Will open context interchange ever be real?
Chapters 13 chapters
About this session
This session brings together the voices defining open context, from Apache Ossie to Open Knowledge Format, to discuss what objects must become portable, what specs we should standardize on, and what protocols we need.
Transcript Click a time to jump there
Bob MugliaLet me start by saying this. Uh, first, let me, uh, we have, I have an incredibly distinguished set of panelists today. I c- I can't imagine a better group of folks to answer some of the tough questions regarding context. Uh, first is Josh Klahr from Snowflake. Product- Prajakta Damle from, from Google. Uh, obviously, they're doing all sorts of interesting things. Uh, Dipti Borkar from Microsoft. Uh, and, and Microsoft's introduced a whole lot of things in this space. And on the model side, we have, you know, one of the leading model providers, Aman Khurana, who's joining us to see the model perspective. So lots of incredible depth. And, uh, I, you know, to me, I, I wanted to, to make this a set of questions that are really useful to everyone and, and frankly are useful to me.
And I, I'm, things that are interesting from my perspective, and I hope that you find them to be just as interesting to you. Uh, to give you an idea, just background, I mean, to me, this is a very important topic. I've been focusing on context in one form or another. We used to call it semantic view, semantic view, views or, or ontologies or knowledge graphs. But I've been focusing on context and thinking about that as a discipline for at least five, five or six years. And it really came from the work I did at Snowflake from the very early days when we were a brand-new business, the cloud was new, user-based pr- uh, usage-based pricing was completely new.
People didn't understand it a lot. And Snowflake had this attribute that really introduced a, a twist to usage-based pri- pricing, which was that our product was very dynamic. It could change in size and change in, in charge costs at any moment in time. And so that's when we introduced the idea of a credit and sort of ex, uh, sort of brought that into the marketplace, and a whole set of attributes about, around a business model associated with that. And frankly, we were learning as we went along. We didn't know exactly what we were doing, so we worked closely with customers who helped us to understand the best way to do this. And I found it to be so darn complex, I needed to write it down.
So my head of finance and I, we created what we called the Snowflake Bible, which ultimately became about a 60-page document that described all of the core business practices we had, the different edge conditions, the so, the, what, how we charge, how we discount, uh, what happens when a customer doesn't renew, what happens when they renew early. All these hard questions that had really never been answered before, we wrote it all down. And, uh, we actually put equations in there. I was very proud of those equations. It talked about the attributes of our business, the core aspects of it. Now, of course, this was a document that people used, and it became the way we ran the company, and it, we kept adding to it over time.
And when I left Snowflake, I said, "Geez, you know, that was incredibly valuable what we did with this bible. I know that other companies don't really have that. What would it be like if, if all companies had some kind of canonical knowledge about their business that was more than just a bunch of Slack conversations, a doc, a few documents here and there, uh, but was actually stored in some sort of centralized repository?" And at the time, we thought about that as a knowledge graph or a semantic model, and I started working with my, my, with, with one company I work with, Relational AI, and my friend Moham Arif on that, and that was early days of thinking about this.
This was before there were models and before the, the ChatGPT moment of 2022. And, uh, and what we discovered pretty quickly is people had a hard time writing all of this stuff down and, and, and, and, and conceptualizing it. And then when the models came out, and certainly by '24, '25, everything changed in at least two dimensions. One, models are great at helping people establish this context and going through all the different sources in a business and, and, and writing it down, and two, models need this information. So with that background, which I think shows you how important this is for everyone, and now taking and internalizing this into core parts of a business process, let me open things up to my distinguished panelists to ask a few questions here.
So the, the first question is, is, you know, context is, is the buzzword of, of 2026. I mean, certainly, certainly, uh, Prukalpa said that at the very beginning today. Uh, but, but P- Palantir talks about ontology as their secret sauce, and then you talk about semantic models, I just mentioned that. How do you define context, and what is the difference between context, a semantic model, and an ontology? Uh, Josh, why don't we start with you? What do you think the differences are?
