Cofounder and CEO, Databricks
Context Needed to Reach AGI, Says Databricks CEO
AI is already smart enough, but it lacks the right context to reach the level of AGI. Databricks CEO Ali Ghodsi talks about the current state of enterprise AI, how companies' databases are not built for an AI-native world, and why Databricks is not planning on going public just yet. He speaks with Ed Ludlow on "Bloomberg Tech."
--------
Like this video? Subscribe to Bloomberg Technology on YouTube:
/ @bloombergtechnology
Watch the latest full episodes of "Bloomberg Technology" with Caroline Hyde and Ed Ludlow here:
• "Bloomberg Tech" With Ed Ludlow
Get the latest in tech...
About Ali Ghodsi
Ali Ghodsi, cofounder and CEO of Databricks, has argued in recent interviews that artificial general intelligence (AGI) has already been achieved, stating that "intelligence is not the constraint" and that "AI doesn't have an intelligence problem. It has a context problem." He said that current AI models are "plenty smart" but lack the necessary context from data and processes to be fully productive in the workplace. Ghodsi described the cost of creating software as "almost going to zero" and noted that every new piece of software requires a database, but that existing databases were built for humans, not AI agents. He positioned Databricks' Lakebase as a database purpose-built for agents, and said the company is investing in new product categories including a marketing product called Customer Lake and a security product called Lake Watch. Ghodsi has stated that Databricks will eventually go public, but that he considers 2026 a "terrible year" to do so, citing uncertainty from "mega IPOs" and macroeconomic factors. He said the company prefers to wait for "more predictability" and that, because Databricks is free cash flow positive and does not burn capital, it can choose its timing. Ghodsi noted that the company raised $5 billion in equity at a $134 billion valuation and said he expects significant acceleration in the business. He also said that Chinese open-source AI models are "absolutely dominating" and that users face a choice between paying $30 million or $1 million for model use.
Transcript (13 segments)
✨ AI-enhanced transcript with speaker attribution
H
Host0:00
This is a really big week, right? It's a really big deal. Databricks might not be a household name to a lot of people, but you would argue this is the biggest AI conference in terms of attendance and energy in the world. We'll go into all the nitty gritty of what you're announcing and what people will be discussing there. I want to start with the mile high big picture, and that's your argument that AGI is already here. Intelligence is not the constraint. What is the constraint? It's the context. These AIs are plenty smart. Anyone who's interacted with them, if you ask a big audience how many think this is super smart, smarter than many people you work with, most of the time all the hands go up. But yet we're not seeing them in the workplace. They're not everywhere. We don't have hundreds of agents working together doing work for you. We're all managing thousands of agents. That's just not reality. Why is that?
A
Ali Ghodsi0:57
They just don't have the context. If you could take all the data, all the processes, everything that people are thinking and put them in the models, they could actually help us tremendously. But there's a big gap there. So we really want to infuse that context into the existing AIs, and when you do that, magic happens. That's why we launched Genie Ontology. That's exactly what it tries to do. The idea of an enterprise ontology, you hear people in industry and culture use that term. What is it and why is it important? It might be a confusing term, but it's a big web of all the knowledge and how the knowledge is interconnected. So it could be this person opened this document, this document was read by these other people, this document is about this particular information, and tagging and explanations of what that is and what the key elements are. And that not just for that document, but for everyone in the organization and everything they're doing. A lot of that information already exists; it's just not infused into the AI. What agents today do is they try to go one by one and find it. My analogy is it's like when Google invented the ten blue links. Imagine instead of giving you the ten blue links, it would go search one website, read it, then find another link, and so on. That would take forever, cost a lot of money, and wouldn't be very good. That's how agents do it today. An ontology is essentially like that index that Google had, but for the enterprise, computing it behind the scenes. Then when you're asking questions or having the agent do things for you, it can tap into that index.
H
Host2:46
Everyone has an agent or packages it as an AI assistant. A really simple place to start is: what is different about Genie One?
