Interview, Fireside Chat, Other
Building AI Systems for Capital Markets
- The firm's platform is projected to evolve into a core component of client investment by shifting from software that requires user learning to systems that learn user intent, enabling natural expression of goals.
- Future systems are planned to decompose user queries into research topics, aggregating data from articles, trading commentary, and Market View widgets into cohesive, auditable thought documents where every sentence can be verified.
- When calculations are necessary, the architecture intends to retrieve data from expert widgets, execute Python scripts, and deliver well-grounded answers without manual intervention.
- The development roadmap is expected to progress from prompt and context engineering to "agentic engineering," followed by "environment engineering" for secure operations, with a future "mandate engineering" phase defining agent authority once liability markets exist.
- Significant progress in near-term reliability is anticipated, though confidence intervals are expected to never reach perfection due to the inherent probabilistic nature of the technology.
- Future AI products will focus on institutional-specific capabilities such as entitlements, mandates, and data connectivity rather than general training data, while infrastructure investments are justified by scaling laws and synthetic world models.
- The firm plans to move beyond automating legacy processes to "reconcepting" approaches that answer previously unanswerable questions, including wrapping derivatives visual tools in agents to interpret outputs based on intent without requiring users to learn complex interfaces.
- To prevent "cognitive atrophy" and the loss of tacit knowledge, the firm intends to foster an apprenticeship culture ensuring the next generation retains intuitive trading reasoning that cannot be codified.
- Building around specific model deficiencies like context window limitations is anticipated to be futile as models are expected to evolve to handle larger data volumes natively.
- AI researchers are predicted to prioritize automating their own R&D to run experiments at scale, potentially removing bottlenecks in discovery speed, though persistent hallucinations remain a challenge due to models' inability to distinguish facts from interpolations.
- A "glass ceiling" is feared regarding self-learning capabilities due to catastrophic forgetting, and limitations currently exist due to the absence of a native "form factor" beyond command-line interfaces.
- Goldman Sachs disclaims warranties regarding the accuracy of these forward-looking statements and notes that past performance does not indicate future results.