Dwarkesh PodcastNat Friedman (Github CEO) — Reading ancient scrolls, open source, & AI
CHAPTERS
- 0:00 – 1:12
Why reading the Herculaneum scrolls could double surviving ancient texts
Nat frames the Vesuvius Challenge as a rare chance to recover unknown writing from ~2,000 years ago. If the intact-but-unopenable scrolls can be read at scale, classicists estimate it could roughly double the corpus of surviving texts from antiquity.
- •Hundreds of carbonized scrolls remain roughly intact but unreadable by physical unrolling
- •Potential impact is unusually large: a major expansion of known Greco-Roman literature/philosophy
- •Motivation: high-stakes historical discovery rather than incremental academic progress
- •Nat’s bias: assume opportunities exist because the world isn’t fully efficient
- 1:12 – 3:13
From Vesuvius (AD 79) to a buried library in Herculaneum
Nat recounts how the eruption buried Herculaneum under deep mud and ash, preserving a villa with a papyrus library. The villa’s status, location, and burial conditions created the strange outcome of scrolls surviving but becoming extremely fragile.
- •Herculaneum as a wealthy town; villa linked to Julius Caesar’s family network
- •Burial under ~20 meters of volcanic material—more like entombment than ashfall
- •Papyrus normally decays quickly in Mediterranean climates without repeated copying
- •The library’s survival is an extreme historical anomaly
- 3:13 – 5:44
1700s discovery, looting-era excavation, and the tragedy of destroyed scrolls
The first modern excavations were tunnel-based and treasure-oriented, leading to widespread damage. Early attempts to open the scrolls destroyed many, though a few were successfully unrolled and revealed Greek philosophical works.
- •1750 well-digging leads to tunneling discoveries; period ethos resembled looting
- •Scrolls initially mistaken for charcoal; many discarded or ruined
- •Physical unrolling attempts (cutting, prying) destroyed hundreds
- •A monk-built unrolling machine slowly opened a few, revealing Philodemus (Epicurean tradition)
- 5:44 – 11:02
Brent Seales’ virtual unwrapping and the birth of the Vesuvius Challenge prize
Nat discovers Brent Seales’ effort to scan and virtually unwrap scrolls, then collaborates with him to launch an open competition. The prize aims to attract many approaches and push the problem over the finish line.
- •High-resolution CT scanning at Oxford’s Diamond Light Source (micron-scale voxels)
- •Goal: read text without physically opening scrolls
- •Nat reconnects with Seales’ team; progress exists but not yet full passages
- •Decision: create an XPRIZE-like contest (with Daniel Gross) to accelerate breakthroughs
- 11:02 – 13:43
What else might still be buried: the unexcavated villa and a possible Latin library
They explore why the villa remains only partially excavated and what that implies for the number and diversity of scrolls. The most tantalizing possibility is a still-buried main library, potentially containing Latin literary or historical texts.
- •Most exploration occurred via 1700s tunnels; the villa was never fully exposed/dug out
- •1990s work suggests additional floors and that only ~a third has been excavated
- •Philodemus texts likely came from a ‘working library’ room; other scrolls found in hallways/crates
- •Classicists suspect a larger, unexcavated main library—possibly Latin—and recount instances of accidental destruction
- 13:43 – 18:13
Why physical unrolling fails: papyrus structure, carbonization, and destructive “solutions”
Nat explains how papyrus is made and demonstrates how carbonization turns it into dust-like charcoal. Historical ‘fixes’ (oils, rose water, mercury) highlight why a non-destructive digital approach is necessary.
- •Papyrus is layered reed strips; scrolls are glued sheets formed into long rolls
- •Carbonization without oxygen produces extremely brittle, blistered material
- •Nat’s experiments with a Dutch oven replicate how easily scrolls crumble
- •Past restoration tricks were often ruinous—underscoring the need for virtual methods
- 18:13 – 23:50
The technical bottlenecks: ink detection, segmentation, and training from fragments
Nat breaks down virtual unwrapping into subproblems and explains why Herculaneum is ‘expert mode.’ The core difficulty is that the ink’s X-ray signature is close to the papyrus, forcing subtle ML-based detection trained via fragment ground truth.
- •Seales pioneered virtual unwrapping with the En-Gedi scroll, where ink had high X-ray contrast
- •Herculaneum ink is near-indistinguishable in X-ray, requiring CNNs to detect subtle patterns
- •Scroll geometry is severely distorted; layer segmentation is hard at full-scroll scale
- •Training data comes from broken fragments: align CT scans with infrared images to label ink
- 23:50 – 30:00
Why a prize (not a grant): exploring a big solution search space and triggering a funding cascade
Nat argues a competition is faster because many teams can try diverse approaches in parallel. The broader strategy is catalytic: prove one scroll is readable, and the money and urgency to scan/excavate the rest will follow.
- •Contest format increases parallel experimentation across approaches (ML or otherwise)
- •Key risk: models trained on fragments may not transfer cleanly to ink pressed against papyrus layers
- •Confidence drivers: demonstrated ink recognition; scans are extremely high resolution (large TIFF slices)
- •Catalyst thesis: once one scroll works, funding to scan remaining scrolls and excavate more of the villa becomes inevitable
- 30:00 – 33:21
Finding leverage points: enthusiasm, stubborn follow-through, and “the world isn’t efficient”
Dwarkesh zooms out to Nat’s broader pattern—seeding high-leverage projects like YIMBY and now the scrolls. Nat explains his decision style: assume neglected opportunities exist, follow persistent enthusiasm, and accept that projects become harder after commitment.
