Databricks CEO Ghodsi Warns Enterprise AI Payoff Will Take 15 Years

Databricks CEO: AGI Is Already Here, Why Companies Are Behind, and the Superintelligence Debate — All-In Podcast

All-In Podcast

Enterprise AI adoption will take up to 15 years to move macroeconomic growth because regulatory compliance and legal liability prevent models from accessing the internal corporate context required for true autonomy.

AI-assisted summary of Databricks CEO: AGI Is Already Here, Why Companies Are Behind, and the Superintelligence Debate — original source. Check the original for full context.

Source published:

Key ideas

  • Databricks has scaled from $1.5 million in revenue in 2015 to an estimated $7 billion run rate compounding at 80% year-over-year, driven by early contrarian bets on public cloud, AI workloads, and open-source foundations. (0:00)
  • Databricks CEO Ali Ghodsi argues artificial general intelligence (AGI) has already arrived under original academic definitions, but frontier labs moved the goalpost toward superintelligence while enterprise adoption lags behind. (7:22)
  • Enterprise AI implementation faces severe enterprise context bottlenecks, with strict compliance rules, legal liability fears around meeting recordings, and security attack surfaces preventing models from accessing critical organizational context. (12:12)
  • Ghodsi rejects imminent recursive self-improvement (RSI) doom scenarios, noting that frontier models currently face exponential compute walls and manual data curation bottlenecks rather than an autonomous recursive flywheel. (45:49)
  • While frontier labs predict complete economic disruption within months, enterprise AI adoption is projected to require 10 to 15 years to materially alter macroeconomic GDP due to organizational inertia and regulatory hurdles. (60:00)

Detailed notes

## Background Published on YouTube by the All-In Podcast on November 17, 2026, this episode features Databricks co-founder and CEO Ali Ghodsi in conversation with host Jason Calacanis. According to the show notes, the interview covers Databricks' early foundational bets on cloud and open source, why enterprise AI is lagging despite AGI's arrival, a breakdown of internal agentic workflows, the strategic rationale for remaining private, and the debate surrounding superintelligence and cyber risks.

## TL;DR Databricks CEO Ali Ghodsi argues that Artificial General Intelligence has already arrived by historical academic standards, yet enterprise adoption lags significantly due to compliance restrictions, liability concerns, and missing organizational context. Highlighting Databricks' internal tooling, Ghodsi details how agentic systems like Genie map organizational ontologies to streamline decision-making, optimize engineering output, and audit company culture. He outlines why Databricks chooses to stay private during the volatile AI platform transition, relying on structured secondary tenders rather than subjecting its engineering culture to public market gyrations. Addressing AI safety, Ghodsi dismisses immediate existential extinction and recursive self-improvement claims as detached from reality, urging the industry instead to prioritize immediate, asymmetric cybersecurity threats and long-term enterprise integration.

## Atmosphere & style The discussion has a sharp, pragmatic, and insider tone, contrasting frontier lab hyperbole with the unvarnished realities of enterprise software deployment. Calacanis and Ghodsi interact with high-bandwidth Silicon Valley familiarity, trading operational war stories and analyzing internal engineering data with systems-level precision. Listening to the episode feels like sitting in on a candid executive strategy session that methodically demystifies tech hype in favor of durable business architecture and data governance.

## Bulletin - [00:00](https://www.youtube.com/watch?v=1bR0gtecoHU&t=0s) Databricks has scaled from $1.5 million in revenue in 2015 to an estimated $7 billion run rate compounding at 80% year-over-year, driven by early contrarian bets on public cloud, AI workloads, and open-source foundations. - [07:22](https://www.youtube.com/watch?v=1bR0gtecoHU&t=442s) Databricks CEO Ali Ghodsi argues artificial general intelligence (AGI) has already arrived under original academic definitions, but frontier labs moved the goalpost toward superintelligence while enterprise adoption lags behind. - [12:12](https://www.youtube.com/watch?v=1bR0gtecoHU&t=732s) Enterprise AI implementation faces severe enterprise context bottlenecks, with strict compliance rules, legal liability fears around meeting recordings, and security attack surfaces preventing models from accessing critical organizational context. - [45:49](https://www.youtube.com/watch?v=1bR0gtecoHU&t=2749s) Ghodsi rejects imminent recursive self-improvement (RSI) doom scenarios, noting that frontier models currently face exponential compute walls and manual data curation bottlenecks rather than an autonomous recursive flywheel. - [01:00:00](https://www.youtube.com/watch?v=1bR0gtecoHU&t=3600s) While frontier labs predict complete economic disruption within months, enterprise AI adoption is projected to require 10 to 15 years to materially alter macroeconomic GDP due to organizational inertia and regulatory hurdles.

## Key takeaways - [01:43](https://www.youtube.com/watch?v=1bR0gtecoHU&t=103s) Databricks was founded by UC Berkeley AMPLab researchers who commercialized open-source distributed frameworks like Apache Spark after legacy enterprise vendors dismissed machine learning as academic. - [05:00](https://www.youtube.com/watch?v=1bR0gtecoHU&t=300s) The company's commercial trajectory rested on three contrarian 2013 convictions: committing 100% to public cloud architecture, prioritizing AI and ML workloads, and anchoring the core engine in open source. - [07:22](https://www.youtube.com/watch?v=1bR0gtecoHU&t=442s) Under original academic definitions established in the late 2000s, AGI has already been achieved, providing sufficient computational intelligence to reshape global industry even if foundational model scaling stopped immediately. - [12:12](https://www.youtube.com/watch?v=1bR0gtecoHU&t=732s) Most Global 2000 companies remain trapped at superficial AI adoption levels like basic search and coding assistants, delivering modest 20% to 30% productivity gains rather than structural transformation. - [13:48](https://www.youtube.com/watch?v=1bR0gtecoHU&t=828s) Enterprise AI deployment is fundamentally constrained by missing organizational context, as liability concerns prevent companies from recording meetings, while sovereign data risks and expanded attack surfaces restrict agentic autonomy. - [22:04](https://www.youtube.com/watch?v=1bR0gtecoHU&t=1324s) Databricks uses its internal Genie agent to construct an enterprise ontology across communication streams, analyzing HR surveys, org-chart spans of control, and software engineering output across 2,000 developers. - [26:45](https://www.youtube.com/watch?v=1bR0gtecoHU&t=1605s) Top engineering performers follow a power-law distribution by running automated AI software factories and optimizing costs through routing routine queries to smaller open-source models. - [32:00](https://www.youtube.com/watch?v=1bR0gtecoHU&t=1920s) Internal workflows reveal that agentic AI velocity stalls at human consensus handoffs, including legal sign-offs, finance checkpoints, and cross-functional approvals. - [41:57](https://www.youtube.com/watch?v=1bR0gtecoHU&t=2517s) Databricks remains private to insulate its engineering culture and strategic investments from short-term public market volatility during the ongoing AI platform transition. - [40:00](https://www.youtube.com/watch?v=1bR0gtecoHU&t=2400s) Predictable annual secondary tenders allow private companies to provide liquidity and use equity for major acquisitions without exposing operations to quarterly earnings sentiment shocks. - [45:49](https://www.youtube.com/watch?v=1bR0gtecoHU&t=2749s) True recursive self-improvement requires super-linear speedups, sub-linear resources, broad generalization, and a closed loop—conditions modern LLMs violate due to exponential compute walls and manual data labeling dependencies. - [50:00](https://www.youtube.com/watch?v=1bR0gtecoHU&t=3000s) Defending enterprise infrastructure is structurally asymmetric, requiring defenders to secure 100% of attack surfaces against offensive agent swarms that only require one breach. - [51:00](https://www.youtube.com/watch?v=1bR0gtecoHU&t=3060s) Automated pre-deployment code scanning and shadow patching will ultimately reduce the global latent bug surface to a resilient steady state despite massive surges in AI-generated code volume. - [54:02](https://www.youtube.com/watch?v=1bR0gtecoHU&t=3242s) Existential AI doomer narratives distract from practical immediate risks such as targeted infrastructure cyberattacks, parasocial child manipulation, and automated disinformation campaigns. - [56:32](https://www.youtube.com/watch?v=1bR0gtecoHU&t=3392s) AI acceleration in healthcare is bounded by sequential physical reality, as drug discovery models cannot bypass multi-year animal and human clinical trial requirements. - [01:00:00](https://www.youtube.com/watch?v=1bR0gtecoHU&t=3600s) The timeline for enterprise AI to impact macroeconomic GDP is 10 to 15 years, mirroring the multi-decade adoption curve of the early internet.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=1bR0gtecoHU&t=0s) Teaser, Intro, and Databricks' Meteoric Growth

The episode opens with a rapid-fire montage of the core themes to follow: starting Databricks out of raw frustration when no one believed in AI, whether artificial general intelligence (AGI) has already arrived while enterprises remain stuck using it as a “glorified Google search,” the mechanics of agentic workflows, the runaway costs of exponential token consumption, and the looming question of an initial public offering. Following a brief presentation on modernizing financial infrastructure, Jason Calacanis welcomes Databricks CEO Ali Ghodsi to the *All-In* interview, pointing out that top global founders, scientists, and leaders frequently tune into the show.

Jason sets the stage by highlighting Databricks’ financial trajectory, framing it as one of the most remarkable private software businesses in tech history. In 2015, the company was booking just $1.5 million in revenue; by mid-2026, market rumors and reported benchmarks put the business at roughly a $7 billion run rate while compounding at 80% year-over-year. Facing off against entrenched rivals like Snowflake, Databricks has established itself as an indispensable data and AI platform, prompting Jason to press Ghodsi on how AI is utilized internally, the mechanics of recursive self-improvement, and the lingering mystery of why Databricks has chosen to remain private for so long.

- **2015 Revenue:** $1.5 million - **Reported 2026 Revenue Run Rate:** Estimated $7 billion, growing roughly 80% year-over-year - **Primary Market Competitors:** Legacy cloud data warehouses and platforms such as Snowflake

---

[01:43](https://www.youtube.com/watch?v=1bR0gtecoHU&t=103s) The Genesis of Databricks: UC Berkeley's AMPLab and Starting Out of Despair

Reflecting on the company’s roots, Ghodsi traces the story back to around 2009–2013 at UC Berkeley, where a tight-knit cohort of roughly 20 researchers formed AMPLab. At the time, AMPLab was arguably the largest academic AI research lab in the world, focused squarely on bridging large-scale data processing with advanced algorithms. Ghodsi and his co-founders developed open-source distributed frameworks (such as Apache Spark) and tried desperately to get enterprise software vendors and corporate IT departments to adopt them. The corporate world was entirely indifferent: machine learning was considered a niche academic curiosity, and the term “AI” was largely relegated to robotics.

