Interview
Nat Friedman (Github CEO) — Reading ancient scrolls, open source, & AI
- Reading all 600+ intact scrolls could double the total text available from antiquity, with a 50% probability of solving this challenge within the current year if current techniques generalize.
- A successful proof of concept on a single scroll may catalyze low-million-dollar funding (potentially $1M) to scan remaining scrolls, potentially leading to the excavation of thousands of additional papyri within the year.
- AI lab secrets are predicted to have a relatively short half-life due to cultural norms and weak non-compete enforcement, potentially reducing hedge fund profits by 80% after 12 months if shared widely.
- The current year is forecast as a period of "proliferation" featuring four or five major AI announcements involving new models, APIs, platforms, and tools.
- GitHub's "100-day" initiative aims to ship daily fixes to build trust and overcome shipping hesitation, with specific community-identified issues targeted for resolution on the CEO's first day.
- Software productivity is expected to rise as AI tools like Copilot increase output for all engineers, potentially reaching a state where 95% of code is AI-generated.
- As software creation becomes cheaper, the market may shift toward special-purpose custom code for individual companies rather than general-purpose tools, though open-source success will likely remain strongest where large-scale coordination is unnecessary.
- The AI landscape is predicted to reach a stable equilibrium of wide-scale proliferation rather than a tight oligopoly, driven by accessible data, commodity hardware, and the existence of a "compute overhang" with millions of GPUs available.
- GPT-4 is noted as not being two orders of magnitude more capable than GPT-3 despite higher training costs, suggesting that narratives on diminishing returns from scale may be over-optimistic.
- Future foundation models are expected to develop very similar skill repertoires due to memetic forces chasing marginal improvements, while data sources like LibGen and Archive will likely become harder to scrape as their value for trading newer models is recognized.
- Technology "ceiling raising" (e.g., cheap cell phones) is viewed as more effective for expanding internet access to Africa than targeted "floor-raising" projects.
- Large companies are expected to continue acquiring products due to distribution costs exceeding invention costs, even if internal dynamics complicate execution.
- The speaker anticipates that the probability mass for AI safety is best served by broader public tinkering with current models, despite acknowledging risks of industrial accidents or misuse.
- Incumbents are becoming more adept at avoiding disruption, with every leading CEO currently adopting AI technology aggressively.
- Open-source technical alignment communities are currently missing but are expected to emerge as the field matures.