Transcript
Transcript prepared by
Bob Therriault
Ashok Reddy
00:00-00:12
Architecture also becomes important. It's not just the functional coding and generating the language is one thing, but how do I actually generate the right architectures to optimize for any type of particular execution.
Music
00:12-00:23
Conor Hoekstra
00:23-00:42
Welcome to episode 112 of ArrayCast. I'm your host, Conor. And today with us, we have special returning guests and a special guest panelist along with three of our regular panelists. So we'll go ahead and do panelist introductions. We'll start with Bob, we'll go to Marshall, then to Adám, and we'll finish with our guest panelist, Alex.
Bob Therriault
00:42-00:50
I'm Bob Therriault, and I am a J enthusiast. But for this episode, I'm going to be very interested in q and KDB-X.
Marshall Lochbaum
00:51-00:55
I'm Marshall Lockbaum. I'm known for BQN worked with various other A languages.
Adám Brudzewsky
00:56-01:00
I'm Adám Brudzewsky. I work mainly with APL, Dyalog APL these days.
Alexander Unterrainer
01:00-01:06
I'm Alexander Unterrainer. I'm a KDB/q enthusiast and the founder of DevConQ.
Conor Hoekstra
01:07-01:28
And as mentioned before, my name is Conor, host of ArrayCast, massive fan of all the Array languages and excited to talk about the big announcement that we mentioned last episode, but we've got a special guest to talk about that. Before we get to that, though, we have, believe three announcements we'll start with Marshall then go to Adám and then finish up with Alex.
Marshall Lochbaum
01:28-03:49
So mine is less of announcement and more story time, [01] so you may remember a while ago our episode with James Price where he talked about the video games that he'd done for the Commodore 64. I found very interesting that he kind of unearthed recently another APL implementation from 1979. So this is called APLS, and it was the programming language for the Umtek Video Brain, which was kind of pioneering. It was one of the very first kind of attempts at a personal computer. And the obvious choice for anything like this at the time would have been to allow people to program it if they wanted to would have been BASIC. But they decided they didn't want to do BASIC. They instead wanted to do APL. So they made their own kind of simplified APL. It was like redesigned to work with this simple computer. So first they didn't have a big character set. So they used ASCII for everything. Multiplication is actually the letter X. They couldn't, I mean, this was implemented in 13 kilobytes of read-only memory. So they didn't have a lot of, they had a lot of constraints. And they only did arrays of rank zero or one. So you couldn't do bigger arrays. and they had a very limited set of functions for working with arrays. But the interesting thing about this is, I mean, this was not, nobody really programmed this because the company folded and they only finished implementing it like just before they did. So it barely even exists. But two years ago, somebody managed to get a hold of an actual cartridge and dump the ROM. So what James did was he took this and he took an emulator for the video brain and everything. and he was able to run programs on it and see how it worked, which was really cool. And now he's working on this repository, 21 APL games and one-liners for your Umtek video brain. So I found all that really interesting. I wrote up an APL wiki article on this APL implementation. So we'll link that article in the notes. And that article has links to everything you need to know, the history of the machine and the implementation and also James' repository with his code and screenshots of it running.
Bob Therriault
03:50-03:55
And we can link also to the episode we did with James, which I think was a Draculark was the game that he'd made.
Marshall Lochbaum
03:56-03:59
Yeah. And he made another one before that.
Bob Therriault
03:59-04:03
Oh, he's done a lot of cool stuff, but this is really, that's amazing.
Conor Hoekstra
04:03-04:08
Yeah. Awesome. Over to Adám.
Adám Brudzewsky
04:08-05:26
So I'll announce a new round of the APL challenge, which is our quarterly intro APL course slash competition at Dyalog. And while it might at first glance look much like the other rounds we've done, this one is an especially good one to get into if you're interested in this array-programing, but you're not really much into the mathematics because it does not deal with any mathematics more advanced than multiplication of integers. And it especially focuses on some text processing and some patterns and things like that. And actually, this round has gotten some praise from various people for actually doing something interesting at the very end. The last problem is a bit harder. It takes everything you've learned, which is all you need. So the composition itself teaches you everything you need to know as you go along. You don't need to go off and read anything else, learn anything else. And then eventually it implements basically a variation over Stephen Wolfram's formalized automatons, one-dimensional patterns like that, that can implement all kinds of interesting things, including being Turing complete. So it actually allows you to build up a whole computer just for what you learn in this competition. And then you come in cash prices as well, which is also a plus. So challenge.dialect.com is that.
Conor Hoekstra
05:26-05:30
Link will be in the description. And I think that takes us to our final announcement from Alex.
Alex Unterrainer
05:31-06:31
Yes. So those of you who have been following me and DevConQ, they will have seen that we had a community meetup in London in May this year. And as I will be in New York later this year in September, I'm going to have host a community meetup in New York for all KDB developers, kindly sponsored by Data Intellect. So you can find all the details on my blog. I'm going to link the post. Free drinks. So come along. It's going to be fun. We had a lot of people in London show up, around 60 KDB developers. So it would be great if you're listening from New York, make sure to come along. And there will be another one in Belfast, but I still have to confirm the details. So for now, the New York one is live. And if you're in New York and you're interested in KDB, please come along.
Conor Hoekstra
06:32-08:11
Awesome. We will obviously link to that in our show notes as well. You can find it at DefConQ, but you can also find it in the show notes of this episode. And I think with that, it brings us to introducing our third time guest, Ashok Reddy, CEO of KX. If you haven't heard his first two interviews on this podcast, you can check them out. Links will be in the description. I believe it's episode 45 back in 2023. And I think I've got it here. It says episode 94 in 2024. So now it's 2025. Clearly the pattern is we have to have you on at least once a year. And there's no better time to have you on Ashok. On the tail of the announcement that KDB-X, which if you're looking to check this out, I think the URL, although once again, it'll be in the show notes, is kdb-x.kaox.com. you can go and try this new community edition. I myself have installed it. It was flawless. I can't speak for all operating systems, but if you happen to be on Linux, it was very painless. They have a curl command. You hit it, boom, download it, done. I think you have to check your email at one point for some confirmation thing, but no more painful than any other process that works. And then you have access to this new community edition, which is what we have Ashok here to talk about. So maybe we'll throw it over to you. tell us what went into making this happen. I know in past interviews, definitely the last time we had you on, you alluded to some kind of community edition that was going to be more usable and permissible than past versions. So I'll just throw it over to you and you can expound or expand on this announcement.
