ENTERPRISE

Databricks' AI edge starts with research

Jonathan Frankle describes his job at Databricks as being "the risk budget" for the company. 

As chief AI scientist, he leads the research team to push the boundaries of what Databricks' future products will do. Some of those bets landed on the keynote stage at this year's Data + AI Summit, and some quietly didn't. 

"That's the whole point of research," he said.

Frankle joined Databricks three years ago through its acquisition of MosaicML. It's the only job he's ever had since finishing his PhD. I sat down with Frankle at the Data + AI Summit to discuss the company's new products, why Databricks' no-lock-in strategy keeps the company sharp, and why agent governance is "one of the key problems of our time." This interview has been edited for brevity and clarity.

Jason Hiner: Tell us what you do at Databricks.

Jonathan Frankle: My official title is chief AI scientist. What that really means is I'm a researcher here. My job is to work with the folks in our research team to push the boundaries of what our products can do, what this technology can do, and to be, in some sense, the risk budget for Databricks. My job is the tip of the spear. I'm supposed to try to do new things, and some of them have worked out so spectacularly they were announced at the keynote in various ways. And some of them did not work and were not announced at the keynote. And that's the whole point of research.

Jason Hiner: The promise for data analytics for years, especially since chatbots and LLMs came around, was that you could just ask, "What did we do two quarters ago in this business unit compared to the same quarter last year?" Previously, there wouldn't be a single version of the truth. You'd have to hit up one of the BI people to put together a spreadsheet. Does Genie One deliver on that promise?

Frankle: I think the whole suite of Genie products actually delivers that. At least, that's been my personal experience of using it on our internal Databricks workspace, which is the biggest Databricks workspace in the world. And it's really difficult: that's probably tens or hundreds of thousands of SQL tables, tons and tons of dashboards, lots of notebooks, and huge amounts of unstructured data. We have internal versions of Genie that are able to answer really sophisticated questions about this and get it right. It's been nothing short of extraordinary to watch this happen.

It's been in the past few months where this has really started to work, and it's been exciting that now this is available to our customers. It's not perfect, far from perfect, but it's no longer AI that just looks for the answer somewhere obvious, copies the answer out of the document and reads it back to you. This is now AI that is computing on your data. It's writing notebooks or running queries and putting the pieces together. It's searching really deeply and finding all sorts of things where I didn't know what that table was, or what it was called, or where to find it.

Jason Hiner: How do you test it?

Frankle: I'll ask about the products that I've worked on, where I know the numbers really well, and where the story of the numbers is actually quite complicated because the product went through multiple iterations. It's something that, if I gave it to a staff data scientist at Databricks, would take some effort and some thinking to figure out the appropriate context. And models are doing really well on that question.

I also love whenever one of our account executives reaches out and says, "Hey, can you please meet with this customer?" There's a little questionnaire I usually give them: Are they using AI? What products are they using? Have I talked to them before? I usually ask it to both the account executive and to our internal Genie at the same time, because again, I can get ground truth. Then I'll give the salesperson the answer from Genie and ask, "Was this right? Was this wrong? Can you help me debug it?" That's another evaluation data point that we have now. I'm using it in my day-to-day life. It's become a key part of my workflow. Whenever I have a question, it's the first thing I do.

TOGETHER WITH CODER

Most engineering teams don't know where they are in the AI journey. Here's a map.

85% of developers are already using AI tools. Most organizations are governing almost none of it.

A new operating model from Coder reframes the problem with a two-dimensional maturity model: where your development infrastructure sits and how deeply your builders actually trust and delegate to AI.

The two dimensions evolve independently. Each of the four resulting profiles has a distinct risk posture, a distinct set of gaps, and a distinct 90-day action plan.

Whether you're working to contain shadow AI, scale what's already working, or build the foundation that makes agentic development safe, the model tells you precisely where to start.

Read the whitepaper

ENTERPRISE

Why Databricks rejects AI lock-in

Jason Hiner: There's this interesting phenomenon we've seen this spring, with all of the tech and AI companies doing their conferences. Everybody has their AI agent solution, because there are these same problems to solve (compliance, security, privacy, costs) and agents aren't delivering the value they promised. [Databricks CEO] Ali [Ghodsi] acknowledged in the keynote that it's a bit of a quagmire, because now everyone is offering solutions, but they're all essentially siloed, locked-in solutions. My understanding is that Genie is meant to be the open solution that works across these different platforms. Why is that?

