you something. I am not convinced that, for example, for financial market, we need these large language models.
I think there is a huge place for small and medium language models and even more need for small, medium and large quantitative models.
So I can see in some of these AI companies either a bold direction or a bad decision on the management by they are starting very much with the agenda to go and develop also small and medium models, that you will be able to take them into your enterprise, into your own environment, into your own private cloud, train them for what you need with your own data in a very easy way that it should be provided for you.
And people are not doing not that, not giving you the provision for that. They want to come and do for you solutions. And then it’s expensive, it’s prohibitive.
Gary
And consulting services. They have built consulting companies.
Kfir
Everybody’s building consulting services. The reason for that, that you need them. You really need them, Gary. People don’t have this kind of expertise inside. I I encourage people, all these big guys to kind of pay attention to the enterprise.
Take a learning from the book of Google. Take a look where TK started there in the cloud business and take a look where it is right now and just learn. What was the what is the story there?
You know, when they pay more attention to the client and that’s let us think biggest thing, right? Always, you know, be a student of the clients. That’s what I learned from him. Lesson number one. And I believe into that one thousand percent. And it’s a mantra.
Gary
So I want to ask, mainfulness, what does the business do? So tell me a little bit about what you created.
Kfir
It is just a platform for me to kind of continuing my ideation sessions to wherever I want to go around the world and to talk with people about problems that they have. But this time, know, when I’m not, you know, I’m allowed to sit on a board. I’m allowed to be on an advisory board. And if this is not…
Gary
So everyone knows the number of times I’ve asked Confer to join us in that respect. And the answer was, I can’t.
Kfir
Yeah, but yeah, you know, the story is, you know, if I don’t do this or that, because of the spectrum of the topics that I’m touching, and maybe that’s related to what we did not discuss at the beginning of what was my journey in my career, I see a lot of things I see sometimes deal flow almost equivalent to a to a VC and not because of my last piece of job at BlackRock, believe it or not.
So I can actually go and invest from our family office. And I’m not doing funds. I’m not doing any kind of just, you know, our small piece into that.
So I do find myself sometimes investing a small amount of money next to Nvidia or to other corporations because of because I what I’m able to bring to the table or because maybe technical side. I don’t know.
Gary
I can see that. This is amazing. By the way, we could go for another five hours. That’s kind of the way I feel here.
One question. So two businesses, one looking at the world from an agentic side, the other one looking at it as, let’s actually break down the problems deterministic, define the deterministic process, run.
Which is a better direction in your mind? Financial services, let’s narrow it down, or healthcare.
Kfir
I don’t think that there is a better direction there. You kind of need to mix and match. I have no doubt about it. Good. You have to mix and match. You will need to basically touch things that are low-hanging fruits. So you’re going to attack them.
If you have, you know, I had… a data factory of billions of data points. You put on that whatever you want, Six Sigma on top of that, you’re still in your hand with 500,000 mistakes. But the data needs to go to a process. You’ve worked in financial services. It needs to go through a certain risk procedure, some Monte Carlo simulation or other kind of things.
Now, you cannot replace the process for replicating the data or giving data to a client that is not going through this process because that’s is your fiduciary responsibility. But you can use a very straightforward process using probabilistic, literally advanced stats, AKA machine learning and basically create a mega table for all these billions of data points to create imputation with a certain level of confidence that you have for that and can serve you for QA, QC.
And then, suddenly with a bunch of people sitting in the middle, you can close these things very, very fast rather than tackling 500,000 mistakes early on.
So the story that remains right now between these kind of approaches is if I’m able to put smart comps in the middle, I can use some AI that has a little bit of more mistakes, but I know how to clear them up on the way.
There are things that it’s zero risk. So if it’s zero risk, guess what? No AI. And I’m familiar till nowadays with a lot of financial services and banks around the world that they say no agent is coming into my perimeter.
And I cannot blame them as long as we even talk like Dave beforehand, he brought even the story of hallucination for companies like this. This is the end of the discussion. What do you mean? I have zero risk. What hallucination are you talking about?
