Databricks CEO: AGI Is Already Here, Why Companies Are Behind, and the Superintelligence Debate
- Ghodsi says AGI is already here, but expects enterprises’ context gap to take 10–15 years to close. Existing intelligence could “completely transform the planet”; the bottleneck is getting it into actual workflows with the necessary information. A bank that refuses to record meetings cannot give AI that context, and he expects GDP effects to unfold over decades.
- Databricks’ commercial bet is that agents increase consumption of governed enterprise data. Ghodsi identifies four deployment hurdles: context, security, cost, and ownership of IP. He describes a $7 billion-plus business growing 70–80%, with outside agents sending more queries rather than necessarily displacing the platform: “we are the governed data repository.”
- Engineering’s productivity frontier is automated factories, rather than simply generating more code. Across 2,000 Databricks engineers, the host reads the leading contributors as shipping roughly 10 times more; Ghodsi says they run factories others do not. Some leaders spend fewer tokens, supporting his emphasis on routing simpler tasks to cheaper models rather than treating spending as productivity.
- Genie’s demonstration makes company context the foundation for better management decisions. Its persistent ontology connects people, reporting lines, meetings, documents, and decisions. An employee survey involving more than 10,000 respondents becomes a department-level diagnosis of stalled action, while the host sees a broader opportunity for “management Moneyball.”
- AI writing 98% of code does not establish recursive self-improvement in Ghodsi’s framework. He requires repeatedly faster training, fewer resources, improving general accuracy, and a loop that sustains all three. Today’s resource-intensive runs and domain-specific human intervention fail that test: “That flywheel doesn’t exist right now.”
- Ghodsi puts extinction risk “close to zero” while treating cyberattacks as a material threat. Defenders must stop every intrusion; an attacker needs one success. His optimistic endpoint rests on scanning new code and cleaning up a finite stock of legacy vulnerabilities—a specific claim about security’s eventual trajectory.
- Databricks will go public, but Ghodsi gives no timetable and wants insulation from AI-era market volatility. Annual tenders provide liquidity, remain undersubscribed, and support billion-dollar acquisitions. Public markets offer capital and quarterly marketing, but he wants to avoid the employee-compensation “rollercoaster” created by share-price swings.
- Healthcare provides concrete evidence of AI’s usefulness without validating a ten-year cancer-cure deadline. Ghodsi cites Regeneron’s chronic liver disease work and Merck’s TEDDY model for drug discovery. Human trials remain a constraint, and his answer on timing stays unresolved: “maybe in 10 years, maybe not in 10 years.”
1. Databricks began with three bets the industry resisted
- Ghodsi traces the 2013 launch to roughly 20 researchers at UC Berkeley’s AMPLab. Netflix recommendation work suggested that every organization might eventually need a data science team, rather than AI remaining an ad-tech specialty. Industry adoption lagged, and the founders themselves debated the market’s potential: starting Databricks was “kind of out of despair.”
- Commercializing the research exposed what the lab had overlooked: security, governance, and organizations carrying 30–50 years of software and processes. Cloud infrastructure separated storage from compute, allowing companies to retain data without continually buying combined hardware. Ghodsi emphasizes core-product data—the podcast itself, insurance, or a drug—rather than just information held in CRM or HR systems.
- Databricks bet on cloud, AI use cases, and open source. An influential analyst firm called the cloud decision “a strategic blunder”; Ghodsi credits those unpopular bets with creating its head start. The host opens with reported figures of $1.5 million revenue in 2015 and a $7 billion run rate in August 2026; Ghodsi later describes $7 billion-plus and 70–80% growth.
2. Ghodsi separates achieved AGI from hypothetical superintelligence
- Using the definition his Berkeley colleagues held around 2009–2010, Ghodsi believes AGI arrived “last year.” He checked with former lab members, who agreed but said “we’ve moved the goalpost.” His claim concerns sufficient capability: even if progress stopped, “we have all the intelligence we need” to transform the planet.
- He distinguishes that from the superintelligence described in Bostrom’s 2014 book: something performing 1,000 years of human thought in a minute or producing a novel PhD thesis instantaneously. Current AIs are “kind of slow”; obtaining more throughput requires parallel agents collaborating. Frontier labs’ pursuit of difficult mathematics is therefore a different project from Databricks’ enterprise deployment work.
- Customer conversations remain grounded in “a glorified Google search” and coding tools delivering perhaps 20–30% speed improvements. Ghodsi asks companies whether thousands of agents are doing core work; generally, they are not. The host questions whether even that scale is practical with humans still correcting drift and slop. Ghodsi says Databricks is further along, while acknowledging the broader deployment gap.
3. Four obstacles prevent intelligence from becoming enterprise output
- Context comes first. Imagine capturing the information inside a 200,000-person organization, including its meetings. One bank CEO told Ghodsi that recording them would happen “not in a million years,” because of liability. His analogy is a brilliant mathematician excluded from the discussion: even Terence Tao cannot help with decisions he never hears about.
- Security becomes the next obstacle once agents receive broad access. Connecting something like OpenClaw to all company data creates an “insider threat” alongside external attack risk. The host points out how dramatically the exposed surface expands in a large organization. Having the information available and being willing to grant agents access are separate problems.
- Economics follows: “if the token cost is going up exponentially, is your revenue at your company going up exponentially?” Ghodsi says he knows no company whose AI-driven revenue acceleration matches that cost curve. Initial CEO enthusiasm eventually encounters a CFO confronting higher OPEX, potentially higher COGS, and an approaching earnings call.
- Finally comes avoiding lock-in and retaining ownership of IP. The host adds a competitive concern: model providers are releasing applications that encroach on established software markets, citing Anthropic’s moves against Excel and Photoshop. Ghodsi praises the labs’ contribution while maintaining that enterprises must solve all four constraints before realizing the broader opportunity.
4. Genie turns organizational context into inspectable analysis
- Genie maintains an ontology—a graph of people, reporting relationships, work, emails, Slack messages, meeting transcripts, decisions, and documents. It runs continuously and prepares an inbox of background work. Ghodsi distinguishes this accumulated knowledge from searching everything afresh: one demonstrated analysis drew on 19 pieces of information contributed by five people, with inspectable definitions and authorship.
- An employee survey with more than 10,000 responses becomes the first substantive example. Executive leadership scores 88%, while work-life blend is 60%. Ghodsi is particularly frustrated by the roughly 60% “action” score: whether employees believe feedback produces change. “Every year I promise the company we’re gonna fix this, and every year it doesn’t happen.”
- His CFO suggests poor work-life blend causes low action scores. Ghodsi initially disagrees, briefly reconsiders after seeing correlation, then describes Genie rejecting the causal explanation. One manager’s work-life blend is 44 while action is 77%. The host’s reading: that leader pushes people hard while responding to their concerns.
- Unsatisfied with the initial explanation, Ghodsi asks for department-level analysis. Product management objects to too many decision-making stakeholders; engineering dislikes tools and systems; sales operations cites career growth. He says obtaining comparable analysis from the data team would previously have taken about a week. The demonstration moves from a company average toward distinct problems requiring different responses.
5. Agents create consumption demand and uneven engineering gains
- When the host asks whether Anthropic’s dashboards threaten Databricks, Ghodsi replies that Databricks does not charge for dashboards. It charges consumption when customers access governed data beneath them. Outside agents can connect to that repository and generate more queries; he identifies this as one reason the business is accelerating, despite adoption still being early.
- A chart of 2,000 engineers’ production pull requests shows an L-shaped distribution. The host reads the leaders as contributing around ten times more. Ghodsi’s explanation: “they’re running automated factories, the rest are not.” His organizational objective is to move the whole curve to the right.
- Adding token costs reveals that some leading contributors spend less than others. Ghodsi points to open-source models and routing simpler questions to simpler models as ways to reduce costs. He wants productivity gains to support more products, market-share gains, and growth “way over 100,” rather than settling for 70–80%. That is an ambition, not guidance.
6. Management shifts toward finding blockers and testing culture
- Ghodsi contrasts hierarchical org charts with work that crosses product management, engineering, marketing, and sales. Beyond Dunbar’s number of 150, coordination becomes harder. Genie also identifies unusually deep management chains with small spans of control. The desired AI understands responsibilities across the company and shortens the paths work must travel.
