Listen on
Overview
Snowflake and Databricks built their June 2026 keynotes around the same sentence: the bottleneck for enterprise AI is context, not the model. They held flagship summits within a fortnight of each other, and the message was identical. When direct competitors converge that fast, the agreement tends to be part genuine and part sales narrative. This episode separates the consensus from the positioning and finds the larger shift underneath the launches.
In this episode of Eventual Consistency, Ross Katz (Principal and Data Science Lead at CorrDyn) and Jason Bradwell (founder of B2B Better, host of Pipe Dream) work through the announcements that matter. They cover Databricks’ LTAP and its promise of one copy of data for transactional and analytical work. They also weigh the race to own the ontology through Genie Ontology and Snowflake’s Horizon Context, and what Genie One says about how mature these agentic platforms are. Ross’s throughline is older than any launch: data work has always been the curation of intelligence, and this is the next chapter of that problem rather than a clean break.
The deeper shift is that data gravity is no longer the moat it used to be. Moving data between platforms and formats has never been easier, so both vendors are racing to become the system through which a business understands itself. That turns the headline question around. The choice now depends on where value comes from in your data, what your data looks like, and who will use it. Most of these announcements are still in preview, so treat the vision as hype until it ships, and treat sitting still as the bigger risk.
Key Takeaways
The context consensus is real and manufactured at once
The market has agreed on where enterprise AI value now comes from: agents that read structured and unstructured data and act without a human drawing every conclusion. The identical keynote messaging is also positioning aimed at existing customers. Signing a contract with either platform does not make your data ready for those agents. Most enterprises bring in only the structured slice, and even that is often modeled for reporting, not reasoning.
Databricks’ LTAP is a bet on where the architecture is heading
LTAP (Lake Transactional/Analytical Processing) runs transactional and analytical workloads on a single copy of data, built on Lakebase serverless Postgres and the Lakehouse under Unity Catalog governance. The data sits once in open formats like Delta and Iceberg, and the query engine decides whether you get transactional or analytical processing. The decades-old OLTP/OLAP split will not vanish for most organizations soon. The real signal is that Databricks is designing for companies that treat the intelligence their production systems throw off as the point, so they architect analytics-first from the start.
The ontology race is why 20 years of failed semantic layers might work this time
Both platforms now lean into an ontology or semantic layer. Databricks has Genie Ontology; Snowflake has Horizon Context, the semantic layer inside its Horizon Catalog. What is different from 20 years of mixed results is the pairing of structured and unstructured data with agents that compile context on demand. Genie Ontology is a self-improving knowledge graph that watches how data gets used and ranks what is likely to answer a question, closer to Google’s approach to search than to a hand-built metrics layer. The prize is becoming the business’s system of record, intelligence, and engagement at once, the strongest lock-in in enterprise data.
Data gravity is no longer the moat, so the platforms compete on capability depth and governance
For years, wherever your data lived was the platform most likely to keep you. That has stopped being true. Moving data between platforms, formats, and SQL dialects has never been easier, so holding data is no longer a sufficient advantage. Snowflake and Databricks now compete on capability depth, operational simplicity, and governance, and above all on the race to become the ontology for your business.
The two platforms evolved from different places, so the choice is persona-driven
Snowflake grew from governed business intelligence and well-structured data warehousing for analysts and business users. Databricks grew from high-volume, high-complexity data science, machine learning, and engineering. Both have since layered on the other’s capabilities and now chase the same new buyers: the CISO, the CFO, and the P&L owner. The practical decision comes down to which of those two starting points your team and data resemble, and where your applications already run across GCP, AWS, or Azure.
No agentic platform is mature yet, and your context defines maturity
No vendor’s agentic platform is mature. Maturity means a system reliably and repeatably produces the output you expect across the full range of tasks you assign it. Reaching it is human work: build a baseline interface, define a benchmark of ground-truth answers, then tune the semantic layer and workflow until outputs match. CorrDyn saw this with a hundred-year-old manufacturer running databases 30 to 40 years old, where moving the data into a warehouse and adding Claude and ChatGPT access still required benchmarking to catch wrong answers.
