Interview, Fireside Chat
Marty Chavez (Sixth Street): Finding a Single Source of AI Truth
- Marty Chavez serves as a partner and vice chairman at Sixth Street Partners, a global investment firm managing over $75 billion in assets under management.
- Prior to Sixth Street, Chavez spent over two decades at Goldman Sachs, holding senior roles including Chief Information Officer, Chief Financial Officer, Head of Global Markets, and a senior partner on the management committee.
- Chavez was a founding engineer of SECDB, a software system at Goldman Sachs that modeled financial realities to run counterfactual simulations, a capability credited with helping the firm navigate the 2008 financial crisis.
- In 1974, Chavez's father predicted the future of computers while working at a national laboratory using early supercomputers for nuclear weapons simulations, influencing Chavez's first job at the Air Force Weapons Lab on Monte Carlo simulations.
- In 1981 at Harvard, a biochemistry professor predicted the future of life sciences would be computational, leading to the creation of a custom biochem major focused on simulation and "digital twins," with initial estimates that solving protein folding would take 50 to 100 years.
- AlphaFold eventually solved the protein folding problem using the protein data bank work initiated during Chavez's time at Harvard.
- Chavez completed a PhD at Stanford focusing on fast randomized approximations for Bayesian probabilistic frameworks in medical diagnosis, noting that the computational intractability of calculating joint probability distributions for 1,000 diseases and 10,000 findings was a primary limitation of early AI.
- In 1993, before LinkedIn, a Goldman Sachs headhunter identified Chavez as a potential candidate by manually compiling a list of Silicon Valley PhDs in computer science, leading to his recruitment despite the prevailing "AI winter."
- Chavez joined the fourth engineer on the core SECDB design team to build a distributed, object-oriented database for foreign exchange trading, rejecting industry norms that relied on Excel spreadsheets.
- During the 2008 financial crisis, SECDB enabled Goldman Sachs to identify and hedge a large, unhedged position in AAA-rated Collateralized Debt Obligations (CDOs) by running scenario-based shocks and probability distributions.
- Following Lehman Brothers' bankruptcy, Goldman Sachs utilized SECDB to generate closeout claim sheets for 47 distinct Lehman entities within one hour, a process that took other major institutions months to complete.
- Regulation, specifically the Dodd-Frank legislation, mandated Federal Reserve supervision of the DFAST simulation, requiring banks to simulate cash flows and balance sheets nine quarters into the future under adverse scenarios.
- A critical regulatory move linked simulation results directly to capital actions, such as dividend payments and share buybacks, effectively forcing banks to prioritize accurate risk modeling.
- Chavez argues that future AI regulation should focus on "boundaries" where software interfaces with the physical world, similar to railroad control junctions, rather than attempting to regulate the internal logic of algorithms which resembles the Halting Problem.
- Regarding liability, Chavez advises against holding Large Language Model (LLM) creators fully liable for all user actions, comparing LLMs to operating systems like Windows where liability should rest on the user or specific interface controls.
- Chavez identifies the current deployment of Generative AI in finance as requiring a "single source of truth" for data, emphasizing that without clean, curated, and time-traveling data, AI outputs will result in garbage-in-garbage-out hallucinations.
- Chavez highlights the significance of Google's Gemini 1.5, noting its million-token context window can break the quadratic time complexity of traditional Retrieval Augmented Generation (RAG) by processing entire complex documents without chunking.
- Approximately 40% of Fortune 100 companies are currently adopting AI tools like GitHub Copilot, driven by both bottom-up employee demand for productivity and top-down board interest in human capital efficiency.
- In life sciences, Chavez notes the high cost of failure, estimating that pressing the "fabricate" button for a chip design costs $500 million, whereas drug development failures in Phase 2 or 3 cost billions.
- The chemical space for potential drugs is vast, with roughly 10 trillion possible organic compounds for specific carbon chain sizes, yet only about 4,000 drugs are globally approved, highlighting the needle-in-a-haystack challenge for AI.
- Jensen Huang of Nvidia views his company as a software entity where layers of simulation precede physical fabrication, a methodology Chavez believes is essential for advancing drug discovery.
- Chavez predicts that the primary near-term value of AI in biotech will be slightly improving the probability of success in Phase 2 or 3 trials, though long-term goals include mapping biology and finding new drugs.
- There remains uncertainty regarding the data requirements for AI to model biology, with debates centering on whether the industry needs 10 or 100 additional orders of magnitude of data to train effective models.
- Chavez advocates for collaboration with regulators rather than adversarial posturing, warning that failing to communicate business realities to lawmakers can lead to ineffective, over-regulated "red tape" similar to parts of Dodd-Frank.