- Customer-meeting analysis can start with “The meeting went poorly,” then identify precisely what was said and which features were requested. Ghodsi describes aggregating requests into engineering factories. The remaining delays are often “human blockers”: legal permission, finance approval, or repeated meetings to resolve disagreement between teams. Databricks measures those blockers to reduce them.
- The host’s parallel experiment, “Major Decisions,” reads Slack, Notion, and email to identify consequential decisions, participants, resolutions, and unresolved questions. It reveals who actually contributes to decisions and who responds quickly. His framing is “management Moneyball”: making organizational contribution more visible rather than relying on the founder’s detective work.
- Ghodsi asks whether leaders enact principles such as truth-seeking, and what culture their teams actually practice—including hundreds of colleagues in Japan whose language he does not speak. Divergence does not automatically indict the local leader: “Maybe they’re right” and he is wrong. He shares analyses with employees and uses them to establish a common direction.
7. IPO timing depends on whether public markets help the next decade
- Ghodsi evaluates an IPO against making Databricks successful over the next decade or two, including expansion into security and marketing. Public markets offer capital, visibility, and a quarterly opportunity to retell the company’s story. They also provide an established liquidity marketplace, replacing the “heavy lift” of arranging private financing and annual tenders.
- Annual liquidity has not exhausted the money allocated to tenders; Ghodsi interprets that as employee bullishness. Predictable liquidity has also supported billion-dollar acquisitions. At 70–80% growth, he says employees reconsider how much to sell after contemplating several years of compounding at 1.7 times annually—sometimes deciding on only 5%.
- His objection is public-market volatility flowing into perceived compensation. An employee expecting $1 million, perhaps $2 million later, might see equity worth $200,000 and leave before sentiment reverses three months afterward. “We will go public,” he says, but the current AI transition creates a distracting “rollercoaster.” He supplies no date.
8. Recursive self-improvement requires four conditions operating together
- Ghodsi understands why programmers find AI-written software existential: coding defined much of his own life. But a model contributing to its successor does not establish runaway improvement. “Just because 98% of the code is being written by AI, we don’t have RSI.” Automating the remaining 2% does not resolve the physical and economic constraints.
- His first condition is materially faster training of successive models; he describes the required acceleration as “super linear.” Second, each improvement must require fewer resources. Buying ten times more hardware cannot sustain the loop indefinitely because hardware and ultimately atoms run out. He characterizes prevailing scaling laws as demanding exponentially more compute for comparatively linear improvement.
- Third, accuracy must improve broadly across tasks, rather than only in mathematics or another reinforcement-learning target. Fourth, the process must repeatedly sustain all three improvements. His illustrative loop moves training from three months to one month, two weeks, then one week, while costs fall from $100—or perhaps $100 million—toward $1 million. These are conditions for the hypothetical flywheel, not forecasts.
- “That flywheel doesn’t exist right now.” Ghodsi instead sees longer training, more resources, data cleaning, skilled human input through firms such as Mercor, and domain-specific reinforcement-learning “whack-a-mole.” Even meeting all four conditions would not guarantee a conscious entity taking over. He treats them as preconditions, with the means to act physically posing another question.
9. Cyber risk is real, but Ghodsi forecasts a finite cleanup
- The host distinguishes sophisticated intelligence from exploiting refrigerators protected by “admin, admin.” Ghodsi agrees cyber is a genuine risk and points to Lakewatch, Databricks’ product for detecting intrusions and monitoring agents. Later, he makes the stakes concrete: a hospital going down would be more concerning than a refrigerator.
- The host’s pushback is that defenders possess far more compute and should discover vulnerabilities first. Ghodsi counters with the asymmetry: thousands of attacking agents can fail, but one successful intrusion can still win. A defender who protects every boundary except one may suffer a disaster: “I need to be 100% safe all the time.”
- Ghodsi claims scanning newly produced code will prevent the vulnerabilities under discussion: “all that code is gonna be secure.” He also predicts that less than the next 12 months will produce more code than all prior human history. In his account, that leaves legacy software as the principal stock requiring remediation.
- His analogy is a fruit garden: agents can reach vulnerabilities higher than humans previously could, but there is a finite amount of fruit. Rewriting legacy systems or using “shadow patching” can remove those opportunities. He predicts an eventual steady state with potentially fewer vulnerabilities than before; the host emphasizes humans as a continuing source of compromise.
10. Healthcare progress does not justify a fixed cure deadline
- The host challenges Dario’s stated belief that cancer and Alzheimer’s could be gone in under ten years, arguing that families may receive false hope even if the belief is sincere. Ghodsi declines to endorse a deadline. Progress should accelerate, but human trials remain necessary; his possibilities range from six months to 30 or 50 years: “I don’t know.”
- Ghodsi cites Regeneron researchers using Databricks to identify a genetic driver of chronic liver disease and develop drugs, using an earlier generation of AI. With Merck, he describes TEDDY—Transformer Enabled Drug Discovery—a transformer working on gene regulatory networks, or GRNs, rather than English tokens. Its purpose is to distinguish disease-causing factors from responses, helping researchers avoid targeting what is merely reacting.
- His examples extend beyond discovery: Crisis Text Line uses Databricks, language models, and agents to detect self-harm risk and assist prevention. Zipline used the platform for drone operations, including routes and batteries; the host adds blood delivery. Ghodsi presents these as existing useful applications rather than contingent on superintelligence.
11. The disagreement is over evidence and diffusion speed
- Ghodsi treats the safety community’s concerns as sincere: people are “the heroes in their own book,” and Bostrom’s account frightened him too. His objection concerns the inferential jumps connecting better models to runaway superintelligence: “there’s a few dot, dot, dots there that are not clear.” He allows for missing information: “I don’t know what I don’t know.”
- His own extinction-risk estimate is “close to zero,” while cyber remains real and manageable in his view. The host adds deployment concerns involving children’s parasocial relationships, harmful advice, and algorithmic manipulation of news. Their shared distinction is between concrete harms requiring attention and the paperclip-maximizer scenario.
- Closing the enterprise context gap, Ghodsi expects, will take 10–15 years; GDP will “move very slowly” over decades. He contrasts that with labs expecting transformation in six months, 12 months, or two years. The bank refusing meeting recordings remains his simplest explanation: without access to what organizations know and decide, AI cannot perform the work demonstrated in Genie.
Full transcript
It was out of despair that we started Databricks because no one else was listening. At that time, there was zero interest in AI. Now I think we have all the intelligence we need to completely transform the planet.
It’s not possible to have 1,000 agents doing stuff realistically. Am I correct?
They’re just using AI in very rudimentary ways. They’re using it as a glorified Google search.
We’re actually getting into management Moneyball, aren’t we?
If the token cost is going up exponentially, is your company’s revenue going up exponentially? But I actually think GDP will move very slowly, and it’s going to take decades.
Are they a little drunk on their own success?
Just because 98% of the code is being written by AI, we don’t have RSI. If it automates the last 2%, who cares?
All right, everybody. Welcome back to the All-In Interview. Ali Ghodsi is one of those amazing leaders in the world—CEOs, politicians, scientists—who watch the show and are part of the community. He’s the CEO of Databricks, a private company that has been doing incredibly well. I don’t know why it’s not public yet. I think a lot of people are asking that question, and we’ll talk about it today.
In 2015, the company was doing $1.5 million in revenue. The rumor I heard is that by August 2026, Ali, you’ll be at a $7 billion run rate, growing 80% year over year. Obviously, you have a lot of competition—the Snowflakes of the world, et cetera—but this is a tremendous business. We’re going to talk about how you’re using AI to run it and about recursive self-improvement. Welcome to the program. Maybe you could tell the audience a little about what you do at Databricks and how it became so successful.
1. Databricks Began in Despair
Yeah, happy to do that. Thanks for having me on. I’m excited for this. We started the business in 2013, but if you rewind a little bit before that, we were all at UC Berkeley. There were about 20 of us, and we were part of a lab called AMPLab, which I think was the largest AI research lab in the world at that time. We were trying to figure out how to combine data processing with AI. We came up with really interesting things, but the industry just wasn’t ready for it.
Mm.
