newsfilter
← Interviews
All-In Podcast1h 2m

Databricks CEO: AGI Is Already Here, Why Companies Are Behind, and the Superintelligence Debate

Ali Ghodsi, Chamath, Jason, Friedberg, Ollie

Interview, Fireside Chat
  • Company Background & Financials

    • Databricks was founded in 2013 by Ali Ghazi and a team from UC Berkeley's Amplab (approx. 20 researchers) out of "despair" that the industry was ignoring AI/ML integration.
    • The company generated $1.5 million in revenue in 2015 and reached a $7 billion run rate by August 2026, growing 80% year-over-year.
    • Ghazi states there is "zero interest in AI" (then called machine learning) at the time of founding, with the industry viewing it as limited to robotics or ad tech.
    • Databricks was one of the first companies to bet exclusively on cloud infrastructure, open source, and AI use cases, which were initially viewed as "strategic blunders" by major analysts.
  • Current State of AI Adoption in Enterprises

    • Ghazi asserts that most enterprises currently use AI only in "rudimentary ways," such as "glorified Google search" or chatbots, rather than deep workflow integration.
    • While AI assists with coding, enterprises report only 20-30% speedups, not the exponential efficiency gains often predicted.
    • Ghazi argues it is "not possible" for companies to deploy "a thousand agents" realistically to do core work without human oversight.
    • The primary barrier to broader AI adoption is the "context problem": organizations fail to feed AI comprehensive data because they refuse to record meetings due to liability concerns.
    • Secondary barriers include cybersecurity risks (increased "surface area" for attacks), exponential token costs outpacing revenue growth, and data sovereignty/IP concerns (fear of vendors incorporating proprietary data into public models).
  • Databricks "Genie" Platform & Operational Use

    • Databricks uses a proprietary system called "Genie," which maintains an "ontology" (a graph) of the entire organization, including org charts, Slack messages, meeting transcripts, and documents.
    • Ghazi demonstrated Genie analyzing a 10,000-employee employee survey in real-time, identifying that "work-life blend" and "action" were negatively correlated but not causally linked without human nuance.
    • The system automatically identified specific departmental pain points (e.g., Product Management disliked stakeholder overload; Sales desired career growth).
    • Genie detects "management blockers" by analyzing unresolved decisions across Slack and emails, identifying who participates most in high-level decisions versus those contributing to "whistleblowing" or "drift."
    • Databricks uses AI to visualize engineering output, revealing an "L-shaped" distribution of code commits where a small top tier is significantly more productive due to "automated factories."
    • The company uses this internal data to create a "North Star" for culture, measuring leader adherence to core principles like "truth-seeking" and identifying cultural divergence across global regions (e.g., Japan vs. US).
  • Strategic Outlook: Public vs. Private

    • Ghazi has no immediate plan to take Databricks public, preferring to remain private to avoid the volatility of quarterly earnings calls and stock price fluctuations distracting from long-term product vision.
    • The CEO cites the emotional toll of public markets on leadership ("depressed" CEOs) and employees (perceived pay cuts when stock dips) as reasons to delay.
    • Secondary markets provide liquidity for employees, with most staff choosing to hold rather than sell given the company's 70-80% annual growth trajectory.
    • Databricks intends to go public eventually but views the current AI transition period as too volatile for an IPO.
  • Thesis on Superintelligence (RSI) and Risks

    • Ghazi believes humanity possesses "all the intelligence we need to completely transform the planet" already, distinguishing this from "Superintelligence" (RSI) as defined by Nick Bostrom (a mind capable of millennia of thought in minutes).
    • He argues that "Recursive Self-Improvement" (RSI) is unlikely because four specific conditions are not currently met:
      • Speed: Training time is not super-linearly decreasing.
      • Resources: Hardware constraints prevent infinite scaling (we are running out of atoms).
      • Accuracy: Gains are domain-specific rather than generalized across all tasks.
      • Looping: The flywheel of faster, cheaper, more accurate training cycles does not yet exist.
    • Ghazi dismisses the "paperclip maximizer" scenario, stating that even if code is 98% AI-generated, automating the final 2% is not a systemic risk.
    • He identifies cybersecurity as the primary immediate risk, noting the asymmetry where attackers only need one successful breach while defenders must be 100% secure.
    • Ghazi predicts the solution to cybersecurity lies in AI scanning for and patching legacy code vulnerabilities, suggesting a "steady state" where fewer bugs exist than in the human-driven era.
  • Future Economic & Social Impact

    • Ghazi projects that while AI technology advances quickly, macroeconomic GDP growth will move "very slowly" over "decades" due to organizational friction and legacy infrastructure.
    • He compares the current AI rollout to the early internet, where the vast majority of users were on dial-up while a small minority maximized the technology's potential.
    • Specific high-impact use cases mentioned include:
      • Healthcare: Accelerating drug discovery (e.g., Regeneron's work on liver disease) by modeling Gene Regulatory Networks (GRN) rather than just English text.
      • Safety: Crisis Text Line using AI agents to detect and prevent self-harm in teenagers.
      • Logistics: Zipline using Databricks for drone delivery path optimization.
    • Ghazi warns against "false hope" regarding rapid cures for diseases like cancer within a decade, citing the slow timeline of human trials and regulatory hurdles.
  • Competitive Landscape

    • Ghazi views Databricks' main competitors not just as data warehouses (Snowflake) but as companies building AI-native platforms (Anthropic, Replit, Cloudflare) that are "bolting AI onto broken infrastructure."
    • He argues that companies built on 50-year-old legacy rails will be at a disadvantage compared to those designed for the "intelligent era" from day one.
    • Databricks' strategy involves capturing the "governed data repository" layer, charging consumption rates as AI agents query their internal data ecosystems.
    • The company faces a decision on how to deploy efficiency gains: whether to reduce headcount or invest the massive developer surplus into building new products to increase competitiveness.