Ashok Reddy
08:12-13:17
All right, thanks, Conor. It's great to be back here. So I wanted to start off with the community edition, which you mentioned. And the basis for this was really, you know, the feedback we've gotten from the entire community. Since I joined, you know, we, KX, had a KDB Plus version, for example. It's called Personal Edition. It wasn't allowed for people to build commercial products. The second, we had something with the 32-bit version, which is equivalent to about 4 gigs. And you could do anything with it, but it wasn't sufficient. So it really started by focusing on what are the needs of the customers and the community, especially developers and quants. So it started with that premise. And based on the feedback and everything else, the three big areas of focus was it's not just that you need to give something. One is, you know, how do we make something which is truly free for commercial use and with no limitations in terms of the capabilities? So that's one which we have taken to heart. So the KDB-X, it includes all capabilities, and it's limited to 16 gigabytes of RAM, which is almost four times as 32-bit version what we had. The second part of that is we didn't want to just bring out the next generation KDB+. We focused on it not only supports time series, but actual vector or text, both structure, unstructured, support for AI workloads, And the number one requested thing, which is open data formats, supporting for Parquet natively. And we layered in both Python and SQL. So not only do you have the q, now you have SQL and Python natively available. So instead of having different things. And then the third part was, you know, I call this as more like an iPhone moment for KDB. People think about, okay, it's a time series database. It could be analytics. It's a streaming, it's batch, and you have dashboard and visualization. And then we have a KDB.AI vector database and vector search. But really, KDB-X is really all of these together. So you have a one unified platform. You can use the same array-first approach of q workspace in memory. One memory workspace for structured, unstructured AI workloads. And you don't need to bolt on vector databases separately, or you don't need a separate streaming plus batch. So it's really coming together, all of it. So it's free. The second part is extensible. So we wanted to make sure that the module management is something which we have introduced, and there's so much excitement about that. Because what it does is you're able to bring libraries, everything you want people have built. They can bring it, first of all, if you have Python library, any library, you are able to load them up. I think, Conor, you mentioned that your install, everything was so easy. So it's the same thing. If you have an AI library, we used to have something called Fusion libraries. You're able to just load them up, and then you can bring in code, what people have written. I know Alex talks about it in a lot of our communities. People have written a lot of q code over the years. You can refactor and now bring them into as libraries. So that's a fantastic capability. That's why I think about it as an iPhone, means it's like an app store. And instead of having some experts trying to do this, we allow that. And then the third part is really the interoperability and the ecosystem support, which, you know, people, everything with KDB was you had to only support KDB format, and it supported high performance. But most of the people have Parquet data, open formats, and iceberg. So we wanted to make sure that it's supported natively for read, write, as well as analytics. And the add-on part of this is, with the Conor, I know with you being in NVIDIA, [02] we are making GPU workflows native also. So there'll be a KDB-X version for supporting GPU. So we have basically taken things already like as of joins and other things. It's extreme performance. So X is, to me, was extreme, accelerated, and scaling. So it's really about taking any data, just like NVIDIA talks about accelerating computing and CUDA-X. That was kind of, to me, was an analogy, was just like CUDA-X is accelerating compute for a lot of things. KDB-X accelerates data and analytics for anything with time series, unstructured data, and being able to bring that to the entire community. So that's the kind of way it came across was not just bring out KDB Plus and make it free, but really address the community needs. And I know Alex and others already downloaded some papers on it. So we are getting a lot of feedback and we are still in a public preview right now. I expect it to be GA in the next month or two.
Conor Hoekstra
13:17-14:34
Okay. So this is announcements about the current state of affairs and the future, too, because it sounded like the GPU support is something coming in the future. one of my questions is is this a kind of new version of q that's being released because one of the things I did notice when I just ran a couple simple tests is I think in the previous q versions you had some support for the underlying k meaning that if you wanted to do a summation you could still use plus slash as well as the keyword sum which is what's highlighted primarily in the the q, uh, reference in the new KDB hyphen X, um, install, at least in the version I had, you couldn't do the plus slash anymore. I'm not sure if Alex, if you've run into this, but with playing around. So it, it, my first thought when that happened was, Oh, potentially this is kind of a, maybe not from, from scratch rewritten, but it's a, this is kind of a different, um, q. Um, or maybe this is a question for a different person, Ashok. But anyways, that's my first kind of question is this is, is this a kind of different, q than we had downloaded before in the past on the personal edition? Or is it something else? Like it's the same thing, but just with a different paint job or something like that?
Ashok Reddy
14:35-16:16
Yeah, no, I think it's a great question. You know, one fundamental principle for this was we wanted to make sure everything with whatever people are written in q, it's backward compatible. It's actually, it doesn't break anything, right? And the KDB is always known for it. Don't break anything. So that's the number one rule is to make sure that we don't break. If you build something with the new modules or libraries, something that may cause compatibility issues. But generally, one of the things I would say is that q, over the years, actually, we've added a lot of things like any map for storing unstructured data. That came out after the 4.0, enhanced debugger. We've got unlimited connections and pattern matching, a new dictionary notation, JSON encoding, all these things we kind of introduced. But what we have done in this release in the KDB-X is actually brought all of them together, make sure that everything is represented. And so happens, I'm going to announce this in the show here. So Jeffry Borror, who wrote q for models for 4.0, we are working with him to update a version of that to support this latest version of q in KDB-X. So it will have not only all these things brought in, you know, everything from unique sockets and pipes to everything about the algorithms we added. And with previous examples, we are going to update all of them. So it's really to your point that way people will not only be backward compatible, but it's bringing all the different elements and different versions of things to the latest. Plus, people love that book. And so we are fortunate Jeffry is working with us. So that should also become available both online as well as part of the documentation.