Josh KlahrOh, you had to pick me first. Um, I, I'll give you kind of maybe, uh, a specific answer, and then I'll generally say maybe it doesn't matter. Uh, but at least kind of at, the way we've been thinking about things at Snowflake is the, the thing we call a sema- semantic model or semantic view, uh, is a definitional object, and it, it has execution. It defines how you calculate something. This is the formula to calculate a metric. Um, uh, often it has execution associated with it, with it, so you can say, "Hey, calculate this metric for me. Select metric from semantic view." Uh, I would say increasingly, um, we're seeing the idea of ontology, which I think also has, is a buzzword and has lots of use cases, uh, at least when it comes to agentic analytics.
The, the ontology is more the roadmap for the business. It's conceptual. It describes ideas, concepts, how they link together. Maybe like the business processes that you were talking about, Bob, around how do you calculate a refund, a refund goes to a customer. Um, and then I would say the, the macro bucket of this, of, like, context is it could be semantic models or views, it could be ontologies, and it could be a bunch of other stuff. It could be, you know, unstructured data, it could be policies, it could be documents. Um, uh, and if you glue that all together ... I actually love your example of, like, the, the bible of how things work for, you know, Snowflake billing.
That bible is context. It has some details about here's how you actually specifically calculate it, it has an ontology that maybe talks about concepts and their relationships, and then maybe there's documents and policies that say, "Hey, this is, this is how we think about things in this particular domain." Um, and so I think that, that, that blob is kind of the overall context story.
Bob MugliaOkay, let me open it up to my other panelists. Any different, any i- any additional things you'd add to that?
Prajakta DamleUm, so plus one to, uh, what Josh said. Um, I think, uh, from the previous session there was a lot of talk about, like, feedback and, you know, there was a lot of talk about capturing the human feedback, capturing, like, how new, new ways of doing things, et cetera. So I think context is certainly broader. I agree with Josh. Like, I, I love the analogy to the Bible, um, uh, Bob, that you mentioned. But context is broader than ontologies and, um, semantic models. Uh, it captures essentially the why and the ever-evolving sort of, like, tacit and tribal knowledge that exists in enterprises. So, uh, it has to be self-learning, it has to learn from how humans are, you know, working, um, in an enterprise, but also how agents are interacting, uh, with humans and with each other.
So, um, uh, that's an important part. Uh, it has to be a ever-living, uh, ever-evolving layer.
Bob MugliaGreat.
Dipti BorkarYeah, completely agree, Bob, with, uh, what they've said. Uh, data, metadata, structured, unstructured. The video recording of this, uh, this panel is context. Uh, and, um, and it's, it's Outlook emails, Teams chat. Uh, context is the new king, and it's different because it's targeted at agents versus humans, and in that, uh, we will have to evolve and innovate.
Bob MugliaWell, yeah. That's a great ... Aman, do you have anything to add at this, a- a- at-
Amandeep KhuranaYeah, I think, uh, uh, the, the, the one piece... So that, that's a broader context. I fully agree with everything that's been said about the broader context, right? So ontologies and semantics kind of play into the context. That's one component of it. I view context, again, we're talking about in the context of agents, I view it as what is the set of things that the model sees the moment it is given a task, right? So it has to draw from a broader pool of data, of information. Um, and semantic layer and ontologies, et cetera, are a component of it.
Bob MugliaGreat. That's really great. Well, I have to say I learned something from that, from those answers. Uh-
Dipti BorkarMm-hmm
Bob Muglia... I, I wasn't sure what the differences were and, and I now have a much better idea, and that's honest, honest, honest and true. Um, second question. Um, uh, everybody's talking about context. I, I joke that, you know, I, I talk to a lot of small companies. A lot of small companies reach out to me and say, "Hey, I'd like your opinion on this, Bob," and I meet with a lot of different CEOs and a lot of different people, and I joke that everybody and their dog is, has got a context solution. And, uh, and I actually think that's true in, in the, in the, in the technol- in the technology space. But one of the things that's unclear to me, and I'm curious, but I've got a bunch of platform vendor companies, so maybe that'll, uh, put the perspective on this.
But will the context live in the platforms that exist? You know, will it i- live in the apps, or will there be independent players that will actually drive context? How do you see this evolve? How do you see this evolving? Aman, well, why don't we just start with you? What do you, what do you think? You guys, you guys are, you guys are in, uh, you know, you are a model pro- provider, but in some senses you're, you have a different role than the other, than the other players do.