Related: AI Strategy | Technology Strategy | Episode 23: Credence Goods, Junior Cuts, and the Audit Value Chain
Full Transcript
Jason: Two of the biggest names in data held their flagship conferences within a fortnight of each other this June. Snowflake went first, then Databricks, and both of them independently built their entire keynotes around the same sentence: the thing holding back enterprise AI is not the model, it is the context. When two competitors land on an identical message that fast, you are entitled to be skeptical. Is that the industry agreeing on something real, or is it the same marketing slide rendered in two different brand colors? I come at this from the marketing side, so a message that travels that cleanly always makes me want to check the workings. That gap between the keynote claim and what is actually shipping is where this episode of Eventual Consistency lives. To pull it apart, I am again joined by Ross Katz, principal and data science lead at CorrDyn. We get into the announcements that matter: L-TAP and the promise of one copy of data, the scramble to own the ontology, and what Genie 1 says about how mature these agentic platforms really are. But the through-line is bigger than any single launch. Ross makes the case that data work has always been the curation of intelligence and that this is the next chapter of a very old problem, rather than a clean break. And the real shift is happening underneath the announcements. Data gravity is no longer the moat it used to be, and both platforms are now racing to become the single system through which a business understands itself. For the data leader, that turns the headline question on its head. It is less about which platform wins and more about the question we keep coming back to on this show, which is where does the value in your data actually live? Ross, let’s get into it.
Jason Bradwell: So Ross, Snowflake and Databricks both held their summits over the last couple of weeks, and both built their keynotes around the same claim, that the bottleneck for enterprise AI is context rather than model intelligence. When you’ve got these two direct competitors landing on almost an identical message that fast, is that the industry reaching a genuine consensus, or do you see it as the same marketing pivot dressed in two different ways?
Ross Katz: It’s a little bit of both. I think the market’s reached a consensus about where businesses expect the value to come from from their enterprise data. Snowflake and Databricks both play in the arena where companies are getting value from their data, and it used to be in the not-too-distant past that simply unifying that data and making it available to people in a structured way to create dashboards, to send alerts, to build machine learning models, was enough to drive value from data. But then with the emergence of generative AI and the availability of Claude and ChatGPT and Gemini, and the open models and the proprietary models that Databricks and Snowflake offer on their platforms, there is a new set of opportunities that are available to companies from their data. Broadly speaking, the intelligence does not require a human to draw conclusions from structured and unstructured data anymore. If you construct an agentic workflow to a sufficient degree of clarity for an LLM to work with, then you’re able to do things that you couldn’t do before. But that is not easy to do for a variety of reasons. One of those reasons is that your data lives all over the place. Even after you have contracted with Databricks and Snowflake, you’re probably only bringing a portion, likely the structured portion of your enterprise data into those platforms. The other problem is even if it’s structured or unstructured, it’s likely not clean and ready and in a position where these agentic workflows can take advantage of that data and drive value for your business. So, what both Snowflake and Databricks are doing is saying to their existing customers, you’re already getting value from your structured data primarily inside of our platforms, and you’re starting to experiment with the agentic capabilities that we’ve brought to market and starting to see values with those from primarily your structured data. Now, let’s think about how we expand the aperture of the data that we bring into that ecosystem in a way that automatically keeps the context up to date and available for the agents that you’re creating on our platforms to be able to drive value in your business. That sounds easy in principle and in practice, but over the long horizon of data work, data work in a sense is really just curation of intelligence. And so when you look over the long horizon of curation of intelligence, everything that we’re talking about right now is not that distinct. The number of vendors historically who have launched an internal enterprise search platform is relatively high. It’s a hard problem because very few people were able to do it well historically, and even the vendors who were able to do it well were not able to do it well sufficiently cheaply that they could gain a giant market by doing it. And so the bet right now is that the introduction of these agentic capabilities makes the value of having that sort of enterprise search capability across structured and unstructured data that much higher, and also that the tooling that’s available now can do a better job of organizing the information inside of your business to make it semi-automated in the way that that data gets presented. The other problem with curating intelligence is that it requires a lot of deep thought and expertise from your data team or from people who have a combination of technical knowledge and business domain knowledge and know how your business is organized or should be organized in order to present that data to agents. And so Databricks and Snowflake already have data teams, although they’re different kinds of data teams and we can talk about that if you want, and now what both platforms are trying to do is introduce the capabilities that allow those teams that are already doing curation of intelligence to a certain degree to expand the value that they can drive by semi-automating them, by bootstrapping them to a place of: oh, we can answer questions we could never answer before, we can take actions or execute workflows that we could never do before. And for the primary stakeholders that are on those platforms, the selling point is: we’re giving you these new tools, now you can drive more value internally, which makes you better and also makes your organizations better.