Faced with an industry that flatly refused to engage, the researchers realized that if they wanted their technology to see production outside of academia, they would have to build the commercial vehicle themselves. “So it was kind of out of despair that we started Databricks because we thought, you know, if we don't start a company around this, no one else is listening to it,” Ghodsi explains. When Jason asks what the Databricks suite actually does for companies that do not interact with developer tooling daily, Ghodsi explains that early machine learning was dominated almost exclusively by consumer tech giants—Google indexing search, Facebook targeting ads, and Uber building dispatch and routing algorithms. Databricks was created to democratize those same predictive data capabilities for the rest of the enterprise economy.

```mermaid flowchart LR accTitle: Origin of Databricks accDescr: Flowchart tracing the path from UC Berkeley's AMPLab research to the founding of Databricks. subgraph Berkeley["UC Berkeley (2009-2013)"] A["AMPLab Research Team"] -->|"Developed"| B["Data Processing & ML Engine"] end subgraph Market["Industry Friction"] B -->|"Attempted Transfer"| C["Zero Enterprise Interest in ML"] end subgraph Commercialization["Commercial Launch (2013)"] C -->|"Founded out of Despair"| D["Databricks Platform"] end ```

---

[03:41](https://www.youtube.com/watch?v=1bR0gtecoHU&t=221s) The Netflix Prize, Enterprise Legacy Realities, and Core Product Data

Ghodsi reveals that the very first test case that sparked their architecture at Berkeley was the famous Netflix Prize. Around 2008 to 2009, Netflix put up an open $1 million bounty for any computer science team that could improve their collaborative filtering recommendation algorithm by 10%, effectively crowd-sourcing tens of millions of dollars in research. Inspired by that challenge, Ghodsi’s team aimed to build algorithms that predicted movie ratings, which convinced them that every company on Earth would eventually need its own internal data science group to run predictive models.

When they launched Databricks, however, they quickly ran into the harsh reality of corporate inertia. Most established enterprises were weighed down by technical architectures, operational practices, and compliance standards built over thirty to fifty years—or, in some cases, an entire century. The Berkeley researchers had solved mathematical and computational challenges, but they had completely overlooked security, data lineage, and regulatory governance. Furthermore, companies misunderstood where their real value lived: executive teams assumed critical data was confined to transactional platforms like CRM or HR systems. Ghodsi stresses that the most valuable data an organization owns is the operational data generated by its actual core product—whether that is listener engagement on a podcast, underwriting metrics for an insurer, or clinical readout data for a pharmaceutical firm.

> “The most important data any company has is the data on the core product that they have.” — Ali Ghodsi

---

[05:00](https://www.youtube.com/watch?v=1bR0gtecoHU&t=300s) Three Contrarian Bets: Cloud, AI, and Open Source

Jason recalls the early days of internet publishing, when hosting providers would offer to save detailed server access logs, only to deliver an unexpected $8,000 monthly storage bill that forced founders to immediately purge the data because storage was too scarce and expensive. Ghodsi notes that prior to 2013, companies operated their own on-premise data centers with servers that physically bundled storage disks and compute processors together. If an IT department ran out of disk space, it was forced to rack more server blades, tying up capital and engineering time.

The cloud fundamentally broke that paradigm by decoupling compute from storage, creating the practical illusion of infinite storage where stale or “cold” data could be warehoused for pennies, ready to be spun up against on-demand compute clusters or GPUs. When Databricks launched in 2013, it bet its entire future on three distinct premises: operating exclusively in the public cloud, prioritizing AI and machine learning workloads, and anchoring the core engine in open source. Every single one of those bets was mocked at the time. Ghodsi recalls a venture capitalist who held a standing bet with him, warning that abandoning on-premise enterprise deployments was corporate suicide. Yet because the consensus was completely wrong, Databricks enjoyed years of uncontested runway before the rest of the enterprise landscape caught up to the cloud and data explosion.

> “And it turned out actually the only reason we succeeded was because of these bets, because they were contrarian, because other people did not agree with them.” — Ali Ghodsi

- **On-Premise Paradigm:** Compute and storage were tightly coupled; expanding storage required purchasing and racking complete machines. - **Cloud Paradigm:** Storage decoupled from compute; storage costs dropped toward zero while compute became dynamically scalable on demand. - **The Three Core Bets:** 100% public cloud native, AI-first workloads, and foundational commitment to open source.

---

[07:22](https://www.youtube.com/watch?v=1bR0gtecoHU&t=442s) AGI Is Already Here: Shifting Definitions and the AMPLab Standard

Jason revisits the naming of UC Berkeley's AMPLab, noting that while the “A” and “M” stood for Algorithms and Machines, the “P” actually stood for “People.” Ghodsi confirms that the lab intentionally incorporated human-in-the-loop feedback because the underlying statistical models of 2009–2010 routinely stumbled on edge cases and required human guidance to navigate the final mile. Jason jokes that injecting people into the middle of machine intelligence felt distinctly Berkeley, noting that co-host David Friedberg is an alumnus of the university.

The conversation turns to existential AI debates and public anxieties surrounding recursive self-improvement and catastrophic outcomes ("p(doom)"). Jason mentions that venture capitalist Ben Horowitz recently appeared on the podcast and credited Ali Ghodsi with providing the single most coherent, grounded explanation for why humanity will be able to manage and govern the transition from artificial general intelligence to superintelligence. Ghodsi accepts the prompt and pushes back against prevailing media rhetoric, stating flatly that the threshold for AGI has already been crossed.

> “On the one hand, I actually think we already have artificial general intelligence. I've been saying it since last year. And I based that on the kind of consensus we had in that AMP lab where it was, I think, the largest AI lab at the time in 2009, 2010.” — Ali Ghodsi

[10:00](https://www.youtube.com/watch?v=1bR0gtecoHU&t=600s) AGI Has Arrived, but the Labs Moved the Goalposts to Superintelligence

Ali Ghodsi argues that by historical definitions, artificial general intelligence (AGI) has already been reached. In fact, Ghodsi felt so strongly about this that he consulted researchers inside the leading frontier AI labs to see if he was simply hallucinating the milestone; they confirmed that by original definitions, AGI was achieved the previous year, but admitted, “we've moved the goalpost.” Ghodsi contends that even if all scientific progress in foundational models were frozen today, humanity already possesses all the raw machine intelligence necessary to completely transform the planet and perform cognitive tasks on par with humans. The hosts agree, noting that AGI is essentially here—it just has not been equally distributed across the economy yet.

However, the frontier research labs are no longer satisfied with general intelligence; they have become fixated on superintelligence. Ghodsi references Nick Bostrom’s 2014 book *Superintelligence* to clarify the distinction: superintelligence is not just a smart assistant, but an entity capable of compressing a millennium of collective human thought into a single minute, thinking across 11 dimensions, operating thousands of times faster than any human agent, and producing novel, unique PhD theses instantaneously. While the hosts ponder whether current global compute could unleash such an entity via an unimaginable mega-prompt, Ghodsi clarifies that today’s models are still relatively slow. True scaling will require massive parallelization and collaborative communities of decentralized agents rather than a single monolithic brain. In the meantime, while labs obsess over solving the Clay Millennium Prize problems and Navier-Stokes equations, Ghodsi argues that Silicon Valley is neglecting the practical reality of bringing existing intelligence into operating businesses.

> “So I actually think we have all the intelligence we need. And actually, if you happen to stop all progress right now, I think we have all the intelligence we need to completely transform the planet.” — Ali Ghodsi

[12:12](https://www.youtube.com/watch?v=1bR0gtecoHU&t=732s) The Enterprise Reality: Glorified Google Search and Modest Coding Gains

When Ghodsi speaks with enterprise executives across the Global 2000 and Fortune 500, he finds that actual enterprise AI deployment looks nothing like the breathless narratives of total automation. Instead of thousands of autonomous agents running core business workflows, most companies are using AI in exceptionally rudimentary ways. Primarily, organizations deploy chatbots as a “glorified Google search”—sparing employees from clicking through ten blue links by synthesizing an answer—and adopt coding assistants. Even in software development, executives report modest efficiency gains of only 20% to 30%, rather than exponential leaps.

The hosts point out that having thousands of autonomous agents operating unsupervised remains unrealistic because models still hallucinate, drift off course, and produce low-grade “slop,” necessitating a human in the loop. While Ghodsi notes that Databricks is already running internal agentic workflows extensively, he acknowledges that the broader enterprise world lags dramatically behind. This raises a fundamental paradox: if modern foundation models are sophisticated enough to solve elite mathematical theorems, why are the world's most well-resourced enterprises still stuck at the starting line?

[13:48](https://www.youtube.com/watch?v=1bR0gtecoHU&t=828s) The Four Obstacles to Enterprise AI: Context, Cyber, Cost, and Sovereignty

Ghodsi breaks down the enterprise adoption bottleneck into four distinct, compounding barriers that every large organization inevitably encounters:

```mermaid flowchart TD accTitle: Four Barriers to Enterprise AI Adoption accDescr: A progressive sequence of four major hurdles enterprises face when implementing agentic AI workflows. subgraph S1["1. Context Capture"] A["Organizational Silence<br/>(Refusal to record meetings / legal liability)"] end subgraph S2["2. Security & Surface Area"] B["Massive Attack Surface<br/>(Insider threats & data breaches)"] end subgraph S3["3. Token Economics"] C["Exponential OpEx & COGS<br/>(Token maxing without matching revenue)"] end subgraph S4["4. Strategic Sovereignty"] D["IP Leakage & Vendor Lock-In<br/>(Frontier labs launching rival software)"] end

A --> B B --> C C --> D ```

1. **Context and Legal Liability:** To make an AI genuinely helpful, it must sit inside the room where decisions happen. Yet when Ghodsi asked a major bank CEO about recording their executive sessions, the CEO flatly responded: “no way, we're never recording the meetings, not in a million years” due to liability and regulatory discovery fears. Ghodsi stresses that even if an enterprise had the world’s greatest mathematician like Terence Tao on staff, that genius would be completely useless if left out in the cold without knowing what hundreds of thousands of employees are deciding day-to-day. 2. **Cybersecurity and Attack Surface Expansion:** If a company does grant an autonomous agent (like an OpenClaw-style tool) sweeping access to corporate data stores, the attack surface balloons. While connecting tools across a tiny startup entails manageable exposure, doing so across an enterprise with 10,000 to 200,000 employees multiplies the security perimeter a thousandfold, creating unprecedented insider threats and data-exfiltration risks. 3. **Runaway Token Costs vs. Flat Revenue:** Once companies figure out data connectivity and security, they encounter what the hosts term “token maxing.” Employees and autonomous agents begin burning tokens at an exponential rate. However, Ghodsi observes that no company's top-line revenue accelerates exponentially alongside its token consumption curve. Inevitably, the CFO intervenes before an upcoming earnings call, panicking over ballooning operating expenses (OpEx) and cost of goods sold (COGS). 4. **Data and Intelligence Sovereignty:** The final existential hurdle is avoiding platform lock-in and safeguarding intellectual property. Enterprises fear that feeding their proprietary workflows into third-party foundation model providers like Anthropic or OpenAI will cause their IP to leak or be absorbed into the vendor's next base training run. This fear is exacerbated by frontier labs rapidly encroaching on enterprise SaaS territory—shipping major standalone products every two weeks, from spreadsheet tools aiming to displace Microsoft Excel to image-editing suites targeting Adobe Photoshop.

> “If they're in the cold and they don't know what decisions or what you're doing, you're meeting with hundreds of thousands of employees. There's no way it can help you.” — Ali Ghodsi

[17:35](https://www.youtube.com/watch?v=1bR0gtecoHU&t=1055s) Data Governance as the Tollbooth and the Rigidity of Enterprise Org Charts

When asked whether frontier labs releasing built-in features like business analytics dashboards threatens Databricks’ core business, Ghodsi clarifies that Databricks does not sell standalone visualization software. Instead, dashboards are only as reliable as the underlying data governance. Without rigorous data integrity, agents generate incorrect figures that propagate chaos throughout an organization. Databricks positions itself as the secure, governed data foundation to which any third-party agent can connect. Because Databricks operates on a consumption-based pricing model, external agent activity actually accelerates its revenue: whenever external agents query governed enterprise data, Databricks charges for that compute and access.

Beyond technical infrastructure, Ghodsi and the hosts explore how corporate human behavior must adapt when teams shift to working with autonomous agents on an hourly basis. Large enterprises remain constrained by rigid, hierarchical tree-structured organizational charts where all decisions must painfully travel up and down the chain of command. Reaching alignment across deeply entrenched departmental processes remains one of the steepest cultural obstacles preventing legacy companies from adopting agentic velocity.

- Enterprise coding tools are currently delivering efficiency improvements of roughly 20% to 30%, falling short of full-scale workforce replacement. - Major enterprises view meeting recordings as significant legal liabilities, systematically starving internal AI models of executive decision-making context. - Expanding agent access across a 200,000-person enterprise increases the enterprise attack surface by up to 1,000x compared to smaller startup deployments. - Databricks monetizes the rise of third-party agentic tools through consumption rates charged whenever external agents execute queries against its governed data repositories.

[20:00](https://www.youtube.com/watch?v=1bR0gtecoHU&t=1200s) Dunbar's Number, Cross-Functional Latency, and the Company World Model

Up to Dunbar’s number—around 150 people—organizational communication remains relatively straightforward because everyone can know everyone else. Once a company grows beyond that threshold, it inevitably runs into a massive coordination problem. Ali Ghodsi explains that while traditional companies try to solve this with hierarchical organizational charts, the actual flow of value across a business does not follow top-down reporting lines. In practice, customer feedback enters the product management organization, product managers hand specifications over to engineering, engineering builds the software, marketing packages and promotes it, and sales closes deals with customers. Because value creation cuts horizontally straight across the org chart, companies suffer from immense latency at every handoff. To eliminate that drag, an AI system must understand the internal org chart, company intellectual property, and exactly who is responsible for what.

```mermaid flowchart LR accTitle: Cross-Functional Value Creation Loop accDescr: Diagram illustrating how real-world customer value creation cuts horizontally across departmental silos rather than following hierarchical org charts.

subgraph OrgLoop["Cross-Functional Flow"] C["Customer Feedback"] --> PM["Product Management"] PM --> ENG["Engineering"] ENG --> MKT["Marketing"] MKT --> SLS["Sales"] SLS --> C end ```

When Jason Calacanis asks what percentage of companies actually grasp this dynamic—wondering if it is even 5% or 10%—Ali admits he does not know anyone operating this way at scale other than Databricks and possibly Jack Dorsey at Block (formerly Square). Jack Dorsey refers to this exact concept as building a "company world model." Ali observes that digital natives founded within the last ten to fifteen years are inherently further along than legacy incumbents. By contrast, a 50-year-old Fortune 500 giant with 100,000 loyal employees and exceptional products is weighed down by sprawling legacy software, heavy regulatory compliance, and rigid security protocols. Databricks' ultimate product strategy is to package and provide the exact AI infrastructure required for those legacy enterprises to operate with the agility of modern digital natives.

- “At scale, I don't know how many. I don't know the answer to that. At scale, I know that Databricks is doing it. I'm doing it.” — Ali Ghodsi - “He calls the same thing I kind of described. I think, you know, a different way of viewing it is calling it the company world model.” — Ali Ghodsi

---

[22:04](https://www.youtube.com/watch?v=1bR0gtecoHU&t=1324s) Databricks Genie: Building an Enterprise Ontology Graph

Jason shares his own obsession with systems thinking, describing how he is currently deconstructing his entire work across early-stage investing and podcasting into distinct agents and precursor metrics. He likens the mental effort required to coordinate autonomous agents to “playing five chess games at once while riding a horse.” Ali then shares his screen to demonstrate how he runs Databricks internally using their product, Databricks Genie.

```mermaid flowchart TD accTitle: Databricks Genie Ontology Engine accDescr: Architecture of Genie ingesting multi-modal enterprise data into a continuous knowledge graph that powers autonomous agents.

subgraph Inputs["Enterprise Data Sources"] M1["Calendar & Meetings"] M2["Slack & Emails"] M3["Docs, Sheets & Slides"] M4["HR & Org Hierarchy"] end

subgraph Core["Genie Core Engine"] G["Enterprise Ontology Graph"] A["24/7 Agent Inbox"] end

subgraph Outputs["Executive Output"] D["Real-Time Query Dashboards"] C["Causal Analysis & Metrics"] end

Inputs --> G G --> A A --> D A --> C ```

Ali emphasizes that while the market is flooded with chatbots and generic assistants, Genie differs fundamentally because it operates over a living enterprise ontology. This ontology is a continuous graph constructed across the entire company. It monitors reporting structures, ongoing projects, incoming emails, meeting transcripts, Slack channels, and executive decisions alongside traditional static documents like spreadsheets and slide decks. The agent runs continuously, 24/7, maintaining an active inbox and retaining direct visibility into Ali’s calendar and corporate communications.