Conor Hoekstra
16:17-16:27
Okay. I mean, that is super exciting news. Is there a timeline for the refreshed edition or potentially it's going to be under a new title, TBD? Or is that just coming soon?
Ashok Reddy
16:28-17:21
It's, well, it'll be, I think we are going to probably, you know, this Q4 model sits for 4.0, right? So we'll probably have a slightly updated version of it. But I think it's being reviewed right now. And as it becomes GA, we have got that coming up. And also we are launching a really, I'm really excited about this KDB-X developer center built from ground up. And people are going to love it because it's completely everything we have done. It's documented and it has examples, recipes, including this book. So we're going to use things which makes it easy for people to bring it all together versus right now, if you had to do some Python, There's a PyKx and KDB.ai somewhere else. And so we really focus on developers. I think we really want to make sure that this experience is world class and also the documents, what people are used to, update them so that way you don't have a new paradigm.
Conor Hoekstra
17:22-17:52
Awesome. And we should have mentioned, I mean, thank goodness we have Alex here because clearly my qFoo is not, I guess I haven't been programming enough in q. It was a user error. The expression that I had typed when I was testing it a couple hours ago was plus slash space till space 10, which till is the q keyword for Iota in APL and BQN. What does BQN call it? Index?
Marshall Lochbaum
17:53-17:56
What does APL call it? It is index generator usually.
Conor Hoekstra
17:56-18:00
Technically, that's true. Index generator. But everyone refers to it as Iota in APL.
Marshall Lochbaum
18:00-18:03
BQN actually calls it range the same as Python.
Conor Hoekstra
18:03-18:22
Right. Range in BQN. And apparently that's a syntax error in q. You need to parenthesize the fold expression, the plus slash, which in my memory, I have no recollection of that. But like I said, my QFU is clearly not up to snuff, but Alex has clarified. I'm not sure, Alex, you just unmuted if you want to add anything to that.
Alex Unterrainer
18:23-18:38
Yeah, maybe you were thinking of k. In k, you're right. You don't need the parenthesis. In q, you need them to use the plus operator in a unary form in this case. Okay. I didn't want to call you out.
Conor Hoekstra
18:38-19:23
No, no, no. You got to call. I, of all people, am happy to have my mistakes corrected because, one, I learned. But then, two, the listener doesn't have to be misinformed by my – I mean, that's the worst case. I mean, even in a non-podcast setting, I'm happy to be corrected. But like the most essential place that I'm corrected is while we're being recorded because the harm that is caused is much greater than one person. So, yes, thank you so much for that correction. And it's good to know that this is basically like you said, Ashok, it's compatibility. So you don't need to worry about if you've written some q previously that you download KDB-X and it's no longer going to work. It's the exact opposite. You should expect it to work. And if you don't, you know, probably let someone know because that shouldn't be the case at all.
Ashok Reddy
19:24-19:24
Yeah, yeah.
Conor Hoekstra
19:25-19:46
Yes, on that note, like, the amount I spent with KDB-X and from playing around with it, I had no... I run all my blog posts on the current KDB-X community version to see whether anything didn't work that I showed previously. But it seems to be fine and I haven't found any incompatibility so far.
Conor Hoekstra
19:47-20:19
Yeah, I mean, and do you want to comment maybe? So it sounds like you've been playing around this for a lot longer than the basic research that I did for this interview. What is your experience with the KDB-X and not just the executable, but they've kind of got a new workflow install and, you know, website that's linking to some of the existing docs. Like if you click on, you know, count and whatnot, it'll bring you to the reference documentation before. But they've also got some new things about AI libraries and stuff like that. So maybe, yeah, I'm sure the listener would be interested to get your take, seeing as you've played around with it a lot more than me.
Alex Unterrainer
20:19-21:38
Yeah, sure. Absolutely. As you mentioned before, like the installation process, it just worked out of the box very smooth. Currently, there is an expiry date on because it's the public release version for now, public preview. But as soon and I have been told by the core KDB development team, this is confirmed as soon as they go into general release. this will be removed and there's no more updates needed and the great thing is like from my experiences like you can use it completely offline so you don't need the internet connection which is great so those couple of times I wanted to try something on a plane I couldn't really use it because there was no internet on the plane unfortunately. Installation process very smooth and anything else just works out of the box. For now, the version KDB-X is very similar to what Q4.1 was. The reason for that is what I've been told KDB-X, they want to release a version and then get an iterative flow from the users to feedback what their experience is before they're going for general release. And I think it's a great approach.
Ashok Reddy
21:39-22:41
Yeah, I mean, I think, you know, one of the point you made, Alex, is that it's the transparency part. And, you know, in the past, we didn't share the roadmap or we didn't actually have a community working with the community. So what we've done here is being able to actually say these are the things we are working on, the module management, for example. We have a few customers we worked initially, and now this week, end of this week, we'll have that available for the broader public preview. And then to also support Alex's point, we have some sort of license which expires end of this year. But this is a forever license for people who get KDB-X, the GA version. It will not expire, right? So the whole goal is to give somebody who doesn't have to worry about the license. And we had never done that before. So the people used to always worry about license, finding something. And so we will make sure that the initial release, we didn't want this public tech preview license to be around so that we'll make sure that everybody will have access to the GA without worrying about licenses.
Bob Therriault
22:41-23:10
So this is really a, like, this is a huge step from, like, when we first interviewed you a number of years ago, Ashok, to where you are now, where you're releasing the language for free. I mean, there are limitations. There's a 16 gigabyte limitation, but you've opened up everything else. That's got to be like, without getting into corporate direction and things like that, you've changed the course of the company, essentially.