Jason Bradwell: One of the big announcements to come out of the Databricks conference was L-TAP, right? And I think the idea behind that is it’s meant to unify transactional and analytical processing on a single copy of data. And their claim, at least, is that it should collapse this decades-old split between OLTP and OLAP. So do you see that as a real architectural shift, or have we already heard versions of this one-copy-of-data promise before? And if it is real, what actually does that mean? What does it actually change for the data teams that adopt it?
Ross Katz: The fundamental split between OLTP and OLAP—that’s online transactional processing and online analytical processing for people who aren’t indoctrinated—has existed for a long time. What this basically means is that the types of things that you need to do when you’re running a business or an application on a database are very different in kind than the kinds of things you need to do when you are analyzing or attempting to see over the long horizon of your business how things are going and generate insights from the business data that your database might have available. And so the entire emergence of the data ecosystem is in many ways a result of this split, because databases like DB2, SQL Server, MySQL, Postgres—these have existed for a long time, and they’ve been the backends of business systems for decades—Oracle too. But they are not the right tool for analytical questions. And so this is a problem that many people have worked on, that people have thought about for a long time. Originally it was solved by: okay, let’s take a transactional database and let’s move it into a system that is designed to answer the analytical questions. And there have been a variety of systems that have been designed to answer analytical questions, that I won’t go through here, but they’ve evolved over time and become easier to use over time. Then with the emergence of things like federated querying in BigQuery or zero-copy ETL in AWS, you started to have this ability to query your transactional database from within your data warehouse, which is the analytical system, without having to necessarily create a separate data movement intermediary that got the data prepared for analytical work. And recently over the last several years, there’s been an idea that with Iceberg and Delta and emerging data formats that sit in cloud storage, you can do both on top of the same data set, using a new set of technologies that resolves many of the tensions that led this to be a problem in the first place. And so Databricks is offering an exciting new version of this with L-TAP. The idea that you can take your analytical data that’s living in Delta or Iceberg format inside of the Databricks ecosystem and then you can do transactional processing on top of it, and the same—it’s the same data, it just sits there, the difference is the query engine that you use to access that data. So you can do transactional processing on that data using its serverless Postgres approach. You can do analytical processing using its Spark-based platform or Spark SQL. And you can increasingly do low-latency reads and writes on that data as well. So is this problem now solved? I find it to be unlikely that it is fully solved right now given how early the technology is. The nature of the problem is unlikely to go away entirely, but my guess is that the number of organizations that have a business structure that allows them to operate this way will expand over time. And what you’re really giving people is the opportunity to centralize your application development and hosting on the same platform as your analytical work. This is a bifurcated ecosystem right now. You’ll have application teams speaking one language, and you’ll have data and analytical teams or machine learning and AI teams speaking another language. And so this opportunity to unify them so that there aren’t so many arguments about how data should be managed or stored is really beautiful and exciting. But it’s been beautiful and exciting for a long time, and I don’t expect that to be resolved for the vast majority of organizations for the foreseeable future. But what you should see this as is where Databricks sees the puck going, which is that they are going from—it used to be that the reason you made an architectural decision internally was because you needed to make sure that your production applications worked in the best way possible. And so you said: application development team, you make all the choices, then analytical team, you figure it out from there. You get the data out and then you do with it what you need to do. Increasingly the value of an organization is the intelligence that it can build and generate based on the data that’s coming out of those production systems. And so if you’re designing from the beginning or redesigning from the beginning and you firmly believe that the production applications are not the thing, the intelligence that they throw off or have the potential of throwing off are the thing, then you might choose to architect your application with that idea first, with analytical work first, with agentic work first. And Databricks is positioning themselves to capture that emerging type of organization.