Amandeep KhuranaSo the way I think about where does the context live is where does it live today, right? So context, like we talked about, it's a much broader thing than what exists in one system. So the video recording, for instance, is context. Uh, the document is context. Uh, stuff living in SharePoint is context. Google Drive is context. Slack is context, right? Uh, and then you have a system that kind of indexes it all and enables you to find the relevant pieces. Uh, I don't think context is going to live inside a single platform, right? It is spread across the enterprise. The, the thing that enables you to find it and then access it might be a tool, a set of tools, right?
So, but that's not going to own the context. I would, like, I, I would reframe this as the enterprise context is spread across warehouses, apps and chat and all of that, um, and no single platform holds it in my opinion.
Bob MugliaOkay. How about one of our platform ven- ven- ven- vendors? Josh, what do you think?
Josh KlahrUh, yeah, I, I mean, I agree. I think there's ... Kind of there's the physical asset, which is a piece of context or a piece of information, um, and yet ... And that, that could come from anywhere, and it, and ideally might live e- anywhere. I think the interesting, like, innovation and evolution I see is, uh, kind of what does the system look like to, um, evolve that context? And I think this c- this idea of context being something that is evolving and evergreen, uh, is really important. And so the, you know, h- how do you connect to those sources of context? How do you characterize them, group them, make them accessible to an agent, um, and have, like, a curation workflow that actually works to make it usable and always up to date?
I think the systems that do that are, are very interesting. Um, I think a lot of us are probably thinking about how do we make sure we have good systems to help with that kind of context evolution.
Bob MugliaA- and do you think that's Snowflake's role? Or do you think that, that, that that's something that third parties will add, and then you'll interoperate with them?
Josh KlahrYes. I th- I, I think there's always a little bit of, like, like we want to provide that kind of capability to our customers because they're asking for it, but I also think that we are not unique in wanting to solve that problem for, for a customer. So, like, I think, uh, it may happen in, in different lo- different locations.
Bob MugliaOkay. Thank you. Additional perspectives, Prajakta or D- Dipti, what would you, what do you guys think?
Dipti BorkarYeah, I can ju-
Bob MugliaIf you have one
Dipti Borkar... I can jump in. Um-
Bob MugliaOkay. Go ahead
Dipti Borkar... We, we might be, I, I should be highly biased, given Microsoft is such a broad breadth of, uh, everything, right? From Office, Outlook, uh, Teams, uh, data, right, f- uh, uh, AI. Uh, but I do believe that, uh, it's very important for us actually as a, a, a cloud, uh, provider leader to be very, uh, open. It's very, very important for our customers that we actually interop with all the other data platforms across data, metadata, and context layers, right? Um, and we do that a- already today with, with, with OneLake and with we partner with Snowflake, with, with others who are not Databricks, others, right? We interface with, uh, uh, Google and S3, and, and, uh, and it's really important for our customers because they do not want, uh, you know, uh, uh, context only from one system.
They really want, uh, all of it as, as, uh, Josh and, uh, Amandeep kind of called out. It's, it's going to be important to do it for our customers.
Bob MugliaGreat. Well, it's sort of classic Microsoft. You guys have always built a solution- ... That's a bespoke solution for your customers, but then also work very broadly in the industry. How about Google? What do you think, Prajakta? What do you think? What's your perspective?
Prajakta DamleYeah. So, uh, I think, uh, plus one to what Dipti said. I think time and again, we've seen, like, open ecosystems sort of, uh, win, and that's the best thing for our customers. So, uh, similar to, like, data and AI sort of, uh, ecosystems, having context be open, interchangeable, interoperable is, is, is important. Uh, and with that in mind, like, Google certainly has Knowledge Catalog, which is our enterprise-wide self-learning context layer. Uh, but we've also invested and added, uh, the open knowledge format, which is, uh, open format, uh, that is an opinionated way for customers to create their own LLM wiki-like, uh, context bundles and have those be readily available. Uh-
Bob MugliaWell, that's great. We'll, we'll get to that question in just, in just a se- in just a second because it's a real, it's a really important topic when we think about interoperability, which is really a core aspect of this panel. But before we get there, I want to ask my favorite question, and this is the one that I, I ask everybody, and I get different answers from just about everybody. A lot of people I really respect have strongly varying opinions on this question, which is, what's the winning approach to context? Is it going to be stored in markdown? Is it text, or is it in some kind of complex knowledge graph? How do we think about how this is going to be structured?