Ashok Reddy
23:11-25:17
Yeah, I mean, I think, you know, the last time we've talked about, I think one of the things we've announced since then is we got taken private. So we are now privately owned with DA Associates Investing. They're the number one investor in private equity and some of the things what we can do is really focus on long term instead of quarterly you know release and what we needed to for a public company when you have to worry about what our earnings will be so we couldn't do things like this because we thought like if you can't take chances like you know suddenly if it impacts our customer our revenue or whatever so here you know I think it's still we still have to deliver on some of the you know business expectations but generally we do it in the long term so that I think we are partnering with our customers and also being able to do this sort of a thing with the type of talent we can hire now with being able to offer equity for people who can bring on board. So that we couldn't do before. When I talked about initially about separate things, we did Python, but it was a separate product, PyKx. [03] And then we had KDB.ai. So to some extent, we also had things like Insights, which was doing analytics. And it just became more complex in terms of a portfolio versus going back to the roots and saying, let's work with existing customers, make them highly successful, but in the process, get to the new customers, new developers through this community addition and help people to onboard. So that I think is the focus. And then of course, make it, for the next generation AI, I think the type of things we are doing in AI is amazing right now. You know, agentic workflows and people are able to, we have come up with MCP server with almost most of the functions here. so people can incorporate this in the context of the LLM workflows and still get structured, unstructured. I think we are getting a lot of that because it's really truly going back to the functional programming era where we have functions now, except they are, you can also have deep learning and you can learn those functions in a sense, right? So that's part of the flow what we're going with.
Bob Therriault
25:17-25:37
And in the two weeks, I think it is maybe three, you've had this released. Do you see changes in the people? Are there new people coming on in terms of users? Is that something that's a big swing? And that's probably a question for Alex as well, because he's going to have to deal with this in-depth on q, but I'll go to you first Ashok.
Ashok Reddy
25:37-27:47
Yeah, I mean, I think, you know, I just was looking at earlier, we had like 4,400 downloads in two or three weeks. And, you know, we used to get like a few hundred here and there, mostly existing, you know, somebody trying it out. I think we have to get more analytics in some of it because we went from, you know, the reason why people are able to now use it on the plane or whatever is we don't have what used to have call home type of thing. People used to, you know, we used to get more insights into what people are using, but we kind of took the chance. Let's trust the users, but we know who, you know, if they're a new customer or whatever, they may be downloading it. But we'll have more information on it. But generally, you know, the type of things we're seeing is both existing customers. Because what it does for one of the concern in the community, right? It's the community itself is a risk for your existing customers. We have the world's biggest, they take every largest bulge bracket banks, for example. They all use KX, KDB. But their concern is that we don't have enough people with people who have the q and I think the knowledge of that. And so this is expanding that. It's expanding the pool. And then what I'm really excited is our academy program. We have some of the world-class universities, the top. Carnegie Mellon, for example, is a reference for us. So they are teaching this, and they are now promoting it. So we have the MITs of the world, the Cambridge and Imperial College and Princeton PhD classes. They want to use this, so we are making it very easy for them. So I think that's where we wanted to make it easy for the experience because if they have to figure out, you know, how to install and make it work, the recipes, the things what we're doing with tutorials, the examples, the samples at GitHub, but they can load them and they can create classes around it. People can build their own libraries. So that's the, you know, where we think, you know, this needs to go. And that's what I'm excited about is not just trying to commercially try to build something, but let's make it where it really solves problems or if it can do for the world's best universities, the top arts and others, it obviously can help everybody else.
Bob Therriault
27:48-27:54
And Alex, DefconQ, what does this mean for what you've been working on there?
Alex Unterrainer
27:55-29:19
Well, I posted about KDB-X release, like, you know, a summary, [04] what it is, what we can expect, what it means for the community. And I've definitely got like, you know, mainly positive feedback I had people reaching out wanted to know more how to get started where where to find resources and and learn so like from a community perspective I do see like it's it's a very positive aspect um even even from myself you know like like a community additional like a free license that you could work with like has been I don't want to say demanded but like it the kdb developers I know they always wish that we're going to move back in towards that direction or at least you know give the developers something they can play around with in their free time or you know try out and now KX seems to seems to take a step into the right direction so this is It's definitely positive. In terms of whether they're going to get new clients or not. Like I wouldn't be the right person to ask because, you know, I'm my blog is just for educational purposes. And it's more like trying to teach people how to learn the language rather than, you know, attracting new clients.
Bob Therriault
29:19-29:22
Is it going to change the way you approach teaching people?
Alex Unterrainer
29:24-30:50
I wouldn't say it's going to drastically change. solely because, as we know, the community edition comes with certain restrictions and limitations. So it doesn't mean now that if you're a hedge fund and you want to use KDB for free in an enterprise setup, you can now do that with KDB-X. So the limitations are in place to, I assume, keep the business model running the way it's running right now. But it will lower some entry barriers for people who want to, you know, play with it, explore, have a, you know, UIT environment where they are not running production code. And then it has to be seen, like, what's going to be done with the community edition and if there is really a way to build something commercially viable. Like the drawback I see is that the limitations would require you to build and design around those limitations, which at the point where I'm then ready to upgrade, I will basically redesign my system from scratch if I go down the paid version. but we have to see if KX is going a step further or, you know, that's not me to decide.
Ashok Reddy
30:51-33:00
I think to add to that, right, I guess, you know, one of the things people always think about KX and KDB as expensive, and this is kind of the narrative for many years. Part of this is, apart from making it easier and having that transparency on pricing, and people don't understand, people talk about it as like it's a Ferrari and must cost like a Ferrari. So I'll guarantee you that that's part of the things what we'll be announcing is we'll have it on our website for like a 32 gig, for example. It'll be the 20s and a list price. And so it's something where people want to do that. And I would almost challenge anybody in the industry. Like if you just take the entry level pricing for any of the time series databases, vector databases and streaming, add on to it, analytics will be cheaper even just a list price perspective. So it's something that people just have this perception that we're expensive. So I would definitely challenge if it is, as I said, it's like an iPhone. If you're buying something, like I'm buying a camera separately, I do this. And very few people can actually take a batch analytics. Some of the things like ClickHouse can do just batch analytics, and it cannot do streaming analytics. It cannot do things like what we do in AI, and you need to add on vector DBs. and people buy time series databases, vector databases, vector search, streaming, and then you've got visualization. It all comes together, and I'm saying that it'll be something which is, then there's a high performance with upgrades don't kill things, so it breaks. I think that's the value. So it depends. If people have to redesign to work within the free version, we would have done a poor job because if you want enterprise capability support outside the community and it does it give you 20k worth of this thing compared to everybody else yes that's what we need to evaluate and i'll alex you'll be able to tell us that as we kind of go into it and I don't think you know it's not about hundreds of thousands of millions and I think that's what I want to make sure that transparency working with the community and people think that you know oh we once we call the salesperson it costs so much now we will put it on the website and that's the promise I'm making on this call.