Jason Bradwell: I think another one of the big themes to come out of both of these conferences was that seemingly both platforms are now leaning quite hard into this idea of an ontology or semantic layer. So you’ve got Databricks with Genie Ontology, you’ve got Snowflake with a context layer on its Horizon Catalog. And then there’s this argument that these agents are only as reliable as the business context grounding them, which we’ve already touched on. We’ve been trying as an industry to build semantic and metrics layers for 20 years and we’ve had mixed results. So why would building an ontology work this time? And from your perspective, what makes the agentic use case different from the BI use case that came before it?
Ross Katz: What makes it different this time is the opportunity that’s available if you combine your structured and your unstructured data and you allow agents to access both of them in a way that enables them to, either on the fly or in advance, compile the context they need to answer different types of questions. So the promise of these evolving ontologies is that the system will recognize from the questions that people ask and from the information that they use to answer those questions, how the information needs to be organized in order to answer similar questions in the future. So for example, Databricks’ Genie Ontology, they’re connecting to all of these different enterprise SaaS data sources and they’re connecting to all of the data that’s already in your data lake in Databricks and it’s monitoring the way that all of these different sources are used and it’s using a kind of PageRank, a similar algorithm to what Google has historically used in Google search, to construct a way of knowing which results from within your entire business ecosystem are likely to be useful given the question that you’re asking. And then they’re exposing an interface for not just data users, but business users, to ask questions on top of that. Because if you think about these systems, there’s sort of three kinds of systems: there’s systems of record, systems of intelligence, and systems of engagement. So originally Snowflake and to a certain extent Databricks were systems of record—or sorry, they were taking data from the systems of record and going into being the systems of intelligence. Now they are increasingly layering on top the systems of engagement, the agentic interfaces that both data and business users can use. But ultimately if they’re unifying all of that data and they’re constructing that single pane of glass through which all of that business data can be queried, then they’re making an end run for all three—becoming the system of record, the system of intelligence, and the system of engagement. Now, I can tell you that the hyperscalers want to do that too. Claude and ChatGPT and Google via Gemini want to do that too. All of these companies are looking to be that. Because that is the ultimate lock-in and the ultimate source of pricing power in the enterprise space. So the why now is: opportunity is great and the urgency is felt by the business community, the tooling has evolved so you can see on the horizon the possibility of being able to capture that opportunity, of being the ontology for the business, being that structured understanding of how the business works. And also from a competitive dynamics perspective, you can see how this is a road that if you’re able to capture it, if you’re able to build that platform, you’re able to basically make Snowflake or Databricks the firm center of the organization. Not a consumer of system-of-record data, but the most important system of record that the organization has.
Jason Bradwell: If you’re a VP of data and you’re watching these two vendors start to converge on the same context story, the same Iceberg support, the same agentic ambitions, how do you make a platform decision now? Is this convergence making the choice harder, or does it mean that the platform matters less than it used to?