Now, when you’re talking about anything to do with operations and things that have to do with research… You know, I think an augmented approach of people, know, deterministic, non-deterministic type of applications or what is coming next, these layers of agents, AKA workers with, you know, it doesn’t matter how you there is a topology to that, right? So either it’s, you know, plugins, skills, whatever you want to call these things that they are coming in.
And even then, you know, like take the example of legal, you know. Why legal applications are still, companies are very hot. You saw Harvey, Legora, all these. Why? In spite of the fact that you have Claude dropping the legal skills in amazing way and people. Because there are some people in the legal offices they know how to use a little bit of prompting coding, basic things, and they can interact with the Claude environments and they get some very good answers tailored to what they want.
For the majority of the lawyers, they need a workflow. They need a structure. Is this the new SaaS for? I don’t know. But maybe it is. Because if you think about the philosophy of it, it works in the same way, just in the modern topology, if you want. And very successful.
You know, if you ask people in the law offices, they will tell you, this software is nothing. But also, you know, the Claude skill, I don’t know how to use. So what is the answer? you know, I’m here. I need to be here. That’s people that they just are, you know, they read the newspaper of a couple of years ago. So that’s a problem.
So the reality is that this kind of a combination, when you think about it. the introduction of new workers treated, real workers in the organization, that they do not do self-learning in the process of inference. And, you know, we talk a lot about these things, but, you know, Dave, you brought this point of the inference side. You know, how much inference is right now into the market? Not a lot. Therefore, the experience for majority of the people is theoretical.
It’s, I believe that. in general with AI, in all the layers of the cake that Jensen was putting there, we see nowadays a lot of people that hallucinate and not AI that hallucinates – that basically, you know, people are going with interdisciplinary claims that they might not be correct 100%. And we might do this kind of mistakes too from here and there.
So, you know, we just try to speak from experience. That’s the most important thing. But I see into the question that you ask the most important piece of it. We didn’t touch it a lot today. And that was something that I dealt, you know, trying to run forward and put something that is holding me off running.
Gary
What is that?
Kfir
Governance.
Governance, know, cultures of companies need to change. You need to allow people to touch AI wherever they meet AI, wherever they need to meet AI. You cannot stop this from happening. From one hand. From the other hand, you cannot find yourself like the story of majority of the companies, they got hit with, you know, I have a million APIs, I don’t know what they are doing, what’s going, you know, it’s a disaster. Imagine that with agents that actually do things more than an API. that have connections to more than one API.
Gary
And they learn.
Kfir
God forbid, you know, or, you know, or feedback into the into the training models, the results might be catastrophic. So platforms.
So, you know, you need to write the governance. You need to be there before. you know, things are happening. Otherwise, you’re running biography of, you know, you’re running, you know, we had this incident and stuff. Go explain that to the board. It’s not a good story to tell.
All I want to say, there is a big need for, you know, automatic governance that is going into all structure of an enterprise nowadays. You need governance structures that they are live, very important. and documents are okay. That’s the way that we used to run investment committees, you know, all the years.
For AI governance, that’s not good enough. That’s not good enough. It needs to live like, you know, in cybersecurity is the best example, you know, you cannot protect, you know, from agents attack with regular environments that you have, you need to have, you know, the same level of sophistication.
When we are building right now enterprise at the power of agents, at the power of AI workers, we need to have governance that runs at the same power.
Gary
Dave, what do you think?
Dave
No, I I couldn’t agree more that I think it’s not one or the other, you know, because they’re just, we’re addicted to solving a class of problem that requires there’s sort of inference that LLM’s make that very hard to write deterministic programs for. Very hard, if not sometimes impossible because they just didn’t process the data as efficiently as LLM.
So I think enterprise applications are going to be hybrid. And I think that line of saying, should I be using? Where should I allow for or tolerate a probabilistic process where I couldn’t solve it any other way? And where do I really rely on a deterministic process? And just how do you balance those two things? And you kind of have to know what your tolerance level is. have to know what kind of investment you’re making, but I think in the end, it’s going to be hybrid.