- “The key thing is that it has an ontology. The ontology is a graph that it constantly is constructing for your company.” — Ali Ghodsi - “You have to be a real systems thinker. It's like kind of playing five chess games at once while riding a horse.” — Jason Calacanis

---

[24:30](https://www.youtube.com/watch?v=1bR0gtecoHU&t=1470s) Real-Time Survey Analysis: Testing Hypotheses and Slicing Department Friction

To demonstrate Genie's capabilities, Ali pulls up a real-world scenario: an internal Databricks HR survey completed by more than 10,000 employees. The dashboard revealed high marks for executive leadership at 88%, but two metrics flagged issues: "work-life blend" stood at 60%, and "action on feedback" sat at 60%. Ali explains that the low action score was particularly frustrating because every year he promises the company leadership will act on employee feedback, and every year the score stagnates. During an executive staff review, Databricks' Chief Financial Officer speculated that poor work-life blend was the root cause driving down the action score.

Rather than waiting a week for the internal data analytics team to run regressions, Ali typed the question directly into Genie during the executive meeting. Genie instantly examined the dataset and disproved the CFO's hypothesis: the correlation was weak, and the relationship was non-causal. To prove it, Genie highlighted a manager whose team reported an abysmal 44% work-life blend score, yet registered an impressive 77% action score—demonstrating a leader who pushes his team hard while actively listening and executing on their feedback. Ali then prompted Genie to break down the low action scores by department to uncover the actual drivers: - **Product Management:** Frustrated by sluggish decision-making and too many competing stakeholders. - **Engineering:** Bogged down by cumbersome internal tooling and bureaucratic systems. - **Sales Operations:** Dissatisfied with career progression and internal promotion paths.

Genie was able to deliver these answers instantly because it drew upon an existing knowledge graph containing 19 pre-indexed data assets and definitions contributed by five distinct employees, complete with authorship metadata and org chart context.

- “And this action one really bothers me because, you know, it's action, by the way, it means that when we ask them for feedback, they give us feedback, improve this and that, they feel like you're not listening to us.” — Ali Ghodsi - “It would have taken me like a week to get the data team to do this analysis for me. But that's just me getting information. That's useful.” — Ali Ghodsi

---

[26:45](https://www.youtube.com/watch?v=1bR0gtecoHU&t=1605s) Org Chart Spans of Control and Measuring Engineering Velocity

Ali transitions to another internal analysis conducted by Genie regarding managerial span of control. Without custom prompting, the system automatically identified organizational anomalies—flagging a deep, inefficient hierarchy where managers had narrow spans of control (chains of individuals managing only three, four, or five direct reports each). This immediate visibility enabled leadership to spot structural bloat and evaluate engineering costs without subjective bias. Jason remarks that this shifts executive culture away from HiPPO (Highest Paid Person's Opinion) leadership—where executives default to “if we're going to have an opinion, let's go with mine”—toward an intellectually curious, data-driven environment. Jason envisions that within a year or two, AI agents will act like an internal "Nest thermostat" for corporate culture, detecting rising stress or burnout across an organization in real time.

Finally, Ali shares one of his favorite internal metrics generated by Genie: an engineering productivity distribution mapping code shipped to production by Databricks' 2,000 engineers. Displaying weekly pull requests, the distribution takes the shape of a classic power-law "L curve," laying bare the stark variance in output across the engineering organization.

- “And what's great about this, too, is instead of you as the leader saying, hey, listen, if we're going to have an opinion, let's go with mine. You've now created a forum where people can be data first, test theories, and you don't have to be dogmatic about who came up with the theory.” — Jason Calacanis - “So this is our 2,000 engineers, and this is how many pull requests or how much code they've actually shipped into production. Look at the middle here. It's maybe per week, roughly, but it has this L shape.” — Ali Ghodsi

[30:00](https://www.youtube.com/watch?v=1bR0gtecoHU&t=1800s) Automated Software Factories and Cutting Meeting Detective Work

Ali Ghodsi examines an internal performance distribution shaped like an "L curve," where the top engineering performers commit vastly more code than everyone else. When Databricks investigated why this disparity existed, they discovered that these top engineers were running automated factories using AI, while the rest were not. Interestingly, higher output did not correlate with higher compute costs; several top contributors spent less on model inference than peers lower on the curve. Ghodsi points out that open-source models make this possible: “You can route simple questions to simple models.” Databricks built a dedicated routing product that dynamically directs simpler queries to lightweight, cheaper models, driving down operating costs.

Beyond code generation, Databricks records all company meetings to pull ground truth from internal discussions and customer calls. Ghodsi recounts wanting to know the status of a recent customer engagement; instead of scheduling syncs, he simply prompted the AI system, and “it literally started by saying the meeting went poorly,” itemizing who said what and listing the exact features the client felt were missing. The model continuously conducts background deep dives while he sleeps:

```mermaid flowchart TD accTitle: Automated Feedback Loop from Calls to Code accDescr: Customer meeting transcripts are parsed for feature requests and routed directly into automated engineering factories.

subgraph DataIngest["Call Ingestion"] A["Customer Calls & Transcripts"] end

subgraph Intelligence["Agentic Parsing"] B["AI Extraction of Feedback & Feature Demands"] C["Executive Morning Deep Dives"] end

subgraph Execution["Automated Factory"] D["Direct Dispatch to Engineering"] E["Automated Feature Builds"] end

A --> B A --> C B --> D D --> E ```

Jason Calacanis compares executive leadership and founder life to detective work, likening it to Peter Falk’s character in *Columbo*: wandering into rooms asking "just one more thing" to piece together who dominated the call or what a client actually thought. Employees rarely spin leadership out of malice, but human perception is inherently biased. AI delivers objective, transparent first principles so leadership can address reality without weeks of bureaucratic hand-wringing.

- Top engineers leverage automated software factories, outproducing peers while keeping inference spend low via intelligent model routing. - Meeting transcripts feed automated pipelines that extract demanded features and feed them straight into engineering repositories. - Executives receive synthetic morning briefings detailing blockers and customer dissatisfaction without scheduling multiple alignment meetings.

---

[32:41](https://www.youtube.com/watch?v=1bR0gtecoHU&t=1961s) Rooting Out Blockers, the "Spider-Man Meme," and Moneyball for Management

A primary operational metric at Databricks is the aggressive reduction of "human blockers." While internal AI tools operate almost instantaneously in the background, progress frequently stalls the moment human handoffs occur—whether waiting on legal compliance, finance approvals, marketing sign-offs, or inter-team consensus meetings that end with scheduling follow-up meetings. Calacanis observes that when milestones slip, organizational inquiries typically degenerate into the classic meme of three Spider-Mans pointing fingers at one another.

```mermaid flowchart LR accTitle: Organizational Blocker Identification accDescr: Fast background AI loops frequently hit human bottlenecks such as legal, finance, and cross-team meetings.

subgraph FastLoop["Agentic Velocity"] A["Iterating AI Agents"] end

subgraph Bottlenecks["Human Bottlenecks"] B["Legal Review"] C["Finance Budgeting"] D["Cross-Team Consensus"] end

A --> B A --> C A --> D ```

Calacanis describes building an internal bot named "Major Decisions" designed to shadow company interactions across Slack, Notion, and email to parse who participates in crucial debates, who drives closures, and who responds fastest:

> “Join every room, read every Notion page, even if I was invited to it or not, get every email, and find for me for the month and every couple of days, give me a report on the most important decisions, who participated in them, and what was the resolution? What's not resolved?”

The tool helped expose velocity metrics across venture decisions—such as exercising pro-rata rights or accepting secondary offers—as well as day-to-day media operations and personnel management. Both Calacanis and Ghodsi dismiss Performance Improvement Plans (PIPs) as an inefficient, prolonged ordeal: “Let's torture them for two quarters,” Calacanis quips, comparing poor effort to an athlete missing practice and showing up late for games. With AI parsing actual contributions, management shifts from subjective appraisals to an objective framework reminiscent of the book and film *Moneyball*.

---

[35:46](https://www.youtube.com/watch?v=1bR0gtecoHU&t=2146s) Cultural Drift and Capturing Corporate Alpha

Ghodsi uses internal LLMs to audit leadership alignment with Databricks’ core operating tenets, such as truth-seeking and first-principles thinking. By indexing all recorded communications, the system evaluates whether individual leaders abide by these tenets or inadvertently establish an isolated micro-culture:

> “And by the way, you could even ask, hey, this leader, what is the actual culture they're implementing? And it'll come up with different, you know, they'll tell you, you know, actually, the culture there is this, not the company culture.”

Calacanis describes this unmonitored deviation as "cancerous" drift, though Ghodsi counters that divergence can also serve as an early signal that the rogue unit has discovered a superior operating model that executive management ought to adopt. In an organization of roughly 12,000 employees, using natural language queries to immediately contrast cultural dynamics and decision-making cadences across regions like Japan, EMEA, and LATAM was previously impossible.

Transparency is central to Ghodsi's implementation: all employees can access these metrics. By showing the entire company the performance curve, Databricks aims to push the long tail to the right, fueling market share expansion. Ghodsi notes that he is not satisfied with 70% or 80% annual growth; he wants Databricks accelerating well north of 100%. While this operational playbook will eventually diffuse across the corporate world, Ghodsi emphasizes: “Right now, this is big alpha. If you can do this really, really well, better than any other organization.”

---

[38:25](https://www.youtube.com/watch?v=1bR0gtecoHU&t=2305s) The IPO Debate: Private Secondaries vs. Public Markets

Calacanis raises the central tension facing late-stage private companies: with secondary liquidity platforms functioning efficiently and capital available from mega-funds, long-lived giants like Stripe, SpaceX, and historically Uber have elected to stay private for extended horizons. He asks Ghodsi how he weighs continuous shareholder pressure to take Databricks public against staying private.

Ghodsi frames the IPO question strictly through the lens of maximizing Databricks’ competitive posture over a ten-to-twenty-year horizon as the business expands beyond data platforms into cybersecurity and marketing:

> “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?”

Evaluating both sides of the ledger, Ghodsi highlights distinct structural benefits of public listings: - **Capital Access:** An evergreen pipeline of deep public equity and debt markets to fund large strategic bets. - **Amplified Corporate Marketing:** The predictable rhythm of quarterly earnings reports functions as an unmatched global marketing vehicle to repeatedly broadcast product strategy and enterprise momentum to the market. - **Administrative Relief:** Staying private imposes its own massive internal burdens; Databricks' finance department must orchestrate complex company-wide tender offers year after year to facilitate employee liquidity across tens of billions in valuation.

[40:00](https://www.youtube.com/watch?v=1bR0gtecoHU&t=2400s) Secondary Liquidity, Employee Tenders, and Competitive Dynamics

Addressing the emotional tension around employee equity, Ali Ghodsi discusses the delicate balance of offering liquidity while running a late-stage private company. While an interviewer observes that letting employees cash out millions could cause them to lose focus or leave the company, Ali argues that predictable annual tender offers actually stabilize retention. Because employees know a liquidity window opens reliably every year, it removes panic. Counterintuitively, employees rarely exhaust the capital Databricks reserves for these secondary tenders. Once staff realize the company is performing at over $7 billion in revenue and compounding at 70% to 80% year-over-year, their calculus changes: when presented with the math of $1.7^3$ or $1.7^4$ compounded growth, people who initially demanded liquidity often decide to sell only 5% of their holdings. This established liquidity track record also empowers Databricks to execute billion-dollar-plus acquisitions using stock, because target founders and employees trust they will be able to monetize their equity without waiting for an initial public offering.

The conversation pivots to the evolving competitive landscape. While Wall Street and industry analysts reflexively pit Databricks against Snowflake, platform boundaries across tech are dissolving. AI labs like Anthropic, infrastructure players like Cloudflare, and development environments like Replit are aggressively releasing adjacent developer tooling. Furthermore, as coding agents elevate developer productivity, Databricks faces a strategic dilemma: either downsize low-end engineering to harvest margins, or redirect massive developer surplus toward building an expansive suite of peripheral products.

- Databricks sets aside dedicated pools for annual tender offers, but bullish employees consistently leave capital on the table. - Financial baseline cited: business revenue at $7 billion-plus with annual growth rates running between 70% and 80%. - Predictable private liquidity enables Databricks to complete billion-dollar acquisitions with equity that targets treat as near-cash.

---

[41:57](https://www.youtube.com/watch?v=1bR0gtecoHU&t=2517s) The Case for Staying Private During the Volatile AI Transition

Ali explains that Databricks deliberately avoids the public markets right now to protect company culture from short-term public volatility and sentiment swings. In a public company, a sudden quarterly dip or macroeconomic wobble plunges employee stock grants underwater, causing developers to feel misled and threaten to quit, only for revenue to accelerate three months later. Ali emphasizes that Databricks will inevitably go public, but doing so in the middle of the current AI platform shift would inject toxic noise into long-term product execution. Public market swings demoralize chief executives, who regularly call him in distress simply because their share price dropped for reasons they cannot fathom. One of the hosts jokes that obsessing over daily stock fluctuations is just like wearing a Whoop band and constantly stressing over microscopic shifts in sleep recovery scores.

