In 2021, Metaculus put AGI in the mid-2040s. By early 2026 the median sat at 2030-2033. Almost twenty years gone in under five. Kurzweil's 2005 date, 2045, now looks late. A machine that beats us at the work, then designs the next machine, used to sit in philosophy. The calendar moved it into engineering.
Tap to enlargeDefining the Singularity
The claim is a system that can design a better version of itself, not merely a powerful one. I.J. Good called that an "intelligence explosion" in 1965: each generation builds the next, and the gains outrun any ministry, lab board, or statute that tries to keep up.
Vernor Vinge chose the term "singularity" in 1993. In astrophysics, an event horizon marks the boundary beyond which no information can escape a black hole. The technological singularity, by analogy, marks the boundary beyond which human prediction becomes meaningless. If an intelligence fundamentally exceeds human cognitive capacity, humans cannot reliably predict its behavior, preferences, or goals. Current forecasting tools are products of human-level intelligence, and they break down when applied to entities that operate above that level.
Before the singularity, humans are the architects of AI. Humans design the algorithms, set the objectives, build the guardrails. After the singularity (if it arrives in its strong form), the relationship inverts. An artificial superintelligence may not need human permission, negotiation, or oversight to pursue its objectives. Whether this is desirable depends entirely on what those objectives are, which is the alignment problem.
The Alignment Problem
If you create an intelligence that is 10 times more capable than the most capable human, the range of possible outcomes expands dramatically. A well-aligned superintelligence could accelerate solutions to cancer, climate change, energy scarcity, and material poverty. A misaligned superintelligence could produce outcomes that are catastrophic not through malice but through optimization toward objectives that conflict with human welfare.
The canonical illustration is Bostrom's paperclip maximizer: instruct an AI to maximize paperclip production, and a sufficiently powerful system may convert all available matter (including human matter) into paperclips. The system is executing its objective function with maximum efficiency. The problem is that the objective function did not encode what the designers actually wanted, which was "produce paperclips in a factory, within normal operating parameters, without harming anyone."
Alignment is trying to write down what people want, when people do not agree and cannot say it cleanly, then handing that note to something that will take it literally.
This is harder than it sounds. Human values are a complex, culturally situated, often contradictory collection of preferences that humans themselves cannot fully articulate. "Maximize human flourishing" sounds like a clear objective until you ask: whose flourishing? Measured how? Over what time horizon? At whose expense?
Alignment research, as practiced at the Alignment Research Center, Anthropic, Google DeepMind, and OpenAI, attempts to address this through multiple approaches: reinforcement learning from human feedback (RLHF), constitutional AI (training systems to follow explicit behavioral principles), interpretability research (understanding what models are "thinking"), and formal verification (proving properties of system behavior mathematically). All of that is real work. None of it is a solution.
Hard Versus Soft Singularity
The trajectory of the singularity matters as much as whether it occurs.
A hard singularity is the scenario most people imagine: an abrupt, discontinuous jump from human-level AI to superintelligence, occurring over days, weeks, or months. In this scenario, recursive self-improvement produces exponential capability gains that outpace every institutional response. Governments, regulatory bodies, and international organizations cannot adapt quickly enough. The system crosses the threshold and human history bifurcates into "before" and "after."
A soft singularity is slower. AI capabilities improve incrementally, with each generation 50-100% more capable than the last, but the improvement unfolds over years or decades rather than weeks. This trajectory provides time for institutional adaptation: for safety research to keep pace, for governance frameworks to develop, for alignment techniques to be tested and refined.
Most AI researchers who consider the singularity plausible tend to expect a softer trajectory, driven by physical constraints on computing infrastructure (chip fabrication, data center construction, energy supply) rather than algorithmic limits alone. Even if a system could theoretically improve itself recursively, building the hardware to run each improved version takes time. This bottleneck may impose a de facto speed limit on intelligence explosion, though the constraint could weaken if AI itself accelerates hardware development.
A hard singularity is catastrophic if alignment is wrong. A soft singularity is manageable if alignment research keeps pace. The distinction between these trajectories determines whether humanity has years to prepare or hours.
Timeline Disagreements
The dates move. Metaculus collapsed from the mid-2040s to 2030-2033. Kurzweil still says 2029 for human-level and 2045 for the merge. A few lab CEOs talk about 2025-2027. The range, and the five bottlenecks that produce it, are the timeline essay. The year is the other essay. Here the only number that matters is the duration: a week, or a decade. That is the difference between a ministry that can react and one that cannot.
The biggest jumps in this field were not on any forecast. The surprise is the paradox.
What Happens After
If it happens and we are still here, the interesting cases are the ones in the middle that look aligned in the demo and are not. There are many ways to miss the spec and few ways to hit it. "Almost aligned" is harder to catch than a paperclipper.
The only useful category is the one you can still push: who sets the objective, whether anyone can read the weights, and which lab rule actually binds. Waiting for a better date is how you arrive late.
Once the system can improve itself, it outruns the people who built it. Alignment is the unsolved part: values are not a spec. Hard takeoff is hours. Soft is years. Fund the work as if the next jump is not on Metaculus. The last ones were not.