AI Exchanges: The Role of Data

Career Context and Data Genesis

• Nima Raphael joined Goldman Sachs over 20 years ago, transitioning from software engineering to data strategy following the 2008 Lehman Brothers collapse.
• The "Copter" database was built to aggregate front, middle, and back-office data to calculate end-to-end exposure to Lehman, proving data could be a business enabler rather than just exhaust.
• This initiative demonstrated that a single, centralized data source could drive innovation for traders, salespeople, and quants simultaneously.

Technical Shifts in AI

• The evolution of computer science has moved from a 50–60 year era of deterministic, rule-based coding to a "learn by example" paradigm driven by machine learning.
• Generative AI represents a step change from pattern prediction to the creation of images, audio, and language, though it remains a continuation of the "more data, more learning" continuum.
• A fundamental organizational mindset shift is required to accept probabilistic machine outputs, contrasting with traditional deterministic workflows where inputs yield repeatable, traceable results.
• The financial sector may be more acclimated to non-deterministic models due to the inherent stochastic nature of pricing models, derivatives, and market behavior.

Hype, Enterprise Potential, and Data Constraints

• While consumer applications (e.g., image recognition, research queries) validate AI's reality, the full enterprise potential remains to be seen and depends on harnessing proprietary data.
• Raphael identifies "agent coding" as a definitive use case that transformed skepticism into belief, offering "superhuman" problem-solving capabilities.
• The industry faces a constraint regarding the exhaustion of high-quality public training data, evidenced by newer models (e.g., DeepSeek) reportedly training against existing model outputs.
• Despite public data saturation, significant value remains locked in "trapped enterprise data" behind firewalls that has not yet been harnessed for differentiation.
• Future data generation will rely heavily on synthetic data, creating a need to distinguish between low-quality "AI slop" and high-value insights.

Data Engineering and Quality

• Enterprise model quality is directly dependent on the rigor of cleaning, normalizing, and linking internal data semantics.
• The role of data engineering has evolved to treat data as "software for data," requiring architectural practices to ensure correctness and contextual relevance.
• A feedback loop is emerging where AI agents are increasingly used to automate data cleansing, normalization, and linkage tasks.
• Disparate data must be organized into a coherent structure to navigate from one fact to another, enabling accurate business insights.

Future Horizons and Personal Application

• New data frontiers include video data and virtual environments where robots generate their own data to understand the world.
• Philosophical concerns exist regarding a potential "creative plateau" if AI models increasingly train on synthetic data rather than new human intellect.
• Raphael applies AI personally to facilitate learning with his three-and-a-half-year-old son, using the technology to answer complex "why" questions and encourage independent research as he ages.
• The conversation highlights a strategy where businesses leverage proprietary data to create value distinct from the generalized capabilities of consumer-grade models.