Alex Unterrainer
33:01-33:55
Yeah, that's something very great to hear because that's another thing the community is actually keen on getting the details. I, for myself, can only say that after being 10 years in the industry, I would not know what's the price tag of KDB, which makes it hard when you get approached because I've been approached by people who want to start with it and they're looking for a senior developer who can build their platform infrastructure And then when I speak with people and I ask whether they have a license or not, or if they're aware about how much the license costs, like often I've been told they wouldn't know because they haven't spoken or approached KX yet. And I'm not able to comment because I actually don't know what it costs. So this transparency is going to be very helpful for the entire community and definitely the right step to do.
Ashok Reddy
33:55-37:44
Yeah, I mean, I think one of the things, it's also about how we kind of look at people want to build all the things, what the accessibility is one thing. But from a developer perspective, we have put all the qCode linter, for example, the performance profiler, all the visualization libraries and the unit testing is all built in now. And so that's what part of the thing is like, it's not just that. The second part is the scale. So people start with a lot of things. oh, let me get started with something which is free and then maybe it's cheap. But as you try to scale, that's fair. I mean, I'll tell you that one of the largest bank in the world, for example, processes 80 billion records per day in digital records and the volumes of it. And they're our customer. But then, you know, recently a crypto customer who has gone with KDB and KDB-X, they process 800 billion records per day, 800. And scaling from that sort of, if you just look at what it takes, there's nobody even comes close to even processing 80 billion records per day. And then if you want to do both historical batch and real-time analytics together, and then you are also doing. So that's the thing where the idea is that if you start with something like this, a free one, you don't have to change your stack. You can upgrade and scale up. That's the benefit. Just like NVIDIA GPUs, you can scale vertically up first and you can scale out. And that's what we are trying to do here. If you can do that in the more, everybody thought it was memory, but we also help you to scale out. And so we are making a lot of the scale out capabilities also open source. So as we look at the future roadmap, there's not just vertical scaling, but scaling out. And then if you really want to use the GPUs, part of this is the version of the KDB-X, which is supporting. It's not just saying, let me run some things on the GPU, but we are taking the next generation architecture like Grace Blackwell, where there's a memory bus problem between the GPU and CPU. So that's what the Grace Blackwell has solved. Bringing the data to the GPU is the bottleneck. Now, people used to do simulation in the GPUs, but now some of the hedge funds are saying, let me do both structured, unstructured way of training on inference and take the right things you can take to the GPU, but then there's no memory bus problem. So those are some of the things what will be supported in the GPU enhanced version of it. It's not just GPU, but it's a CPU plus GPU with the shared, again, NVLink. So that, I think, is where, again, if you want to scale out and do that, you don't have to change systems in your stack. It's the same unified stack, right? So that's the design principles. And we have tripled the team which is working on the core. I'm telling you that other thing. It's like people thought, hey, we have some of the best core team. The KDB-X team has been tripled, and we've got more outer core people working on it, way of open sourcing it. So that, I think, will allow us to say people are just looking at the cost. It's more about, can it get to me to, I get started today, but I want to go to the larger scale. Whether you're a small hedge fund or a bigger one, you still have to deal with the market data. It's increasing by petabytes, right? It's the amount of data they have to process every day. It's only increasing. And with the tariffs and other things, I get emails from the customers saying that their systems didn't break in the last few months. Everybody has tripled their volumes. And you want that sort of a thing, right? I mean, then it's not only capital markets, so we're being used in aerospace and defense. And all those things, we make sure that it doesn't break. And that's what we need to develop and test. And of course, then our business model needs to support that. So giving something free and getting people to get started, but hopefully that will give them enough value than trying to put entry-level systems on their own.
Conor Hoekstra
37:44-37:50
It looked like you were leaning into your mic, but I'm not sure if my eyes deceive me.
Bob Therriault
37:50-37:53
I felt like I was stepping on your toes.
Conor Hoekstra
37:54-37:57
No, no, no. Go ahead. Go ahead.
Bob Therriault
37:57-38:29
Well, I guess the question I would have would be... well, obviously, if you're expanding your user base by quite a bit, hopefully a lot of people will take it on and start to learn about KDB-X. Customer support, your existing customers, of course, you're going to continue to support the same way you have. And I'm assuming they're happy because they're using it and everything. There must be, I suppose it's the Jeffry Borror book you're talking about, maybe better support to people who are using KDB-X in the public version.