Ross Katz: It’s easy from the outside to think that the platforms are sort of copies of each other, but the truth is that organizations are just like people, which is to say that how they evolved influences what their strengths are and who their customers are, and who their customers are influence the kind of organization that they’re striving to become, because the people they’re talking to are their customers, and those are the people they’re trying to serve. Ultimately Snowflake and Databricks, even though they are held up as in direct competition, they evolved differently. Snowflake’s evolution comes from structured business intelligence, a simple interface for really well-governed data warehousing that analysts and business users can use to answer their business intelligence questions. So governed analytics was sort of the first use case for Snowflake and the primary driver of their growth, and they have been a market leader, if not the market leader, in that aspect of the data space for a long time. Databricks has been much more focused on high-volume, high-complexity data science, machine learning and engineering use cases that are different in kind from business intelligence. So it has been more focused on giving people who are more on the data engineering and data science side a single place to build and manage, relatively easily, these very complex and high-volume, high-throughput workflows for data. And then both of them have layered on top of those machine learning and AI capabilities, and data governance capabilities and AI and agentic AI capabilities. Because the argument is over this emerging AI engineer and where they’re going to do the AI engineering. The argument is also over the new buyers that are coming into the space as a result of where the demand is coming from for these business capabilities. So Databricks would have been selling to the chief data officer or the head of data engineering, Snowflake would have sold to the business intelligence leader or the head of the business line. Now both are trying to sell to the chief information security officer, the chief financial officer, the profit and loss owners at the business level, and what they’re trying to sell is not the technical thing, it is the capability to drive the value in the business. So what would I advise to a platform team that is looking at these two organizations and trying to make a platform decision between them? I would say you have to ask the question, and this is a question we’ve asked consistently on this podcast: where does value come from when you’re using your organization’s data? What are the three things that you really need to be able to do very well with your data in order to drive value from it? And then what does your data look like? How high-volume is it, etc. Where is it coming from, what’s the velocity of the data that’s coming off, etc. And then, based on what you’re trying to accomplish, you look at the capabilities of the two platforms. I can tell you that Snowflake, if you don’t want to have a highly technical platform team managing your data warehouse and you want middle-of-the-road business intelligence and analytics with some machine learning and AI capabilities layered on top, Snowflake is going to be a really strong contender. If you’re looking at more complex, higher-volume, higher-throughput data engineering, machine learning and AI type work, if you’re doing lots of machine learning experiments, if you want to be able to turn the knob of how much you want the platform to own versus how much you want your technical team to own, Databricks is going to offer you more of that kind of environment. So it’s hard because although both platforms can ostensibly do the same things and they have not feature parity but overlapping capabilities, the persona of the people who are going to be using it and the type of organization that’s driving value from it is really going to determine which way you go. And also, it’s not only going to be a decision between Snowflake and Databricks, it’s going to include where your data’s already located. Are you in GCP? Are you in AWS? Are you in Azure? Where are your applications already hosted? What is the best way to get the data from your production systems and into a place where it can drive the value that you’re getting from it? These are all of the reasons why you have a conversation with someone like me and my team, but broadly speaking, that’s where I would point you from the perspective of Snowflake and Databricks.
Jason Bradwell: As we look at wrapping up the interview, a couple more questions. I want to dig into a headline announcement from the Databricks conference specifically, which was around Genie 1. We’ve been talking about that internally as almost like a maturation of their agentic platform. Something that’s moving from something that answers questions to something that produces reports and runs workflows and schedules its own monitoring, things like that. So from where you sit, what separates a genuinely maturing agentic platform from a well-produced demo that you show on stage? And what would you need to see and your clients need to see in a live deployment before you’d officially call it mature?