> “That’s what I want to avoid so that we can focus on that long term, you know, building the company towards the vision that we have in the long run without being distracted.” — Ali Ghodsi

> “We will go public. We are going to be public. But right now, I think during this transition period of AI, there’s just too much up and down.” — Ali Ghodsi

---

[43:45](https://www.youtube.com/watch?v=1bR0gtecoHU&t=2625s) The Superintelligence Debate: Nick Bostrom and Recursive Self-Improvement (RSI)

The hosts prompt Ali to unpack his framework for superintelligence, a perspective previously highlighted by venture capitalist Ben Horowitz. Ali differentiates Databricks’ practical enterprise work—solving the enterprise context problem using products like Genie and domain ontologies—from the frontier labs chasing artificial general intelligence. Frontier researchers remain captivated by Nick Bostrom’s 2014 book *Superintelligence*, which popularized the concept of Recursive Self-Improvement (RSI). Under Bostrom's thesis, an AI begins iteratively rewriting and upgrading its own architecture, triggering an uncontrollable "fast takeoff" or singularity where humanity ends up as pets or eradicated by a paperclip maximizer.

Ali points out that this narrative is creating a severe existential crisis in Silicon Valley. Veteran technologists who spent ten to twelve hours a day coding from the third grade through their thirties are watching AI agents write 98% of company codebases. Because the system writes code and subsequent iterations write better code, engineers immediately jump to the conclusion that Bostrom's recursive feedback loop has already begun. However, Ali pushes back against the doomer consensus, noting that automating the final 2% of software engineering does not constitute recursive self-improvement, and even Bostrom himself has moderated his apocalyptic views in recent years to emphasize practical cyber risks over immediate extinction.

> “Boatstrom in his book also said, how do we get there? He actually in 2014, I believe, defined RSI recursive self-improvement, the AI starts improving itself and then it gets better and better and better and better and better.” — Ali Ghodsi

> “Just because 98% of the code is being written by AI, we don’t have RSI. If it automates the last two percent, who cares?” — Ali Ghodsi

---

[45:49](https://www.youtube.com/watch?v=1bR0gtecoHU&t=2749s) The Four Conditions for an RSI Flywheel vs. Today's AI Reality

Ali articulates the four strict conditions that must simultaneously hold for true Recursive Self-Improvement to occur, demonstrating why current frontier training runs do not constitute a runaway intelligence explosion. Today's reality is the exact opposite of an automated loop: pre-training runs are governed by brute-force scaling laws where extracting modest linear gains requires an exponential 10x surge in GPUs, data centers, and power. Frontier labs only run a handful of massive training runs per year, leaning heavily on human engineering to avoid bricking multi-hundred-million-dollar clusters and relying on data-labeling vendors like Mercor to manually curate domain-specific training data.

```mermaid flowchart TD accTitle: The Four Necessary Conditions for Recursive Self-Improvement accDescr: A flow diagram showing the four technical requirements needed to trigger a runaway RSI loop compared to today's bottlenecked reality.

subgraph Conditions["The Four Necessary Conditions for RSI"] C1["1. Super-Linear Speedup<br/>(Training runs get faster)"] C2["2. Sub-Linear Resource Usage<br/>(Requires less compute & energy)"] C3["3. Broad Generalization<br/>(Accuracy improves across all domains)"] C4["4. Closed Autonomous Flywheel<br/>(Model trains model continuously)"] end

subgraph Reality["Current AI Training Bottlenecks"] R1["Longer Pre-Training Runs<br/>(Months per major model)"] R2["Exponential Compute Wall<br/>(10x hardware for linear gains)"] R3["Narrow Domain Tuning<br/>(RL gyms & whack-a-mole accuracy)"] R4["Manual Human Engineering<br/>(Vendors like Mercor clean data)"] end

C1 -.->|"Not happening"| R1 C2 -.->|"Violated by scaling laws"| R2 C3 -.->|"Currently brittle"| R3 C4 -.->|"No autonomous loop"| R4 ```

To illustrate how disconnected popular doom fears are from physical reality, Ali highlights that posing a genuine existential threat would require physical agency—such as autonomous killer robots capable of breaking into the room and physically subduing people, which do not exist. He contrasts genuine superintelligence with recent overhyped cyber panics, pointing to an incident where an AI model scanned the internet and found commercial French supermarket refrigerators left exposed with default `admin` / `admin` login credentials. Remotely triggering defrost mode on grocery coolers and spoiling French butter and milk is a basic script-kiddie vulnerability caused by poor router and hardware hygiene, not the arrival of an all-powerful, sentient superintelligence.

- **Condition 1 (Super-Linear Velocity):** Each successive model generation must train in a dramatically shorter timeframe (e.g., three months, then one month, two weeks, one week) rather than merely achieving incremental halving. - **Condition 2 (Decreasing Physical Resource Footprint):** Successive runs must use strictly fewer physical resources and atoms, rather than depending on infinite 10x hardware expansions. - **Condition 3 (Generalized Across-the-Board Accuracy):** Performance breakthroughs must apply uniformly across all generalized reasoning tasks, rather than isolated reinforcement-learning benchmarks or narrow domains like math. - **Condition 4 (Continuous Autonomous Closed-Loop Recursion):** The first three criteria must feed continuously into one another without human intervention, driving costs down in an exponential decay from $100 million to $50 million, $10 million, and down to single dollars. - **The Physicality Gap:** Even satisfying all four software conditions would merely provide a precondition for superintelligence; without physical robotics and kinetic agency, real-world catastrophic scenarios remain speculative.

> “That flywheel doesn’t exist right now. The models are taking longer to train, right? And there’s more data cleaning and there’s more manual whack-a-mole happening where you have vendors like Mercor that has to actually go get the data from very skilled labor.” — Ali Ghodsi

> “You can remotely script kitty into them and turn it into defrost mode and make a bunch of butter go bad and milk, which is a tragedy for the French, I understand, but it is not super intelligence, I can assure you.” — Ali Ghodsi

[50:00](https://www.youtube.com/watch?v=1bR0gtecoHU&t=3000s) Asymmetric Cyber Risks and Securing the Global Codebase

While the prospect of attackers targeting physical infrastructure used to be limited by physical barriers—such as having to physically compromise a supermarket to spoil butter by photographing refrigerator thermostats—the modern attack surface is expanding rapidly through automated software agents. Addressing whether defenders have the upper hand because white-hat cybersecurity teams possess vastly more compute than malicious actors, Ali Ghodsi acknowledges the theoretical merit of that argument but explains that cybersecurity is defined by profound structural asymmetry. An attacker can deploy thousands of autonomous AI agents to probe for vulnerabilities, and every single one can fail except for one; if a single agent breaks through, the attack is an absolute success. Ghodsi compares this dynamic to a suicide bomber or a catastrophic terrorist attack: “I need to be 100% safe all the time. They need to just get one thing through. It's very asymmetric that risk.” Databricks itself operates in this domain with LakeWatch, a cybersecurity product that uses automated monitoring agents to analyze platform data and detect intrusions in real time.

Despite this asymmetry, Ghodsi maintains an optimistic view that society will neutralize the threat because software vulnerabilities are finite, latent bugs in source code rather than an endless generative hazard. In the past, human hackers could only pluck low-hanging fruit; AI agents can now reach much higher into the branches of the code garden to exploit deeper vulnerabilities. However, the software development ecosystem is undergoing an unprecedented shift. In the coming year, humanity will write more code than has ever been written in history, and nearly all of that newly generated code will be scanned automatically for security flaws before deployment. As AI tools handle legacy code refactoring and shadow patching, the total pool of software vulnerabilities will steadily shrink toward a hardened steady state. Ultimately, the lingering point of failure will not be the machine layer, but social engineering and human fallibility. As Ghodsi points out, echoing lessons from his days at the UC Berkeley AMPLab, “the humans do have vulnerabilities. I'll leave it at that.”

```mermaid flowchart TD accTitle: Codebase Security Evolution Under AI accDescr: Visualizing how AI agents probe vulnerabilities while automated scanning cleans legacy and new code.

subgraph Attackers ["Attacker Landscape"] A["Malicious AI Agents"] -->|"High-volume attacks"| B["Probe Exploits"] end

subgraph Asymmetry ["Structural Asymmetry"] B -->|"Only 1 probe needs to succeed"| C["System Intrusion"] D["Defenders"] -->|"Must defend 100% of perimeter"| C end

subgraph Remediation ["Codebase Hardening"] E["AI-Assisted New Code"] -->|"Automated pre-deployment scanning"| F["Secure New Codebase"] G["Legacy Source Code"] -->|"Rewriting & shadow patching"| F F -->|"Finite bug surface"| H["Resilient Steady State"] end ```

- Databricks developed LakeWatch as an agentic monitoring layer that audits platform data to identify active intrusions and supervise autonomous internal agents. - The global software industry is on track to author more source code over the next twelve months than across the entire prior history of computing, with integrated AI scanners filtering out syntax and architectural bugs at generation time. - Security remediations for legacy infrastructure currently involve systematic code rewriting as well as "shadow patching," which wraps protections around unedited legacy applications.

---

[54:02](https://www.youtube.com/watch?v=1bR0gtecoHU&t=3242s) Effective Altruism, Bostrom’s Superintelligence, and the AI Doom Debate

The conversation pivots to the philosophical schisms within the San Francisco and Berkeley technology communities, where effective altruists (EAs), AI decelerationists ("doomers"), and accelerationists clash over the future of artificial superintelligence. Asked whether leaders in the EA ecosystem are genuinely insightful or simply intoxicated by their own wealth and prestige, Ghodsi offers a measured, empathetic assessment: “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.” Having spent decades embedded in the Berkeley academic environment, Ghodsi explains that many of these researchers read foundational texts like Nick Bostrom’s *Superintelligence* years before modern large language models existed and deeply internalized those existential warnings. When modern models began demonstrating rapid capability leaps alongside novel cyber exploits, these thinkers connected the dots to conclude that a runaway intelligence explosion was imminent.

Ghodsi rejects this fatalistic trajectory, noting that he sees zero empirical evidence for Recursive Self-Improvement (RSI) fulfilling the technical prerequisites necessary to cause an uncontainable takeoff. Furthermore, even if such software systems were created, they operate exclusively in digital compute clusters without physical control over society: “Like it's just digital. Like you can just shut down the computers right now.” Instead of worrying about hypothetical science-fiction tropes like humanity being transformed into paperclips, Ghodsi argues that society should focus on concrete, tangible near-term harms: hospital and utility infrastructure shutdowns from cyberattacks, the psychological fallout of children forming parasocial bonds with synthetic characters, and algorithmic manipulation distorting public information and shared truth.

- Prominent figures like Anthropic CEO Dario Amodei have made sweeping public claims during investor roadshows, suggesting that conditions like cancer and Alzheimer's disease could be eradicated within ten years. - Critics worry that overconfident public timelines regarding medical breakthroughs risk peddling false hope to families dealing with terminal diagnoses. - Ghodsi separates far-fetched existential doom from immediate societal risks, noting that society already struggled through major disruptions with social networks and Dunbar’s number, and is now actively grappling with automated information warfare.

---

[56:32](https://www.youtube.com/watch?v=1bR0gtecoHU&t=3392s) Real-World AI Impacts: From TEDDY Drug Discovery to Global Good

Evaluating Amodei’s assertion that AI will cure cancer within a decade, Ghodsi explains that computational speedups alone cannot bypass empirical biological reality. While digital modeling will dramatically compress the discovery phase, real treatments must undergo rigorous, sequential animal and human clinical trials, which inherently take time. Nevertheless, AI is already achieving transformative scientific breakthroughs. At Regeneron, researchers used Databricks to pinpoint the exact genomic mutations responsible for chronic liver disease, paving the way for targeted therapeutic drugs. More recently, Databricks partnered with pharmaceutical giant Merck to develop TEDDY (Transformer Enabled Drug Discovery). Rather than predicting the next natural-language word in a sentence, TEDDY operates on Gene Regulatory Networks (GRN), mapping whether a specific gene is actively driving a pathology or merely reacting to it, preventing pharmaceutical companies from wasting years developing drugs against downstream biological symptoms.

Beyond biopharma, Ghodsi points to several humanitarian applications already scaling on Databricks infrastructure. Crisis Text Line uses large language models and autonomous triage agents to detect subtle markers of self-harm in messages from distressed youth and route them to immediate intervention. Similarly, organizations like the Virtue Foundation and drone delivery company Zipline rely on Databricks to process massive data streams, allowing Zipline to map and execute automated aerial drops of critical food and medical supplies across challenging terrains. Ghodsi contends that these tangible benefits are frequently obscured by researchers who understand abstract math but lack a grasp of macroeconomic reality, relying instead on sensationalized science-fiction tropes: “And the idea that there's a 10 percent chance that seven 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.” For Ghodsi, existential doom is virtually zero, while the pragmatic cyber risks remain real, manageable, and worthy of focused engineering.

- **Regeneron:** Utilized earlier-generation Databricks machine learning tools to decode the genetic mechanisms behind chronic liver disease and design targeted treatments. - **Merck and TEDDY:** The Transformer Enabled Drug Discovery model applies transformer architectures to Gene Regulatory Networks (GRNs) to distinguish causal disease drivers from reactive genetic expressions. - **Crisis Text Line:** Employs agentic LLM pipelines to analyze inbound text messages in real time, detecting self-harm intent to prioritize lifesaving human counseling interventions. - **Zipline:** Processes massive volumes of telemetric and operational data on Databricks to plan flight trajectories and coordinate commercial drone logistics for remote medical and nutritional deliveries.