Um, uh, why don't I st- I don't... Let me start with Dipti on, on this one and see what, what, what you think, Dipti, on that. What, what do you have to say?
Dipti BorkarYes, I might, I might have to plead the fifth because we, we have, uh, Ignite coming up in a couple of weeks, uh, uh, to, you know, we'll have some big announcements. But short answer is, look, MD files continue to stay extremely important, right? Uh, they are necessary but may not be sufficient in getting to the latencies and the performance that, that, uh, our agents need. You need, uh, uh, active curation. You need, uh, in-depth indexing, uh, more semantics, right, to get the, the agent the right answer with tokens, uh, being kept, uh, in mind, right? So if you want it cheaper, better, faster, you're going to need more than MD files.
Bob MugliaOkay. Okay. So let me get ... Let me, let me, let me go to the, the, the, the source itself. Aman, what do you got? What's your perspective? What's the mo- what do the models want?
Amandeep KhuranaWhat the models want. Models love text files. Uh, they're really good at parsing through text files, and lots of them, and pretty quickly. Uh, at the same time, to what Dipti said, certain use cases just need lower latency than reading a bunch of text files, and so preparing in some ways is actually helpful. Uh, I'll tell you a, a, a personal experience. So I was building a context layer for, again, my own personal use case, family context layer that my wife shares, and so forth, and I started with MD files, put them in Git, very quickly ran into security problems. Then I ran into isolation problems. Then I ran into, uh, consistency issues because, you know, I would write things a certain way, she would write things a different way, and then before we knew it, this was a hairball to manage.
Now, this is just two people. Uh, what I found myself doing after that was, well, I'll build this MCP layer on it, and then I'll build this kind of, uh, right side, uh, uh, kinda for differential integrity, I would build constraints on the right side. I was kind of building a database. After, you know, three weeks of doing this, I realized I'm kind of replicating everything that a database does. So I moved it into Postgres, and all of that went away, but then I had to give up on all the flexibility of files, right? Um, I don't think this is settled, a- and I, I... It, it kinda comes down to you have to run the evals for that particular use case, um, and then see what happens.
And even with that, the upkeep is where the hard work is, right? So you'll start with something, and very quickly, if you don't maintain it-
Bob MugliaMm-hmm
Amandeep Khurana... The evals start to go down. And so that's what I've found personally.
Bob MugliaI, I agree with you, it's not settled. It, I... If it's, if it's settled, I keep... Everybody I talk to has it seems to have a different opinion about how it's se- settled. So it's, it's good to see, it's good to see that you guys are thinking somewhat consistently in that sense that we're still figuring this out. Prajakta, what do you, what do you think? Where's Google's perspective?
Prajakta DamleUh, so Google's perspective is in line with what Dipti and Amandeep are saying. I think it's a hybrid. Um, and going back to your financial Bible, um, example, I think, you know, you don't want the agents to read a revenue definition from a set of markdown files. Like, you want that to be pretty authoritative, governed in a semantic model. But, you know, I'm sure this is true everywhere else. At least at Google, we at least have five different revenue definitions. So you also need, like, the ontology and, uh, the, you know, the enforcement around, like, which definitions to use, what are the permissible actions, uh, maybe how to calculate discounts, et cetera. So those things are highly governed, highly curated, and in addition to latency, it actually does impact your accuracy and also your token costs, like tokens that agents spend not only finding the right information, but also reasoning about the information that they find.
So having a more structured part of your context is important, but then, you know, not everything can be captured in that structured, governed, uh, structure. Uh, so there, that's where, you know, your markdown files, your documents, uh, your instructions come into play. So I think it's, it's a combination, um, that, that's, that's useful.
Bob MugliaDo you take s- is... I've got a couple questions to ask you on top of this. Do, do you take the data that's in the unstructured environments and move them into the structured environments as they get curated? Is that one way to think about this, as they become formalized across the organization?
