Timelines and Uncertainty

Ask ten serious AI researchers when AGI or transformative AI will arrive, and you may get answers ranging from five years to never. This is not because the question is meaningless — the answer matters enormously for how individuals, organizations, and governments should act today. It is because the question is genuinely hard, the relevant evidence is ambiguous, and the underlying mechanisms are poorly understood. The right response to this situation is not to pick a number and commit to it confidently, nor to dismiss the question as unanswerable. It is to reason carefully under deep uncertainty — treating timelines as probability distributions rather than point estimates, being explicit about which assumptions drive your estimate, and updating as new evidence arrives.

What Expert Surveys Reveal

Several organizations have systematically surveyed AI researchers on their timeline beliefs. The most comprehensive are the AI Impacts surveys of machine learning researchers, conducted in 2016, 2022, and 2023. The results are striking for several reasons. Wide disagreement: even within the research community, median estimates for AGI timelines vary enormously. In the 2022 survey, the median estimate for when AI would perform any task as well as humans was around 2059 — roughly three decades away — but the standard deviation was enormous, with many researchers estimating less than ten years and many estimating over a century. Timelines have shortened: comparing the 2016 and 2022 surveys, AI researchers gave shorter timelines in 2022. This is consistent with the rapid capability improvements observed in the intervening years (GPT-3, AlphaCode, DALL-E, and their successors) but also reflects the risk of anchoring to recent progress. Substantial probability on short timelines: even researchers who give a median estimate of thirty years often assign 10-20% probability to AGI within ten years. When aggregated across the research community, the expected probability of AGI within a decade is non-trivial — perhaps 20-50% depending on the survey and the definition used. This matters: a 20% chance of a transformative event in ten years is not a remote possibility; it is a substantial probability deserving serious preparation. Heavy uncertainty on both ends: the surveys also reveal substantial probability mass on very long timelines — and a meaningful fraction of researchers assign significant probability to AGI never being achieved via current approaches.

Sources of Deep Uncertainty

Timeline uncertainty for transformative AI is not ordinary forecasting uncertainty — it is what researchers sometimes call 'deep uncertainty' or 'Knightian uncertainty.' In ordinary uncertainty, you know the probability distribution over outcomes (like the probability of a coin landing heads). In deep uncertainty, you do not know the distribution itself. Several factors make AI timelines deeply uncertain. Algorithmic unpredictability: the history of AI is punctuated by unexpected algorithmic innovations — backpropagation, convolutional networks, attention mechanisms, reinforcement learning from human feedback — that dramatically shifted what was achievable. The rate of such innovations cannot be forecast from their history; they are, by definition, not predictable in advance. Unknown distance to goal: unlike forecasting how long it will take to climb a mountain when you can see the peak, estimating AI timelines requires knowing how far away the goal is — and the goal itself (AGI or transformative AI) is poorly defined and its requirements are poorly understood. You do not know whether current approaches are 10% of the way there or 95% of the way there. Social and political contingency: timelines depend not just on technical progress but on investment levels, regulatory decisions, geopolitical competition, and talent availability. Any of these could change rapidly in response to events that are themselves unpredictable. Feedback between timelines and outcomes: belief about timelines affects investment, which affects development speed, which affects actual timelines. Widely-held short timelines can be partially self-fulfilling if they attract resources, and widely-held long timelines can be self-fulfilling if they reduce urgency.

How to Reason Under Deep Uncertainty

The appropriate response to deep uncertainty is not paralysis — it is disciplined probabilistic thinking and robust decision-making. Use reference classes: instead of reasoning only from first principles about a unique situation, find reference classes — historical cases that share key features. How long did it take for major new computing paradigms to move from research demonstrations to transformative deployment? Reference classes do not determine the answer, but they anchor estimates in observable history. Elicit and make explicit your assumptions: every timeline estimate rests on assumptions about algorithmic progress rates, compute availability, and what capabilities transformative AI requires. Making those assumptions explicit allows you to identify which ones, if wrong, would most change your estimate — the 'cruxes' of the disagreement. Assign probability distributions with multiple scenarios: rather than committing to a single point estimate, assign probabilities to a range of scenarios. This forces explicit consideration of both optimistic and pessimistic paths and produces actionable uncertainty: you can ask 'given this probability distribution, what decisions are robust across most scenarios?' Update on evidence: when new evidence arrives — a new benchmark result, a new paper on algorithmic efficiency, a new survey of researchers — update your distribution. Good forecasters update frequently and in proportion to the evidence, without overreacting to single data points or dismissing evidence that conflicts with prior views. Distinguish what you believe from what you are willing to act on: under deep uncertainty, the gap between your best estimate and your uncertainty range is critical. A 50% probability of AGI within twenty years and a 20% probability of AGI within ten years may have different implications for which preparations are urgent.