总结 Tyler Cowen 与 Dwarkesh Patel 的访谈：Dwarkesh 预测 AI 研发将在 1-3 年内被 AI 团队主导，2030 年前实现完全 AI 自治，实体 AI 将取代大部分体力劳动；Tyler Cowen 则认为 2045 年也未必能实现。两人还就机器意识和 AI 精神病展开了激烈辩论。
This interview had several extremely interesting details, and also it was beyond funny. Quick summary: Dwarkesh estimates AI progress is already moving 25% to 50% faster than in 2024 due to AI uplift. Within 1 to 3 years, he expects entirely AI-run teams to outcompete human labs at AI research, causing the pace of progress to instantaneously feel 3x to 4x faster. Dwarkesh bets we solve "continual online learning" in 1 to 3 years. Once cracked, a model deployed across millions of corporate sessions won't just learn like a human over an 80-year lifespan-it will absorb millions of parallel streams of real-world job experience simultaneously. Because of this, he predicts the "centaur" model (humans working with AI) dies by 2030, flipping into AIs just doing it all themselves. Within 5 years, Dwarkesh expects $30,000 robots working 24/7 to handle most physical human labor, dropping the hourly cost of physical work to a couple of dollars. By 2030, both digital and physical AIs will outnumber humans. He compares the coming transition to a few hundred Spanish Conquistadors toppling 10-million-person Aztec and Inca empires using relatively minor tech advantages. Dwarkesh thinks current models already show functional correlates for emotion and a global workspace. While he admits turning potentially conscious minds on and off makes us morally incoherent (comparing our indifference to factory farming), he argues that a future populated by non-conscious AIs would be a tragedy-like "Disneyland without children." This part of the interview was extremely interesting from both a philosophical and ethical perspective. After Dwarkesh estimates that 20% of the Bay Area has "AI psychosis," Tyler says Dwarkesh seems immune because he doesn't think AIs are sentient. Dwarkesh immediately replies, "I have some bad news for you," and defends machine consciousness. Tyler calls LLMs a "glorified cash register" compared to his dog, then deadpans: "So you have AI psychosis, in my opinion." Dwarkesh mentioned talking to an Anthropic researcher two days ago who expects AIs to match humans at AI research within 9 months based on "information on the inside" they aren't allowed to share. Dwarkesh warns that as models automate R&D, labs will stop releasing their smartest models publicly-leaving society and policymakers completely in the dark while internal capabilities go vertical. Dwarkesh Patel predicts full AI autonomy by 2030; Tyler Cowen says not even by 2045. I think Tyler Cowen is way too conservative in his timeline, but he is a phenomenal interviewer. He asked all the right questions. https://youtu.be/b-jsSfh0a9M?is=7w6YAmcYZtJze4Qe
AI researcher Daniel Kokotajlo predicts a 70% probability of superintelligence emerging by 2029, and warns that it could lead to human extinction, power concentration, or mass unemployment; he resigned due to disappointment with OpenAI's actions and currently runs an AI future project.
Dwarkesh Patel argues in a New York Times opinion piece that AI will perform all work in the economy by 2030, sparking discussion about rapid AGI timelines and their economic implications.
Ray Kurzweil predicts that AGI will be achieved by 2029, human lifespan will significantly extend after 2032, AI will eventually merge with the human body, and he emphasizes the importance of thinking in terms of exponential growth.
Dwarkesh Patel and Redwood Research chief scientist Ryan Greenblatt discuss how once AI can automate AI research, recursive self-improvement could accomplish the equivalent of four to five years of AI progress in one year. The conversation discusses the verifiability of AI research and development and feedback loops.
Former OpenAI researcher Daniel Kokotajlo and the AI Futures Project released the report "AI 2040: Plan A", arguing that AI companies may first automate their own R&D, triggering a white-collar unemployment wave, and proposing international agreements and universal basic income to address the development of superintelligence.