Prajakta DamleThat, yeah, so that could be one way to think about it, where you can extract certain key information from unstructured sources, as opposed to just indexing them or, you know, indexing everything in them, and pretty much, like, bring that into an ontology or even suggest a new semantic model based on the usage patterns that you're seeing across data and users. So, so I certainly see that as a, a graduation. Similar to your data and medallion architectures, I think context has a similar, uh... I, I see context evolving in a similar way. Yeah.
Bob MugliaAnd, and do you... I don't know, you, I don't wanna ask you for pro- on product announcements that haven't been done, but, but Google has a graph database. You have Graph Q because do you think that's an appropriate structure to, to store these ontologies? Would you use your graph solution for that?
Prajakta DamleUh, certainly. Uh, ontologies and the related knowledge graphs as well, which is, you know, like really bringing your unstructured data and creating a knowledge graph. So, uh, that, that certainly is a, is, is a, a great solution for customers to consider.
Bob MugliaGreat. And Josh, I, again, just with you, I don't wanna... I know the, the, don't wanna cross, twist lines if you have, guys have product announcements coming up. But is, is there some c- context you could a- so to speak, context you could add-
Josh KlahrContext
Bob Muglia... To this point about, about snow, what Snow- what Snowflake thinks?
Josh KlahrYeah. I mean, I, I kind of feel like I'm sitting inside our own, like, product discussions about the types of context and structured versus unstructured. I, I, I think-
Bob MugliaThat was the idea. That was the idea, actually.
Josh KlahrYeah.
Bob MugliaSo.
Josh KlahrI, I think the themes I'm hearing are very consistent with what I'm seeing. Like, i- if something is repeatable and you can write a piece of code to do it, um, that makes sense. Uh, like calculating a metric expression, um, by feeding a document to an LLM that says, "Here, generate a, a long physical SQL string that does a thing," is maybe more expensive than saying, "Here's a metric name. I'm gonna calculate it with a compilation engine." And that's how we think about semantic views as a, as a way to compile and execute queries effectively and deterministically. Uh, but yeah, you can't capture everything, uh, as a structured artifact. I actually love the idea of the medallion architecture for, uh, semantics.
I might, I might steal that. But I think there is kind of this graduation of, you know, structured, uh, to unstructured semantics and, um, and it does make sense when possible to have the ability for your semantics to be executable versus just interpreted and, and kind of generating physical queries.
Bob MugliaGreat. Thank you. So I wanna sort of direct this question specifically to a couple of my panelists. Um, I'll start with Prajakta. Uh, with, uh, there's, there's context interoperability I think is very, very important. It's the, sort of the subject of this panel, is how do we enable interoperability between vendors. Everybody in this panel has, has context as a part of their solution. Certainly, if you look at your applications, your Salesforce, your Workday, whatever it is, you're gonna have context as a core part of that. H- how do we get interoperability in this? Uh- A- Apache OSI is, is a solution that is being proposed by a number of vendors. About 50 vendors have joined that.
Um, Google has, I think you mentioned this already, Google has an open knowledge format. Prajakta, how do you think, how do you think di- the... What's the distinguishing characteristics between those two? How should we think about that?
Prajakta DamleRight. So I think, I think this is the hybrid that I was talking about or we were talking about earlier, right? So the OSI, format largely is covering the more structured, more governed context with the semantic models, and I know there are working groups that are also looking at, like, adding ontology to it. Um, with OKF or Open Knowledge Format, we are really tackling the other, the unstructured or the, the marked down part of the context layer. Uh, and the reason we did that is, like, we saw that gaining a lot of momentum with developers, with, you know, AI builders, and we wanted to put out a standard out there that the community can really come, come together on and, and help evolve.
That being said, Google is, um, considering, uh, Google plans to join Apache OSI. We are in conversation with the community, uh, and we are actively actually-
Bob MugliaWe just made news.
Prajakta DamleSorry.
Bob MugliaWe just made news.
Prajakta DamleWe actively... Uh, I, I think, I think we have mentioned that, that we are, we are, we are planning to, and we are certainly, you know, um, uh, uh, in conversation with many of the members. Um, but we also, one additional thing is we are also actively exploring how OKF and OSI come together, uh, because I think it's important-
Bob MugliaRight
Prajakta Damle... To, uh, bring those two together given that, you know, we see context as a, as a much broader, um, concept.