Frankle: That's our goal. I say "we," but it's really Ali, Matei, Reynold, all of our leaders, who have worked really hard over the past decade-plus to just make sure things are open. The way I frame it to customers is that I have to win your business every day, because if I don't win your business every day, you can leave. You can take Unity Catalog with you. You can take your Delta, your Iceberg with you. You can just take all this data straight out of Databricks. We don't really have much lock-in, and that's kind of the point. It forces us to be on-the-ball every single day, and it makes our customers feel really comfortable being with us, because they know that they're not trapped here.

Open formats, we've always embraced. Open source, we've always embraced. Unity Catalog is now open-source and highly interoperable, and we've worked really hard to get it ready for governance of AI agents. We have Unity AI Gateway, which we just announced. The whole point is cost control and governing agents and helping you keep track of how all of your users are using all of the different models. Genie, same thing: we have MCP support, so you can interface with all sorts of data. Some of the Genie examples you saw during the keynotes were not just using Databricks data sources. We're not the only game in town, and I want to win business every day. If you lock someone in, you've made it too easy for yourself, and you're going to atrophy.

Jason Hiner: Does that make you feel better about working for the company? As an AI researcher, there are a lot of places you could work today. What are the reasons you want to work for Databricks?

Frankle: I've got a lot of reasons. The first is, honestly, the customers. There are very few places in the world where you can be a hardcore research scientist inventing new reinforcement learning algorithms and spend time with customers every day, and nobody asks the question of, why is the customer doing reinforcement learning, and why is the reinforcement learning guy talking to customers? The answer is, of course, you talk to customers. How else do you know what research to do? It's not some benchmark that you're trying to improve. The true measure of success in AI is: did your customer succeed?

I would hate to be a researcher who is kept in the very back, just doing research all day without knowing whether I actually made a difference in someone's life, because otherwise I'm just pushing numbers around. The fact that we're at a company where we've set ourselves up to always have to be at our best to win our customers' business every single day, I love that. It keeps me sharp, it keeps all of us sharp, and it forces us to innovate.

TOGETHER WITH CODER

AI coding agents on infrastructure you already control.

70% of engineering teams are running AI agents on infrastructure never designed for them.

A new operating model from Coder reframes the problem with a two-dimensional maturity model: where your development infrastructure sits and how deeply your builders actually trust and delegate to AI.

Get the whitepaper

ENTERPRISE

How Databricks is betting on custom AI

Hiner: Anything else our audience should know about what was announced this week, or what you're working on?

Frankle: The two things that are top of mind for me personally: Number one, all this work on custom models. We're working really hard to make sure that all of our products have the most cost-effective and the fastest models out there, and the best way to do that is to train our own stuff. For our customers who need that as well, once they're committed to a use case, we want them to have choices, whether it's open-source, closed-source, or custom. The same technology I'm using to build custom models with my team is the same technology that's available to our customers, and we're really proud of that.

The other thing I'm thinking about a lot is governance. To hear a scientist talk to you about governance, again, only at Databricks is a reinforcement learning guy talking to you about governance. But without being really thoughtful about governance, agents can't connect dots that a human would not normally connect, given the same permissions. Agents can do things at scale, and in very aggressive ways, that no human can do.

I think it's a misnomer that if you give the agents the same permissions as a human, and the human didn't do anything silly, the agent's not going to do anything silly. We need to really fundamentally rethink what governance looks like for agents, because an agent with the same permissions as me can do a lot of weird stuff that I would never have the time or creativity to do, and not with bad intentions, just because things happen. We're thinking really, really hard about how to get that right. I won't promise we have the answer. I'd be lying to you if I told you we had found the silver bullet. But I think this is, in some sense, one of the key problems of our time as AI really gets adopted in the world.

Hiner: What's the next big frontier you're thinking about?

Frankle: I'll let you know when I release them. And I mean this, it's a very Databricks culture thing: we speak through our work. I'd hate to tell you that I'm working on something and then come back six months later and say, "Well, that was a disaster." But a lot of the stuff I work on is like that. That's research. Science is about taking risks, so I plan to take some really cool risks in the next six months with my team, and I hope I'll have something really cool to show for it that we can talk about next time around.

Disclosure: Jason Hiner's travel to Databricks Data + AI Summit was paid for by Databricks. The Deep View's coverage is editorially independent from the companies we cover.