Interview, Fireside Chat, Conference Presentation
The 2045 Superintelligence Timeline: Epoch AI’s Data-Driven Forecast
- The AI sector is currently not a bubble, though a sudden burst remains a possibility; catastrophic forgetting is an unsolved technical issue that has not yet slowed model capabilities.
- A 20 to 30 percent chance exists for a 5% unemployment increase within six months of the next decade, while a 10% job automation rate is strongly expected over the next decade alongside the creation of high-end new roles.
- AI-driven code generation is projected to reach 90% of total code by March 2026, though this prediction has not yet fully materialized as of the discussion date; 90% of future code is considered likely eventually.
- Anthropic projects "country of geniuses" equivalent systems by 2026–2027, while the median timeline for AI to perform any remote work task is estimated at 20–25 years from the discussion.
- GDP growth is forecast at roughly 1% by 2030 assuming no AGI, whereas achieving human-level remote work capability could result in 30% growth or a 100% collapse.
- A major unsolved math problem may be solved within five years, whereas an autonomous AI "co-scientist" breakthrough in biology or medicine is deemed less plausible in the near term.
- Robotics faces significant hardware and economic hurdles, with training compute currently 100 times smaller than frontier models, and software benchmarks lag due to vision capabilities and long-context coherence issues.
- Anthropic's "New Carlisle" data center will come online in January followed by Colossus 2, while Microsoft's "Fairwater" project will consume power comparable to over half of New York City's usage.
- Companies are starting data centers like Colossus One before grid connection to bypass infrastructure delays, and energy is not expected to be a durable bottleneck due to willingness to pay for solar and battery solutions.
- Revenue trends will clarify labor market impacts within one to two years, and pre-training focus is shifting toward post-training for reasoning without indicating a scaling plateau.
- Government attention to AI is forecast to grow exponentially, doubling or tripling annually, likely triggering rapid regulatory consensus resembling the COVID-era stimulus.
- Public reaction to rapid unemployment could shift from mild concern to strong consensus for unimaginable actions, with student focus on prompt engineering discouraged due to model improvements.
- Computational scaling continues with experimental spend on compute rather than researchers, making a software-only singularity currently unlikely, and AI automating almost all non-manual jobs remains plausible.
- If AI matches human capability in remote work within the next 10 years, the economic outcome is polarized between massive growth or collapse; otherwise, a 1% GDP increase is expected by 2030.