DOGE's AI-powered audit of government spending: hosts break down what the team is actually doing

Summary

The hosts discuss what DOGE (Department of Government Efficiency) is actually doing with AI to audit federal spending, focusing on Luke Ferriter's role and the technical problem the team is solving.

The core mission: AI to parse government documents at scale. Ferriter, working under Elon Musk, is leading an effort to use large language models to extract and analyze data from government PDFs, forms, JSON files, Excel sheets, and other unstructured documents that have historically been accessible only through FOIA requests but never systematically analyzed. The hosts note that when journalists or lawmakers have dug into government spending before, they've found anomalies only by accident—the New York Times or Wall Street Journal occasionally "hack through" a few documents and surface a headline. What DOGE is building is the infrastructure to do that at scale, across every line item in federal spending, not through manual investigation but through automated parsing.

A concrete example of waste. Josh Steinman surfaced data showing the University of California San Francisco received $815 million in research funding, with the school retaining approximately 40 percent—roughly $300 million annually—for administrative costs. The hosts use this to illustrate both the scale of the problem and why it matters: $300 million in overhead for a single institution is striking when compared to a commercial company's cost structure. The broader point is not that 40 percent is definitively wrong, but that without systematic visibility into spending patterns, lawmakers and the public cannot evaluate whether it's justified or how it compares to historical benchmarks (e.g., administrative overhead during the Apollo program or Cold War missile development).

Why this requires a founder-type hacker. The hosts argue the technical work is non-obvious. Ferriter recently shared a post about LLMs built specifically for format conversion and document parsing, suggesting he's already thinking about the infrastructure needed. One host describes manually using Google OCR and Adobe tools to extract text from a scanned New Yorker article, then feeding the corrupted output to multiple LLMs before ChatGPT o1 Pro could produce clean, flawless text—a process that took an hour for a single document. Doing that "millions and millions of times" across government records requires both the right AI models and someone who understands how to pipe data through them at scale. The hosts characterize this as a task for "a generational mind" and note that Ferriter's earlier work with Nat Friedman (at Anthropic, implied context) has positioned him well for the job.

Why Ferriter is suited to this role. The hosts frame Ferriter as someone with founder instincts who has wisely avoided rushing into his own startup. Instead, he worked for Nat Friedman and has now taken on the DOGE role—a path that, in hindsight, will likely appear inevitable when he eventually launches his own company. The hosts also note he's received death threats since his work became public, a grim reminder of the political intensity around DOGE, and publicly commit to defending him.

Operational support from founders. Matteo Franceschetti, co-founder of Eight Sleep, sent sleep pods to DOGE staff to support the team working long hours. The hosts cite this as an example of founders recognizing an opportunity to help and acting immediately.

The segment does not provide quantitative goals for DOGE's audit work, timelines, or estimates of potential savings identified. The focus is on explaining the technical problem—making government spending data legible and analyzable for the first time—and why Ferriter's skill set makes him the right person for it.