Ashok Reddy
38:29-42:12
Yeah, no, I think the support model for the community edition, we actually have a very good Slack community now for, you know, they've got several thousand people who are signed up. [05] And so it's not just the community, but including myself and the product and R&D team, everybody answers those questions. We're providing samples and other things. So if you don't buy any support from KX, that's one first level of things. And then, of course, the documentation. And we're also working on what we call aq-LLM. So we'll have the accuracy of it. We need a lot more code examples and other things to train and improve accuracy. So we're doing that. The MCP server is another one where it kind of takes all the capabilities we have and is able to bring them as functions in the context of a tool chain. So you can do it in agentic AI. So if you're a quant or a front office or anybody who is not familiar with the array based programming and q, well, you can do it. The best way is to do it in q, but if you want to use some of these things. So that's, and then within the q, I think that's where, you know, I'm looking for customers also who want to participate in this. If they have code, which they can use and train the model and make that available to everybody. And so that's something which I think will be a game changer because I think part of this is array based is, it's got so much of value, but, you know, getting people to train and if everybody has to, not just the learning the language, but you need examples, you know, for example, you know, we kind of have people, everybody has to build an order book. And if you want to bring something where you're taking structured, unstructured, something happened in Japan, a couple of tariff happened, right? So what happens, somebody has to come in and look at all those things to bring the structured, unstructured data and quickly develop the model for it. And it takes a lot of effort to really have all this knowledge and then which LLMs are you going to use and which one has the best backtesting capability. So this is some of the things what I think with MCP, imagine a headline breaks out and then KDB-X, a sales trader can ask, what happened? How did that conversion between US dollars and Japanese yen? And who are my exposed clients? Get that answer in milliseconds. And then it goes and looks at the vector search, joins the pricing data and checks the client positions they have and you do it all automatically. And this is sort of, if you think about it, anything else you can do in any other language or any other system, it takes a lot of things to do. And so that's the, some of the power we want to build into this MCP and being able to bring into this, take advantage of LLMs for what they're really good at and also not allow the hallucinations. So we are bringing that. So one of the largest bank we're working with is got a banker agent. So we are able to bring unstructured and structured together and then try to find what happened and combine structured and unstructured information. But then the documentation, other things will be built in because you're not learning new things, but you're using agentic way of doing things, which all of us are using now the chat GPT 5.0 has come out. It can be part of that workflow also. So without knowing the underlying things, how do I process market data? How do I process sensor data in semiconductors or aerospace defense? So that's something which we're working on. So it's not just the traditional documentation to answer your question, Bob, but also bringing in more easier way for people to understand because they're all moving so quickly today in terms of the gendered AI context.
Bob Therriault
42:12-42:45
And with the LLM q you were talking about, would that be you're taking the corpus of q coding and you're entering that into an LLM? So it's got that corpus to be able to draw from. So hopefully it's a bit more accurate. And I know all of us in the array languages have tried different things with a variety, and they're getting much better at it. I know from working with some of the J stuff, they're getting better, but at fundamental levels, they're still missing the mark by a mile. But I think it's because of the amount of code that's out there.
Ashok Reddy
42:45-43:48
No, it is. It is actually a good point. You know, we looked at, I think Claude [06] is supposed to be the best in terms of the code generation today. So what we found is even for q, Claude seems to be the best in terms of all the LLMs for generating. While they were not trained on q, but the paradigms of how they train, because you've got different aspects of the languages all comes in, so their models seem to be pretty good. But one challenge with the q is, remember, we go from right to left. Most programs, languages are not that way. So that's one of the things is that how you parse and how you do this. It's training, you know, looking at the other languages versus q is different. So that's part of the things what we are trying to make sure that, you know, we build into how we process that and then be able to get enough. You need enough sample code and good samples, not somebody who's just generated, you know, some part time. So we are trying to look for those things, patterns which we can use to train the models and then be able to improve accuracy.
Adám Brudzewsky
43:49-43:55
Do you have any problems with actual, important, useful qCodes being proprietary then?
Ashok Reddy
43:56-46:10
I mean, it is, you know, I think the nature of the customers we work with, there's aerospace and defense or hedge funds and, you know, investment banks. And their secret is the code or some of the things what they've done, right? So what happens is we actually have code, what we have developed for our own products, in what q code we have used and the samples we have created. But it's something which customers, if they are finding some of the companies are starting to, they can actually use it themselves to train and give us the model weights or whatever we can incorporate. So I'm fine with that. We can do that. But it helps them also because it's a risk for people if they don't have the right people and they don't get enough skills. So it's kind of good for them from a risk management. but also as they expand into other areas in the companies. So we start in equities and most of fixed income and derivatives and now in wealth management. Not everybody has those skills. So we kind of have to make it easier. So, yeah, no, I think it's the, like people like Alex probably, you know, he just quickly found, you know, looking at it. So we need like an Alex GPT, right? Somebody who can help us kind of as a human in the loop to not only the code examples, but make sure that finding out what's really good looks like, really great looks like. And it's also people talk about vibe coding, but you really need vibe architectures because I think the part of the problem is that we have, KDB is architected. You've got real-time historical. You've got how you process what is in RDB, HDB, and intraday. And how do I use the GPUs? And so there is a architecture also becomes important. It's not just the functional coding. and generating the language is one thing, but how do I actually generate the right architectures to optimize for any type of particular execution, liquidity management, or if I'm trying to do situational awareness in aerospace and defense context. So those things, architectures, have to be very highly optimized. So you need a q-Vibe coding and a q-Vibe architecture to go along with it. So we need to train based on the architectures also.
Conor Hoekstra
46:11-47:22
Speaking of training, You mentioned module management a couple of times. And when I searched it on the docs.kx.com, the top result was the roadmap as one of the things that you're working on in the future. Can you speak to that a little bit in terms of, I mean, I think not just in the q ecosystem, but in all of the different array language ecosystems, library management, or just how do you share code, even if you don't need a module system? You know, I know Uiua has the ability to kind of do a Go-like, the Go the language-like thing where you can just kind of include a GitHub hash or something like that. And then it's able to import code. But anyway, so it seems like this is something that's on the roadmap for the KX team. Is there like a vision for that? Is it going to be kind of like pip installable stuff or you're just going to be making different currently existing libraries that are out there or maybe internally at KX? What's the vision for the module management? Because that is definitely a frequently talked about topic, but there aren't too many solutions in the array language space at the moment.
Ashok Reddy
47:22-49:09
Yeah, absolutely. So first of all, I think it's not it's available. It'll be available in this next week. I think, in fact, week of August 11th, we have a soft public preview with a few customers right now. So we were getting the code and readme expert feedback for it. So it becomes starting next week. So the module management should become available for the rest of the public preview users. The vision for it really, you know, we have some libraries already. For example, AI libraries. It's called Fusion libraries. And we also release this fuzzy filter, there's BM25, DTW, hybrid. These are all AI libraries. So people can actually bring them. And anything with Python, you can actually load them up, right? So that's the idea is, you know, any of these things people can build or they can bring. So we'll provide some examples of documents. So it's something which we had a few customers working with it. Conor, to your question, for array-based things, I think this is really the way that I think I think about like an app store, right? So the concept is to me, it's like you want anything which built in. We want people or, whether it's Julia or, you know, some of the things what people have done. So we are going to provide some libraries to bring all those things into it. So if you have written any of those things in MATLAB or Julia, [07] we already have code and some of the libraries, but we need to develop and test it. But I ask also the communities who want to participate in it are able to bring that in. And so we want to provide the best surface area for any of the array-based things to be brought in. And hopefully we provide that mapping in the memory and make it easy for people to execute.