We were trying to get the industry to adopt this technology, and they weren’t. So it was out of despair that we started Databricks. We thought, “If we don’t start a company around this, no one else is listening to it.” At that time, there was zero interest in AI. It was actually called machine learning, and AI was just robotics. That’s the beginning of the story.
For people who don’t use the product day in and day out, or aren’t in the industry, what’s the goal of the Databricks product suite? What do you do? I know you released some new products today, so we can talk about those as well. Conceptually, you got there so early. Obviously, machine learning was used at Google, Facebook and its ad network, and a lot of interesting services, like Uber’s routing algorithms. Could you talk a little about where Databricks started, what it’s become, and how it’s generating billions of dollars in revenue?
The first use case—actually, the first use case ever—was at UC Berkeley, when we wanted to help Netflix recommend better movies so that people would watch more movies.
Ah, yes.
It was a recommendation algorithm, an AI use case, that started the whole thing. We thought more organizations would probably need this. Probably every organization would need a data science team that could do this kind of AI work, and that was a bold idea at the time. A lot of people said, “No, no, AI is just for ad tech. The ad companies need it.”
Mm.
But we thought, “No, this is going to be a bigger market.” We had a lot of internal debates and weren’t sure it was going to explode. Then, when we started the company, we realized why organizations weren’t doing it and why it was so hard. Organizations around the world are bogged down by legacy software they’ve had for 30, 40 or 50 years. Some companies have been around for 100 years. Things move slowly. They have a lot of processes, compliance, regulation and security. All these things have gotten worse with generative AI in recent years.
What was missing from our research at UC Berkeley was security and governance. In the first few years, we figured out how to take this technology and let you process all your data in an enterprise. Organizations have a lot of data. The most important data any company has—and people don’t talk about it—is the data on its core product. For you, it’s this podcast: the data on that. Someone else might be selling insurance; someone else might have a drug. It’s the data that comes out of that. That data doesn’t sit in CRM or human resource management.
In fact, it probably wasn’t even tracked. I remember in the early days of the internet, they’d say, “Hey, would you like this log file?” I’d say, “Okay, yeah, keep everything.” They’d say, “Well, you’ve got an $8,000 storage bill a month.”
Exactly.
I’d say, “Okay, delete it.” They’d say, “Well, we could roll it up and keep some of it.” I’d ask, “Who’s ever using the open rate on our emails? Who cares about that? We send the emails. That’s it.” But then that changed. Was storage and compute, and cloud computing, a major accelerant to the big data movement and then eventually machine learning? Were those the waypoints in this movement?
2. Betting on the Cloud
Before we started in 2013, people were running their own data centers. When you ran your own data center, you had to buy machines, and the machines came with storage and compute combined. If you ran out of storage, you had to buy more machines, and IT departments at every organization were busy with that.
The cloud enabled you to completely decentralize this and separate compute and storage. You could imagine having infinite storage and dumping everything there. If the data gets old, it knows how to archive it so it becomes almost free to store. Then you could get the illusion of as much compute as you wanted, and, of course, you could start running machine learning on it. Eventually, GPUs showed up.
We bet everything—the whole farm—on being a cloud company. We bet everything on AI use cases, and we bet on open source. At the time, all three of these bets were controversial. People said—I had a standing bet with one of the most important analyst firms out there—“It’s a strategic blunder. You’re going to go under because you’re betting on the cloud.” It turned out that the only reason we succeeded was because of these bets: they were contrarian, and other people didn’t agree with them. We got a head start for many years. Eventually, when we happened to be right, we were so far along and had gotten too big that there was too much momentum to stop it.
It’s so funny you mention this Netflix algorithm. Netflix started in the Bay Area as DVDs mailed to you in an envelope, and then you had this queue. The Netflix Prize was in the 2008–2009 time period. They offered a million-dollar prize to any computer science team that could make the algorithm, I think, 10% better at predicting which movie you’d give four or five stars to. They released all their data and probably got $50 million worth of work from computer scientists around the world.
It was the first time there was an XPRIZE. I think the XPRIZE and this were around the same time, and they were all inspired by those old prizes: Who could circumnavigate the globe? Who could solo-fly over the Atlantic or the Pacific? It was a very interesting moment in time. And that hit AMPLab, right? AMPLab stood for Algorithms, Machine Learning, and People. I don’t know what the P stands for.
It was for people, actually, because we thought the AI wasn’t good enough yet.
Mm.
So it needed a little human nudging.
Got it.
Human feedback was already there. We thought we needed it because this last mile is too hard for AI to figure out by itself, so we had people there as well.
Putting the people in there makes sense because Berkeley’s a little hippy-dippy—Friedberg’s alma mater, obviously. They were like, “We need some people in the middle of this.”
We had Ben Horowitz on the other week, and he said the best explanation he’d ever heard for why we’re going to be able to control the rollout of superintelligence—and the jump between artificial intelligence or artificial general intelligence and superintelligence that gets recursive, runs off with it, dooms us all, and means we’re all going to die—came from you. So please explain your thesis here. Then I’d like you to go one step further and talk specifically about the culture in the Bay Area, perhaps even the East Bay, and this group of people.
3. AGI Has Already Arrived
I’m happy to get into it. I think there are 2 different things happening at the same time. On the one hand, I actually think we already have artificial general intelligence. I’ve been saying it since last year, and I base that on the kind of consensus we had in that AMPLab, which I think was the largest AI lab at the time, in 2009 and 2010.
By the definition of AI that most of us had in that lab, we already achieved AGI last year. I wasn’t actually sure about this, so I checked with some of the people who were in the lab. I asked, “Hey, is it just me? Am I hallucinating?” And they said, “No, no, no. That definition, for sure, but we’ve moved the goalpost.” So I actually think we have all the intelligence we need. If you stop all progress right now, I think we have all the intelligence we need to completely transform the planet. But that’s AGI.
That can do all the intelligence work humans do, right? If it solves this latest cohort of math problems, and we’re just starting to build harnesses and build out the infrastructure, I guess your statement would be correct. I’m kind of in line with you. We’ve hit artificial general intelligence; it just hasn’t been distributed equally yet.
Yeah. The labs and a lot of researchers are kind of obsessed with superintelligence. When I say superintelligence, I mean it the way it was defined by Bostrom in his 2014 book, Superintelligence. He’s talking about something completely different: something that can do a millennium’s worth of thought—1,000 years of human thought—in a minute. All of mankind in a minute. It thinks in 11 dimensions. It’s so much faster. A single agent is many orders of magnitude faster. Imagine 1,000 times faster than any of us.
That’s just a different thing. It can write a novel PhD thesis instantaneously, and so on. I think the labs are on a quest for that.
Okay. So, just to summarize: the ability to write an entire, unique PhD thesis instantly. Do we even have enough compute to do something like that today? If everybody got together and said, “Today, we’re going to give all our compute—or, for the next hour, just as a thought experiment, give it to this incredible mega-prompt”—do you think it would actually happen? Is it a compute issue, or a model issue?
The AIs are kind of slow. What we have to do is parallelize them and have many of them, and then have them collaborate with each other. So it’s not really the way he wrote it. He also had intelligence in the book that could be communities of intelligence, but I’m just saying the labs and many of the other companies are focused on this.
Yes.
That’s why they’re so obsessed with the math problems, the Clay problems, the 7 Millennium Prize Problems, and so on. But I argue that the biggest challenge facing us is how we actually make this AI applicable inside organizations. That’s what Databricks is focused on. We just want to help all these organizations.
When I talk to enterprises, I see that they’re using AI in very rudimentary ways. They’re using it as a glorified Google search. They’re using chatbots all day. Instead of typing into Google Search, getting 10 blue links, and looking at them, the AI just does that for them.
The second thing they’re doing is coding. With coding, they’re not seeing enough speed-ups. They’re seeing maybe 20% or 30%. I ask them this question every day. I’m in meetings with customers every day, and I ask them these questions. That’s kind of what reality looks like. I always ask them, “How many of you have thousands of agents running around, automating all of your work and doing core work for you?” It just hasn’t happened yet.
They might be running half a dozen agents, or a dozen agents. Somebody at the forefront might actually be coordinating that many, but a human still needs to be in the loop, just like AMP. There’s a person who still needs to be in there because it will return slop, return weirdness, and drift off course. It’s not possible to have 1,000 agents doing stuff realistically. Am I correct?
