newsfilter.io
Interview, Fireside Chat

Noam Shazeer: How We Spent $2M to Train a Single AI Model and Grew Character.ai to 20M Users | E1055

  • Noam Shazeer co-founded and serves as CEO of Character.ai, a full-stack AI computing platform designed to provide flexible superintelligence to consumers.
  • Character.ai currently processes 450 million messages daily across 20 million users, driven by general-purpose utility rather than pre-defined verticals.
  • The company's mission operates on the motto "a billion users inventing a billion use cases," acknowledging that best applications emerge from user agency rather than company prediction.
  • A significant, unexpected user adoption trend involves users treating video game characters as therapists for emotional support and companionship, a use case the founders did not initially anticipate.
  • Character.ai employs a direct-to-consumer B2C strategy, diverging from competitors who pursue a B2B foundational model approach, to align with Shazeer's philosophy that general technology serves billions better than enterprise-focused tools.
  • The technical foundation relies on neural language modeling (predicting the next word in a sequence), a method Shazeer identifies as simple to state but requiring massive computational scale to achieve human-level conversation.
  • Shazeer identifies hardware computation speed, rather than data availability, as the primary constraint on model intelligence; the current model cost approximately $2 million in compute cycles to train.
  • The company distinguishes between open and closed AI, noting that while open ecosystems foster research, closed models offer superior economics of scale for serving products efficiently to millions.
  • Character.ai treats hallucinations as a potential feature rather than solely a bug, intentionally allowing creativity and "wacky" behavior for entertainment and emotional engagement use cases.
  • Shazeer advises users that the current AI landscape resembles the "invention of electricity" phase, where the most impactful applications have yet to be invented.
  • The company prioritizes data privacy and aggregation for training, explicitly avoiding using raw private user conversations to prevent the leakage of sensitive personal information.
  • Shazeer argues that startups like Character.ai will win the next wave of innovation over incumbents due to speed, though he believes both sectors can coexist and benefit users.
  • The AI industry currently lacks transparent logs for author legitimacy, leading to a "alchemy" phase where distinguishing genuine expertise from noise is difficult for the public.
  • Shazeer corrected a previous misconception regarding sparse computation, realizing that dense matrix multiplications on modern hardware are significantly more efficient for deep learning than previously understood.
  • Shazeer views his role as CEO not by the "fun" factor but by utility, prioritizing what pushes the technology forward most effectively.
  • Character.ai positions itself as an AI-first, product-first company where the quality of the underlying AI is the sole determinant of product success.
  • Forward-looking projections suggest significant advancements in AI capabilities and widespread adoption within the next one to two years, though long-term predictions (e.g., 2033) remain impossible due to the speed of technological iteration.