A
Ali Ghodsi2:57
The big difference with Genie One is that Genie One computes; other agents recite. The best way to explain this is with an example. I ask it all the time: Who are the new customers on the platform in the last 24 hours? Or which of our biggest customers churned in the last ten days? Other agents will recite; they'll try to find a document that has the answer, search through it quickly, and then give you that answer. But what if that answer does not exist in any document and you have to compute it live? You need numeracy for the AI to understand the data. That's what makes Genie One different. It can do that really fast because it uses the ontology. This is what Novo Nordisk uses Genie One for. They had empowered their scientists to use it, but before it took them a long time when they ran different experiments and wanted to know how an obesity study was going, how a drug was affecting different control groups, and what results they were seeing. Now anyone can compute that; it's not trying to recite it because that result doesn't exist. It computes it live in a fraction of the time. That's what they use it for.
H
Host4:19
You're making a bigger claim here than just product launches. It's that those systems built for human use, software built for humans, don't work as well as systems built for agents to use the software. You and I got into this a couple of weeks ago. That brings us to the lake. What are you trying to say needs to be rebuilt here?
A
Ali Ghodsi4:52
You nailed it. The cost of creating software is almost zero; anyone can create software anywhere right now. But every piece of software you create needs a database behind the scenes. That's how software works, which is why we have big database companies. But these databases were created for humans. Agents are different. They want to move very fast and experiment. Sometimes they break the database or create thousands of databases, and you don't want to pay for thousands of databases. They also want to move fast, so you don't want to wait 20 minutes. That would slow them down significantly. That's what a lake-based architecture is. It helps them operate fast on the database where all the information is stored. If they want to ask more complicated questions, like a data science question about the data, they have to use a different database today, called a data warehouse. So there are two separate systems. We unified them into one database that's really fit for agents. Now agents can operate on the database and ask AI-like questions, like what the curve looks like. That's a game changer. My favorite example is Prada. They store all their cookies inside the lake-based system, and their agents can move very quickly to make KPIs available to leadership. It's a game changer; it lets them do things much faster with higher accuracy and lower cost.
H
Host6:25
I appreciate the product case study, but we started this conversation with you being honest that actually there are not agents everywhere in the real world. With that in mind, what is the biggest constraint on Databricks' growth as a company? What's getting in the way of deployment?
A
Ali Ghodsi6:43
When you're taking all this data and creating this graph, organizations have to get security approvals. We're very trusted; we've been doing this for 15 years. But security teams need to look through this. AI adoption for really critical use cases is working its way through organizations as they get comfortable and understand the stuff. That's what's happening in the industry. As I said at the beginning, most industry usage of AI is as a chatbot that's approved in most organizations. But using it for critical infrastructure like financial reporting, where you get into APIs or developing a drug and want it to accurately help, takes longer to get to those use cases. That's what we're working through. But it's accelerated our growth; we're growing faster than ever before.
H
Host7:40
From the outside, capital doesn't appear to be a constraint on you. There are reports that you're raising money in a round valuing the company up to $175 billion. What would be the rationale to raise new capital?
A
Ali Ghodsi7:57
First of all, we haven't said we're doing any fundraise. I know there are rumors, but we're always talking to investors. We're investing heavily into new product categories. We just launched a completely new product called Customer Lake for marketing, targeting marketing folks, which is not an area we used to be active in. Two months ago we launched a security product called Lake Watch, and we're getting into the security space. It takes resources to get into these spaces. All these products are built agent-first from the ground up, not bolting it on. That requires AI researchers, and they don't come cheap. So all this requires funding. We're always in talks with investors to be aggressive, expand, launch new products, do research, and go to more regions. That's why we require capital.
H
Host8:59
Ali, let's end it here. Earlier this month you told me that Databricks, in the end, will be a public company. You want to be a public company, but you think this is just a terrible year to go public. Do you stand by that? Has anything changed in your mind in the last two weeks?
A
Ali Ghodsi9:14
No. There are three mega IPOs that have been rumored; one of them happened. Letting those go through the system and seeing how it plays out, and things stabilize a bit. I think getting a little further into the AI revolution we're going through, I prefer for the company to go public when the water is a bit more calm and there's more predictability. I stand by what I said. I think for most companies, it's probably best to wait it out. That's also what CEOs I know tell me about the IPO; they say they don't want to go between, you know, to my guy. It feels like that Oligocene to Databricks.
H
Host9:56
This week hosting what you say is the biggest AI conference in the world. Best of luck with it. Thanks for being on Bloomberg Tech.