Yeah. Somewhat further along, I’m actually using this all throughout Databricks, and I’m happy to show you. But the reality is that they’re not. You have to ask the question: why is that? If it can solve Navier–Stokes, very difficult math problems, how come enterprises and organizations are so far behind?
Why is that?
There are many reasons, but the main reason—or maybe one way to look at it—is this: if we could gather all the information inside a large organization with 200,000 employees and record all the meetings—which, by the way, one of the big bank CEOs told me the other day, “No way. We’re never recording the meetings. Not in a million years”—but assume you had that.
Because of liability.
Liability. I was already kind of shocked, because if you can’t do that, how can AI ever help you? The AI simply isn’t in the meeting. If it’s not in the meeting, it’s useless. It doesn’t matter if you have Terence Tao or the greatest mathematician in the world. If they’re in the dark and don’t know what decisions you’re making or what you’re doing, and you’re meeting with hundreds of thousands of employees, there’s no way they can help you.
If you capture all that context and feed it to the AI, it would nail the problems. It could help us right now. That’s really the context problem. You might say, “Okay, let’s figure that out and do that,” and we’re doing that at Databricks. But that’s not enough.
Hmm.
Even if you do that, you have a second problem: you could unleash something like OpenClaw and have it access all your data everywhere. Then you get into the problem that you’re going to get hacked. Cyber is real, and all the big organizations with successful businesses are terrified right now of getting attacked. That’s a big concern, a top concern.
It certainly creates a lot of surface area when you embrace this. If a CEO at a small company wants to connect everything, that’s a certain amount of surface area. But for a 200,000-person organization, or 10,000 people, my Lord, the surface area has now gone up 1,000 times.
Yes, exactly. That’s the second problem you run into. We can access all the data, but we might not want agents to access all of it because now we have an insider threat, and we might get hacked. So what do we do about that?
Let’s say we solve that too. We’ve solved the cyber problem, which, by the way, we have ways to do at Databricks, and I’m happy to show you. Then you still have another problem: the token cost is going up exponentially. If the token cost is going up exponentially, is your company’s revenue going up exponentially? I don’t know any company that started using AI and then saw its revenue accelerate exponentially to match its cost curve.
First, CEOs say, “Go get the AI. Get more AI. Get it in there.”
Yeah.
“We need to move faster.”
Token matching.
“We’ve got to…”
Yeah.
Exactly. But then at some point the CFO shows up and says, “Our OPEX is going up significantly here, and even the COGS. What are we doing about it? And we have an earnings call soon.” Then they freak out. So how do we fix that? You also need to do this efficiently, which we have ways to do at Databricks.
Now you’ve done all 3 of these, and there’s still a strategic last question that remains: okay, fine, we figured it out.
We access all the data, secure it, and figure out the cost situation. How do we avoid getting locked in, and how do we own our own IP? Otherwise, there’s no business to be made in the future.
Data sovereignty is what you’re referring to: intelligence sovereignty. If you use Anthropic, OpenAI, or any of these platforms, there’s a risk that your data will leak and get incorporated into their next version. As we’ve seen, Anthropic seems to be releasing half a dozen products every two weeks. Now they’re competing against Microsoft, one of their giant partners, I think.
Yeah.
They just came out with basically an Excel killer. They’re killing Photoshop. They’re going after everybody wholesale.
Yeah, so these four—
For free. For free.
Yeah. By the way, these companies are amazing. They’re doing amazing work, and without them, we probably wouldn’t even have this intelligence. I’m thankful. This is awesome. I’m just pointing out why—
Yes.
—when you look around at all these successful organizations, all the Global 2000, all the Fortune 500s, NASDAQ, and so on, how come it’s just a glorified chatbot and a little bit of coding? This is the reason. You need to tackle all four of those. Only then can you start getting something out of it.
One of those issues as well, I believe, is the autonomy in an organization to come up with a groundbreaking idea. You may have captured a large portion of your market. You may be getting incremental benefits. If I tell you, “We can put another 500 horsepower into your Ford Taurus,” it’s like, “Okay, my Ford Explorer now has 500 extra horsepower. It’s going to rip the chassis off. This thing is not capable of going 500 miles an hour.” So who decides we’re launching three more product lines? At Anthropic—
Yeah.
—somebody is deciding, “Why don’t we just go after every single business in the world?” And they started doing dashboards. That’s a little close to your business, is it not?
Well, we don’t have a product where we charge for dashboards.
Mm.
Dashboards are usually backed by data, and you typically want that data to be governed. You don’t want just numbers that are incorrect. This is a big problem in organizations. That gets back to the second thing I said—
Yeah.
—the governance and security aspect. Maybe we can get access to all the data and get all the context, but we have to make it secure. You can connect all of these agents from outside to Databricks, but we are the governed data repository. When you access that data, we charge consumption rates. That’s one of the reasons you see acceleration in our business: all these companies have agents, and when they use them, they send us more queries. I think we’re just at the beginning. As I said, I still think it’s mostly Google search and coding, but the agents will start using that.
How does behavior change when people actually embrace agentic technology and start working with agents every hour, every other hour of the day, like you and I do?
I think the organization and the org charts are built this way. It’s a tree structure, with the CEO at the top, and decision-making going up and down these org charts. You have all these old processes, and getting agreement across the org chart is very hard inside an enterprise or any sizable organization. It’s somewhat easy up until Dunbar’s number, which is 150. You can know everyone. Small startups are a different story, but above 150, this becomes a coordination problem.
If you look at how any organization sells a product or service to its customers, it’s not this hierarchical thing. It cuts across the org chart. Here at Databricks, we have customers who give us feedback. That feedback goes to product management, which hands it to engineering. Engineering builds products, which go to marketing. Marketing markets them, then gives them to sales, which sells them to customers again. You want to go across the organization and cut out that latency so you can be really fast. That requires an AI that understands your whole business and your IP. It needs to understand your org chart and know who’s who and who’s responsible for what.
What percentage of companies understand what you just described? Would you say 10% of organizations understand this? 5%?
At scale, I don’t know the answer to that. I know Databricks is doing it. I’m doing it.
Yeah.
I haven’t seen anyone else do it at this kind of scale. I think probably Block and Square—
Mm-hmm.
—with Jack Dorsey, since he’s pushing it. He calls it the company world model. He describes the same thing I did, I think in a different way. So I think they’re probably pretty far ahead. In general, I think digital natives are further ahead. Digital natives are companies that have only existed for 10 or 15 years.
Yeah.
If you’ve been around for 20, 30, 40, or 50 years, that means you’re an amazing company. Maybe you’re a Fortune 500 company, one of the best in the United States, with 100,000 employees. They’re loyal, they love your product, and you know what? You’re the best in the world at that. You have so much software, so much regulation, and so many processes around security that it’s very hard for you to adopt this. That’s the very problem Databricks wants to solve for them.
Yeah.
We want them to adopt the technologies so the AI can help them do what I do at Databricks.
Let’s see it. You said you’d be willing to show us how you run Databricks with agents and AI. I’ve been looking forward to this all day because I’m taking my entire life in podcasting and early-stage investing, looking at each component as a service or an agent, defining the goal, and working backward to the signs of success that show you’re heading toward the goal. Achieving the goal is binary: you made the sale or you didn’t. But what are the precursor signals? Then you try to have each of these bots hand things off back and forth. You have to be a real systems thinker. It’s like playing five chess games at once while riding a horse. I find it intellectually super stimulating. I can tell you do, too. You can share your screen.
Yeah. Let’s jump in. The key thing here is that we have a product called Genie. There are so many chatbots, assistants, co-work products, and agents running in the background. The key thing is that Genie has an ontology. The ontology is a graph that it’s constantly constructing for your company. That graph understands all the people, who reports to whom, what they’re working on, what emails are coming in, what meetings have happened, the transcripts of those meetings, the decisions being made, and what’s on Slack and being communicated.
It has access to all of those things, and also all the documents you’re creating, all the Excel sheets, and all the PowerPoint presentations. That’s what we call Genie here. You can see here: I wake up, and it has an inbox. In the morning when I come in, it has done a bunch of background work for me. It’s running 24/7. As you can see, it has access to my calendar. In fact, I could just ask it, “Hey—”
Which I’m going to tell Jason on All-In today. Here it goes.