[01:00:00](https://www.youtube.com/watch?v=1bR0gtecoHU&t=3600s) The Enterprise Context Gap and the Decades-Long AI Rollout

Reflecting on the pace of real-world adoption, Ali Ghodsi pushes back against the ultra-compressed timelines touted by frontline AI research labs, arguing that closing the enterprise "context gap" will be a multi-decade journey rather than an overnight transformation. While frontier labs often claim massive disruption will arrive within six months, a year, or two years, Ali believes macroeconomic indicators and GDP will shift far more gradually. As Ali puts it: “I actually think this is something that’s going to take 10, 15 years. The labs think this will happen in the next six months, 12 months and two years. But I actually think GDP will move very slowly and it’s going to take decades for us to actually get through it.”

Ali grounds his skepticism in the operational realities of heavily regulated corporate environments, circling back to his earlier anecdote about the bank CEO who revealed that his company is legally prohibited from even recording internal meetings. If enterprise systems cannot capture basic day-to-day corporate communication, an AI model has no raw material to build context or understand workflows, making advanced tools like Databricks Genie or autonomous ontologies impossible to deploy. Jason Calacanis agrees, noting that this pacing mirrors the early trajectory of the commercial internet: early adopters and tech pioneers sprinted ahead and recognized that the technology would transform the world, yet 99% of users were still constrained by slow dial-up modems and could not fathom high-bandwidth applications like YouTube or infinite cloud photo storage.

```mermaid flowchart LR accTitle: Enterprise AI Adoption Barrier accDescr: Diagram illustrating the bottleneck between raw AI frontier models and enterprise deployment due to missing institutional context.

subgraph Labs["Frontier Labs Outlook"] A["Frontier Model Breakthroughs"] -->|"Expect rapid impact (6–24 mos)"| B["AGI Expectations"] end

subgraph Enterprise["Enterprise Reality"] C["Strict Corporate Constraints<br/>(e.g., No meeting recording)"] --> D["The Context Gap<br/>(Missing organizational data)"] D --> E["Delayed Implementation<br/>(10–15 year GDP rollout)"] end ```

The hosts draw an analogy between early AI vulnerabilities and the nascent days of cybersecurity on ARPANET and the early web. They recall Robert Tappan Morris, the creator of the famous Morris worm, who inadvertently brought down roughly 10% of the entire internet in 1988, narrowly avoided a severe prison term, and ultimately went on to become an esteemed computer science professor at MIT. With that historical parallel underscoring that transformative technologies require time, maturation, and resilience against systemic risks, Jason brings the interview to a close, thanking Ali for breaking down Databricks' vision and the broader future of artificial intelligence.

- The gap between AI capabilities and real enterprise value is estimated by Ali to require a 10-to-15-year horizon, contrary to research labs predicting total upheaval within 6 to 24 months. - The fundamental bottleneck remains data access and institutional permissioning; models cannot automate or reason about workflows they are not permitted to observe. - Historical parallels like the dial-up era and the 1988 Morris worm illustrate that early network technologies endure growing pains and infrastructure bottlenecks long before mass adoption fully materializes.

## Notable quotes > [10:00](https://www.youtube.com/watch?v=1bR0gtecoHU&t=600s) And actually, if you happen to stop all progress right now, I think we have all the intelligence we need to completely transform the planet. — Ali Ghodsi

> [03:41](https://www.youtube.com/watch?v=1bR0gtecoHU&t=221s) The most important data any company has is the data on the core product that they have. — Ali Ghodsi

## Mentioned Adobe Photoshop, Airwallex, Ali Ghodsi, All-In Podcast, AMPLab, Anthropic, Apache Spark, ARPANET, Ben Horowitz, Block, Cloudflare, Crisis Text Line, Dario Amodei, Databricks, Databricks Genie, David Friedberg, Facebook, Google, Jack Dorsey, Jason Calacanis, LakeWatch, Merck, Mercor, Microsoft Excel, MIT, Moneyball, Netflix, Nick Bostrom, Notion, OpenAI, Regeneron, Replit, Robert Tappan Morris, Slack, Snowflake, SpaceX, Stripe, Superintelligence, TEDDY, Terence Tao, UC Berkeley, Uber, Virtue Foundation, Zipline

## Video description (0:00) Ali Ghodsi joins the show: Databricks’ early bets on cloud, AI, and open source (8:50) Is AGI Already Here? Why Enterprise AI Is Falling Behind (22:04) Live Demo: How Databricks Uses Agentic AI For Efficiency (38:25) Databricks IPO: The Case for Staying Private (44:02) RSI, Cyber Risk & the Doom Debate

Thanks to our partner for making this possible:

Airwallex is a leading global payments and financial platform for modern businesses, offering trusted solutions to manage everything from business accounts, payments, treasury, and spend management to embedded finance. Learn more at https://airwallex.com/allin

Follow Ali: https://x.com/@alighodsi

Follow the besties: https://x.com/chamath https://x.com/Jason https://x.com/DavidSacks https://x.com/friedberg

Follow on X: https://x.com/theallinpod

Follow on Instagram: https://www.instagram.com/theallinpod

Follow on TikTok: https://www.tiktok.com/@allin

Follow on LinkedIn: https://www.linkedin.com/company/allinpod

Intro Music Credit: https://rb.gy/tppkzl https://x.com/yung_spielburg

Intro Video Credit: https://x.com/TheZachEffect

#allin #tech #news

[Source](https://www.youtube.com/watch?v=1bR0gtecoHU)

Reading summary

Databricks CEO Ali Ghodsi contends that artificial general intelligence has already arrived under its original definitions, yet enterprise adoption will take decades rather than months due to severe institutional constraints around data context, security, cost, and governance. AGI Has Arrived While Labs Chase Superintelligence (10:00) Ghodsi argues that the raw machine intelligence necessary to transform the economy already exists, noting that frontier labs shifted goalposts from AGI to superintelligence—entities capable of superhuman reasoning across multiple dimensions. While lab researchers focus on esoteric academic benchmarks, most Global 2000 and Fortune 500 enterprises deploy AI merely as glorified search or for modest 20% to 30% developer efficiency gains. Ghodsi asserts that if model progress froze today, existing technology would remain sufficient to reshape commercial enterprise once properly integrated. Four Structural Obstacles to Enterprise AI Adoption (13:48) Enterprise agent deployment is bottlenecked by four practical barriers: context capture, cybersecurity, token economics, and strategic sovereignty. Enterprises routinely block recording of executive meetings over legal discovery and regulatory liability, starving models of critical institutional context. Additionally, connecting autonomous agents across hundreds of thousands of staff balloons corporate attack surfaces by up to a thousandfold. Meanwhile, runaway inference costs threaten operating margins without concurrent revenue expansion, and companies fear intellectual property leakage to foundational model vendors shipping competing SaaS applications. The Corporate World Model and Objective Operations (20:00) Value creation cuts horizontally across departmental silos rather than vertically through hierarchical organizational charts, producing friction and handoff latency. Databricks addresses this internally by building an enterprise ontology graph using Databricks Genie, which continuously indexes meetings, communications, and workflows. This infrastructure allowed leadership to instantly disprove executive assumptions—such as demonstrating that poor employee feedback was tied to tooling and bureaucracy rather than work-life balance—and track engineering productivity distributions to route simple tasks to cost-effective models. Sustained Private Growth and Capital Strategy (40:00) Reported at an estimated $7 billion revenue run rate compounding at 70% to 80% year-over-year, Databricks elects to remain private during the AI transition. Ghodsi explains that staying private shields engineering culture from public equity volatility and short-term earnings pressures, while structured annual employee tender offers provide sufficient liquidity and stabilize retention. While an IPO remains inevitable to access public capital markets and leverage earnings cycles for corporate marketing, current liquidity frameworks allow the company to pursue major stock-based acquisitions without market distractions. Dismantling Recursive Self-Improvement and Doom Scenarios Ghodsi rejects catastrophic AI extinction narratives, arguing that true Recursive Self-Improvement requires four simultaneous technical conditions that current architectures fail to meet: super-linear training speedups, sub-linear physical resource usage, uniform across-the-board generalization, and closed autonomous execution without human data curation. Instead, contemporary scaling hits resource walls requiring 10x compute expansions for linear returns. Ghodsi views tangible near-term risks—such as asymmetric cyber threats against legacy software and algorithmic disinformation—as manageable engineering problems, while transformative applications advance in fields like genetic drug discovery and emergency response. Ghodsi projects that closing the enterprise context gap and seeing machine intelligence register in global GDP will require a 10-to-15-year rollout, likening today's fragmented adoption to the dial-up era of the early commercial internet.

Source quotes

[10:00] Ali Ghodsi: And actually, if you happen to stop all progress right now, I think we have all the intelligence we need to completely transform the planet. [3:41] Ali Ghodsi: The most important data any company has is the data on the core product that they have.

Original transcript

[0:00] So it was kind of out of despair that we started Databricks because no one else is listening to it. 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 a thousand agents doing stuff realistically. Am I correct? They're just using AI in very rudimentary ways. They're just using it as a glorified Google search. We're actually getting into like management money ball, aren't we? So if the token cost is going up exponentially, is your revenue at your company 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? When are you going public? I'm going all in. If you were building a global financial system from first principles today, you wouldn't build it on 50-year-old legacy rails. You'd build Airwallex, one AI-native platform for global accounts, cards, and payments. It's designed to make the entire world feel like a local market. Others are bolting AI onto broken infrastructure. But Airwallex was built for the intelligent era from day one. Stop paying the legacy tax and start building the future at airwallex.com. Airwallex, built for the future. All right, everybody, welcome back to the all in interview. We're so lucky because all of these amazing leaders in the world, CEOs, politicians, scientists, they watch the show. They're part of the community. And Ali Gatsi is one of those individuals He, of course, is the CEO of Databricks, a private company that has been doing incredible. I don't know why it's not public yet. I think a lot of people are asking that question. We'll talk about it today. In 2015, they were doing 1.5 million in revenue. It's August or the rumor I heard August 2026, Ali. $7 billion run rate growing 80% year over year. And obviously you got a lot of competition, the snowflakes of the world, etc. But this is a tremendous business. And we're going to talk a lot about how you're using AI to run this incredible business. We're going to talk about recursive. self-improvement but welcome to the program and then maybe you could tell the audience a little bit about what you do at Databricks and how it became so successful. Yeah, happy to do that. Thanks for having me on. Excited for this. Yeah, so 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 actually part of a lab called Amplab, which was, I think, the largest AI research lab in the world at that time. And, you know, we were trying to figure out how can we combine data processing with AI. And, you know, we came up it. So we were trying to get the industry to adopt this technology and they weren't. So it was kind of out of despair that we started Databricks because we thought, you know, 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. Actually, it was called machine learning at the time and AI was just robotics. So that's kind of like the beginning of the story. And for people who don't use the product day in and day out, who 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, as well. Conceptually, you got there so early, obviously, machine learning. Google was built on that Facebook and their ad network. A lot of interesting services, Uber and their routing algorithms. Maybe you could talk a little bit about where Databricks started and what it's become and how it's generating billions of dollars in revenue. Yeah. So the first use case, actually, it was the first use case ever was at UC Berkeley. We wanted to help Netflix recommend better movies so that people would watch more movies. So it was like a recommendation algorithm, AI use case that started the whole thing. And then we thought, you know, more organizations, probably, probably every organization will need a data science team that can do this kind of AI stuff. And it was kind of bold at the time. There's a lot of people that said, no, no, no. AI is just for ad tech. The ad companies need it. But we thought, no, no, this is going to be a bigger market. And actually, we had a lot of debates internally. We weren't sure it was going to explode. So we got into it. And then we realized when we started the company, we realized why they're not doing it, why it's so hard, because basically all the organizations out there in the world are bogged down. they've had for 30, 40, 50 years. And some of them, you know, companies have been around for 100 years. You know, things are moving slow. They have a lot of processes. There's, you know, a lot of compliance, a lot of regulatory, a lot of security. But all these things have gone worse with generative AI in the recent years. So basically what was missing from our research at UC Berkeley was that we didn't think about the security, the governance. enterprise. So, you know, organizations have a lot of data. The most important data they have, actually, people don't talk about it. The most important data any company has is the data on the core product that they have. Like for you here, it's this podcast, [5:00] you know, the data on that. Someone else might be selling insurance, someone else might have, you know, a drug. It's the data that comes out of that. That data doesn't sit in CRM, it doesn't sit in human resource management. In fact, it probably wasn't even tracked. And remember, in the early days of the internet, They'd be like, hey, would you like this log file? And, I'd say, OK, yeah, keep everything. And they'd say, OK, well, you got an $8,000 storage bill a month. And I'd be like, okay, delete it. And they're like, well, we could roll it up and keep some of it. I'm like, well, who's ever using the open rate on our emails? Who cares about that? You know, we send the emails. That's it. But then that changed and was storage and compute and cloud computing a major accelerant? to this big data movement and then eventually machine learning. Those were kind of the waypoints in this movement. Yeah. Before we started in 2013, basically people were running their own data centers. And when they're running their own data centers, you have to buy machines and the machines came with storage and compute combined. So if you ran out of storage, you have to buy more machines. And the IT departments of every organization were busy with this. What the cloud enabled was to just completely decentralize this and separate compute and storage so that you could just imagine you have infinite storage, you just dump everything there. Okay. And if it's gone cold, so the data it becomes almost free to store it and then you could get the illusion of as much compute as you wanted and then of course you could start running machine learning on it and then eventually gpus showed up uh so it kind of transformed everything we've got everything the whole farm on just being a cloud company we bet everything on ai the use cases of ai and we bet on open and open source and at the time actually all three of these bets were controversial people i had a standing bet blunder, you're going to go under because you're betting on the cloud. And it turned out actually the only reason we succeeded was because of these bets, because they were contrarian, because other people did not agree with them. So we got a head start for many, many years. And then eventually, when we happened to be right, then we had, you know, we were so far along and we got too big that, you know, there's too much momentum to stop it. It's so funny you mentioned this Netflix show. algorithm. Netflix started, obviously in the Bay Area as DVDs that were mailed to you in an envelope, and then you had this queue. And it was critically important for people who don't know the Netflix prize. This was in the 2008-2009 time period, they put a million dollar prize out to any computer science team that can make the algorithm, I think, 10% better at predicting, what movie you would give four or five stars to. So they just took all their data, they released it. And then they got probably $50 million worth of work from computer scientists around the world. It was the first time an X Prize And I think XPRIZE and this were around the same time, and they were all inspired by those old prices to like who could circumnavigate the globe, who could solo fly over Amelia Earhart, who could who could solo fly over the Atlantic or the Pacific. It was like a very interesting moment at time. And That hit AMP, right? And AMP The group was algorithms, machine learning, I don't know what the P stands for. It was for people, actually, because