And I think with regard, I think there’s just no question about that. And Kofir identified all the different kinds of strategies you have depending on what you’re even trying to compute.
I think the other point on governance, I also couldn’t agree more. I think there needs to be a governance layer. I think there really needs to be, I don’t know if you said this, I didn’t understand the words you said at the last minute, but even a dedicated process, runtime process—that’s constantly governing what’s going on. You want the benefit and the ease of deploying agents, great, but there’s gotta be a traffic cop that knows exactly where the boundaries are.
Gary
I, you know, for me. again, on the non-AI side, I think about every time I had to update a Java virtual machine, a JVM, and what broke. You’re not even changing code, you’re changing the middleware layer, and everything broke, and this was deprecated, this is no longer functioning, this did. And I just can’t picture the models being, even if they weren’t learning models, if they were closed and they knew what they were doing and they did it well, when that model needs to be updated for whatever reason, you’re screwed. You know, you’re going to pray. You’re just going to be like it’s comparing the JVM. It’s like that was easy.
Kfir
So speaking about exactly that topic, it’s connected. You know, you hear more company with AI trying right now to to create self healing startups, you know, in self healing world.
But. you know, with governance, there is less talk about that less companies that they think this way, you know, so I think I want to see how this is going to be solved with self-healing and amount of things that you need to know how to connect. It’s the same problem as the enterprise. I remember with HP even building that kind of call home function for a gen 8 that I own. Not to speak about things that I do not own. exactly. No, because for mission critical, you want…to be called before things are going down, in particular on fault-tolerant kind of topologies.
AI might solve some of these things. On the governance, I think that people did not think about it.But again, you cannot run away with all this without having this solved one way or the other.
Gary
100%. So I think we’re going to do only one interesting lightning round question each. I’m going let Dave pick first and I’ll do second. Dave, pick one. What’s a fast one?
Dave
What’s that? I always like, what did you last binge watch, if anything?
Kfir
I’m, you know, always binge watching Survivor. I enjoy doing that with my daughter, little daughter. It’s kind of… with both of my daughters,. my son is let was less into that, but my daughters they were always into it and it was kind of Hosting.
And lately I actually binge-watch with my wife a series called reset. There is a brilliant comedian by name is our dear Millie and it’s called it’s Reset reset. Yeah, it might be less available for because of language barriers, but Very good for people that they are
Gary
I’m going it I’m assuming there’s captions. Did you ever. My question I’m going to not use one of these. I’m to follow up on. Yeah. Whatever. Have you ever dreamed Survivor would be in the corporate world where you could vote people out in the corporate world. I have. That’s why I’m just curious.
Kfir
No. But what I did with Survivor I created once a management you know exercise that I used Survivor for teams to solve all kinds of problems in order to identify skills for technology and management and I deploy it in all the three regions and I run it myself.
Gary
It’s funny, I always wanted, there was one point when I was at MD at City where we were talking real estate rebuilding buildings and I remember going to the person who ran real estate, sat down and go, I have an idea for you. I said, every conference room should start out where the chairs are all the same height and if you’re speaking and it’s valuable, you lift up with the sound of an angel.
And if you’re speaking and it’s not valuable, the chair goes down to the ground so you’re standing.
Like that’s my only, that’s a survivor, like you should be able to vote people out eventually.
Kfir
I know a couple of companies that will adopt that. Majority of them not so much. I think one day might have worked out before. Yes.
Dave
I think that’s exactly how Ray Dalio imagined it would happen. it would depend on whether or not he liked what you were saying and then it would keep going up. Or you were thinking the way he thought you should.
Gary
I think it’s great. So, Kavir, this is amazing. Thank you for joining us here. Thank you for all the years supporting Unqork and background and just advice. so, congrats on your new business and your family office. And that’s amazing. Thanks everyone for joining us.
Until The next episode of Architecting AI Enterprise, I hope everyone has a great time. Thank you. Thanks. Thank you. Bye.
Dave Bye bye.
Gary Thank you, Dave. We’ll take off the headphones. That was awesome. Thanks.
Dave Thank you. Bye.