Conor Hoekstra
49:10-49:34
So it sounds like definitely a part of this is you have written or are writing libraries that are going to be accessible for users of KDB-X. It also sounded like that if folks have their own q libraries, that they'll be able to, you know, publish or post somehow. That's the goal, is that folks will be able to share and have other people download or use their libraries as well?
Ashok Reddy
49:35-50:50
Yeah, absolutely. I think the developer center, that's part of it, what we want to make it available for people. So in the context of it, there may be some refactoring people have to do just to make sure it fits into that. But we've got libraries like AI libraries. There's a temporal similarity search, TSS, and there is HNSW and things like that we have built ourselves. But we also want the entire community libraries to be contributed. So there is an area where people will have, if they want to make it available, or it's on GitHub. So we will allow people to share. Or if they build proprietary things within the company and they want to reuse across different teams for modularity. So it's really about modularizing what they have and not built in silos. Sometimes companies have multiple teams building it. So more and more we are seeing they're building a machine learning library as a service for real-time things. And so this allows them to do that. So that's the SDK. So it's more not only the KDB-X, it's a platform. So we have a SDK for people to build and publish. And then it's up to customers if they want to provide it internally or they want to share it with the community.
Conor Hoekstra
50:51-51:43
I mean, that's really exciting. I'm excited to, I mean, we're already talking to you right now. And I feel bad whenever I am in the midst of a current thing and I'm like, I can't wait to the next one. You should just appreciate, you know, the current one that's happening. But it'll be exciting, you know, the 2026 when we have you back on to see if there's like a proliferation of libraries that are out in the wild that, you know, folks don't know about. I mean, I'm hopeful that a ton of people will go and start, you know, off with a small ecosystem of libraries, but then potentially will grow as, you know, once the snowball gets rolling, you kind of get some momentum. That would be very encouraging. I mean, only time will tell how willing, you know, all the firms that are making money off of KX and q are willing to share their hopefully not so secret sauce. But, yeah, it'll be exciting the next time we have you on. Bob, I just saw you wanted to weigh in.
Bob Therriault
51:43-52:40
Yeah, in advance of the snowball getting rolling and possibly creating an avalanche, have you thought, like to me, this is potentially a really, really, really big thing. People may not realize the magnitude of what it could become. And you guys have moved from your short-term thinking to your long-term thinking, which I congratulate you on. I think long-term thinking is something that we all require these days. long-term, what do you see the implications of this? Are you able to direct some of the power that you're unleashing into specific areas that might be maybe less financial where it's been used in the past because that could be afforded and maybe into some areas where, well, things like climate change where you're dealing possibly with balancing huge energy systems and it becomes very important to be able to become as efficient as possible.
Ashok Reddy
52:41-54:44
Yeah, no, I think about it. I think the part of the, you know, community edition, making it available, we believe that, you know, the type of technology, what we have can solve so many different problems. And commercially, we are just focused on certain areas now, as part of this is whether it's capital markets or aerospace and defense and others in high-tech manufacturing, but generally there's so many things in life sciences and healthcare and even in the energy we have use cases uh people are using this and uh you know we believe that this allows people to solve and that's the one of the partnerships i'm excited that nvidia also is nvidia has also this whole areas of really big problems the hard problems to solve right that's one of the things what we always have focused on is the biggest and hardest problems is where we can really help and i'm hoping that you know this will although with supporting, for example, the open data formats and others, before even if people wanted to use us, they couldn't use us in this context. Some of these, even the open formats, don't support sub milliseconds, sub nanosecond type of things. We can bring that, and we are able to do that with the most efficient way. We still are in the core KDB. The core technology is less than one megabyte. It uses and whatever. And the amount of power, I mean, I know we have published some things. We'll be doing that more. Energy consumption is a big problem with any of the AI stuff and being able to do that in the most efficient way. And we have, and I am talking to a couple of these different countries. I know we already have the entire country of Poland and Finland using for energy meters. We work with the entire countries. And so we want to be able to help do those things. I'm hoping also that some of the top universities, we can get them to adopt and start using it. And if they graduate and start getting to use it in different problems. And so startups and others also be willing to make it easier for people to adopt this and these sort of problems which matter to the world.
Bob Therriault
54:45-55:49
Yeah, it's a big question. A book I read in the last year was Christoph Becker. I'm trying to remember the title of it. (ed. Insolvent) [08] But essentially it talked about the challenges of technology. and society and the fact that in the past the big issue has been that not had there's not been an element of inclusion so that there were whole groups of people that were ignored as the technologies went forward and that's not a problem I don't think any single person can solve because I think it's something that a society has to solve but I'm not sure how you approach that but if there's any way to make it more available to wider communities and I think your KDB-X is certainly a step in that direction but whenever that can be done I think you're in a much better position to have uh this kind of opportunity this kind of power released to a wider group rather than a small group that might be making decisions right now that not everybody agrees with.
Ashok Reddy
55:49-56:14
Yeah I mean I think for EPAM, which acquired our first derivative, which was our consulting division, they are involved with the World Health Organization and all the United Nations efforts. So we are partnering with them in terms of their working with them to kind of help those organizations. So that's one of the areas where we're looking at in terms of how we can potentially help those more global organizations.
Bob Therriault
56:14-56:15
It's a big deal.
Ashok Reddy
56:17-56:17
Yes, it is.
