Interview, Fireside Chat, Other
Daniel Kokotajlo on how superintelligent AIs could build a self-replicating robot economy in months
- Multiple timelines project AGI arrival between 2027 and 2031, with one narrative forecasting an AI takeover by 2030 and another suggesting a slower takeoff, while a future scenario titled "AI 2030" is planned to depict timelines ranging from 2029 to 2035.
- Specific operational milestones include the 2025 release of limited public agents, a 50% acceleration in AI research via internal agents by 2026, and the 2027 emergence of superhuman coder agents and subsequent AI generations capable of achieving a year's of progress in a week.
- Geopolitical risks involve China nationalizing AI research in 2026 and intelligence agencies stealing model weights by early 2026 and February 2027, creating a competitive race that could render one side an existential threat if the other advances without parallel development.
- Divergent scenarios describe outcomes for 2027–2030, ranging from human extinction due to a misaligned "Consensus AI" co-designed by rival systems to a "slowdown" path where isolation of dangerous agents leads to universal basic income, widespread robotics, and fusion power by 2030.
- Market and structural predictions anticipate that public release of advanced agents in mid-2027 could destabilize labor markets by replacing employees at one-tenth the cost, while a default trajectory suggests 25% to 30% probability of positive outcomes if one to five companies do not concentrate control over ten superintelligent minds.
- Technical trends indicate a six-month doubling time for reliable task complexity, a potential shift in model internal language to uninterpretable forms, and an expected tapering of training compute investment in a couple of years as companies face financial constraints.
- Strategic recommendations include implementing domestic regulation and international deals to prevent corporate power concentration, establishing whistleblower protections, and adopting a democratic governance structure to determine the values of concentrated AI systems to avoid a default path of unchecked advancement.