Ross Katz: This is really hard because nobody’s agentic platform is mature. Anybody who tells you their agentic platform is mature is lying to you. There are degrees of maturity here. There are agentic platforms that are far out ahead of other agentic platforms. But fundamentally, the scope of work that we are assigning to these agentic workflows and agentic platforms is so broad and there are so many potential use cases that it is impossible for me to know what maturity looks like independent of the context from which data is driving value at your organization. For one of our clients, for example, this is a hundred-year-old company with 30, 40-year-old databases underlying their basically manufacturing production floors. We very efficiently were able to move the data into a data warehouse, stand up a semantic layer on top of it, and give them access to Claude and ChatGPT to answer questions coming from that data. Immediately when they’re answering questions from that data, they notice that they’re not getting the right answer all the time. So what do we do? We build up a benchmark for: what is the truth when answering these questions? And this is a simple example. And then what do we do from there? We tune the semantic layer and the agentic workflow to repeatedly and reliably get what we expect from the system based on the questions that we know are going to be asked or the tasks that we know are going to be assigned. That is the work of all agentic AI development. It is creating a baseline interface and access pattern, creating a baseline benchmark, and then comparing what you’re getting out to what you expect. And maybe your expectations aren’t clear, so maybe your benchmark needs to evolve. In fact, your benchmark will evolve. You’re going to get questions you didn’t expect, you’re going to get people asking for tasks that you didn’t expect. The first time somebody asks for a task, unless that task is very much like other tasks that you’ve asked it before, it’s unlikely that your system is going to produce what you expect the first time somebody asks for it, if you did not plan for that in advance. So this is the work that human beings need to do inside of an organization working with agentic AI, is working with the business, understanding the technical components that you can use, and defining: what is the value we expect to get and how does that manifest in terms of the way that the answers get produced and the way that the tasks get completed? There is no one answer to what does a mature agentic AI system look like. A mature agentic AI system is a system that reliably and repeatedly produces the output that you expect and, broadly speaking, covers the entire set of outputs that you expect.
Jason Bradwell: As you take both of these summits together, from your perspective, what did they tell us about where the data industry is heading over the next couple of years? And for the data leaders listening to Eventual Consistency right now, what is the one thing that you think they should be acting on today, and what’s the one thing from these conferences that you think they should treat as hype until it proves itself?
Ross Katz: Okay, let me start with where I think the industry is heading. I think there is emerging consensus about where the opportunities for driving value from data live. And it is increasingly leveraging automated tools to gather and organize information and then answer questions from that information and then execute tasks based on information. When you compose those things together, they start to look like: oh, we can automate work that a human used to have to do. Now it’s still just task-based, we’re not automating jobs, but the number of tasks that we can automate is going to go up, and the more tasks we automate, the more value we drive, the more we can charge for the platforms that we’re developing here. I think that where the industry is going is that we’ve talked a lot about data gravity as being the main moat in the data space, wherever your data is located, that’s going to be the platform that is very, very likely to keep you. But data gravity is no longer the competitive advantage that it once was. It’s easier than it ever has been to move the data around, to change it to different formats, even to change the flavor of SQL or the flavor of transformation that you’re using. And so that’s no longer a sufficient moat. So what these platforms are trying to do is establish new moats that keep customers coming back in the form of capability depth. What are you able to do at a level that no other platform is able to do it at, a level of simplicity in operations, at a level of governance? And these are the ones that are most applicable to Snowflake and Databricks. But when you think about the hyperscalers, the public clouds, you’ve also got distribution and then the availability of capital, which in practice means the availability of compute. And so each organization is playing to its strength and trying to protect its weakness in order to establish a foothold with more of these buying personas while also building the things that will keep customers wanting to stay in their platforms or needing to stay in their platforms because having their data is no longer enough. To your other questions: what should they be acting on versus what should they be treating as hype? I think there’s been this entire pendulum that’s swung from we’re all token maxing to we’re all trying to control costs on token utilization. I would say the thing is and has always been you need to be constantly focused on where the value’s going to come from from your data, and then the action is you need to be looking for opportunities to experiment with platforms and capabilities that can allow you to seize that value more effectively. Databricks and Snowflake are trying to make it easier than ever to experiment with their platforms, which I appreciate. And I think that really the risk is not experimenting enough, not looking out there and understanding what the emerging capabilities are and how they can serve your organization. But with that said, you need to treat the things that you should be experimenting with or watching other people experiment with as hype until proven otherwise. Because once it reaches general availability, you can generally rely on the fact that it is going to do what it says it’s going to do. But while it’s in public or private preview, which is what many of these announcements are, it’s not entirely clear whether the vision that they’ve laid out for where they’re trying to go is a place that they will necessarily get to. And that’s part of the trust you build with the platforms that you’re on, is that they’ll deliver on the promises that they say they’re going to deliver on. But also it is part marketing message and part public positioning and it is part positioning with their customers. And they’re trying to thread the needle between the two, and so it’s up to you within the context of what your business is trying to accomplish to understand: will this platform be able to do what I need it to do, not will it be able to do the sum total of things that they announced at their conferences.