- •Rejecting the reflex: “If it were good, someone would already be doing it”
- •Leverage framing: small push → avalanche; prizes as catalysts for larger mobilization
- •Personal operating mode: low activation energy, impulsive commitment, then stubborn execution
- •Efficiency is uneven across domains; default estimates of ‘someone has this covered’ are too optimistic
- 33:21 – 40:36
AI secrecy vs open culture: hedge-fund analogies, lab norms, and open-source tradeoffs
Nat discusses why AI labs may struggle to keep secrets given academic and California norms, even as training efficiency improvements become economically meaningful. He then weighs open-source proliferation against safety concerns and data access constraints.
- •Hedge fund thought experiment: small disclosures can massively reduce profits; secrets can have long half-lives
- •AI labs face leakage pressures: academic publishing culture, social intermixing, non-compete limits in CA
- •Open-source instinct remains strong; excitement about local models (e.g., quantization running on laptops)
- •Safety worries exist (misuse/industrial accidents), but Nat still leans toward broad tinkerer access in current era
- 40:36 – 44:35
Inside Microsoft’s GitHub acquisition: the pitch, the memo, and becoming CEO
Nat recounts advocating for GitHub soon after joining Microsoft and later sending Satya a decisive memo. The core argument: developers now drive purchasing, and GitHub is the best way to rebuild Microsoft’s relationship with a new generation of developers.
- •Context: Xamarin acquisition brings Nat to Microsoft; he quickly pushes ‘buy GitHub’
- •Memo thesis: developer mindshare is the new distribution; GitHub is the developer hub
- •Leadership alignment: Satya’s responsiveness and decision to run GitHub independently (LinkedIn-style)
- •Rapid execution: term sheet and announced deal within weeks
- 44:35 – 50:18
Earning developer trust post-acquisition, why acquisitions fail, and why Nat left
Nat describes the distrust wave after the deal leaked and how he tried to signal developer-first priorities immediately. He also explains why many acquisitions destroy value (fragile culture, key people leaving) and why he ultimately chose to return to smaller-team work.
- •Leak-driven backlash: calls to ‘evacuate GitHub’ over a weekend before announcement
- •Trust-building tactic: ship a community-requested fix on day one; repeat daily for 100 days
- •Why acquisitions fail: cultural harmonics break; the few load-bearing people disengage or leave
- •Reason for leaving GitHub/Microsoft: coordination overhead in giant orgs became less energizing; preference for nimble teams
- 50:18 – 56:55
Copilot’s origin story: from GPT-3 awe to the UI that made it stick
Nat explains how GitHub and OpenAI iterated from early prototypes (chatbot Q&A, function synthesis) toward an autocomplete experience that delivered consistent value. The breakthrough was pairing the right model latency/quality tradeoff with a low-friction inline UI.
- •Initial instinct: coding Q&A chatbot; great demos but unreliable product experience
- •Exploration: whole-function synthesis with multiple options created heavy cognitive load
- •Key UI/product insight: inline gray-text completions; adapt token-length based on AST/context
- •Validation metric: unusually high early retention among internal GitHub engineers; clear product-market fit
- 56:55 – 1:11:46
GitHub as proof-of-work, open-source governance limits, and Stallman-era critiques
They discuss whether GitHub profiles replace resumes, and how open-source reputation helps outsiders break into elite networks. Nat also covers where open source excels (infrastructure) vs struggles (end-user UX) and addresses free-software ideological critiques.
- •GitHub profiles as partial credentialing; strong for ‘breaking through’ via public contributions
- •Labor market dynamics: scarce top dev talent enables ‘working in public’ and portable reputation
- •Open-source governance burden: entitlement, coordination overhead, and need for editorial control in UX-heavy products
- •Stallman critique: ideological freedom vs what most users practically optimize for; licensing simplicity (e.g., MIT)
- 1:11:46 – 1:23:14
Nat.org principles: ceiling vs floor, EMH skepticism, micromanagement, and AI-driven engineering productivity
Dwarkesh walks through Nat’s compact list of beliefs and extracts the reasoning behind several provocative claims. Nat argues ceiling-raising tech often becomes the best floor-raising strategy, that few actors drive many outcomes, and that strong judgment should sometimes override anti-micromanagement norms—especially as AI accelerates software creation.
- •‘Raise the ceiling, not the floor’: many direct aid-like tech projects underperform; rich-world tech scaling often wins (e.g., mobile internet)
- •‘Model the world as ~500 people’: local optimization, memetic conformity, and correlation effects of the internet
- •Micromanagement taboo: bad managers are the problem; great judgment should be empowered even if it looks like micromanaging
- •AI tools and engineering: Copilot-like systems increase productivity across skill levels; trajectory toward majority/near-total code generation and more bespoke software
- 1:23:14 – 1:38:23
AI’s economic impact, foundation-model equilibrium, incumbents adapting—and audience Q&A on alignment
Nat and Dwarkesh discuss how much of the economy is ‘text-to-text,’ why benchmarks lag real deployment, and where scaling may face diminishing returns. Nat predicts model proliferation (not a tight oligopoly), explains why incumbents now take new platforms seriously, and closes with Twitter questions including Nat’s view that alignment needs more open technical tooling.
- •Economic lens: low double-digit share is directly text-to-text today; impact depends on demand response as costs fall
- •Bear case: diminishing returns from scale (GPT-4 far more expensive than GPT-3, not proportionally more capable)
- •Model market structure: strong forces toward proliferation—public data, commodity hardware, leakable best practices—vs training-cost concentration
- •Incumbents: post–Innovator’s Dilemma world; CEOs take AI seriously rather than dismissing it, unlike past platform shifts
- •Alignment: capabilities and alignment may be complementary; need for open-source technical alignment community, not just closed-door work
- •Wrap-up: how to find the Vesuvius Challenge (scrollprize.org) and what winning enables
