Toby Ord 🇦🇺 🇬🇧
Philosopher and author whose research examines existential risk, humanity's long-term future, and safeguards against global catastrophe.
University of Oxford
- Evidence items
- 5
- Toward pressure
- +0.14
- Away pressure
- −0.35
- Net attributed pressure
- -0.21
0 comments · 0 votes
Sign in to join the discussion →
No comments yet. Start the discussion.
First-person and official sources
These sources guide discovery. A statement still needs a dated, attributable, source-backed evidence assessment before it can affect the index.
- personal siteToby Ord
Assessments involving Toby Ord
Toby Ord argues AI planning should use broad timelines rather than a single date
Toby Ord argued that decision-makers should plan against a broad distribution of transformative-AI timelines rather than a single forecast year. He defined the relevant threshold as systems capable of takeover if misaligned, permanently displacing half of workers, or doubling technological progress, and placed his own median at 2038 with an 80% interval spanning roughly three to one hundred years.
Toby Ord argues reinforcement-learning scaling is approaching an effective limit
Toby Ord analyzed public o1, o3, and GPT-5 performance curves and argued that reinforcement-learning scaling requires orders of magnitude more compute for continued gains and may be nearing an effective limit. He presented this as a quantitative constraint with important uncertainties, not as proof that frontier progress has stopped.
Toby Ord finds most measured reasoning gains came from longer inference
Toby Ord estimated that 63% to 92% of measured reasoning gains on three benchmarks came from using more inference tokens rather than improved performance at a fixed token budget. He argued that this raises deployment cost and moderates conclusions about gains in underlying intelligence.
Toby Ord argues reinforcement learning may impose a steep efficiency bottleneck
Toby Ord estimated that frontier reinforcement learning extracts roughly 1,000 to 1,000,000 times less learning signal per token than pretraining and may trade generality for benchmark performance. He presented the claim as a quantified bottleneck with explicit caveats, not as proof that capability progress has ended.
Toby Ord estimates a 10% century-scale risk of losing control to unaligned AI
In a full interview about The Precipice, Toby Ord estimates a 10% chance that unaligned AI causes humanity to lose control of its future within a century. He decomposes that forecast into a 50% chance of developing systems more intelligent than humanity and an 80% conditional chance of surviving that transition while retaining control.