we thought that, you know, the AI is not good enough yet. So it needs a little bit of human nudging. So it's actually this kind of, you know, human feedback was already there. We were like, we do need the human feedback because this last mile is just too hard for AI to figure out itself. So we had people there as well. Putting the people in there makes sense because Berkeley is a little hippy dippy. Freeberg's alma mater, obviously. They were like, we need some people in. the middle of this. We had Ben Horowitz sat in the other week. And, you know, He was like the best explanation I've ever heard. for why we are going to be able to control the rollout of super intelligence and this jump between artificial intelligence or artificial general intelligence and super intelligence that gets recursive and runs off with it and p dooms us all and we're all going to die sentiment by some people came from you so please explain your thesis here and then i'd like you to go even one step further and talk specifically about maybe the culture in the bay area perhaps even the east bay and this group of people i mean happy to get into it you know i think there's uh sort of two 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 based that on the kind of consensus we had in that AMP lab where it was, I think, the largest AI lab at the time in 2009, 2010. And that definition of AI that most of us had, AGI, [10:00] We already achieved that last year. And I wasn't actually sure about this. So I actually went and checked with some of the people that were in the lab like, 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. And actually, if you happen to stop all progress right now, I think we have all the intelligence we need to completely transform the planet. And do all intelligence work by humans, right? I mean, 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. And I kind of am in line with you. We've hit artificial general intelligence. It just hasn't been equally distributed yet. Yeah. So the labs and a lot of researchers, they're kind of obsessed with super intelligence. OK, so when I say super intelligence, I mean it the way it was defined by Baudström in 2014 in his book called Super Intelligence. And there he's talking about something completely different. He's talking about something that can do a millennia worth of thought. So like a thousand years of human thought, it can do it like in a minute. OK, all of mankind in a minute. It thinks in 11 dimensions. It's so much faster. Just a single agent is just, you know, many orders of magnitude. Imagine thousand times faster than any of us. And that's just a different thing. Like that's it can write a novel PhD thesis instantaneously and so on. So I think the labs are on the quest of that. And in that book, by the way. And so just to summarize that there. The ability to write an entire unique PhD thesis Instantly. which do even have the amount of compute in your mind to do something like that today? If everybody got together and said, Hey, Today, we're going to give all our compute, or for the next hour, just as a thought experiment, to this incredible mega prompt, do you think would actually happen? Is it a computer issue or is it a model issue? Well, actually, the AIs are kind of slow. What we have to do is we have to paralyze them and have many of them and we have to have them to collaborate with each other. So it's actually not the way kind of that he wrote it. He had also intelligence in the book of how it can be communities of intelligence. But I'm just saying the labs and many of the other companies are focused on this. That's why they're so obsessed with the math, the clay problems, the seven millennial problems and so on. Why don't we actually just make this AI applicable inside organizations. That's what Databricks is focused on. We just want to help all these organizations. And when I go talk to organizations out there, enterprises, I see that they're just using AI in very rudimentary ways. They're just using it as a... Glorified Google search. They're using chatbots all day. And instead of typing in Google search and getting 10 blue links and then looking at the 10 blue links, the AI just does that for them. The second thing that they're doing is they're doing coding. Okay. And with the coding, actually, they're not seeing enough speed ups. They're seeing maybe like 20, 30 percent. I ask them this question every day. I'm in meetings every day with customers 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 work and doing core work for you and so on? It's just not happened yet. Now, they might be running a half dozen agents, a dozen agents like. somebody who's at the forefront of, might actually be coordinating against that many, but the human still needs to be on the loop, just like AMP. There's a person who... Still needs to be in there because it will return slop. It will return weirdness, it can go drift off course. It's not possible to have a thousand agents doing stuff realistically. Am I correct? Yeah, well, some are further along. I mean, I'm actually using this all throughout Databricks and I'm happy to show you. But, you know, the reality is that they're not. And you have to ask the question, why is that? Like, if it can solve Navier Stokes, very difficult math problems, how come enterprises and organizations out there are so far behind? Why? That itself, 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 have liability. liable. I mean, by the way, which I already I was already kind of shocked because if you can't do that, how can the AI ever help you? Like the AI is simply not in the meeting. If it's not in the meeting, it's useless. It doesn't matter if you have Terry Tao or your greatest mathematician in the world. If they're in the cold and they don't know what decisions or what you're doing, you're meeting with hundreds of thousands of employees. There's no way it can help you. So if you capture all that context and you could feed it to the AI, they would nail the problems. They could help us figure that out and do that. Then we were doing that at Databricks, but that's not enough. Even if you do that, you have a second problem, which is, you know, you could unleash something like OpenClaw and just have it access all your data everywhere. Well, then you get into this problem that you're going to get hacked. Cyber is a real thing. And all the big organizations that have successful businesses, they're really [15:00] be terrified right now of getting attacked. Right. This is like a big concern. It certainly creates a lot of surface area. when you embrace this i mean if a ceo or in a small company you want to connect everything It's a certain amount of service area, but for a 200,000 person organization, 10,000, My lord, the surface area has now... gone a thousand X. Yes, exactly. So that's like the second problem you bump into, which is, OK, we can access all the data. Well, wait, we don't want to maybe access all of it to the agents because, you know, now we have an insider threat. and we might get hacked. So what do we do about that? Well, let's say we solve that too. So now we've solved the cyber problem, which we have actually ways to do that, that Databricks happy to show you. Then you still have another problem, which is, okay, now the token cost is going up exponentially. So if the token cost is going up exponentially, is your revenue at your company going up exponentially? And I don't know any company who started using AI. And as of their use of AI, the revenue started accelerating exponentially, matching their cost curve. get more AI, get it in there. We need to move faster. Token maxing. Yeah, exactly. But then at some point the CFO shows up and says, "Hey, our OpEx is going up significantly here. And even the COGS, what are we doing about it? And we have an earnings call soon." And then they freak out. So then how do we fix that? So you need to also do this efficiently, which we also have ways to do that at Databricks. Now you've done all of these three, then there's a strategic last question that still remains, which is like, okay, fine. We figured it out. We accessed all the data, we secured it, and we figured out the cost situation. How do we avoid getting locked in and how do we own our own IP? Because, you know, it's otherwise, you know, there's no business to be made in the future. Data sovereignty is what you're kind of referring to intelligence sovereignty. And if you use Anthropic, OpenAI or any of these platforms, There is this risk that it will leak. that it will get incorporated into their next version. And as we've seen, a company called Anthropic, Seems to be releasing... Every two weeks, a half dozen products, that just now they're competing against Microsoft, one of their, I think Microsoft is one of their giant partners, And they just came out with, basically an Excel killer. They're, killing photoshop and they're just going after everybody wholesale Yeah, so these four. For free. Yeah. Yeah. So by the way, these companies are amazing. They're doing amazing work. And without them, we wouldn't even have this intelligence, probably. So, you know, I'm thankful. This is awesome. I'm just pointing out why when you look around you all these, you know, successful organizations all the global 2000 all the fortune 500 nasdaq and so on how come it's just glorified chatbot and a little bit of coding This is the reason. So you need to really tackle all four of those. And only then can you start getting something out of it. And one of those issues as well, I believe, is... the autonomy in an organization to come up with a groundbreaking idea. You know, you may have captured a large portion of your market. You may be getting incremental benefits. So if I tell you like, hey, we can put another 500 horsepower into your four tourists, it's like, OK, my Ford Explorer now has 500 extra horsepower. It's going to rip the chassis off. It's not this thing is not capable of going 500 miles an hour. So then it's, Well, who decides we're launching three more product lines? Well, at Improvix, somebody is deciding. Why don't we just go after every single business in the world? And they started doing dashboards. So that's a little close to your business, is it not? A little dashboards over here. Yeah, so dashboards, we don't have a product where we charge for dashboards. Dashboards are usually backed by data and you typically want that data to be governed. So you don't want just numbers and then those numbers are incorrect. This is a big problem in organizations. So that gets back to the second thing I said, the governance security aspect. OK, so like, yeah, maybe we can get access to all the data and get all the context, but we've got to make it secure. So, yeah, it's great. So you can actually connect all of these agents from outside to Databricks. But then we are the governed data repository. And then when you access that data and the governance, we charge consumption rates on that. So actually, that's one of the reasons you see acceleration of our business, because all these companies, whenever they have agents and they're using it, then I think we're just in the beginning. As I said, I still think it's mostly Google search and coding, but the agents, when they start using that, they just send us more queries. How does the behavior change when people actually embrace? agentic technology and they start working with agents. let's say every hour, every other hour of the day like you and I do. I think that the organization, the org charts are built this way. You know, there's like a tree structure, the CEO at the top and all decision making goes kind of up and down in these org charts. And you have all these old processes and getting agreement across this org chart is very, very hard inside of an enterprise or inside of any organization that's sizable. It's something [20:00] Easy up until Donber's number, which is 150. you can know everyone. So small startups, it's a different story. But above 150, this becomes a coordination problem. But if you look at how actually any organization actually you know, sells a product or a service to its customers. It's actually not this kind of hierarchical thing. It actually cuts across that org chart. You know, like here at Databricks, we have customers. They give us feedback. The feedback goes to product management organization. Product management hands it to engineering organization. Engineering builds products that goes to marketing that markets it, you know, and the marketing gives it to sales that sells it to customers again. So you kind of want to go across. So you want to cut out that latency so that you can be really, really fast. your IP, it needs to understand your org chart, it needs to know who's who, who's responsible for it. What percentage of companies actually understand what you just described, would you say? 10% of organizations understand this? 5%? At scale, I don't know how many. I don't know the answer to that. At scale, I know that Databricks is doing it. I'm doing it. Yeah. I haven't seen anyone else, to be honest, do it at this kind of, you know, I think probably block and square with Jack Dorsey. Since he's pushing, he's calling it the company world model. He calls the same thing I kind of described. I think, you know, a different way of viewing it is calling it the company world model. So I think that they're probably pretty far ahead. In general, I think digital natives are probably further ahead. Digital natives being companies that. have only existed 10, 15 years. You know, if you've been around 20, 30, 40, 50 years, which means you're an amazing company, maybe you're a Fortune 500, the best companies in the United States. You have maybe 100,000 employees. You know, they're loyal. You know, they love your product. And you know what? You're the best in the world on that. You're just you have so much software. You know, you have so much regulation and you have so many processes around security and so on that it's very hard for you to adopt this. But that is the very problem Databricks wants to solve for them. technologies so that they can do what what I do at Databricks, basically. Yeah, let's see it. You said you would be willing to show us like how you run Databricks with agents and and with AI and. I've been looking forward to this all day because I literally am taking the entire while you queue it up. my entire life on podcasting. and on early stage investing and i'm looking at each of the components of those things as services as agents and defining the goal working backwards to what are the signs of success that you're heading towards the goal, not that you've achieved the goal, the achieving the goal is kind of binary, yes or no, you made the sell, you didn't. But what are the precursor signals? And then trying to have each of these bots hand off things back and forth. It's, yeah, You have to be a real systems thinker. It's like, kind of playing five chess games at once while riding a horse i mean you're It's really like... Intellectually, I find it super stimulating. I can tell you do, too. So here you can share your screen. So let's jump in. OK, so the key thing here is that we have a product called Genie. But the key thing that Genie has because, you know, there's so many chatbots out there and there's so many assistants and there's so many co-work products out there and agents in the background. The key thing is that it has an ontology. The ontology is a grasp. that it constantly is constructing for your company. That graph understands all of the people, who reports to who, what are they working on, what are the emails that are coming in, what meetings have happened, and what's the transcripts of all those meetings and all the decisions that are being made there, what's on Slack, what's being communicated. So it has access to all of these things. And then also all of the docs that you're creating, all the Excel sheets, all the PowerPoint presentations. So it has access to all of those. That's what we call genie here, so you can see here. So I wake up, it has an inbox. In the morning when I come in, it basically has It's running 24-7. As you can see, it has access to my calendar. In fact, I could just ask it, like, you know... Thank you. Which I tell Jason on all in today, here it goes. Yeah. So see what this is going to say. But let's look at some of the stuff that I've done. So like actually the other day, one of the interesting things is I was looking at our HR department came back and they were saying, hey, we've done survey on all of Databricks. You know, so we had over 10,000 people employees that responded to the survey. And, you know, I looked at the results and she was going through it. And, you know, I said, hey, make me a dashboard of the results. You know, nothing magical here yet. So it told me like, look, here's how you're doing. You're doing great on some things. Executive leadership 88% you're doing okay. Yeah. Work life blend 60% that you're working people pretty hard over there. But maybe everybody says that, right? Yeah, no, we want to know about this. This has been so we want we wanted to improve these two. Okay. And this action one really bothers me because, you know, it's action, by the way, it means that when we we ask them for feedback, they give us feedback, improve this and that. [25:00] they feel like you're not listening to us. So two out of three feel you are one out of three feels you're not just you're doing okay. It's not terrible, by the way. This could be like 30 percent. 