Let’s see what it’s going to say. But let’s look at some of the stuff I’ve done. The other day, one interesting thing was that our HR department came back—
Mm.
—and they were saying, “Hey, we’ve done a survey across all of Databricks.” We had over 10,000 employees respond to the survey. I looked at the results while she was going through them, and I said, “Hey, make me a dashboard of the results.” Nothing magical here yet. It told me, “Look, here’s how you’re doing.”
You’re doing great on some things, but not so good on others.
Executive leadership, 88%. You’re doing okay.
Yeah.
Work-life blend, 60%. You’re working people pretty hard over there.
Yeah.
But maybe everybody says that, right?
Yeah. No, but we want to know that this has been… We wanted to improve these two.
Okay.
And this action one really bothers me because, you know, “action” means that when we ask them for feedback, they give us feedback about what to improve, and they feel like we’re not listening to them. And this one, I think—
Ah. So two out of three feel you are, and one out of three feels you’re not, just to… You’re doing okay. It’s not terrible, by the way. This could be 30%.
Yeah, yeah. But then again, 60% is not great.
No.
And especially not great if you take into account that every year it’s low. Every year I promise the company we’re going to fix this, and every year it doesn’t happen.
Hmm.
So we’re going through the results and discussing them with the staff. While they’re discussing it, they’re speculating. My CFO says, “Hey, I think it’s because of work-life blend. Work-life blend is bad, and that’s causing action to go down.” So work-life blend is causing low action. I’m like, “That doesn’t make any sense.” But I entered it here immediately as we were discussing, and it immediately pulled up the correlation between these two and showed how it’s looking.
Hmm.
I was like, “Okay, wow, maybe he’s right.” But then it says, “No, actually, it’s not causal. It’s not causing it.” And here’s the exact proof points. We probably have to anonymize these folks.
Yeah.
Here’s a manager with a work-life blend score of 44, which is pretty bad, and an action score of 77%. Things like this—
They’re working hard, and he’s recognizing and taking action on their concerns.
Exactly. So then I said, “Okay, well—”
That’s your guy right there. He’s pushing people hard, and he’s listening to them. I like that guy.
Exactly. So now I’m learning stuff about the organization. This is amazing. It would’ve taken me a week to get the data team to do this analysis for me. That’s useful information. We’re cutting these latencies up and down the org chart a little bit. But then I’m like, “Okay, why is it low?” So I looked through all of the comments.
Ah.
It gives me some reasons. I didn’t like the reasons, so I clarified the question and said, “Hey, can you slice it by department for me?”
Hmm.
Then it figured it out. It said, “Look, product management has low action because they don’t like decision-making. There are too many stakeholders.” Engineering doesn’t like the tools and systems. There’s too much. For sales operations, it’s career growth, and so on and so forth.
That makes sense because salespeople always think they should get a promotion because they’re salespeople, and they should get their commissions raised. Product people think you should just listen to them instead of having six people trying to design a piece of art, you know?
Exactly.
This all starts to track now with the data and the AI. I love it.
Yeah. So how did it do this? It has access to this ontology.
Mm.
It’s not just searching live. It found that these people were already in its graph. Five people had already contributed to this, and there are 19 pieces of information here that you can see it’s leveraging. It already had them; it didn’t search for them live. These are in the ontology. We can look at this particular one. If you click on it, it has a definition, it’s a metric, and there’s a graph.
Mm.
You can see the graph here. There’s a favorability score, and this is the author. You can click on the author, and because it understands the org chart, it understands who this person is. This person actually has 29 snippets. They have lots of different assets that it’s learned from them. So it’s constantly learning the intelligence, or ontology, of your organization. That’s a typical example. Here’s another one that might be interesting. I wanted to know about these new results from the last 59 minutes. I haven’t actually looked at all of those.
Hmm.
I wanted to understand the org chart, what’s going on with it, and span of control. I asked it, and right out of the box, I was shocked by this. It immediately told me, “Hey, there is an org inside Databricks that is very deep.” This person reports to this person, who reports to this person, and so on. They have three people reporting to them, five people reporting to them, four people reporting to them, and five people reporting to them. So that’s not a flat organization. Then you can look at what we can do about that. Again, since it understands all the people, it can understand that right out of the box. Engineering cost is something we look at all the time, and that’s another thing we wanted to see if we could understand.
What’s great about this, too, is that instead of you as the leader saying, “Hey, listen, if we’re going to have an opinion, let’s go with mine,” you’ve created a forum where people can put data first and test theories. You don’t have to be dogmatic about who came up with the theory. You just have to be intellectually curious and open to possibilities. Eventually, if this is where we are today, can you imagine where we’ll be in a year or two? It’s going to be giving you these insights. Instead of a culture report, it’ll be like a temperature gauge in your company. You’ll be looking at the nest in your company and thinking, “Oh, wow, looks a little hot today. What’s going on?” And it’ll be like, “Oh, yeah, we made people come in this weekend.”
Yeah.
“Yeah, maybe we should buy some pizzas or give people a comp day during Thanksgiving.” Yeah.
Yeah. This is my favorite. This one actually came out of Genie.
Ooh.
This is our 2,000 engineers and how many pull requests, or how much code, they’ve actually shipped into production. We look at the middle here, which is roughly per week. But it has this L shape. When you look at this L shape, the top people are committing way more.
Yes. Ten times.
Yeah.
Yeah.
So what the hell is going on? We looked into this, and it turns out they’re running automated factories; the rest are not.
Mm-hmm. Wow.
So we pushed them. An interesting thing: I asked it, “Hey, can you add the cost to this line?” One of the things we found is that cost isn’t necessarily higher for those people. Look at the top. Some of the top folks are actually not spending as much as some of the other folks. That goes back to the fact that there’s a way to spend tokens smartly. You can use open source, route simpler questions to simpler models, and we actually have a product for this. You can cut your costs down significantly. This is the kind of stuff… By the way, recording the meetings is really important.
Mm.
One of the things I heard was that there had been a customer meeting, and I wanted to understand how it went. I asked it, “Hey, was there a meeting with this customer recently? How did it go?” It literally started by saying, “The meeting went poorly.” It listed exactly who said what and what features they need. These are just some of the things you can do with AI. In the background, when I wake up, it has all these deep dives for me. Instead of waking up, calling people, and doing meetings, it’s doing the work for me. It’s pinging them and sending whatever material is needed for the company. It just compresses the time it takes to do things.
Well, a lot of being in management and being a founder is detective work.
I call it being Columbo, if you remember that TV show.
Yeah.
I think it might show our age, but you're kind of like, “Oh, you know, one more thing there. Can you tell me about the client meeting? How'd that go? Who talked most on the phone? What percentage of the time did they talk, and what did the customer actually think?” And then you, as the founder, are getting spun.
Yeah.
You're getting spun.
Yeah.
It's not that an employee necessarily wants to spin you. A team member isn't necessarily looking to spin you. That can happen, but it's not the majority. Most people are good people. But they're human, right? They're the people in the org.
Yeah.
Here, you take the people out of that discussion and say, “Here's, transparently, what happened from first principles.” Now we have ground truth, and we can start the discussion without you, as the CEO, having to be Mr. Detective. I don't know how frustrating it was for you two years ago to spend two hours being a detective. You get three meetings into this, and it's like, if people had just told you what happened, or if you had that report on culture, the problem would've been solved. You wouldn't have had three weeks of hand-wringing, yeah?
Yeah. Or the customer feedback, right? I can immediately find out what all the customers think. I have some cases where I'm having it automatically look at features requested by all our customers across all the transcripts of our customer meetings. It's gathering those, and they're immediately going to engineering. Then we have factories in engineering that can build features automatically. So it's a game changer. It's just that game of whispers we played when we were kids.
Yeah.
Right? When people have to meet, they have to tell each other. One of our main goals is to reduce the number of human blockers. We measure that throughout the organization, and we see—
Blockers. Explain a blocker in this context, please. What's something I can learn from?