Jason Bradwell: So we move into our favorite segment for Eventual Consistency, what we’re watching, where we talk about what we’re watching in the world of data and AI. Ross, what’s been catching your eyes and ears over the last couple of weeks?
Ross Katz: Not too long ago we all had a chance to use Fable 5 from Anthropic. It was a good new model. I was excited about it and using it across a variety of different projects that I have going on. And shortly thereafter it was shut down. But even before it was shut down by the US government, it was already imposing restrictions on the types of questions that you could ask. Two of these types of questions are types of questions that I actually find myself asking. The first type of question was research on LLMs and machine learning. They don’t want you to be able to use their model to recreate Claude, so they make it very difficult to do machine learning or LLM-based research, and the way that they shut it down is in my opinion pretty ham-handed. Basically the entire subject area is shut down. And similarly, I host another podcast called Data in Biotech, and so for a couple years now I’ve been tinkering around with protein engineering. One of the things I get really excited about is the opportunity for these AI platforms to be able to help people who do not have deep expertise in a thing like protein engineering to do meaningful work in that domain. And that entire domain is just a complete no-no from the perspective of Fable 5. So what’s been on my mind and what I’ve been watching and thinking about is: traditionally we’ve treated intelligence as something that we should all have access to. Public libraries, the internet—the implied limitations of the human mind to be able to consume information made it so that the number of people who could potentially gain access to a set of intelligence and capabilities was relatively small, and it would happen over a long period of time, so there was less concern about it. But switching on the intelligence availability to lots of people who could now do things they couldn’t do before is causing a lot of anxiety and concern among the people who create the models, among the governments who oversee the people who create the models. I just find myself wondering, because these models are just a tool. They’re a tool that can be used for good and a tool that can be used for evil. Where do we draw the lines in terms of their use in order to enable this explosion of intelligence to be beneficial to society and prevent people from misusing it? Because I understand why biological use cases would be something that Anthropic would prevent people from pursuing. But is it really true that all explorations of biology in this way are necessarily dangerous? It just seems like the way that it was implemented is counter to the positioning of creating tools that allow people to educate themselves, allow people to do things that they were never able to do before. So I’ve been thinking a lot about that and trying to figure out from my perspective, what is it that I expect from these platforms in terms of what intelligence should be available to the public and what intelligence shouldn’t be available.
Jason Bradwell: Do you expect that we will be seeing Fable 5 return for public use?
Ross Katz: At some point, yes. I think that the answer’s inevitably yes. I just wonder if we will see those guardrails go away. Will, for example, you have to work at a verified lab in order to gain access to the biological capabilities of Fable 5? Or will you have the ability to use the model as a student in biology, as someone who has a passing interest in biology? These are questions I don’t really know the answer to, but I think access to the model is an inevitable yes.
Jason: So where does all of this leave you? If there is one thing to take away from two conferences and two near-identical keynotes, it is that the convergence is mostly on the surface. Snowflake and Databricks evolved from different places, they sell to different people, and they are strong at different things. And the right choice still comes down to where your value comes from, what your data looks like, and who is going to be using it. The platforms have stopped competing on holding your data hostage because moving data has never been easier, and they have started competing on capability depth, on governance, and on becoming the ontology for your business. That is the genuine signal here. The caution that Ross has left us with is worth repeating. A lot of what is announced is still in public or private preview, not general availability, and the gap between a vision laid out on stage and a capability you can rely on is exactly where trust gets built or broken. Treat it as hype until it proves otherwise, but do not let that become an excuse to sit still, because the real risk in this moment is not betting on the wrong preview, it is not experimenting at all. Thanks to Ross for unpacking this with me and with the patience he always brings to these conversations. If you want to talk to someone who actually does this work and who can tell you where the real value of your data is hiding and which of these platforms gets you to it, that is where Ross and the team at CorrDyn do their best work. So you can visit them at their website, www.corrdyn.com. We will see you in a couple of weeks.