60 percent is not great. So but and especially not great if you have to take into account that every year it's low and every year I promise the company we're going to fix this and every year it doesn't happen. So, you know, we're going through the results and, you know, we're discussing it with these stuff. And 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 it's causing, work-life blend is causing low action. And I'm like, that doesn't make any sense. But I just entered here immediately as we were discussing, and it immediately pulled up. Okay, this is the correlation between these two. And this is how it's looking. But then 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. But here's a manager that has work life plan 44. That's pretty bad. Action is 77%. So, you know, they're working hard and he's recognizing and taking action on their concerns. So that your guy right there He pushing people hard and he listening to them I like that So it OK So now I learning stuff about the organization This is amazing It would have taken me like a week to get the data team to do this analysis for me But that's just me getting information. That's useful. So like we're cutting this, you know, latency is up and down the chart a little bit. But then I'm like, OK, why is it low? And look through all of the comments and it gives me some reasons. I didn't like the reason. So then I actually clarified the question. I said, hey, can you slice it by the department for me? And, you know, and then actually figured it out. It said, look, product management has low action because they don't like decision making. It's too many stakeholders. Engineering, they don't like the tools and systems. There's too much. You know, sales operations, it's career growth and so on and so forth. So that makes sense because salespeople always think they should get a promotion because they're salespeople and they should get their commissions raised. And product people think you should just listen to them instead of, you know, six people trying to design a piece of art. Yeah, exactly. 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 its ontology. So it's not going and searching live only. It actually found that it already had in its graph these people. Already five people had already contributed to this And there 19 pieces of information here that you can see that it actually leveraging that it already had it didn search for them live these are in the ontology so we can look at for instance this particular one if you click on it it has a definition it a metric and there's a graph and this graph you can see it here uh you know so there's a favor ability and it came from you know this is the author you can click on the author and because it understands the work charts you understand who this person is this person actually has 29 snippets you know they have lots of different assets you know that they've uh you know that it's learned from them so this constantly is learning the intelligence or ontology of your organization so that's a that's a you know typical thing here's another one that uh you know is maybe uh interesting um i wanted to know actually there's new results here in the last 59 minutes i haven't actually looked at all of those but i wanted to understand the org chart and what's going on with the org chart and you know span of control and i asked it and already out of the box it was actually it I was shocked by this. It immediately told me, like, hey, there is an org inside Databricks, which is very deep. You know, this person reports to this person, reports to this person, the person to this person, but they have three people reporting to them, five people reporting them four people reporting them five people reporting to them So that not a flat organization you know so then you can look at okay what can we do about that Again it understands since it understands all the people it can immediately out of the box understand that engineering you know cost Cost is something that we look at at all. And, you know, so that's that's another thing we wanted to see if we could understand. And what's great about this, too, is instead of you as the leader saying, hey, listen, if we're going to have an opinion, let's go with mine. You've now created a forum where people can be data first, test theories, And 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. And eventually, if this is where we are today, can you imagine like in a year or two, it's going to just be giving you these instead of through a culture report, it's going to be like a temperature gauge in your company. It's just like you're going to be looking at the nest in your company. It's like, oh, wow, looks a little hot today. What's going on? It's like, oh, yeah, we made people come in this weekend. And I mean, we should buy some pizzas or give people a comp day during Thanksgiving. Yeah. Yeah. This is my favorite. So this one actually is it came out of Genie. So this is our 2000 engineers. And this is how many pull requests or how much code they've actually shipped into production. Okay. Look at the middle here. It's maybe per week, roughly, but it has this L shape, you know, and you look at the [30:00] This L shape. The top people are actually committing way more. Yes. you know, so what's going on? Yeah. So what the hell is going on? So we looked into this and it turns out they're running automated factories. The rest are not. So we pushed them. And actually, by the way, interesting thing, I asked it, hey, can you like add the cost to this, add the cost to this line? And one of the interesting thing we found is that it's not even the case that cost necessarily is higher for those people. So look here at the top. The top folks, some of them are actually not spending as much as some of the other folks. So which, by the way, goes back to, open source, you can route simple questions to simple models. And we actually have a product for this and you can cut your costs down significantly. So this is the kind of stuff, by the way, it's also recording the meetings is really important. So, you know, one of the things. I heard that there was a customer meeting and I wanted to understand how that meeting went. And, you know, And I just asked it, you know, hey, was there a meeting with this customer recently? How did it go? And it literally started by saying the meeting went poorly. And it's listing exactly what, who said what, what are the features they need. So, you know, yeah. These are just some of the things that you can do with the AI. And in the background, when I wake up, it has all these deep dives for me. So instead of waking up, calling people, doing meetings, it's doing the work for me. It's pinging them. It's sending whatever material is needed for the company. So it just compresses the time it takes to do things. Well, you know, it's a lot of being in management and being a founder is detective work. I call it like being Columbo. If you remember that TV show, I think it might show our age, but you're kind of like, oh, you know, it's one more thing there. Can you tell me about the meeting, the client meeting? How did that go and who talked most on the phone and what percentage of time did they talk and what did the customer actually think? And then you as the founder are getting spun. You're getting spun. And it's not that an employee necessarily wants to spin you. A team member is not looking to spin you necessarily. That can happen. But that's not the majority. Most people are good people. But they have their humans, right? They're the people in the air. And here you take the people out of that discussion. You say, here's first principles transparently what happened now we have ground truth and we can start the discussion without you. Being the CEO having to be Mr. Detective and you know, it is so I mean, I don't know about 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 have been solved. You wouldn't have had three weeks of hand wringing. Yeah. Yeah. I mean, or the customer feedback, right, that I can immediately find out all the customers. You know, I have some. Some of them are just having it look at features that are requested from all of our customers automatically from all the transcripts of all the customer meetings that we have. And it's just gathering those and they're immediately going to engineering. And then we have factories in engineering that can build features automatically. So it's a game changer. You know, it's just that game of whispers that we played when we were kids. Right. It's like when you have people have to meet, they have to tell. So one of our main goals is reduce number of human blockers. So we measure that throughout the organization and we just explain a blocker in this content. Please. That's something I can learn from. Yeah, I mean, the AI is very fast, right? It's just iterating in the background and it's doing its work and, you know, it's preparing material and so on. But at some point it's like, OK, well, now we need to go to legal and make sure that we are allowed to do this thing. Or now we need to go to finance to get approval for this budget. Or now we need to make sure that marketing is OK with this or actually just within the same organization. We need to talk to a sibling organization, get a meeting with them and see what they think. And we have to get an agreement. And if we can't reach an agreement, we'll schedule another meeting where we look at this again. Maybe we escalated to the managers and then so on. So these are the blockers that slow down organizations. 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. Why isn't this done is what we always ask. You're a leader, whatever your group, another group. Hey, we were supposed to do this by today. We're supposed to hit this target. Who's responsible? And then you get three Spider-Man's pointing at each other. If you know the meme. actual piece of data has not been resolved and your team has run nothing. I had a similar thing. I came up with a bot and I was going to share it with the whole company. And thank God I didn't. But I came up with a bot just called Major Decisions. And I just said, go into Slack. Join every room as me. "Just join every channel as me." Which some people are like, "Wait a second, "how did you even know we had a room?" Join every room, read every Notion page, uh even if i was invited to it or not get every email uh and find for me for the month And every couple of days, give me a report on the most important decisions, who participated in them, and [35:00] what was the resolution? What's not resolved? Then I said, give me the people who participate the most in these kind of discussions. And then I realized, oh, wow, these are the people who are actually making decisions in the organization. These are the people who are quickest. And my next report is going to be is, Who's quickest to contribute to those discussions? Now, in a venture firm, That's going to be 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? And then it picked up. employment discussions. And it was like, should we fire this person or not? Or do we do pips? And like, we don't do pips. And like, I just think pips are ridiculous. Like the performance isn't here. That's like, the person didn't show up Let's torture them for two quarters. Yeah, like you mispractice and you showed up late for the game. That's table stakes. You can't be on the team if you can't do those things. So now we're going to torture you, 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 a CEO. in terms of figuring out who are your all stars on the team, who are the ones who are actually contributing? Because now we're getting to essentially business moneyball like that movie Moneyball. It's fantastic. Brad Pitt is such a good movie, such a good book. We're actually getting into like management money ball, aren't we? Yeah, absolutely. But one of my favorite questions is just look at different or because we have certain culture principles, truth seeking, first principle and so on. And I say, hey, which of these leaders follow these principles? Look in their own or look at all the meetings and it can derive from those. And by the way, you could even ask, hey, this leader, what is the actual culture they're implementing? And it'll come up with different, you know, they'll tell you, you know, actually, the culture there is this, not the company culture, which is. So they're deviating from the core principles. That's 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. We need to be consistent, whatever it is. Maybe they're right, and the rest of the organization, including me, are wrong. But then we need to at least, it tells me that there's this divergence. Wow. So those kind of things were nearly impossible to do. Many of this other stuff you could have done, but a question like that, get the culture of this subgroup in a 12,000 person organization, figure out how the Japan organization at Databricks, What's the culture like? How do I make decisions? How is it different from the EMEA or the LATAM region that we have? So those are kind of game changers. For us, it's really I share all this information with the employees. They have access to all of this themselves. And then I tell them, this is what we want. Like that L curve, we wanted to shift it to the right. So everybody, let's change that. Let's move towards this, the North Star that we have together. And that helps us just be more competitive. and we take market share and we can accelerate our growth further. You know, I'd like to not be 70, 80 percent growth here. I'd like to be 100, you know, way over 100. So then we can put, you know, that's that's how I see it. So this is in this era right now. This is a competitive advantage. Whoever can do this best, you know, at some point, maybe it's all. Everyone is using it's fully diffused. and everybody's doing this. But right now, this is big alpha. If you can do this really, really well, better than any other organization. of big alpha. A lot of people have been very patiently been 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. And 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-san was splashy-cashy. When are you going public? And how do you make that decision? I know you 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 security market. getting to the marketing market uh i'm focused on those then the question is the path to get there is the best way and the fastest way for us to be public. or to be private? And I actually think that's what's on each side of the ledger. Yeah, I actually think that you get access to capital. you get much, much better marketing. You know, it's like everybody knows about you. Like every quarter, you have this big marketing event. You can tell your story again and again and again. Those are huge advantages. We have to organize our own tenders every year. Every year we do, finance has to go through this big process. You know, I think we raised 20, 30 billion dollars. And then we have to arrange, [40:00] that it's a heavy lift. There is a marketplace that just does that. It's also probably a pretty emotional and distracting, I would think, right? Like people have to decide... Then you might lose people that, hey, they take $10 million off the table. Now they're like, maybe I don't want to come to practice. I think the liquidity is a good thing because, you know, people know that they're going to get it every year and it kind of comes and it's like predictable. And that's like the one time they have. And actually in the past, it's turned out since we started doing these a few years ago, it turns out that, you know, the money that we set aside for tenders, they never use up all of it. So most employees are bullish. Yeah. So most employees are like, no, this is good for you. You get signal from that, right? Who's selling? Who's not? It's a little bit of signal. It could be personal reasons. But acquiring companies is a great reason to have equity, yeah? we can I mean, we haven't had a problem doing billion dollar acquisitions because they know because of the liquidity that we will provide. And then I think most people do the calculation is like, OK, if it's, you know, the business is seven billion dollars plus and it's growing 70, 80 percent. They do the math. It's like, OK, I could what most most people say, I want liquidity. But then when you tell them, like, OK, well, how much do you want to sell? We're going to get a different story. you know, can you do compounded growth of, you know, 1.7 to the power, you know, to the power of three or four? And then they start scratching their heads like, OK, maybe I'll sell like five percent, you know. Yeah. Sitting on your hands is a very. important skill to learn as an investor and an employee with this this kind of suck. I mean, listen, you do have to protect downside when you when you get in 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 replete is looking features next to them. Everybody seems to because you have a lot of developers, if those developers keep going up that ramp that you showed, You only had 10% of people at that peak, which means as that 90% climb up there, your developer resources are going to go through the roof. So you have a decision to make. Do we just... have less developers at the low end of the tail, and be more profitable? Or do we invest all that developer time as they move up the ramp? So how do you think about both of those issues, what to do with massive resources that this efficiency is creating. And then who is the competitor? And do you need to just keep making peripheral product moves? yeah we will uh make their full product moves but maybe i should just start by saying okay so those were the upsides of going public the downsides are uh you know is is the reason why you know uh to answer your specific question like you know why why do we compete who do we compete with why do we build these products i don't want the market to force me that okay now it's down and now it's you know uh sentiment is terrible this week and now your stock is down and when the stock is down the me I was going to make a million dollars. Actually, you said this is going to be two million in a couple of years, but I'm looking at it right now. It's under water. 