Well, look, AI is very fast, right? It's iterating in the background, doing its work, preparing material, and so on. But at some point, it's like, okay, now we need to go to legal and make sure we're allowed to do this thing. Or we need to go to finance to get approval for this budget. Or we need to make sure marketing is okay with this. Or even within the same organization, we need to talk to a sibling organization, get a meeting with them, see what they think, and get an agreement. If we can't reach an agreement, we'll schedule another meeting to look at this again.
Yeah.
Then if that meeting goes down, maybe we escalate it to the managers, and so on. These are the blockers that slow down organizations.
Ah. So it's a blocker finder, which is great if you're the CEO, because here's the next thing: one, you have to play detective; the other one is the blame game.
Yeah.
“Why isn't this done?” is what we always ask. When you're a leader, whatever your group is, and another group is involved, it's, “Hey, we were supposed to do this by today. We were supposed to hit this target. Who's responsible?” And then you get three Spider-Mans pointing at each other, if you know the meme.
Yeah.
I had a similar thing. I came up with a bot, and I was going to share it with the whole company. Thank God I didn't. I came up with a bot called Major Decisions, and I said, “Go into Slack, join every room as me. Just join every channel as me.” Some people were like, “Wait a second. How did you even know we had a room?” I was like, “Join every room, read every Notion page, even if I was invited to it or not. Get every email, and find the most important decisions for me for the month. Every couple of days, give me a report on who participated in them, what the resolution was, and what's not resolved.”
Then I said, “Give me the people who participate the most in these kinds of discussions.” I realized, oh wow, these are the people who are actually making decisions in the organization. These are the people who are quickest. My next report is going to show who's quickest to contribute to those discussions. Now, in a venture firm, that's a discussion like, should we invest or not? Should we take our pro rata or not? Should we sell our shares if we got a secondary offer? On the podcast, it's like, should we book this guest or not? Should we use this clip or that one?
Then it picked up employment discussions, and it was like, “Should we fire this person or not?” Or, “Do we do PIPs?” I'm like, “We don't do PIPs.” I just think PIPs are ridiculous. If the performance isn't there, that's like the person didn't show up for—
Let's torture them for two quarters.
Yeah. You miss practice and you show up late for the game. That's table stakes. You can't be on the team if you can't do those things. Now we're going to torture you to try to make you come into work on time or do a workout you don't want to do. So maybe you could speak to what this is going to do for you as a founder and CEO in terms of figuring out who your all-stars on the team are, who are the ones who are actually contributing. Now we're getting to essentially business Moneyball, like that movie Moneyball.
Mm-hmm.
It's fantastic. Brad Pitt. It's such a good movie and such a good book. We're actually getting into management Moneyball, aren't we?
Yeah, absolutely. One of my favorite questions is to look at different orgs, because we have certain culture principles: truth-seeking, first principles, and so on. I say, “Hey, which of these leaders follow these principles? Look in their whole orgs, look at all the meetings,” and it can derive that from those. By the way, you could even ask, “Hey, this leader, what is the actual culture they're implementing?” It'll come up with different things. It'll tell you, “No, actually, the culture there is this, not the company culture,” which is essential, I think.
So they're deviating from the core principles. That drift is cancerous for you as a founder, or it could be something you need to change and could be an early warning sign.
Yeah. We need to be consistent, whatever it is, right? Maybe they're right and the rest of the organization, including me, are wrong. But then we need to know. At least it tells me there's a divergence.
Wow. Yeah.
Those kinds of things were nearly impossible to do. You could have done many of the other things, but a question like that—get the culture of this subgroup in a 12,000-person organization; figure out the culture of the Japan organization at Databricks. I don't even speak Japanese, and we have hundreds of people over there working. What's the culture like? How do we make decisions? How is it different from EMEA or the LatAm region that we have? Those are game changers.
For us, I share all this information with the employees. They have access to all of it themselves. Then I tell them, “This is what we want. That L-curve, we wanted to shift it to the right. So everybody, let's change that. Let's move toward this North Star that we have together.” That helps us be more competitive. Then we can build more products, take market share, and accelerate our growth further. I'd like not to be at 70 or 80% growth year over year. I'd like to be at 100%, way over 100%. That's how I see it.
In this era, this is a competitive advantage: whoever can do this best. At some point, maybe it all becomes fully diffused and everyone is using it. But right now, this is big alpha. If you can do this really, really well, better than any other organization—
Interesting. Speaking of big alpha, a lot of people have been very patiently on your cap table, and we have a big debate between founders and capital right now: stay private longer. What's the point of going public if we have secondary markets that are quite efficient? Stripe has been private closer to 20 years, I think, than 10. They seem to have given the indication they're going to stay private longer. SpaceX took a little while. Uber took a little while. I felt that one, but luckily Masayoshi Son was splashy cashy. When are you going public? When, when, when? How do you make that decision? I know you've got a lot of pressure on you to do it, but you also have a lot of pressure relief valves in the vibrant secondary market.
Yeah. Look, my goal is to make Databricks as successful as possible in the next decade or two. That's how I'm thinking about it. What's the right strategy to get there? I mentioned this alpha of what we're doing here, the products we're building. But we're getting into new markets. We got into the security market.
We got into the marketing market. I’m focused on those. Then the question is: are the paths to get there better and faster if we’re public or private?
Hmm.
And I actually think that—
Take me through that. What’s on each side of the ledger for you?
I actually think you get access to capital and much better marketing. Everybody knows about you. Every quarter, you have this big marketing event, and you can tell your story again and again. Those are huge advantages. We have to organize our own tenders every year. Finance has to go through this big process every year.
Yeah.
We raise $20 billion, $30 billion, and then we have to arrange that. It’s a heavy lift. There’s a marketplace that just does that.
It’s also probably pretty emotional and distracting, I would think, right? People have to decide. Then you might lose people who take $10 million off the table and think, “Maybe I don’t want to come to practice anymore.”
I think the liquidity is a good thing because people know they’re going to get it every year. It’s predictable, and it’s the one time they have. In the past, since we started doing these a few years ago, the money we set aside for tenders has never been fully used. So most employees—
Oh, that’s good.
Most employees are bullish. They’re like, “No, this is…” I mean—
That’s actually good for you. Do you get a signal from that about who’s selling and who’s not? It’s a little bit of signal. It could be for personal reasons. But acquiring companies is a great reason to have equity, yeah?
We haven’t had any problem doing billion-dollar acquisitions because people know we’ll provide liquidity. Most people do the calculation. If the business is at $7 billion plus and growing 70% or 80%, they do the math. Most people say, “I want liquidity,” but when you ask them how much they want to sell, and remind them, “We’re growing 70% or 80% plus”—
Yeah.
It’s a different story. Can you do compounded growth of 1.7 to the power of 3 or 4? Then they start scratching their heads and say, “Maybe I’ll sell 5%.”
Yeah.
No, for real.
Sitting on your hands is a very important skill to learn as an investor and an employee with this kind of stock. You do have to protect the downside when you get into that space. Who do you see as your competition right now? Everybody says something like Snowflake, but I’m watching a market where, as I said earlier, Anthropic is launching feature after feature. Cloudflare is looking at features next to them. Replit is looking at features next to them. Everybody seems to be looking in their peripheral vision. Because you can build software fast and you have a lot of developers, if those developers keep going up that ramp you showed, you only had 10% of people at the peak. As the other 90% climb up there, your developer resources are going to go through the roof. So you have a decision to make: do you have fewer developers at the low end of the tail and become more profitable, or do you invest all that developer time as they move up the ramp? How do you think about both of those issues? What do you do with the massive resources this efficiency is creating? And who is the competitor? Do you need to keep making peripheral product moves?
Yeah, we will make peripheral product moves. But maybe I should start by saying those were the upsides of going public. The downsides are the reason why, to answer your specific question—
Yeah.
Why do we compete with whom we compete with? Why do we build these products? I don’t want the market to force me into a situation where sentiment is terrible this week and our stock is down. When the stock is down, the employees are literally getting a pay cut.
Mm.
They say, “You promised me I was going to make a million dollars. Actually, you said it was going to be $2 million in a couple of years.”
Yeah.
“But I’m looking at it right now, and it’s actually worth—”
It’s underwater.
