Building AI Systems for Capital Markets

Platform Mission and Strategy

• Marquee, Goldman Sachs' digital platform for institutional clients, aims to support decision-making under uncertainty by integrating Gen AI directly into the investment process rather than applying AI as an afterthought.
• The platform functions as a unified ecosystem containing millions of research articles, trading floor commentary, and "MarketView" widgets (expert-curated analytics).
• The paradigm shift involves moving from users learning software to software learning users, allowing individuals to express intent in natural language to access complex capabilities.
• Internal "Marquee AI" aggregates relevant research, commentary, and data to assemble cohesive, auditable answers, grounding every sentence to a specific source or calculation.

Technical Challenges: Hallucinations and Reliability

• Building reliable AI for institutional use requires forcing the model to ground itself, as frontier models inherently cannot distinguish between facts and extrapolations.
• Early technical solutions like vector databases and chunking strategies were rendered obsolete as context windows expanded from 4,000 to 128,000, and eventually to 1 million tokens.
• The "Lost in the Middle" phenomenon, where models fail to retrieve information buried in the center of a context window, was largely resolved by Google's 1-million-token context window updates.
• Hallucinations persist because LLMs are optimized for multiple-choice completion; they are not rewarded for abstaining, leading them to guess rather than admit uncertainty.

Evolution of AI Engineering Regimes

• Development has progressed through three distinct phases: prompt engineering, context engineering (managing data chunks), and agentic engineering (iterative self-checking loops).
• Chris Churchman identifies "environment engineering" as the current critical phase, focusing on securing the operational environments for agents and ensuring they have access to institutional knowledge.
• The next anticipated phase is "mandate engineering," which defines the authority limits and allowed actions for AI agents, potentially stabilized by the creation of insurance or liability markets.
• Developers must build for the model they will have at launch rather than current prototypes, avoiding wasted effort on scaffolds that become redundant as reasoning models and function calling evolve.

Human Capital and Cognitive Preservation

• A primary risk of AI adoption is cognitive atrophy, where humans outsource reasoning to models, leading to a loss of first-principles thinking.
• The platform design prioritizes transparency to ensure human capital can understand, audit, and stand behind AI-generated analysis.
• Goldman Sachs is actively working to preserve its "apprenticeship culture" to prevent the loss of tacit knowledge held by senior traders.
• The goal is to empower the next generation of employees to handle high levels of reasoning under uncertainty, comparable to the skills required of a fighter pilot.

Innovation Frameworks: Automation vs. Reconceptization

• Automation Mindset: Replacing human cognitive bottlenecks (reading, reviewing) with AI to achieve the same outputs faster, which risks locking in legacy processes.
• Reconceptization Mindset: Ignoring legacy processes to redefine the "job to be done" using scalable intelligence, aiming for orders-of-magnitude improvements.
• Example: Wrapping complex derivative visualization tools with AI agents to allow users to simply express intent rather than navigating complex interfaces.

Future Outlook and Scalability

• Churchman remains a believer in the continued scalability of AI, citing the "bitter lesson" that next-token prediction forces the creation of deep synthetic world models.
• Scaling laws (Kaplan, Chinchilla) justify massive infrastructure investments, predicting that performance improvements will hold up to 10 orders of magnitude.
• Current bottlenecks are shifting from data access to experimental design; AI R&D labs are prioritizing the automation of their own R&D processes (running experiments) to accelerate progress.
• Key missing elements for the next stage of maturity include self-learning capabilities (avoiding catastrophic forgetting) and the discovery of a "native form factor" for AI beyond command lines.

Operational Lessons Learned

• Products are judged by their "worst day" performance, not the best-case demo scenarios; building around model deficiencies is unsustainable.
• Institutional-specific knowledge (entitlements, mandates, data connections) represents a defensible moat that models cannot replicate via public training data.
• Trust is a long-term asset that takes years to build and is easily lost; transparency in AI generation is essential for client adoption.