Bob MugliaWell, I couldn't agree more with that. I mean, and I really appreciate that, that answer. Thanks so much. I mean, I, I just think, I think it's incredibly important that the world evolve to a standard that allows for context interoperability. I mean, I've, you know, I've been in this industry for 50 years, and I've watched what SQL has done to-
Prajakta DamleMm-hmm
Bob Muglia... Standardize databases around the world, and the applications and everything that can build. And context clearly is, is critical to everything going forward, and we need a standard that is, is not exactly like SQL, but has some attributes that are similar to that to allow for true interoperability. And so I think it's fantastic that you guys are gonna join. And, and yes, of course, we need to think about the unstructured data, and Google is a tremendous company. I mean, kind of Google def- is defined by unstructured pretty much, right? That's what the whole... Uh, when you think about searching the web, it's all unstructured.
Prajakta DamleThat's right.
Bob MugliaIt's all unstructured, right? So you're the kings of unstructured in some sense. A lot of your input there is great. Aman, what is your, what's your thoughts on this? Uh, love to see Anthropic on that OSI list.
Amandeep KhuranaUm, yeah, so when it comes to standards, I think this matters, right? So certainly we want these standards. At the same time, I don't think there will be one standard to rule them all, right? So keeping it super simple is ... This is our learning from MCP, for instance, right? Keeping it super simple, um, and then working around it has been successful. Um, when it comes to these formats, I think one standard to rule them, rule them all is unlikely to happen. And Bob, you know this better than anybody else. SQL is a standard, you just mentioned, but every database still speaks its own dialect, right? So it's gonna come down to a s- one base standard and then a bunch of, uh, kind of dialects around it which you should be able to translate across.
Now, exactly how that plays out in practice is to be seen. Um, of course, OSI is one, uh, one attempt at it with several companies behind it. Um, yeah, that's, that's kind of my take right now.
Bob MugliaOkay. Well, thanks for that. And, and, you know, your input into how we could bring these things together in a rational way would be great. And I mean, to that point, that really t- that was really a very, uh, good buildup to my next question, uh, which is that, you know, OSI's a ... OSI, which is now, you know, it's being done under the a- Apache, so it's n- it's, uh, fully open, um, is focusing on both dimensional context and full ontologies. It has sort of two di- two aspects to it. And there's... And the other thing you'll find in the OSI spec if you go through, and I do recommend people go through it, by the way, I mean, it's, it's, it's, it's easy to understand at some fundamental level, um, is that there is a canonical model that's being built for both dimensional and for more complex ontologies that is, it, that, that, that has aspects from across all vendors, but then there's an opportunity for every vendor to have their vendor-specific formats and extensions in there.
So to the, to my panelists, my, my two panelists who are, are, are members of O- OSI, I'll start with Josh. How do you think about this? How will it, how is this gonna play out? Is each vendor expected to provide a translator from their proprietary formats into the canonical standard? I mean, this a little bit gets to Aman's point, which is everybody has their own format, and I don't think that's gonna go away, but how do we think about this as we standardize it? Josh, what's your thoughts? You've really been doing a lot of work in driving the standard, so you're the right guy to start with.
Josh KlahrYeah, I think, uh, I'm, I'm glad to hear you've been spending time reading through the, the repo. There is a lot going on there. Um, uh, and it's kind of exciting to see. Actually, since the Apache donation, I think the, the community activity has, uh, has really skyrocketed, which is great to see. Um, and one of the things ... So, uh, y- you called out something that I think is really important for the future of OSI, which is, um, today the expression language for things like complex metrics, uh, is relatively simple in OSI, and what that means is that, um, vendors who are participating, uh, let's, let's talk about, uh, uh, Microsoft, which has, um, complex analytical query languages like DAX or MDX.
Uh, you kind of can, you can create an OSI document, and you can say, "Hey, I'm gonna tell you I have a call-out special expression DAX metric." Uh- What that means for anybody consuming that OSI spec is they would need to write their own DAX translator in order to convert, you know, an OSI document to a semantic view. And I think in the long term, uh, we wanna change that. Um, there's a, a, a working group in OSI focused expressly on, uh, the metric expression language and trying to make it more universal, um, so that there's fewer and fewer vendor-specific dialects in an OSI document, and there's more and more of a standard definition language.