“—$200K. So you lied to me. I’m leaving.” Then, 3 months later, it’s, “Oh, no, no, no. Actually, revenue is accelerating. We were totally wrong. It’s up to the right now.” And they’re like, “Oh, okay. Wow, it’s great.” This back and forth—
A little bit of a roller coaster can get frustrating.
A roller coaster. That’s what I want to avoid, so we can focus on building the company toward our long-term vision without being distracted. We will go public; we are going to be public. But right now, during this transition period of AI, there’s just too much up and down. I think it’s distracting to companies and CEOs, and it’s demoralizing to CEOs. They call me depressed because the stock is down and they don’t even understand why, and then the stock is up. That’s what we’re trying to avoid.
I think it’s an issue people are starting to have with their sleep scores. You get a WHOOP, and I love it. But I got my wife a WHOOP, and then we’re comparing sleep scores. It’s kind of like looking at your Uber stock every day. You get a little bit of volatility in your relationship; it can be challenging.
Yeah, exactly.
In other ways, it’s interesting. We didn’t circle around to what Ben Horowitz said. He said your formulation of superintelligence is some combination of the model, the power, and the time it takes to train and be accurate. Can you walk us through it? Do you have it off the top of your head?
Yeah, I have to do it. It’s pretty simple. I think the main problem is the context problem. We’re solving that with Genie and the ontology. The labs are focused on superintelligence. That’s what Bostrom talked about. In his book, Bostrom also asked how we get there. In 2014, I believe, he defined RSI, recursive self-improvement, which means the AI starts improving itself and gets better and better and better. Then we get this takeoff, maybe a fast takeoff, and reach the singularity. We’re no longer in control. At that point, let’s hope the AIs are well aligned, because if they’re not, we’re all going to be made into pets. That’s where the 10% risk—
The paperclips. Yeah.
Yeah, paperclip maximizers and the 10% risk to humanity and all of that come from there. But I think a lot of this is coming from the fact that most of the people building AI have seen that AI has become really good at coding.
Hmm.
Almost at every company, as I said earlier, coding is being done by coding agents. I think this is causing a big existential crisis in Silicon Valley. For most of us who grew up coding our whole lives, it’s a big deal. I spent 10 or 12 hours a day, 7 days a week, from third grade until my 30s, just coding all the time. I was really good at it, and I knew all these languages. Then suddenly, the AI is doing that. If it’s doing it—
Right. And how much better than you is the AI?
Exactly.
That’s the other part. It’s like—
Now it’s writing itself, and the next version it wrote itself. Wait a minute, isn’t that recursive self-improvement? But then this is going to speed up, and we’re going to get into this loop. The book called it superintelligence and the paperclip maximizer, so it might go off and all these bad things happen. Actually, Bostrom has changed his view now.
Hmm.
If you follow him now, he’s not saying that anymore. He says we should do this. There are risks, of course, with cyber. But I think you need to think this through further. Just because 98% of the code is being written by AI, we don’t have RSI. If it automates the last 2%, who cares? It doesn’t matter if it’s 98% or 100%. Why does it matter? It doesn’t.
We really need 4 conditions to be true. The first is that the next model needs to be trained in a much shorter amount of time. The labs still don’t do that many big pretraining runs per year, and they’re still following what’s called the scaling laws, which means you need exponentially more hardware to get that improvement.
So exponentially more—like 10× more compute—to get a little linear improvement. They’re scrambling to get all the GPUs and put together the biggest data center they possibly can, with the networking, and do all the engineering. They’re having humans do all of this because they really don’t want to screw up the next run. If the next run gets screwed up, the whole thing falls apart.
We’d like that next training run to be faster. That’s condition number one. It can’t just be a little bit faster. If you’re halving the distance every time, you’ll never get there. In some sense, it has to be superlinear. Secondly, we also want it to use fewer resources, because you could say it’s getting faster and faster, but that’s because we’re 10×-ing the hardware all the time.
Yes.
You can’t do that trick because we run out of hardware and we run out of atoms.
Yes.
That’s the second requirement. The third requirement is that accuracy is also improving, at least linearly, or we’re getting better accuracy across the board in a generalized manner. It can’t just be getting better at this one dimension we did reinforcement learning on, or just math. It needs to be across the board on all of its tasks.
Even if you do all 3, it’s not enough. Let’s say you do that 5, 6, 7 times. That’s awesome. You get really awesome models. That’s not that kind of superintelligence. The fourth condition is that you can actually put this in a loop, and all the previous 3 conditions need to be continuously happening. Then we end up in a world where the next model takes 3 months, then 1 month, then 2 weeks, then 1 week, and it’s getting smarter and smarter and smarter. It’s costing us $100—or maybe it costs us $100 million, then $50 million, $25 million, $10 million, $5 million, $2 million, $1 million, and then it goes down and down. That’s the kind of scenario we have.
Yes.
If you have to look at it—
That’s the flywheel.
That’s the flywheel. That flywheel doesn’t exist right now.
Mm-hmm.
The models are taking longer to train, and there’s more data cleaning. There’s more manual whack-a-mole happening, where you have vendors like Mercor that have to go get the data from very skilled labor. Then you have to do reinforcement learning in these gyms, and you have to whack-a-mole in each domain.
Mm-hmm.
So it’s very domain-specific. And then, of course, pre-training requires more resources. That doesn’t look like we’re getting into this loop. By the way, even if you satisfy all 4 conditions I’m saying, that doesn’t guarantee you’ll have a conscious being that’s taking over everything.
It’s the precondition for that happening, and there’s no guarantee that it actually happens.
It’s a precondition. But even if you had that—which we don’t have—to threaten me right now, you would have to have a killer robot come into this room right now and wrestle me down.
Right.
Or something. Those don’t exist yet.
Yeah. Right now, you can have a bunch of hackers take a really capable model, and you can find out that the refrigerators in grocery stores in France have the password “admin, admin.” You can remotely script-kiddie into them, turn them into defrost mode, and make a bunch of butter go bad, and milk.
Yeah.
Which is a tragedy for the French, I understand.
Yes.
But it is not superintelligence, I can assure you.
Yeah.
I mean, did you see that hack and how people were losing their minds over it? I’m like, “Guys.” You know how, when you buy a new router now, it has the serial number and then a random password?
Yeah, admin, admin, yeah.
If they’d done that on the refrigerators, they wouldn’t have been able to hack one of them unless they got into the place, took screenshots of 20 different refrigerators, and physically compromised some supermarket in the 8th arrondissement. That supermarket would have lost, I don’t know, 10 pounds of butter. It wouldn’t have been a big deal.
So look, cyber, I do think, is a real thing.
Yes.
And I think it’s a real risk. We actually have a whole product in the cyber space called Lakewatch, built on all this data, that has automated agents that detect these kinds of intrusions and monitor your agents. This is a real problem, and I do think bad things will happen here, but then we will adjust. Bad things happened when the internet came out as well.
Shouldn’t we be able to solve for this? You tell me if I’m right or wrong here. One theory I heard, which resonated with me for a certain type of hacking—not for script-kiddie hacking, but real hacking—is that you need a lot of compute to execute certain hacks, like the Hugging Face one: 1,000 agents, whatever. The bad guys don’t have that much compute. The good guys have massively more compute, and so if you did your basic job as a white hat, defended properly, and looked for these exploits, let’s say you have 1,000 times the resources, you should be able to get to these problems long before the hackers. Does that make good sense to you? They’re just—
Yeah.
Yeah, it’s a good argument, right?
Yeah. But there’s an issue: you have to weigh that against the following asymmetric scenario that we have—
Yes, yes, yes.
If I’m an attacker, I can launch thousands and thousands and thousands of agents, and all of them can fail except for one. If one of them gets through, I’m completely successful. The asymmetry is that if I’m a defender, I have to protect my boundaries all the time. If one gets through, then it’s a complete disaster.
This is a suicide bomber—not to be too graphic—but you only need to get one 9/11 to make your point. You can try 1,000 times and not get there.
Yeah, yeah. So I need to be 100% safe all the time.
Yeah.
They need to just get one thing through.
Get lucky once.