Conor Hoekstra
56:18-57:10
All right. Well, I see we just crossed the hour mark. So, I mean, a final potential question, and maybe the answer is I've already answered it throughout the whole podcast, is, I mean, we've talked about KDB-X and its release and a few of the things coming up. But you've also mentioned a ton of your roadmap. You've gotq-LLM, the MCP, the book q for Mortals. We just talked about module management. You mentioned that you're working on GPU acceleration as well for some of the, I think you mentioned as of, and potentially some other operations. Outside of all of the future stuff that you're working on that you already have mentioned, is there anything else that you want to detail that you're working on, either in the short term or the long term for KX and KDB-X? Because I don't want to miss anything that you were hoping to chat about or mention while we have you.
Ashok Reddy
57:11-58:20
Yeah, no, I think from this audience perspective, right? I think one of the things with, when we look at the speed and efficiency of our A-based, so we are looking at some just-in-time compilation things to take advantage of next-generation architectures because we always had q was more of an interpreted thing and the performance when you actually look at the new generation thing. So we have made a lot of progress on that. So I'm hoping that some of those things will come out. And the second part, more from a business side, And even where we are, we are looking at not only organically, so hopefully before next time I come, we'll see. We expect to triple or make the company five times in the next three years, so both our acquisitions. So we are looking to acquire companies within this space and to help our customers accelerate their value. So I'm pretty excited about that opportunity in the markets we are in, as well as in other areas. So I won't be able to share that right now, but you'll see that as we announce, you know, as time goes by.
Conor Hoekstra
58:20-58:47
That's very exciting. It must be, yeah, exciting times at KX right now. If that is the vision is that you're looking to grow that much, it would be, I'd be very curious to be a fly on the wall in some of the meetings that are happening behind closed doors. But we, I guess we'll have to wait until the announcements come out for the public. Any other questions while we have, you know, We've got Alex here, but also from our recurring non-guest panelists. Final questions for Ashok while we have him here?
Bob Therriault
58:47-59:11
The only thing I can think of is I'm guessing Nick Pisaris is going to be hot on the typewriter doing new additions for his q-tips and Fun q books, which are, I mean, Jeffry Borror's book is actually well known as well, but Nick's books are as well really good ones for trying to pick up the language. And there may be some amendments made there.
Ashok Reddy
59:12-01:00:12
Nick has been, you know, he's helping us with a lot of the feedback we got around what the KDB-X should be, because it was based also the class he teaches at Carnegie Mellon here in Wall Street area. And so he used to have problems, you know, the students getting started, for example. And, you know, Python, we had lots of different products. So he has been a really integral part of providing feedback. And I'm sure he's looking at how to evaluate and update some of the books maybe. But he also wants to make sure that what he writes is something which is unique and solving real world problems. And so that's his focus. So I think we are going to obviously partner with him to make sure that he still continues to work with Bank of America. But he has a passion for q and KDB. And we are really great to have him as part of the community. But he has definitely been one of the key person to provide the feedback on this team. I think that's what the Carnegie Mellon also was adopting this for their classes and for master's classes.
Bob Therriault
01:00:12-01:00:28
Yeah, we've had Nick on as a past guest and actually another instance of a guest panelist, I think, in the past. I'll check with Marshall on that to see whether Nick qualifies in that area. But certainly he's had input to us as well with the q language.
Ashok Reddy
01:00:28-01:00:29
Yeah.
Conor Hoekstra
01:00:30-01:02:03
Awesome. Well, this has been super exciting. I mean, I'm glad we were able to have, we mentioned last episode that we would get someone on in the next couple episodes. And then I amended that to say few because you never know schedules and whatnot. But we got you on. We're not just someone, but the CEO of KX. And who better to talk about this kind of announcement? And, yeah, I mean, congrats. This is, as Bob mentioned, I think a pretty big deal for folks that, for whatever reason, you know, hadn't gotten interested in q. It's probably, I think, safe to say, easier now than it has ever been before. And it sounds like it's only going to get even easier because you mentioned that the public preview comes with the expiration date. But as soon as that hits GA, it'll basically be a lifetime license. So once you've downloaded once, you'll have it forever or until you delete it off your computer or your computer dies, you might have to download it again. And it's not network-based, which for those that are, I don't know, coding out in the wilderness with no internet connection or flying around frequently. And definitely at least once I was on a plane trying to prepare for a kind of array language YouTube video. And I had to wait until I touched down because at the time I think the flight didn't offer Wi-Fi. Some of them do now, but you have to pay. Anyways, very exciting announcement. If anybody that is listening has questions, either for us on ArrayCast or for Ashok that we can pass along and potentially answer them privately or on air, you can reach us at
Bob Therriault
01:02:04-01:03:00
Contact at ArrayCast dot com. [09] And we've had an uptick in the number of people communicating to us, emailing us their responses and their suggestions, and we very much appreciate them. I think I've replied to everybody who's sent questions in, hopefully appropriately. but also the the I forward them onto the panel as well so people get a chance to see that and react and there have been some really good really good emails sent in and really some really interesting suggestions for future guests that I am now working on but at this point I'm not liberty to disclose but it should be really interesting and and big thanks to Ashok for being so open about a lot of this because I know there's limits to what you can discuss and I think you stay within those limits, at least you have in the past. And we really appreciate the information. This is really very much cutting edge stuff.
Ashok Reddy
01:03:00-01:03:13
Yeah, great. Thanks for having me on here and I look forward to sharing and we can also get some of our key developers to come in and share the time of GA and I'm sure they can get into more technical deep dive than I did.
Conor Hoekstra
01:03:14-01:03:49
And also thank you to Alex for joining us as a guest panelist and be sure folks, I know we mentioned it at the beginning, But if you are interested, not just in KDB-X, but it sounds like there's going to be meetups both in New York and to TBA to be announced, you know, date and time in Belfast. So check out DEFCON q for more of that. And if I was in New York, I would definitely be attending, even if I wasn't interested in q. I mean, it sounds like there's some free snacks and drinks. So, I mean, what better reason to, you know, dip your toe in the array language world than that? with that we will say happy array programming
ALL
01:03:50-01:03:51
happy array programming
Music
01:03:51-01:04:05