Webinar, Interview
What the hell happened with AGI timelines in 2025?
- AGI timelines have generally extended, with predictions shifting from three to five years in early 2025 to a window of roughly 2028–2032 for a make-or-break period, and Metaculus forecasts for strong AGI moving from July 2031 to November 2033.
- Expectations for fully automated AI research and development have adjusted to be more conservative, viewing 2027 as unlikely, 2028 as requiring surprising breakthroughs, and 2029–2030 as plausible only if current trends continue without new roadblocks.
- The anticipated generalization of reasoning capabilities from easily checkable domains (math, coding) to messy, non-checkable real-world tasks (e.g., booking flights) is expected to fail, leading senior AI staff to update AGI timelines to longer horizons.
- Significant performance gains observed in 2024 and 2025 were largely driven by inference scaling (increasing thinking time), which is considered a one-off event that cannot be repeated for normal situations due to global computer chip constraints.
- Future progress is expected to face diminishing returns and slower pacing compared to 2025 due to physical limits on scaling reinforcement learning, which Toby Ord estimates may be a millionth as compute-efficient as pre-training, alongside the inability to scale thinking time tenfold economically.
- Economic constraints limit the scalability of high-cost AI agents, where running time can equal hundreds of dollars per hour (comparable to human engineer costs), making further increases in thinking time irrational without a massive reduction in costs or a 400-fold decline in benchmark prices already seen.
- By 2032, the AI industry is projected to consume a huge fraction of global computer chips, electricity, and tech staffing, potentially creating financing challenges for single model training runs estimated at $1 trillion to $10 trillion.
- Revenue growth for major companies like OpenAI, Anthropic, and XAI has exceeded bullish forecasts, rising fivefold to $30 billion annually, indicating that current business models are profitable on a per-user basis despite skepticism about their long-term viability.
- Recent data from the EPOC Capabilities Index indicates that the performance of frontier AI models increased twice as fast after April 2024 as before, with costs falling rapidly and GPT-5 not representing a cessation of progress.
- While some skeptics previously estimated AGI was decades away, revised forecasts place a revolutionary technology capable of replicating human intellectual work at approximately 10 years out, though this is still viewed as insufficient time for global preparation.
- Continual learning and the ability of models to build knowledge over time with few samples remains a significant gap compared to human capabilities, with little visible progress on this front observed in 2025.
- Even if AI capabilities improve, the rate of progress may become a grind requiring AI assistance just to maintain current research speeds rather than accelerate them, potentially pushing timelines out by an additional one to two years.
- The narrative that AI progress has stopped or that companies face imminent bankruptcy is characterized as misguided, with evidence pointing to a relentless, albeit potentially slower, monthly increase in utility and capability.