Uh, and so I think that's... I think that is one of the things that will kind of make OSI super successful if we can get agreement on that expression language. I think, Bob, your, uh, analogy to SQL is a good one. Um, uh, the more we can have some shared expression language across, uh, our semantic models, I think the better, uh, the better the ecosystem will be and the better OSI will be. Um, so we're, we're trying to move in that direction. We don't want the future to be everybody has, like, a custom dialect, uh, in the OSI spec.
Bob MugliaThat's great. Thanks, Josh. And again, appreciate your work and leadership in this. Um, Dipti, oh, you guys are new. Microsoft is relatively new announcement to OSI. What, what's your perspective quickly?
Dipti BorkarYeah, absolutely. I think, look, interop is... We, we all agree. I'm surprised how much we've agreed on this panel . I thought we were gonna disagree more.
Amandeep KhuranaYeah, where are the spicy takes? Yeah.
Dipti BorkarIt wasn't as spicy as we thought. But, um, look, everyone wants interop. It's important in every layer. Data, we have open formats, right? Uh, semantic models, ontologies. Now we have, uh, OSI, we have OKF, et cetera. But, um, I mean, Power BI just being that leader in the space, we have so much DAX out there. Um, there's a lot of very advanced metrics, as Josh called out, and it's gonna take a long time to cover this entire surface area. If we take Iceberg as a, as an example of how long it's taken, we haven't even gotten to a spec for security yet for Iceberg or indexing. We just got approved, right? So it's a step-by-step.
It's a great start. It's going to take time and reality to get to coverage of, you know, even 80% coverage, right? So maybe we're at 20% today. It gives some level of interop, but it's, it's complicated because when you get to a higher level abstractions, it's code and logic versus just data format conversions. The higher up the stack you go, the more complex it's going to be. Uh, but it's a good thing we agree that we need interop because we will all work towards it like we did with Iceberg.
Bob MugliaLet's just do the standardization faster than it took SQL to do it, which I think was-
Dipti BorkarYes, yes
Bob Muglia... About 30 years. We have these agents. Last I heard, we had these agents-
Dipti BorkarExactly
Bob Muglia... Maybe they can help us figure this stuff out. Okay, really quickly, we're almost out of time. I'm gonna do a 20-second answer to, to the last question for everyone. This is the rapid-fire question. What's next for context? Josh, what's next?
Josh KlahrUh, what is next for context? Um, I think continued investment on the constant evolution of context. So, like, always... This idea of always up-to-date context that is changing as your business changes, um, I think is, is what's next for context.
Bob MugliaPrajakta, what's next?
Prajakta DamleYeah. So Josh stole part of my answer there, but, uh, always on, but also, like, it's, it's really a continuum. Uh, that's the way I see it from the agent memory to, like, a shared enterprise-wide agent, uh, memory, uh, that can serve millions of agents. So yeah, uh, how do we, how do we build that, uh, enterprise-wide layer that, um, you can depend on for every single agentic interaction?
Bob MugliaDipti, what's next?
Dipti BorkarUm, step one, human curated context. Step two, agent curated context. We are getting there very fast.
Bob MugliaI believe that. That's a good answer. Aman, what do you think?
Amandeep KhuranaUh, agree with everything. I would add the memory piece, right? So episodic memory, short-term memory, long-term memory, kind of the feedback loop. I think that is not very well, uh, defined.
Bob MugliaThis is great. Well, I, I will add one thing, which is interoperability between vendors is, is, is, is, is one of the key next things. I wanna thank my incredible, my incredibly distinguished panelist vendors. This was a fun conversation, and I really appreciate all the insight. I honestly learned a lot. I honestly learned a lot in this panel, which was one of the things I was really hoping for, and I hope everybody else did. Thanks a lot.
Dipti BorkarThank you so much.
Amandeep KhuranaThank you.
Prajakta DamleThanks. Thanks for having us.
Josh KlahrThank you.
Austin KronzSo yeah, thank you to everyone on the panel. Bob, thank you for, for moderating. I think I'm in the same boat as you and, and everyone else, learned something on that panel, so that was, that was really helpful. It's good to hear that there is that level of agreement, so everything should be fine in every organization that's making a decision across these platforms. So that's good.