200k. So you lied to me. I'm leaving. Okay. And then, you know, three months later, it's like, oh, no, no, no. Actually, the revenue is accelerating. We were totally wrong. It's up to the right now. It's like, oh, okay. Well, it's great. So there's like back and forth. A little bit of a roller coaster. That's what I want to avoid so that we can focus on that long term, you know, building the company towards the vision that we have in the long run without being distracted. We will go public. We are going to be public. But right now, I think during this transition period of AI, there's just too much up and down. distracting to companies, it's distracting to CEOs, it's demoralizing to CEOs. They call me, they're depressed because stock is down. They don't even understand why. And then the stock is up. So 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 like comparing sleep scores. And now it's kind of like looking at your Uber stock every day. You get a little bit of volume relationship. It could be challenging. It's interesting. Hey, we didn't. circle around to what Ben Horowitz said. He said that Your formulation of super intelligence is some combination of the model, the power and the time it takes to train and be accurate. What can you walk us through it? Do you have it off the top of your head there? Ready to go? I have to do it. I mean, it's pretty simple. So I think the main problem is the context problem. We're solving that with Genie and the ontology. The labs are focused on super intelligence. That's what Boatstrom talked about. Boatstrom in his book also said, how do we get there? He actually in 2014, I believe, defined RSI recursive self-improvement, the AI starts improving itself and then it gets better and better and better and better and better. And then we get this takeoff, maybe a fast takeoff. And at that point, we reach the singularity and we're no longer in control. And at that point, let's hope that the AIs are well aligned, because if they're not, we're all going to be made into pets. And this is where the 10% risk. Yeah, paperclip maximizes. [45:00] and the 10% risk to humanity, and all of that comes from there. But I think a lot of this is actually coming from most of the folks that are building this AI. They've seen that the AI has become really good at coding and almost in every company, as I said earlier, the coding is being done by coding agents. So then they're seeing and I think this is actually causing a big existential crisis in Silicon Valley. And it's actually for most of us who started, you know, who grew up coding our whole lives. You know, it's it's a big deal. Like, you know, it's like I spent 10, 12 hours, seven days a week. from third grade. Until, you know, my 30s, just coding all the time. Right. And I was really good at it. And I knew all these languages and so on. And then suddenly it's like the AI is doing that. Well, if it's doing it, I is how much better than you. That's the other part. It's like, yeah. So now it's writing itself. And the next version, it wrote itself. Wait a minute. Isn't that the recursive self-improvement? But wait a minute, then this is going to speed up. And then we're going to get into this loop. The book, the book called it Super Intelligence and the Paperclip Maximizer. So it might go off and all these bad things happen. Actually, interesting. Bootstrom has kind of changed his view now. If you have to follow him now, he's actually not anymore. He's saying that, you know, we should do this. There are risks. Of course, there are risks with cyber. So I actually think you need to actually 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 two percent, who cares? Does it matter if it's 98 percent or 100 percent? Why does it matter? Actually, it doesn't matter. So we really need four conditions to be true. The four conditions is that we need the next model that's trained to be trained in much short amount of time. OK, and still the big pre-training runs that the labs do, they don't do actually that many of them per year. OK, and it's still following the what's called the scaling laws, which means you need exponentially more hardware to get that improvement. So exponential more, 10x more compute to get a little linear improvement. OK, so they're scrambling to get all the GPUs and put together the biggest data center that they can possibly put together with the networking and do all the engineering. And then they're having humans do all of this because they really don't want to screw up the next run. apart. But, you know, we would like that next training run to be faster. That's condition number one. Okay. And actually, it can't just be a little bit faster. It's like this, you know, it can't be like, you know, you're halving the distance every time. You'll never get there. So in some sense, it has to be super linear. Secondly, we also want it to use less resources, right? Because, you know, you could say, okay, it's getting faster and faster, but that's because we're 10xing the hardware all the time. Well, you can't do that trick because we run out of hardware and we run out of atoms. So that's like the second requirement. Third requirement is that the we're getting better accuracy across the board in a generalized manner. It can't just be it's getting just better at, you know, this one dimension that we did reinforcement learning on or just math. It needs to be across the board on all of its tasks. OK, even if you do all of those three, it's not enough. Let's say you do that five, six, seven times. That's awesome. You get really awesome models. That's not that kind of super intelligence. The fourth condition is that you can actually put this in a loop and all the previous three conditions need to be continuously happening so that we end up in a world where the next model takes, you know, three months. then one month, then two weeks, one week, and it's getting smarter and smarter and smarter. And it's costing us 100 bucks or maybe cost us 100 million, then 50 million, 25 million, 10 million, five, two, one, and then it goes down and down. That's the kind of scenario we have. That's the flywheel. That's the flywheel. That flywheel doesn't exist right now. The models are taking longer to train, right? And there's more data cleaning and there's more manual whack-a-mole happening where you have vendors like Mercore that has to actually go get the data from very skilled labor. And then you in these gyms and you have to kind of whack them all in each domain so it's very domain specific and then of course the pre-training requires more resources and you know so um it's it's that doesn't look like we're getting into this loop by the way even if you satisfy all the four conditions, I'm saying that doesn't guaranteed mean that you'll have, you know, a conscious being that's just, you know, taking over everything and just, you know, it's the precondition for that happening. And there's no guarantee that it actually, it's a precondition. But even if you had that, which we don't have, you would have to threaten me right now, you would have to have a killer robot come into this room right now, right, and wrestle me down. Okay, something right. And those don't exist yet. So, you know, right now, you can capable model and you can find out that the refrigerators in grocery stores are in france have the password admin admin and that you can remotely script kitty into them and turn it into defrost mode and make a bunch of butter go bad and milk which is a tragedy for the french i understand but it is not super intelligence i can assure you it means i mean did you see that hack and now 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 yeah yeah if they refrigerators. [50:00] They would have been able... they wouldn't have been able to hack one of them unless they got into the place, took screenshots of 20 different refrigerators, physically compromised, you know, some supermarket in the eighth and rosman, which would have lost, you know, I don't know, 10 pounds of butter. It wouldn't be a big deal. So look, cyber, I do think is a real thing. Yeah. And I think it's a real risk. We actually have a whole product in the cyberspace called LakeWatch built on all this data that has automated agents that actually detect these kind of intrusions that happening in the day. Monitor your agents that you have. This is a real problem. And I do think bad things will happen here, but then we will adjust. But bad things happen when Internet came out as well. for a certain type of hacking, not for the, you know, Script Kitty hacking, but real hacking, is you have a lot of compute, you need a lot of compute to... execute certain hacks, like the hugging face one, you know, thousand agents, whatever. The bad guys don't have that much compute. The good guys have a massive amount of more compute. And so if you just did your basic job as a white hat and, you know, defended properly and looked for these exploits, let's say you have a thousand times the resources, you should be able to get to these problems long before the hackers. Does that make good sense to you? Yeah. It's a good argument, The issue is that you have to weigh that against the following asymmetric scenario that we have, which is if I'm an attacker, you know, I can launch thousands of thousands of thousands of agents and all of them can fail except for one. And if one of them gets through, I'm completely successful. The asymmetry is that if I'm a defender, you know, I was able to like, you know, protect, you know, my boundaries all the time, except one got through. 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 a thousand times and not get it. So I need to be 100% safe all the time. They need to just get one thing through. It's very asymmetric that risk. But however, the way I would look at it, the reason I think that we're actually going to overcome this and it's going to be totally fine is that these are actually 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 now scanning it. We all have products. I mean, we actually sell a product where we can scan the vulnerabilities of new code. So new code that's being produced is not going to have these vulnerabilities. By the way, 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. Right. Actually, in less than next 12 months. So all that code is going to be secure. So then. And by the way, right now, the AI is sophisticated enough to not make human errors and to look for it. And it's more dogged. Yeah. You just scan those things. So then what remains is all the legacy code, which, by the way, we can rewrite all of it, which is what we have to go through. And people are doing that, right? Yeah, we're doing that. And there are other ways to actually also protect, you know, do shadow patching before without touching the existing legacy code. But so, you know, what's happening is that you can think of it as like, you know, it's like a fruit garden and there are like these bugs that, you know, attackers are leveraging. They're picking the low hanging fruits. And before humans could pick a certain low hanging fruit. Now agents can reach higher and pick higher, you know, hanging fruit, but we're going to clean those up and then they can maybe reach a little bit higher. But there's only a finite number of vulnerabilities in the code. Once we've secured that code base, there is no more. So it's not like there's an infinite thing and this is going to go on forever. Then we're at the steady state. And then, you know, vulnerabilities might actually be fewer than what we had in the past. I think eventually then, you know, back to AMP, where you went at Berkeley, it's the people in the loop that are the vulnerability. And it's just going to be somebody getting compromised, as we saw 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, hippy dippy, EA community, accelerus, decelerus. What's the nature of this group of the ea faction and how they look at the world and how confounding they are 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 really focus on what they're saying and take them at the word. We is a big controversy. Dario, you know, during the roadshow is saying, Cancer and Alzheimer's, I believe, will be gone in under 10 years. Now, That's very beautiful to say, Except that's because everybody's dead, right? Yeah. Except if your mom or dad are dying from cancer and you think, oh, I got a window here that I could save them or my grandmother's got 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 that I actually think he does. Unpack these people because you've been around them for two decades, three decades. Well, you know, I think that most people on both sides, actually on all sides, [55:00] - 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 super intelligence taking over. If you read Bostrom's books, actually, if you read those books and you sort of internalize what he wrote before any of this, then you actually get kind of freaked out. Right. And I was certainly freaked out when I read it. So I think that they're starting with those assumptions and then they're looking at, you know, OK, well, then if it's getting smarter this fast and look at the cyber attacks that are happening, dot, dot, dot. we're going to get this kind of super intelligence that's describing the book. But there's a few dots there that are not clear. But then they would say, well, look, you know, shouldn't we be cautious? Right. So I think we need to take, you know, 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 the biggest risk right now. is cyber 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, yeah, I don't know what I don't know. But based on what I'm talking to here in San Francisco, where all the labs are and all the people I meet, I don't see those four criteria that I mentioned even being fulfilled. You know, so I don't see any risk to be worried about that happening. And then the second thing I said, even if we would get there, I don't see, you know, even the means in the society. Like it's just digital. Like you can just shut down the computers right now. You know, I do think cyber is real. Like if a hospital went down a refrigerator, right? That is concerning. And kids using it and developing parasocial relationships with character, There are concerns to have about how the technology gets deployed. And we just went through this with, you mentioned Dunbar's number four, we went through this with you know, social and the profound impact that had on site. We're going through it now with truth and news and manipulation of news via algorithms. So there are things to be concerned about being turned into paperclips. Maybe not. But I think that, you know, the health care stuff is absolutely real. Like, you know, at Databricks, we actually we were involved with Regeneron, where they found the actual genome response for chronic liver disease using Databricks. There are researchers there. and actually came up with drugs for that. Best case scenario in under a decade, if you were to take the claim we had Dario make, any chance that like 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 mean, I think progress will be faster. But for a lot of these things, it also has to go through human trials. Human trials, like, you know, we can't just test stuff on humans will and nilly. Right. And just because you tested it on animals, it's not enough either. Right. So I don't know, maybe in 10 years, maybe not in 10 years, maybe it takes 30 years and it takes 50 years. I don't know. Maybe it takes six months. use case I gave you with chronic liver disease, that's super important. By the way, that was the previous generation of AI that we used at Databricks. We trained the new one now with Merck, and it's called TEDDY, Transformer Enabled Drug Discovery. Transformer Enabled Drug Discovery is a model, transformer model. But instead of the next token being English, it's the Gene Regulatory Network, GRN. So it actually can say, OK, 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. You know, drug discovery aside, like, you know, I give you another example, like crisis text line. it uses Databricks and uses large language models and these agents, and they actually detect if kids want to do self-harm and it helps prevent that. Right. You know, Virtue Foundation uses Databricks. You know, Zipline. This was a use case that was reported before used Databricks to deliver, you know, food. They had these drones, the drones that would deliver, you know, everything about those drones and so on, like their their path, the, you know, really big data. Yeah. All of that - Great company. - Yeah, we see these use cases all the time with AI, and it is gonna make the world a better place. There are risks, we will deal with those as well. people are nervous because, you know, things are changing. And I think the companies, there's a lot of researchers who maybe, you know, they understand the research very well, but they haven't looked at, you know, how does the economy look? What's going to happen? You know, how are these things going to evolve? And they've watched a lot of sci fi. Let's be honest, like they are being informed by, you know, a certain narrative. And the idea that there's a 10 percent chance that seven 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. If they say it themselves. Right. 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 without. That's why we're focused on it. [60:00] And then I think actually to solve this context gap that I mentioned in enterprises, the thing we're focused on to actually get value everywhere. I actually think it's going to take 10 years. I think this is something that's going to take 10, 15 years. The labs think this will happen in the next six months, 12 months and two years. But I actually think GDP will move very slowly and it's going to take decades for us to actually get through it. Why? I mean, as I said, I mean, just go back to the anecdote at the beginning, the CEO of the bank that 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 kind of things that we discussed when I showed you Genie and the ontology? 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, which is 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, OK, this is going to change everything. And then 99% of people were on dial-up and had never seen a high-speed connection. of YouTube or, you know, unlimited free photography story. It doesn't make any sense. There was a kid called Robert Morris Morris, who took down 10% of the Internet with the worm. It's called the Morris worm. Yeah. You know, he almost got a jail sentence for it. And then he went on and he became a famous MIT professor in computer science. And he's like a very famous guy. But, you know, that took down a whole 10% of the Internet back in the day. to a certain extent. Okay, sure. Let's see what happens. All right, everybody, another great all in interview, Ali, thank you so much for your time. And we'll see you next time on the all in interview. Thank you so much. Great job, Ali. You are awesome.