Why Human Data is Key to AI: Alexandr Wang from Scale AI

Scale AI's Core Mission

• Defined as building the "data foundry for AI," operating across three pillars: models (led by large labs), compute (led by NVIDIA), and data (the focus of Scale).
• Goal is to produce "frontier data" for large labs while enabling enterprises and governments to utilize proprietary data for bespoke Gen AI applications.
• Strategy involves a "marriage" of human experts and algorithmic techniques to generate data that exceeds the current internet's content scale ("internet on steroids").

Industry Phases and Market Structure

• Phase 1 (Research): Characterized by pure research and small-scale tinkering, ending roughly with GPT-3.
• Phase 2 (Execution): Current phase dominated by engineering, scaling, and execution of large training clusters by companies like OpenAI, Google, and Meta.
• Phase 3 (Forecast): Expected shift back to research-led innovation where divergence in algorithmic breakthroughs determines competitive advantage.
• Compute Status: Not currently a constraint; labs have sufficient clusters, and infrastructure is scaling.
• Data Constraints: Industry has exhausted publicly available data (e.g., Common Crawl), creating a "data wall" that necessitates new production methods.
• Market Dynamics:
  • Model pricing has dropped 100-1,000x in two years, suggesting intelligence is becoming a commodity.
  • Pure model renting is forecast to be a "mediocre long-term business" due to open-source proliferation (e.g., Meta's Llama) and lack of pricing power.
  • High-value opportunities identified "below" (Nvidia, cloud providers with logistical barriers) and "above" (applications with strong product-market fit and workflow integration).

Enterprise Adoption and Data Value

• Current State: Enterprises are moving past the initial POC frenzy; few POCs have reached production, with early gains mostly limited to marginal efficiency and cost savings.
• Future Value: Potential for AI to meaningfully impact stock prices via cost reduction, efficiency, and superior customer experiences (e.g., automating manual wealth management interactions).
• Proprietary Data: Enterprise data (e.g., JP Morgan's 15 petabytes) is currently underutilized due to poor organization but holds "hyper-valuable" potential for transforming products, unlike general internet data which is abundant.
• Regulatory Risk: Large tech companies face significant hurdles in utilizing internal data (e.g., Meta's past regulatory issues in Europe with Instagram data), potentially diluting their data advantages.

Capital and Investment Logic

• Big Tech Incentives: CEOs of major tech firms view AI investment as existential; the risk is "under-investing" more than "over-investing" given the potential for trillion-dollar market cap shifts or business disruption.
• ROI Mechanism: Capex investments are easily recoupable through immediate efficiency gains in core businesses (e.g., improved ad targeting for Google/Facebook) before full AI productization.
• Open Source Impact: Surplus from open-source models is described as "insane," broadening access to the fruits of massive private investment.

Leadership and Hiring Lessons

• Growth vs. Culture: Scale AI grew its business 5x-6x while keeping headcount flat to avoid "regression to the mean" in team performance and communication overhead.
• Executive Hiring Pitfalls:
  • Avoiding the "executive fantasy" that a new hire can instantly fix a business; success requires hiring teammates for long-term judgment rather than "magic wands."
  • New executives must first understand existing operational rhythms and prove value with small steps before suggesting sweeping changes.
  • Founders often make the mistake of stepping back after hiring executives, but founder-led innovation is critical in high-growth startups to maintain competitive advantage.
• Scale AI Hiring Philosophy: Adopted a "Merit, Excellence, and Intelligence" policy, committing to hire the most capable person for every role regardless of demographics, without quota-based optimization, to ensure talent density for high-stakes AI work.

Future Outlook: AGI

• Definition: AGI is defined as AI capable of accomplishing 80%+ of digital/computer-focused jobs.
• Timeline: Forecasted to be at least four years away, though algorithmic breakthroughs could accelerate this.
• Key Challenge: Production of "frontier data," specifically complex reasoning chains (e.g., multi-tool usage, agent workflows) that do not currently exist on the internet.