»🛝»«

Отличный подкаст про AI What is it about Dwarkesh interviews Tamay Besiroglu and Ege Erdil (ex‑Epoch AI, now co‑founders of Mechanize, a startup aiming to automate work). They argue for long AGI/automation timelines (≈2045+), reject the popular “intelligence explosion / alignment-doom” framing as the central lens, and instead predict something else: a big, economy-wide automation wave that could drive extremely fast economic growth—but through broad deployment + capital buildup + supply chains, not a “software-only singularity.” Summary Ege Erdil & Tamay Besiroglu (ex‑Epoch AI, now Mechanize) argue for long timelines to “drop‑in remote worker” automation (~2045+), and say the popular “intelligence explosion / alignment” framing over-focuses on raw IQ instead of the many complementary pieces needed for real-world transformation. Why long timelines: jobs aren’t single tasks; they’re bundles requiring long-horizon coherence, autonomy/agency, and robust multimodal + contextual understanding. Past “easy” breakthroughs (chat-style post-training, reasoning) looked simple only after years of scaling, data, hardware, and engineering. They also expect compute scaling to hit practical constraints (energy, fabs, datacenters, GDP share), limiting how quickly new capability “unlocks” arrive. Why they doubt a software-only singularity: automating AI R&D is harder than benchmark tasks suggest—real research is messy, long-horizon, and about choosing directions, not just solving closed-form evals. They think big progress historically is tightly coupled to experimental compute and hardware; scaling “research labor” alone likely faces diminishing returns and bottlenecks. What they do expect: explosive economic growth once AI can broadly substitute for labor. Because AI workers can be copied, labor can scale, which accelerates capital accumulation, learning-by-doing, and deployment-driven data. They expect this to integrate into existing supply chains/markets (not a tiny “AI island”), though adoption and growth may vary sharply by regulatory jurisdiction. Main brakes/risks: regulation or coordinated slowdowns could meaningfully delay deployment; geopolitics may create arms-race dynamics, but short leads may not translate into decisive conquest. On “alignment” and value lock-in, they’re skeptical that misalignment automatically implies war or permanent lock-in; they emphasize humility, adaptability, and stronger institutions. They also note delaying AI has huge near-term costs (health, wealth, lives). Implications: “fully automated firms” could have superhuman collective advantages via high-fidelity replication of tacit knowledge/culture and reduced principal–agent problems. Central planning might improve with higher bandwidth and coordination, but complexity may still favor markets. Career advice: prioritize high-signal literature, write and share thoughts, talk to smart people, and be aggressive about reaching out.