So that risk is very asymmetric. But the way I would look at it, the reason I think we’re actually going to overcome this and it’s going to be totally fine, is that these are latent bugs that exist in the source code. By the way, we’re not creating more of them now, because almost all of us are scanning it. We all have products. We actually sell a product where you can scan the vulnerabilities in new code. New code that’s being produced isn’t going to have these vulnerabilities.
That might sound silly, but think about it this way: In the next 12 months, we’re going to write more code than ever in the history of mankind. Actually, in less than the next 12 months. So all that code is going to be secure. And, by the way—
Because right now the AI is sophisticated enough not to make human errors, and to look for them, and it’s more dogged.
Yeah, and you just scan those things. So what remains is all the legacy code, which we can rewrite. That’s what we have to go through, and there are—
And people are doing that, right? That’s part of what Chamath’s doing at 8090, yeah?
Yeah. There are other ways to protect things, like doing shadow patching without touching the existing legacy code. What’s happening is that you can think of it like a fruit garden, and there are bugs that attackers are leveraging. They’re picking the low-hanging fruit. Before, humans could pick a certain low-hanging fruit. Now agents can reach higher and pick higher-hanging fruit. But we’re going to clean those up, and then they can maybe reach a little bit higher. There are only a finite number of vulnerabilities in the code. Once we’ve secured that code base, there are no more. It’s not an infinite thing that’s going to go on forever. Then it stops, and we’re at the steady state. Vulnerabilities might actually be fewer than what we had in the past.
I think that eventually, back to AMP, where you went at Berkeley, it's the people in the loop who are the vulnerability. It's just going to be somebody getting compromised, as we saw over the last week or so. The humans do have vulnerabilities. I'll leave it at that.
Hey, I want to end on this one thing. You spent a lot of time in Berkeley, in the hippie-dippy EA community, among the accelerationists and decelerationists. What's the nature of this group, of the EA faction, and how they look at the world? How confounding are they to the rest of American society? Who are these people? What do they believe? Are they a little drunk on their own success, or do they actually have really important insights, and we need to focus on what they're saying and take them at their word? It's a big controversy.
Dario, during the roadshow, is saying, “Cancer and Alzheimer's, I believe, will be gone in under 10 years.” That's very beautiful to say, except—
Not because everybody's dead, right?
Yeah, except if your mom or dad are dying from cancer and you think, “Oh, I have a window here where I could save them,” or your grandmother has Alzheimer's. Now you're talking about giving people false hope, and you maybe shouldn't make claims like that. But I don't doubt Dario believes it. I actually think he does. Unpack these people, because you've been around them for two or three decades.
Yeah. Well, I think that most people on both sides—actually, on all sides—they're the heroes in their own book, right? We all wake up thinking, “I'm the good guy, and I'm doing the right thing.” So I think it comes from a good place.
I think they are seriously nervous about this kind of superintelligence taking over. If you read Bostrom's books—actually, if you read those books and internalize what he wrote before any of this, then you get kind of freaked out, right? I was certainly freaked out when I read it.
I think they're starting with those assumptions and then looking at, “Okay, if it's getting smarter this fast, and look at the cyberattacks that are happening... Dot, dot, dot, we're going to get this kind of superintelligence that's described in the book.”
Mm.
But there are a few dot, dot, dots there that are not clear. Then they would say, “Well, look, shouldn't we be cautious?” So I think we should look at safety. Cyber, I think, is a big deal that we should look at.
But the reality is that I think we know how to handle this. We will be able to manage through this. I think we'll be... The biggest risk right now is cyber—
Mm.
—and I think it's manageable. This RSI, I don't see any evidence of it. If there's data that others have that I don't have, I don't know what I don't know. But based on what I'm hearing here in San Francisco, where all the labs are, and all the people I meet, I don't see those four criteria I mentioned even being fulfilled. So I don't see any risk that would make me worried about what's happening.
And the second thing I said: even if we got there, I don't see the means in society. It's just digital. We can just shut down the computers right now, you know? I do think cyber is real. If a hospital went down—
Absolutely.
—instead of a refrigerator, that is concerning. And they—
Kids using it, their kids using it and developing parasocial relationships with Character.AI, or it giving you bad information if you're suicidal or something. There are concerns about how the technology gets deployed.
We just went through this with social media, and the profound impact that had on society. We're going through it now with truth and news, and the manipulation of news via algorithms. So there are things to be concerned about.
Yeah.
Being turned into paperclips, maybe not.
But I think the healthcare stuff is absolutely real. At Databricks, we were involved with Regeneron, where their researchers found the genome responsible for chronic liver disease using Databricks, and actually came up with drugs for that—
Best-case scenario, in under a decade—
—using AI.
—if you were to take the claim we had Dario make. Any chance that we cure cancer in the next 10 years? That's a pretty bold statement.
It's very hard for me to say. I don't know exactly what's needed. I think progress will be faster, but a lot of these things also have to go through—
FDA and trials and—
—you know, human trials. We can't just test stuff on humans willy-nilly, right? And just because you tested it on animals, that's not enough either. So I don't know. Maybe in 10 years, maybe not in 10 years. Maybe it takes 30 years, maybe it takes 50 years. I don't know. Maybe it takes 6 months. Maybe it's already more or less solved. But it is helping already.
The use case I gave you with chronic liver disease is super important. By the way, that was the previous generation of AI that we used at Databricks. We trained a new one now with Merck, and it's called TEDDY, Transformer-Enabled Drug Discovery. It's a transformer model, but instead of the next token being English, it's the gene regulatory network, or GRN. So it can say, “Okay, this genome is actually causing this disease. This genome is just reacting to it, so don't target it for a drug.” So this will accelerate drug discovery. It's a game changer.
Drug discovery aside, I'll give you another example. Crisis Text Line uses Databricks and large language models and these agents. They detect if kids want to self-harm, and that helps prevent it. Virtue Foundation uses Databricks. Zipline—this was a use case reported before—used Databricks to deliver food. They had these drones, right?
And blood.
The drones would deliver everything. All the data about those drones—their path, the battery, all of that—was running on Databricks. So we see these use cases all the time with AI, and it is going to make the world a better place. There are risks. We will deal with those as well.
People are nervous because things are changing. I think there are a lot of researchers at AI companies who maybe understand the research very well, but they haven't looked at how the economy works or how these things are going to evolve.
And they've watched a lot of sci-fi, let's be honest. They are being informed by a certain narrative. The idea that there's a 10% chance that 7 billion people die because of AI seems so farcical and out there that when you pick a number, it makes you seem credible to an average Joe who's like, “Well, these are smart people who make a lot of money, and if they think it, maybe they're right.”
Yeah, if they're saying it themselves, right?
Yes.
That's how people think. But for me, that risk is close to zero. I think the cyber risk is real, and really bad things can happen with that. That's why we're focused on it.
And then, to solve this context gap that I mentioned in enterprises—the thing we're focused on to actually get value everywhere—I think it's going to take 10 years. I think this is something that's going to take 10 or 15 years. The labs think this will happen in the next 6 months, 12 months, or 2 years. But I actually think GDP will move very slowly, and it's going to take decades for us to get through it.
Why? As I said, just go back to the anecdote at the beginning: the CEO of the bank who said, “We can't even record the meetings.” If you can't record the meetings, the AI doesn't even know what's happening in the meeting. How could it possibly do the kinds of things we discussed when I showed you Genie and the ontology?
Yeah.
It's going to be a paced rollout, to be sure, and there are some of us who are running ahead around the corners, doing really cool stuff. That's exactly what we both experienced with the internet. When the internet first came out, the people who were using it to its fullest extent were like, “Okay, this is going to change everything.” Then 99% of people were on dial-up and had never seen a high-speed connection, so their brains couldn't conceive of YouTube or unlimited free photo storage. It didn't make any sense.
Yeah, there was a kid called Robert Morris—
Yeah.
—who took down 10% of the internet with a worm. It's called the Morris worm.
Yeah.
He almost got a jail sentence for it. Then he went on to become a famous MIT professor in computer science, and he's a very famous guy.
Right. Yeah.
But that took down a whole 10% of the internet back in the day.
Which is completely possible. At some point today, I saw somebody say they may have cracked Signal to a certain extent. Okay, sure. Let’s see what happens. Ali, thank you so much for your time.
Thank you so much.
Great job, Ali. You were awesome. I’m going all in. I’m going all in.